Please post your responses to the readings in this forum. For each reading there will be a new topic thread, your responses should be posted as a reply within the topic thread. Thank you.
Hi Maria,
Can you please provide written instructions for our final assignment? We discussed it in lecture verbally on May 8th, but I don't see written instructions on any of the slides. Would be helpful to have them to make sure the assignment is completed thoroughly.
Thank you,
-Luis
Identify a research question of interest to you, the population of interest, and a study design you might use to examine the question.
Are the contraceptive use habits of Haitian women associated with their obstetric history? I would attempt to answer this question using a cross sectional design of reproductive age women.
Explain how you might incorporate a sampling strategy into the study design (you might, for example, use respondent-driven sampling to initiate a cohort study).
A sampling strategy to capture a representative sample of women in Haiti could employ probability sampling (systematic random sampling) using geographic locations (GPS or similar?) selected randomly. Within each community, women can be approached from households randomly selected based on their location from the geolocation such as the 2nd house located north. This approach would be useful for this question because there may be many differences between women in rural vs urban settings so selecting women from even small market centers may not capture those who live very rurally.
Briefly discuss possible logistical/practical advantages and disadvantages to this plan.
Advantages: Because addresses are not used consistently in Haiti, using geolocations is more precise. Because this approach doesn’t rely on use of community resources (health center, market, school) a wider range of women will be selected. The statistical conclusions drawn from data from this sample would be stronger because of the samples representativeness. The geography of Haiti is clear so the ‘list’ of locations from which to draw the sample is available.
Disadvantages: The terrain in the Haitian countryside may make it challenging to find the chosen household even using this technique. This approach is more expensive than recruiting from one location in the community.
Finally, discuss whether you think incorporating the sampling strategy might help
(1) reduce bias in the estimation of univariate quantities (such as disease prevalence) and
(2) reduce bias in the estimation of causal effects.
Both of these concerns would be attenuated with the use of probability sampling which aims to reduce biased estimation because of a biased sample.
One research question of interest to me is whether the relationship between characteristics of education and later-life cognition in community-dwelling older adults differs between rural and urban schools, with the hypothesis that measures such as student-teacher ratio may be indicators of different exposure characteristics across region classifications. To investigate this research question, I would incorporate nonproportional quota sampling to initiate a cohort study. Nonproportional quota sampling involves segmentation (defining the strata), setting the minimum size of the quotas, and selection of participants. The difference between quota sampling and stratified random sampling falls in the selection of participants, as stratified random sampling gives all potential participants an equal probability of being included and quota sampling may utilize techniques such as convenience sampling, with sampling ending once minimum quotas and other factors (such as sufficient power) are attained. Advantages of quota sampling include the ability for researchers to specify the minimum number of samples in each category, which may be necessary for potentially sparse data when investigating participants in rural areas. This approach may also reduce bias in the estimation of causal effects if the categories are well-defined and account for enough variation in regional characteristics. However, if there is selection bias into the categories used to define the quotas and factors related to educational exposures and later life cognition predict residence in a particular area, univariate quantities may remain biased. Quota-based cohort studies also may still rely on retrospective reporting of circumstances or early life exposures, making them subject to recall bias as well.
Research question of interest:
In a population of HIV-uninfected adolescent girls and young women (AGYW) in what is the enacted adherence (i.e., pill taking) and persistence (i.e., continuation of PrEP over time) in four study conditions: 1) peer mentor intervention, 2) text message reminders and 3) combined peer mentors and text message reminders intervention 4) control in Kisumu county in western Kenya.
The population of interest: HIV-uninfected adolescent girls and young women (AGYW) (aged 15-24 years) initiating Pre-exposure prophylaxis (PrEP) in Kisumu county in western Kenya.
Study design: clustered 2X2 factorial design as seen in the figure below. Five sites will implement both interventions, five sites will implement Peer mentoring intervention only, five sites will implement the text message intervention only, and five sites will implement the existing standard of care only.
|
Figure 1: 2X2 Factorial design |
|||
|
|
|
Peer mentor intervention |
|
|
|
|
Yes |
No |
|
Text messaging intervention |
Yes |
Both Peer Mentoring and Text Messaging (5 sites) |
Text Messaging only (5 sites) |
|
No |
Peer Mentoring only (5 sites)
|
Neither Peer Mentoring nor Test Messaging (5 sites) |
|
Explain how you might incorporate a sampling strategy into the study design
The study intervention is most efficiently applied at the level of the community-based drop-in centers (DiCE) as this approach circumvents potential problems of individual-level implementation such as provider compliance with the protocol, including crossover. Within Kisumu County, there are 43 registered DiCEs. The DiCE will be the unit of randomization.
Sample size: This study requires 20 DiCE and a total of 1280 adolescent girls. I have powered this study based on differences in proportions in a cluster randomized trial for our key outcomes of enacted adherence (i.e. % of AGYW who report at least 95% pill taking) and persistence (i.e., the proportion of women who continue using PrEP over two years). This sample size and power estimates are based on both clinical and logistic factors
Sampling: For this study, I will use multistage sampling.
Simple random sampling of clusters: I will develop a sampling frame of all the 43 registered DiCE in Kisumu County and select using random tables the 20 required.
To select individuals: I will use the respondent driven sampling (RDS) method to find access to potential study participants who meet the inclusion criteria. Seeds who reflect a diverse social demographic profile will be purposively selected and once enrolled, will be given up to 3-5 coded coupons to refer members of their social networks who in turn will become recruiters. Participants will receive 250ksh for every eligible person they recruit into the study. This process will continue in recruitment “waves” until the sample size is reached.
Possible logistical/practical advantages and disadvantages to this plan
The logistical advantages of this plan are:
Disadvantages
- The RDS can lead to potential selection bias of participants especially if the study population is not adequately “networked.” As such this can lead to the systematic exclusion of sub-populations of AGYW who are less networked than others, for example, AGYW who are female sex workers may be more networked than AGYW who are injection drug users
- Some participants will know more people than other participants as such not everyone has the equal chance of being included in the study. (This can potentially be managed through weighting techniques)
- There is a risk of non-independence if participants are more likely to recruit people who are like themselves from their own in-group (homophily)
- The combined multistage sampling has a greater risk of a nonrepresentative sample
Discuss whether incorporating the sampling strategy might help
(1) reduce bias in the estimation of univariate quantities (such as disease prevalence)
In general, the RDS sampling method can reduce bias in the estimation of univariate quantities after several waves of recruitment when the total sample is no longer influenced by the initial sample (seeds). This is because after several waves, subsequent waves begin to represent the underlying population. However, this statement only true if the following assumptions are met:
1) Sampling is with replacement, in which selected peers may be recruited multiple times;
2) Network size: that participants can accurately report personal network size;
3) Random recruitment: peer recruitment is a random selection from recruiter’s network.
Unfortunately for this proposed study, we violate some of these assumptions and as such this method may not reduce bias in the estimation of univariate quantities e.g.
- we will do sampling without replacement and preventing individuals from participating more than once.
- The 250ksh incentive that we provide incentives may reduce the randomness of peer recruitment as participants may preferentially recruit peers based on their relationship and/or assumptions of who will participate.
(2) reduce bias in the estimation of causal effects.
The multistage cluster and RDS sampling are more likely associated with an increased risk of a non-representative sample which may bias the estimate of causal effects. In particular, the RDS method as earlier alluded to has inherent problems such if the study population is not adequately “networked,” average network size and homophily which adversely affect effect size estimation
I’m interested in the relationship between air pollution and birth outcomes. I’m looking at California specifically, and we are proposing to use birth records for outcome classification, and living (and being pregnant) within 5km of a power plant that was retired within 2000-2015 as the exposure variable. Our hypothesis is that women living and pregnant within 20km when the power plants were operating will more likely have pre-term births or babies that are small for gestational age compared to women living within 20km of power plants after power plants were retired. From this we could do a multi-stage sampling scheme. First, we would identify our clusters consisting of women who gave birth with residential addresses within 20km of a power plant that was operating between 2000 and 2015. Then we would stratify our sample to women who were pregnant before versus after the power plant was retired. Random sampling reduces the likelihood of selection bias and minimizes the potential, helping identify causal effects than using a non-probabilistic sampling. This method is advantageous over simple random sampling because it ensures that women living near the retired power plants are included in the study. However, because we are limiting our study to women living within 20km of a power plant, it could reduce the representativeness of our study population to our target population (California).
Research question: Does a community-based dog rabies vaccination program increase the rabies vaccination coverage rate and reduce the incidence of human rabies?
Population of interest: Rural and urban counties of southern Kenya
Study design: This study would essentially incorporate a version of convenience sampling, since participating in the program would be voluntary. I propose using a stepped wedge design randomized cluster design, with the unit of randomization being village for rural counties and divisions for Mombasa. The intervention would entail offering free rabies vaccination clinics for domestic dogs at several time points during a calendar year, with villages or divisions randomized to the year the intervention took place. Vaccination coverage and rabies incidence would be compared for those in the intervention time period vs. control. Advantages of this plan are the relative ease and lower cost of simply offering a clinic rather than actively seeking out domestic dogs (e.g. going house-to-house) and enforcing vaccination. It is a more realistic intervention that has a better chance of being taken up by the ministry of health and continued as a regular government program.
Would this reduce bias in the estimation of univariate quatities? No, because this intervention will only be taken up by those who respond to the advertisement of the clinics. This is a population that is unlikely to be representative of the population of these villages or divisions as a whole, and different sub-populations are likely to have different prevalences of disease.
Would this reduce bias in the estimation of causal effects? Yes, because of the randomized stepped wedge design, we should be able to draw a causal inference on the effect of the intervention, assuming no other temporal trends are affecting the dog vaccination and rabies rates.
Sampling Assignment:
It is not unusual to encounter epidemiological or medical studies that use a non-probabilistic sampling scheme. For example, many randomized controlled trials use convenience sampling. Identify a research question of interest to you, the population of interest, and a study design you might use to examine the question. Explain how you might incorporate a sampling strategy into the study design (you might, for example, use respondent-driven sampling to initiate a cohort study). Briefly discuss possible logistical/practical advantages and disadvantages to this plan. Finally, discuss whether you think incorporating the sampling strategy might help (1) reduce bias in the estimation of univariate quantities (such as disease prevalence) and (2) reduce bias in the estimation of causal effects.
I am interested in studying the real-life effect of a metformin intervention on reducing risk of developing type 2 diabetes among high risk low-income Mexican adults living in Mexico City using a clustered randomized control trial design. We would randomly select a number of clinics who serve the low-income population and randomize half for the intervention and half for the control. Within clinics, we would offer the intervention to all eligible participants using a pre-determined definition (age, pre-diabetes, BMI). From the eligible participants, we would randomly select a number of them and invite them to enroll in our study. Controls, would receive standard of care (weight loss counseling, which those on the intervention would also receive).
Some of the logistical issues with this is that it will take much longer to recruit a sufficient sample size since we first need to identify a sufficiently large eligible pool from which to randomly draw. Since the population of study is also of low SES, it may be particularly challenging to be able to successfully enroll all those randomly selected to be in the study, and thus potentially introducing some selection bias.
Depending on how successful we are on enrolling all randomized participants, we may be able to draw valid inferences on the estimation of univariate quantities (such as disease prevalence among those who receive medical care at Seguro Popular), but if there ends up being a large self-selection then we would not be able to present valid estimates. Regardless of the final sample, we would be able to estimate causal effects of Metformin for the prevention of Type 2 diabetes among high-risk low-income Mexican adults.
Dr. Chen's reflection question:
It is not unusual to encounter epidemiological or medical studies that use a non-probabilistic sampling scheme. For example, many randomized controlled trials use convenience sampling. Identify a research question of interest to you, the population of interest, and a study design you might use to examine the question. Explain how you might incorporate a sampling strategy into the study design (you might, for example, use respondent-driven sampling to initiate a cohort study). Briefly discuss possible logistical/practical advantages and disadvantages to this plan. Finally, discuss whether you think incorporating the sampling strategy might help (1) reduce bias in the estimation of univariate quantities (such as disease prevalence) and (2) reduce bias in the estimation of causal effects.
Research question: What is the effect of district-wide indoor residual spraying (IRS) of insecticide on the risk of low-birthweight infants?
Population of interest: Ugandan infants born to mothers that protected by the Uganda IRS Project, a large-scale IRS campaign initiated in late 2014. (Ideally, we would like to make this generalizable to areas affected by high malaria burden, where the majority of the vectors are indoor-biters).
Study design: Difference-in-differences analysis. We collect data between 2012-2016 from the maternity wards of 28 district-level hospitals, 14 of which were treated with IRS starting in 2014.
Sampling strategy of districts: For treated districts, we will collect data from all 14 districts that were sprayed (n=14). Control districts (n=14) would be sampled by propensity score matching of malaria prevalence and health facility scores (i.e. rigor of data collection) assigned by the District Health Information System (DHIS2).
Advantages/disadvantages of the study design:
Advantages:
Disadvantages:
Sampling strategy questions:
(1) Reduce bias in the estimation of univariate quantities? Disease prevalence or other univariate estimates using this sampling strategy may not be the most accurate portrayal of disease prevalence in the district, since we’d only be collecting data from district-level hospitals, which are often the most saturated with women referred from lower-level health centers who may be at high risk for adverse birth outcomes. Thus, we may expect that the prevalence estimated in each district is an overestimation. That being said, low birthweight and other adverse birth outcomes, including maternal mortality, is severely underreported in Uganda and most of the women who experience these adverse birth outcomes may deliver outside of the district-level hospitals (i.e. come from rural areas where malaria incidence is much higher). In sum, disease prevalence estimated from our data would probably not be accurate, and it’s challenging to know what direction the bias would be.
(2) Reduce bias in the estimation of causal effects? I think by sampling control districts using propensity scores to mitigate confounding would reduce bias in the estimation of causal effects. By using this method, we are asking a different question which is that we are interested in the ATET and not particularly the ATE. In my opinion, estimating the ATET, in this particular case is of public health importance and would help reduce violations of the positivity assumption (there was a selective process in determining which districts would be sprayed—particularly driven by the malaria prevalence reported in each district).
Research Question: Prospective longitudinal case-cohort study in adults aged 50 or over designed to assess the absolute risk of atypical femoral fracture (AFF) and provide important data on impact of bisphosphonate use and other risk factures for AFF.
Background: Studies to date have generated inconsistent estimates of the incidence of AFF and the magnitude of the association with bisphosphonate use, perhaps a result of differing study designs, comparators and definitions of AFF. Because there are few adequately powered large prospective studies, critical areas of uncertainty include the relationships with bisphosphonate exposure, particularly the duration of treatment, and importantly, the resolution of risk after discontinuation. Uncertainty also exists regarding additional risk factors, such as the effects of age and gender, severity of osteoporosis, concurrent medical conditions (such as rheumatoid arthritis), and other medication use (such as corticosteroids and proton pump inhibitors). Given the relative infrequency of AFF, randomized designs are not feasible and very large population-based cohorts with complete drug exposure and objective fracture ascertainment are needed.
Kaiser Southern California (KPSC) is very large, about 4.7 million members, providing high power for an analysis. KPSC has an ethnically diverse population and can explore the interesting suggestion of a strong relationship of race. Moreover, unlike almost all previous studies of AFF, there will be detailed information that will include BMD measurements and drug exposure information on most women who have been treated, which will allow for control of confounding by indication. Additional advantages of the Kaiser data not available in previous studies include access to digital radiographs for reported fractures, comprehensive information about previous fractures and other comorbidities, and availability of pre-treatment bone density measurements.
Sampling strategy: In addition to the AFF cases, it would be necessary to obtain a stratified random sample of subjects from Kaiser Southern California (KPSC) 50 years and older at the start of the study. The cohort will include a good number of bisphosphonate users and should provide good covariate overlap between bisphosphonate users and non-users. In addition, random sampling of typical hip fractures (THF) and typical femur fractures (TFF) for analyses comparing bisphosphonate effects across fracture types will be completed.
The comparison group for the specified fracture outcomes is the unique subcohort consisting of the randomly selected individuals from the cohort minus those in the cohort with the outcome of interest. Thus, the comparison group (subcohort) for the AFF case-cohort analyses are all members of the randomly selected cohort, minus any individual with verified AFF.
Advantages: For this proposed study, it would be appropriate to conduct a case-cohort study, as the aim is to achieve the same goal as in cohort studies, but more efficiently, using a sample of the denominators of the exposed and unexposed cohorts (and to ensure inclusion of all AFF cases). The case-cohort design is chosen rather than the stratified cohort design to ensure inclusion of the percentage (%) of AFFs expected to occur among those with no history of bisphosphonate use.
In the case-cohort design, one would include the same cases and classify them as exposed or unexposed (one would start by choosing the cases which is by design a case-control study). Instead of getting exposure information from all individuals constituting the denominators of exposed and unexposed cohorts, you only use a sample of them. The purpose of this sample is to estimate the relative size of exposed and unexposed components of the source population (the proportion of exposed in the source population at the beginning of the cohort).
Disadvantages: Case-crossover designs would control confounding by fixed factors such as pre-treatment bone mineral density (BMD), but are not suitable for exposures with potential carryover effects. Active comparator designs would be problematic and potentially subject to uncontrolled confounding; also, this design does not ensure inclusion of all AFF cases. Propensity score matching could be utilized, which effectively controls confounding if there are no unmeasured confounders, but this would not ensure inclusion of all AFF cases.
Does living in a deprived neighborhood increase colorectal cancer mortality among men?
Population of interest: men, 50 years and older, living in Northern California Bay Area
Study design: cohort study of men 50-75, living in Northern California Bay Area diagnosed with colorectal cancer over a 4-year-period.
Sampling strategy: systematic random sampling; men will be identified from the Northern California Cancer registry and will be chosen at a pre-determined interval (every nth diagnosed case of colorectal cancer).
Advantage: the registry receives a detailed record of patient diagnosis and includes demographic information including address at the time of diagnosis. Patient address can be geocoded to construct a neighborhood deprivation score for each case.
Disadvantage: there may be extra security measures and hoops to jump through to obtain information on patient residence and other identifiable information. Will need to determine a strategy to link patients in the study to mortality records. One approach is to obtain either the death certificate or the cause of death from the National Death Index. This process may be burdensome.
1) Incorporating this sampling strategy will minimize selection bias into the study although Kandola et al point out that the sampling intervals themselves can coincide with systematic variation in the sampling frame. Unclear what they mean by this, clarification would be helpful. Minimizing selection bias will help reduce bias associated with estimating mortality rate ratio. Without a random sampling strategy there is a chance that I would end up selecting for the sickest patients/patients with more aggressive form of colorectal cancer or the healthiest patients.
2) Incorporating this strategy should ideally help reduce bias in estimation of causal effects but there would still be issues with confounding both measured and unmeasured. One challenge in estimating a causal relationship between neighborhood deprivation and cancer mortality is teasing apart neighborhood factors and hospital services. Patients in poor neighborhood tend to have access to hospitals that provide worse services (e.g. less skilled providers, non-optimal treatment options, etc.). Patients in poorer neighborhoods also tend to report at later stages. So, I would need to control for these factors in order to estimate a causal effect that is minimally biased.
My area of interest is the experience of obstetric care, specifically the mistreatment of women in health facilities. It is not terribly useful to get national prevalence estimates of mistreatment, as the rates and types of mistreatment tend to vary by health facility. Obstetric practice patters also tend to vary greatly from hospital to hospital. So, in order to determine facility prevalence of different types of mistreatment I would use a cluster sampling method by dividing the geographic area into sensible units (for instance, by hospital catchment area), and then sample 100 women from each hospital prior to discharge. I would vary the day of data collection randomly, as practice patterns can vary significantly depending on day/night shift and weekend.
Logistically, this approach would be much easier than trying to determine the national sampling frame of women who have recently given birth; these women are more easily accessed at the hospital level. A disadvantage is that women may not want to answer sensitive questions about their care prior to hospital discharge, or it may take time to reflect on the experience in order to develop opinions about the quality of care. For this reason, women could be contacted in 6 months time to see how their opinions have changed. Also, ideally research staff would observe care on labor unit to compare rates of observed mistreatment to reported mistreatment.
Incorporating a sampling strategy that varies by time of day would greatly reduce (though not eliminate) bias of univariate quantities (such as, prevalence of mistreatment).
However, in terms of causal effects several issues arise. For instance, if we wanted to ask this question: Do higher levels of mistreatment predict postpartum mental health problems? I think there would need to be some combination of rates of mistreatment gathered by observers and reported by women. For instance, there may be unmeasured selection bias in that women who accept a certain level of care (however low that standard may be) may select to go to the same hospital. Or more likely, women who seek a higher standard of care select to go to hospitals with higher quality care (ideally also lower rates of mistreatment). Also, women may interpret even bad experiences in a positive light after the baby is born. Thus, the better sampling strategy would occur prenatally (rather than at hospital discharge), so women’s preferences and expectations can be measured prior to delivery and rates of mistreatment could be followed over time.
For my research question, I propose an RCT to assess the effects of a behavioral counseling intervention on reducing the risk of opioid overdose among individuals who inject drugs.
People who inject drugs (PWIDs) can be a difficult population to reach because many PWIDs live in fringe locations, do not access health services and may have a general distrust of government systems. Probabilistic sampling therefore is not an ideal approach to reach this high-risk yet hard-to-reach population.
To enroll PWIDs in this study, I would use respondent driven sampling because it will allow me a greater chance of reaching otherwise hard to reach individuals through their own social networks. We would enroll individual seeds who have large social networks (family, friends etc). We would recruit seeds through convenience sampling at the local syringe exchanges and/or through snowball sampling from other participants/seeds. The seeds would then reach their own networks to encourage individuals to participate in the study.
The advantages of this method is that it would allow us to reach some of the most difficult-to-reach PWIDs who may not have otherwise accessed the study. However, RDS also has its challenges. First, it may introduce selection bias – as seeds and their recruited participants may be similar to one another yet not representative of the larger PWID population. We would try to capture a diverse group of initial seeds (geography, race/ethnicity/gender/age/drug of choice etc.) to try to mitigate the problem. For instance, we would ensure that at least one seed would come from the Bayview/Hunters Point area as that is a location geographically isolated from most PWID services centered in the TL and Soma. However even with directly seed identification selection bias certainly remains a concern.The second problem is that RDS can be very difficult to implement. RDS means you have to rely on the seeds for recruitment which may be challenging if seeds are not recruiting as quickly or as many individuals as they said they would. There is less control over the recruitment process than when a study’s own recruitment team does the recruiting directly.
RQ: who are the asymptomatic cases of malaria?
In malaria endemic countries, especially in low-prevalence settings, malaria infections are detected and treated passively, meaning that only symptomatic cases of malaria that turned out in the health facilities are treated. Asymptomatic cases are therefore never treated and sustain the transmission cycle of malaria. To eradicate malaria it is therefore important to be able to also identify and treat asymptomatic cases.
In low malaria prevalence settings, probabilistic sampling scheme are likely to be highly ineffective. Plus, in low endemic settings, research has identified a couple of high-risk populations which often time are hard-to-reach populations such as migrants workers, forest-goers,… In that context I think a respondent-driven sampling (RDS) strategies could help us capture more effectively the targeted population. The seeds would be symptomatic cases and they would be asked to refer friends or family members sharing some of the known epidemiologic risk factors for malaria (spend time outside at night, bednets owners,…).
In terms of advantages, I think this sampling strategy would give us access to a high-risk population that might otherwise be hidden and enable to find more of those asymptomatic cases. In terms of disadvantages, this RDS strategy is likely to introduce uncontrolled selection bias.
I think incorporating RDS strategy would reduce bias of univariate quantities such as prevalence simply by capturing a more accurate representation of the infection levels. On the other hand, the selection bias induced by RDS is probably hard to account for when making causal inference both in terms of the point estimate but also in terms of the confidence intervals as inference methods for the variability of an estimate highly depends on independence assumptions and others.
It is not unusual to encounter epidemiological or medical studies that use a non-probabilistic sampling scheme. For example, many randomized controlled trials use convenience sampling. Identify a research question of interest to you, the population of interest, and a study design you might use to examine the question. Explain how you might incorporate a sampling strategy into the study design (you might, for example, use respondent-driven sampling to initiate a cohort study). Briefly discuss possible logistical/practical advantages and disadvantages to this plan. Finally, discuss whether you think incorporating the sampling strategy might help (1) reduce bias in the estimation of univariate quantities (such as disease prevalence) and (2) reduce bias in the estimation of causal effects.
Currently, HIV prevalence in low-income countries is often estimated by the use of district health surveys. These surveys are nationally representative, conducted approximately every five years, and have large sample sizes (over 5000 respondents). The sample is usually based on a stratified two-stage cluster design first selecting a number of geographical areas and then using simple random sampling to select a number of households. This technique can lead to a non-representative sample because the participants are chosen in the geographic strata.
An example of a group that is frequently not captured adequately using DHS instruments, is nomadic pastoralists and mobile populations in general (migrants, homeless, etc.). Because these individuals and communities are not sedentary, they are not accurately captured (if captured at all) on the lists of households or individuals that can be used for simple random sampling. Therefore, even if the geographic unit in which they live is selected for the DHS, they may not be included in the sample.
Using respondent-driven sampling (RDS), as in the Firestone article, may be a good way to ensure that this population is captured. Identifying a number of “seeds” as initial respondents, who each identify additional respondents and so on through multiple waves may produce a useful cohort. However, given the close social and familial ties of the communities that travel together, even though RDS adjusts for small to moderate levels of network clustering using post-sampling weights, this approach may lead to significant sampling bias.
Hmmm. I gave this a crack based on other people's code. But considering that I did not see a change in my coefficients, I must not be coding the selection bias correctly. Not sure where it went wrong.
see attached files
Please find attached
Please see attached!
Refer to the attached files, and let me know if you have any questions (I was able to run a bootstrap to obtain the # of observations with p<0.05 for the four different scenarios).
Best,
Stephen
My apologies for posting this later than expected. Please refer to the attached files, and let me know if you have any questions.
Best,
Stephen
See attachment.
Modify your data generation code from last week to match the following causal DAG (I can't figure out how to post the DAG - it's in the attached word document... if anyone else can figure out how to post directly please do so):
Where S1 and S2 are indicators for whether an observation is selected into the analysis (ie, when S1=0 or S2=0, the observations are excluded).
Write out data generating rules to match this causal DAG for at least 2 values S1 (e.g., S1 = 1 for 90% of observations vs S1=1 for 50% of observations) and 2 values for S2. Please be sure that your data generating model entails that U influences X, Y, and S2.
For each, write a simulation model generating data with N=500 individuals following these data generating rules. Report the structural causal model for each variable in the DAG, including the distributions of the error terms. Use the statistical method of your choice to estimate the effect of X on Y in each scenario, assuming that U is unmeasured (ie, your estimate will likely be biased because you cannot account for U).
Now assume the null, that X has no effect on Y. Under each of your four scenarios above, what is:
- The average estimate of the effect of X on Y (ie, if you repeat your chosen statistical analysis across repeated iterations, what is the average value of the parameter estimate for the association between X and Y)?
- the type 1 error rate using an alpha threshold of 0.05 (ie, the percent of iterations under which you would find a statistically significant association between X and Y, using a p<.05 threshold)?
Please find attached my responses. I apologise for the delay in posting
Please see attached.
see attached
I'm pasting everything below, but also attaching a word document that is formatted for clarity.
Specify a hypothesis regarding a particular exposure and outcome and a binary effect modifier including specific measures of association (specify the magnitudes of that association you anticipate: I suggest making everything cross-sectional). Using the software of your choice, generate a population with 1000 people under a causal structure consistent with this hypothesis. Draw a simple random sample 100 individuals from this population and estimate the population average exposure-outcome association and the association stratified by your modifier of interest within this subset. Repeat this 10 times and write the parameter estimates and CI each time.
Hypothesis: Poverty during childhood increases the probability of being obese/overweight in young adulthood (Hernandez and Pressler, 2014), but exposure to poverty increases risk of obesity more for females than for males.
Hernandez DC, Pressler E Accumulation of childhood poverty on young adult overweight or obese status: race/ethnicity and gender disparities J Epidemiol Community Health 2014;68:478-484.
Exposure: Household income (in $10,000) (average from years 2-10)
Outcome: BMI at age 20
Modifier: sex
Modeled effects:
BMI 24.73 (STD 5.13)
Effect in females: -0.25 BMI/$10,000
Effect in males: -0.17 BMI/$10,000
Difference (females = 1): -0.07
/////generating random exposure variable household income and played around to make plausible//////
. drawnorm house_income_norm, n(1000) means(16) sds(4) clear
(obs 1,000)
. generate house_income = house_income_norm - 2.7
. replace house_income = house_income/2.3
(1,000 real changes made)
////generating random error term for outcome BMI////
. generate error_BMI = rnormal(0,1)
////generating random bilevel mediator variable female/////
. generate female = 1
. generate randomsex = runiform()
. sort randomsex
. replace female = 0 in 501/1000
(500 real changes made)
/////generating outcome BMI using regression equation////
. generate BMI = 24.73 + error_BMI*3 - house_income*(0.25+0.07*female)
(1000 missing values generated)
. generate income_female = house_income*female
///running regression models to confirm that the sample reasonably approximates my input parameters and IT DOES!!///
. regress BMI house_income income_female
Source | SS df MS Number of obs = 1,000
-------------+---------------------------------- F(2, 997) = 13.80
Model | 224.315473 2 112.157737 Prob > F = 0.0000
Residual | 8101.32545 997 8.12570256 R-squared = 0.0269
-------------+---------------------------------- Adj R-squared = 0.0250
Total | 8325.64092 999 8.3339749 Root MSE = 2.8506
-------------------------------------------------------------------------------
BMI | Coef. Std. Err. t P>|t| [95% Conf. Interval]
--------------+----------------------------------------------------------------
house_income | -.1805483 .0553692 -3.26 0.001 -.2892018 -.0718947
income_female | -.0918186 .0297546 -3.09 0.002 -.1502075 -.0334297
_cons | 24.92822 .3230297 77.17 0.000 24.29433 25.56212
-------------------------------------------------------------------------------
. regress BMI house_income female income_female
Source | SS df MS Number of obs = 1,000
-------------+---------------------------------- F(3, 996) = 9.20
Model | 224.478177 3 74.8260591 Prob > F = 0.0000
Residual | 8101.16275 996 8.13369754 R-squared = 0.0270
-------------+---------------------------------- Adj R-squared = 0.0240
Total | 8325.64092 999 8.3339749 Root MSE = 2.852
-------------------------------------------------------------------------------
BMI | Coef. Std. Err. t P>|t| [95% Conf. Interval]
--------------+----------------------------------------------------------------
house_income | -.1869639 .0715987 -2.61 0.009 -.3274656 -.0464621
female | -.0920911 .6511219 -0.14 0.888 -1.369819 1.185637
income_female | -.0772236 .107401 -0.72 0.472 -.2879818 .1335346
_cons | 24.96872 .4317816 57.83 0.000 24.12141 25.81603
-------------------------------------------------------------------------------
/////running bootstrap option in stata for repeated sample size estimation/////
. bootstrap, reps(1000) size(100): regress BMI house_income female income_female
(running regress on estimation sample)
Bootstrap replications (1000)
----+--- 1 ---+--- 2 ---+--- 3 ---+--- 4 ---+--- 5
.................................................. 50
.................................................. 100
.................................................. 150
.................................................. 200
.................................................. 250
.................................................. 300
.................................................. 350
.................................................. 400
.................................................. 450
.................................................. 500
.................................................. 550
.................................................. 600
.................................................. 650
.................................................. 700
.................................................. 750
.................................................. 800
.................................................. 850
.................................................. 900
.................................................. 950
.................................................. 1000
Linear regression Number of obs = 1,000
Replications = 1,000
Wald chi2(3) = 2.59
Prob > chi2 = 0.4587
R-squared = 0.0270
Adj R-squared = 0.0240
Root MSE = 2.8520
-------------------------------------------------------------------------------
| Observed Bootstrap Normal-based
BMI | Coef. Std. Err. z P>|z| [95% Conf. Interval]
--------------+----------------------------------------------------------------
house_income | -.1869639 .2423548 -0.77 0.440 -.6619705 .2880428
female | -.0920911 2.223167 -0.04 0.967 -4.449418 4.265236
income_female | -.0772236 .3665592 -0.21 0.833 -.7956664 .6412192
_cons | 24.96872 1.453286 17.18 0.000 22.12033 27.81711
-------------------------------------------------------------------------------
Here it would appear that the effect estimates are quite accurate after 1000 runs, but the confidence in the estimate is low with nothing achieving the level of statistical significance.
Repeat the data set construction, setting the causal effect to the null. Again repeat this 10 times and write the parameter estimate and CI each time (if you figure out how to automate it, run it 1000 times and post the histogram of the parameter estimates and p-values).
. replace BMI = 24.73 + error_BMI*3
(1,000 real changes made)
. bootstrap, reps(1000) size(100): regress BMI house_income female income_female
(running regress on estimation sample)
Bootstrap replications (1000)
----+--- 1 ---+--- 2 ---+--- 3 ---+--- 4 ---+--- 5
.................................................. 50
.................................................. 100
.................................................. 150
.................................................. 200
.................................................. 250
.................................................. 300
.................................................. 350
.................................................. 400
.................................................. 450
.................................................. 500
.................................................. 550
.................................................. 600
.................................................. 650
.................................................. 700
.................................................. 750
.................................................. 800
.................................................. 850
.................................................. 900
.................................................. 950
.................................................. 1000
Linear regression Number of obs = 1,000
Replications = 1,000
Wald chi2(3) = 0.02
Prob > chi2 = 0.9990
R-squared = 0.0003
Adj R-squared = -0.0027
Root MSE = 2.8520
-------------------------------------------------------------------------------
| Observed Bootstrap Normal-based
BMI | Coef. Std. Err. z P>|z| [95% Conf. Interval]
--------------+----------------------------------------------------------------
house_income | -.0169639 .2395083 -0.07 0.944 -.4863915 .4524638
female | -.0920912 2.195122 -0.04 0.967 -4.394451 4.210268
income_female | .0027764 .3578001 0.01 0.994 -.6984988 .7040517
_cons | 24.96872 1.435228 17.40 0.000 22.15572 27.78172
-------------------------------------------------------------------------------
Everything is pretty much null with p>0.9 for all effect estimates
Use your code above and also a canned software command to estimate statistical power to detect the difference in means under the settings below:
*n=100, μ0=.02, μ1=.12, SD=1, α=.05
. power twomeans 0.02 0.12, n(100)
Estimated power for a two-sample means test
t test assuming sd1 = sd2 = sd
Ho: m2 = m1 versus Ha: m2 != m1
Study parameters:
alpha = 0.0500
N = 100
N per group = 50
delta = 0.1000
m1 = 0.0200
m2 = 0.1200
sd = 1.0000
Estimated power:
power = 0.0785
*n=100,μ0=.02, μ1=.12, SD=2, α=.05
. power twomeans 0.02 0.12, sd(2) n(100)
Estimated power for a two-sample means test
t test assuming sd1 = sd2 = sd
Ho: m2 = m1 versus Ha: m2 != m1
Study parameters:
alpha = 0.0500
N = 100
N per group = 50
delta = 0.1000
m1 = 0.0200
m2 = 0.1200
sd = 2.0000
Estimated power:
power = 0.0570
*n=500, μ0=.3, μ1=.3, SD=1, α=.05
. power twomeans 0.3 0.3, n(500)
the control-group mean and the experimental-group mean are equal; this is not allowed
r(198);
I will say that
For each of the 3 settings above, what is the power to detect whether the ratio of the means=1?
Well, as everything above is set up for detecting a ratio that is not null, the power to detect a null ratio is zero. If on the other hand you are running a test for equivalence rather than a test for difference, then there are alternative tests to run, I however couldn’t figure out how to perform this calculation, as everything I tried gave errors. I tried also using an online calculator:http://powerandsamplesize.com/Calculators/Compare-2-Proportions/2-Sample-Equivalence
n=100, μ0=.02, μ1=.12, SD=1, α=.05
ð power = 0.763
*n=100,μ0=.02, μ1=.12, SD=2, α=.05
Please see attached
Response with figures attached:
Please identify a quantitative research article evaluating mediation in your field and provide the citation.
“The impact of human development on individual health: a causal mediation analysis examining pathways through education and body mass index” by A.Wang and O.A.Arah
What is the primary discipline of the authors?
The authors are both professors at UCLA in the department of epidemiology and Health Policy research.
Draw a DAG representing the implicit or explicit causal model explored in this paper (you do not need to post your DAG, but we will try to discuss in class).
Attached
What is the exposure of interest?
Human development Index (HDI), at the country level.
What is the outcome of interest?
Health, summarized by a factor extracted from the answers to 16 questions in 8 health domains. The factor is at the country level while questions were at the individual level.
What is the hypothesized mediator of interest and how is it measured?
Two mediators were considered – Education and BMI. Individual education was measured by the years of schooling and averaged at the country level. Individual BMI was calculated and averaged at the country level.
Describe the modeling approach and briefly report the estimated total, direct, and indirect effects (if these are reported).
The authors simply used regression to calculate the estimated total, direct and indirect effects (subdivided per gender): Attached
If the direct effect is reported, would you describe this as a natural direct effect, a controlled direct effect, or something else?
I think the authors reported the controlled direct effect.
Do you think there is potential measurement error in the mediator and how would that affect the results?
The exposure, HDI is determined at the country level and, although probably over-simplifying the construct it is trying to measure leaves little room to measurement error. The mediator on the other hand is obtained by self-reported years of education and measurement error could be substantial. If that is the case, the indirect effect via the mediator might be underestimated a little be.
Do you think there are unmeasured confounders of the mediator-outcome association and how would that affect the results of the mediation analysis?
Certainly but even more concerning, the authors considered, in their analysis of mediation via BMI, education as a confounder of the mediator-outcome relationship, while also being a potential mediator of the exposure-outcome relation. That DAG prevents the estimation of the direct effect with simple regression and requires more advanced estimation methods such as marginal structural equations by the use of probability weights.
Paper: Peasant C, Sullivan TP, Weiss NH, Martinez I, Meyer JP. Beyond the Syndemic: Condom Negotiation and Use among Women Experiencing Partner Violence. AIDS Care. 2017. April. 29(4): 516-523.
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5291821/pdf/nihms781142.pdf
Primary discipline of authors: AIDS research, preventive medicine and community research, psychiatry, school of medicine, epidemiology
DAG:
Indirect Effect (a x b) = -6.41*
Syndemic Severity--------> Condom Use
\ c= -1.22 ^
a=-8.33* \ / b=0.77***
\ /
v /
Condom Negotiation
Exposure of Interest: Syndemic Severity of Intimate Partner Violence (scale 0-4) comprised of problematic drug use, hazardous drinking, depression, and posttraumatic stress disorder
Outcome of Interest: Condom Use – percentage of total sexual episodes in which a condom was worn.
Mediator of Interest (how measured): Condom Negotiation was defined as the percentage of instances of sexual intercourse in which the woman asked their partner to wear a condom.
Modeling approach (total, direct and indirect effects): They used PROCESS, ordinary least squares regression and bootstrapping to estimate the direct, indirect effects of the hypothesized associations AFTER they controlled for the covariates and the confounder of fear of condom negotiation. Total effect was -7.63. (SE=2.33, 95% CI: ), Indirect effect was -6.41 (SE = 2.13, 95% CI: -10.71, -2.31) and the Direct effect was -1.22, SE = 2.11, 95% CI: -5.40, 2.99)
Potential for measurement error in mediator: The mediator is being self-reported instances of negotiating condom use, and this is highly susceptible to reporting bias. It is likely socially desirable to report that you attempted to negotiate condom use and to blame the partner for failing to use condoms. Alternatively, women might be hesitant to admit that they asked, but failed to receive consensus on condom use as exemplary of their own lack of agency. It may well be that condom negotiation is under/over reported, but it is not clear in what direction, but it would likely bias the results towards the null as long as they are non-differentially misclassified.
Unmeasured confounders of the mediator-outcome association: They did include fear of condom negotiation, which was my primary thought for confounders of the condom negotiation-condom use association. Other factors might be having the condoms available, so if you know they aren’t available the women probably wouldn’t even ask to begin with.
Please identify a quantitative research article evaluating mediation in your field and provide the citation.
Jackson JW (1,2), VanderWeele TJ(2,3), Blacker D(2,4), Schneeweiss S(1,2). Mediators of first- versus second-generation antipsychotic- related mortality in older adults. Epidemiology. 2015 September; 26(5): 700–709.
What is the primary discipline of the authors?
(1)Brigham and Women’s Hospital and Harvard Medical School, Division of Pharmacoepidemiology, Department of Medicine, Boston, Massachusetts, 02120
(2)Harvard School of Public Health, Department of Epidemiology, Boston, Massachusetts, 02115 (3)Harvard School of Public Health, Department of Biostatistics, Boston, Massachusetts, 02115
(4)Massachusetts General Hospital & Harvard Medical School, Gerontology Research Unit, Department of Psychiatry, Boston, MA 02114
What is the exposure of interest?
New user of first generation antipsychotic versus new user of second generation antipsychotic (binary). In other words, the authors defined exposure as a binary variable comparing first-generation to second-generation antipsychotic initiation (reference).
What is the outcome of interest?
Mortality with 180 days (binary)
What is the hypothesized mediator of interest and how is it measured?
Mediators selected based on previous literature. Authors compared the separate and combined contributions of medical events previously studied in the literature: stroke, ventricular arrhythmia, myocardial infarction, venous thromboembolism, pneumonia, bacterial infection (other than pneumonia), and hip fracture. These were defined as binary variables indicating their occurrence between the index prescription date (inclusive) and the end of follow up (180 days) or death, and were classified using diagnostic and procedure codes based on the International Classification of Diseases. Death during follow-up was defined as a binary variable.
Describe the modeling approach and briefly report the estimated total, direct, and indirect effects (if these are reported).
Using causal mediation analysis, the authors sought to decompose the total effect of antipsychotic- type (exposure) on mortality (outcome) into natural direct and indirect effects through various individual medical events (mediators) on the risk ratio scale, and the proportion of the total effect mediated by each medical event on the risk difference scale.
Using logistic regression, the authors estimated the crude risk at 180 days and the covariate-adjusted relative risk comparing first- and second- generation antipsychotic use. Within groups defined by antipsychotic type, the authors used g- computation (i.e. model-based standardization with 95% confidence intervals (CI) obtained. The authors then used a regression-based approach for causal mediation analysis involving a binary exposure, mediator, and outcome to estimate crude and adjusted direct and indirect effects of antipsychotic-type on 180-day mortality through each medical event on the risk-ratio scale. For each medical event, the authors estimated two models: (1) a logistic regression model for the mediator’s occurrence conditional on antipsychotic type and main effects for all baseline covariates, and (2) a Poisson regression model for mortality conditional on antipsychotic type, the mediator’s occurrence, a product term for their interaction, and the same baseline covariates. The parameter estimates from these models were combined to estimate the direct and indirect effect risk ratios using closed form estimators, which were then used to compute the proportion mediated on the risk difference scale.
In bias analyses the proportion mediated ranged from 6% to 16% for stroke, 3% to 9% for ventricular arrhythmia, 3% to 11% for myocardial infarction, 0% venous thromboembolism, 3% to 9% for pneumonia, 0% to 1% for other bacterial infection, and 1% to 3% for hip fracture.
The crude and covariate-adjusted analyses accounting for exposure-mediator interaction were similar to those ignoring such interaction. Crude indirect effects were close to the null and were lower after covariate adjustment. Although covariate adjustment attenuated the total effect from RR=1.23 (95%CI 1.16 to 1.32) to RR=1.14 (95%CI 1.06 to 1.22), the direct effects in crude and adjusted analyses were similar to the total effect in both cases (the proportion mediated was highest for stroke at 5%).
If the direct effect is reported, would you describe this as a natural direct effect, a controlled direct effect, or something else?
I would describe the direct effect reported as a natural direct effect.
Do you think there is potential measurement error in the mediator and how would that affect the results?
To avoid false positive medical events during follow-up, the authors used restrictive classification algorithms with high positive-predictive values. False negatives may occur more often among those who die, especially with events where pre-hospital mortality is common (e.g. ventricular arrhythmia) or do-not-resuscitate orders dictate whether aggressive treatment is pursued (e.g. pneumonia). To avoid misclassification, the authors performed a bias analysis to explore how results would change under various scenarios of non- differential and differential misclassification.
Do you think there are unmeasured confounders of the mediator-outcome association and how would that affect the results of the mediation analysis?
The authors acknowledged that the approach allows for possible exposure-mediator interactions, it also requires that all confounders of the exposure-outcome, exposure-mediator, and mediator-outcome relationships are measured and adjusted for, and prohibits the existence of any mediator- outcome confounder that is itself affected by exposure.
The authors stated how residual confounding may have influenced their results. Delirium is a strong predictor of mortality in older adults and when it is detected, it is frequently treated with haloperidol (a widely used first-generation antipsychotic) which could lead to exposure- mediator or exposure-outcome confounding. Delirium is also poorly captured in claims data, so residual confounding at baseline could bias the total and indirect effects upwards. Moreover, unmeasured behavioral risk factors (e.g. smoking, physical activity) could also bias the indirect effects for several mediators through mediator- outcome confounding. The authors also stated how their results may be subject to a subtle form of length bias, where associations between antipsychotic-type and medical events are under-estimated
Do you have any critiques of the paper?
This paper was relatively well written, and the authors clearly addressed any misclassification or confounding that may have been present during the analysis.
Please identify a quantitative research article evaluating mediation in your field and provide the citation.
Kahler, Christopher W., Tao Liu, Patricia A. Cioe, Vaughn Bryant, Megan M. Pinkston, Erna M. Kojic, Nur Onen, et al. 2016. “Direct and Indirect Effects of Heavy Alcohol Use on Clinical Outcomes in a Longitudinal Study of HIV Patients on ART.” AIDS and Behavior, July. doi:10.1007/s10461-016-1474-y.
What is the primary discipline of the authors?
Christopher W. Kahler – Center for Alcohol and Addiction Studies, Brown University School of Public Health; PhD in clinical psychology.
Tao Liu – Center for Statistical Sciences, Brown University School of Public Health; PhD in biostatistics.
Patirica A. Cioe – Center for Alcohol and Addiction Studies, Brown University School of Public Health; PhD in nursing.
Vaughn Bryant – Department of Clinical and Health Psychology, University of Florida; graduate student.
Megan M. Pinkston – Department of Psychiatry and Human Behavior, Alpert Medical School of Brown University; PhD in clinical health psychology.
Erna M. Kojic – Department of Infectious Disease, Brown University; MD (infectious disease).
Nur Onen – Washington University School of Medicine; MD (infectious disease).
Jason V. Baker – Division of Infectious Diseases, University of Minnesota, Minneapolis; MD (infectious disease).
John Hammer – Denver Infectious Disease Consultants; MD (infectious disease).
John T. Brooks – Division of HIV/AIDS Prevention, CDC; MD (infectious disease).
Pragna Patel – Division of Global Health Protection, CDC; MD (infectious disease), MPH, DTM&H, AAHIVS.
What is the exposure of interest?
Past 30-day frequency of heavy drinking (consuming 5+ drinks on one occasion)
What is the outcome of interest?
HIV-related (detectable viral load and CD4+ T cell count) and non-HIV-related (hemoglobin and biomarkers of kidney function and liver fibrosis) clinical outcomes.
What is the hypothesized mediator of interest and how is it measured?
The mediator of interest is ART adherence. It is measured by self-report (participants were asked the number of missed doses in the past 3 days), and then dichotomized to either no missed doses or at least one missed dose since missing more than one dose was rare.
Describe the modeling approach and briefly report the estimated total, direct, and indirect effects (if these are reported).
The authors used structural equation models to model the exposure-mediator-outcome relationship, including confounding variables. They adjusted for time-dependent and time-independent confounders and clinical outcomes at the last visit in all models. Heavy drinking and ART adherence were allowed to have an interacting effect on the outcomes. Outcomes were modeled with either a log-linear model (for the dichotomized viral load outcome), or a linear model (for the continuous outcomes).
Just listing results for the HIV-related clinical outcomes (with 95% CIs) for simplicity:
Detectable viral load total effect 1.16 (1.00, 1.31), direct effect 1.13 (0.99, 1.27), indirect effect 1.03 (1.00, 1.05).
CD4+ T cell count total effect -11.33 (-17.80, -4.85), direct effect -10.61 (-17.10, -4.12), indirect effect -0.72 (-1.28, -0.15).
If the direct effect is reported, would you describe this as a natural direct effect, a controlled direct effect, or something else?
Natural direct effect: it is the effect of heavy drinking on health outcomes if ART adherence was kept at the “natural” level in the absence of heavy drinking (i.e., what adherence would be expected to be in an individual had no heavy drinking).
Do you think there is potential measurement error in the mediator and how would that affect the results?
Yes, I think there are two main potential reasons for measurement error in the mediator: people may not accurately report their ART adherence due to social desirability bias – there is clearly a “right” answer to the question about whether someone has skipped medication, and people may not respond honestly due to embarrassment; and/or they may not accurately recall whether they skipped medication. Since researchers were only asking about the past 3 days, this seems less likely but still possible. With social desirability bias, people would tend to over-report adherence, which would tend to increase the indirect effect. With inaccurate recall, it depends on whether people were more likely to over-report adherence (increase the indirect effect), more likely to under-report adherence (increase the direct effect), or if people were neither more likely to report one or the other (which I think would weaken the association between mediator and the outcome, and so would also increase the direct effect?).
Do you think there are unmeasured confounders of the mediator-outcome association and how would that affect the results of the mediation analysis?
I think there are at least a few unmeasured potential confounders – social support, health insurance status, and other comorbidities (HBV infection, HCV infection, and depression were included which are arguably the most important, but other comorbidities could also confound the relationship).
Do you have any critiques of the paper?
The authors did a good job of fulfilling the first three assumptions for controlling confounding in mediation analysis, but the fourth assumption is that there should be no mediator-outcome confounder that is itself affected by the exposure. I think it is certainly possible that heavy drinking could affect the time-dependent confounders employment, drug use, and depression. Given the complexity of the system, it also would have been beneficial to perform a sensitivity analysis for unmeasured confounding. Finally, it would have been helpful if the authors had been more specific/provided more description about how they conducted the mediation analysis itself.
Please identify a quantitative research article evaluating mediation in your field and provide the citation.
Bucagu, M., Bizimana, J. de D., Muganda, J., & Humblet, C. P. (2013). Socio-economic, clinical and biological risk factors for mother - to – child transmission of HIV-1 in Muhima health centre (Rwanda): a prospective cohort study. Archives of Public Health, 71(1), 4. https://ucsf.idm.oclc.org/login?url=http://doi.org/10.1186/0778-7367-71-4
What is the primary discipline of the authors?
Public Health, Statistics, Applied Mathematics, and Medicine (Obstetrics & Gynecology and Maternal, Newborn and Child Health)
Draw a DAG representing the implicit or explicit causal model explored in this paper (you do not need to post your DAG, but we will try to discuss in class).
What is the exposure of interest?
Marital status of mother
What is the outcome of interest?
Infant HIV-1 status at 6 weeks of age
What is the hypothesized mediator of interest and how is it measured?
Disclosure of HIV status to partner.
All data was collected from the women themselves and logbooks in the health center using a structured questionnaire.
Describe the modeling approach and briefly report the estimated total, direct, and indirect effects (if these are reported).
A series of multivariable logistic regression models were used to quantify the relationships between 1) marital status and disclosure, 2) marital status and infant HIV-status, and 3) disclosure and infant HIV-status.
Marital status has shown no effect on infant HIV status when disclosure was controlled. The risk of infant HIV infection at 6 weeks was higher in unmarried (adjusted OR of 1.42 with a 95%CI 0.29-4.08) vs. married (reference) women with undisclosed HIV status as the mediator.
If the direct effect is reported, would you describe this as a natural direct effect, a controlled direct effect, or something else?
The direct effect is not reported quantitatively, but it is stated that the independent variable (martial status) has no effect on the dependent variable (infant HIV-status), therefore this is a controlled direct effect.
Do you think there is potential measurement error in the mediator and how would that affect the results?
Given that the mediator is binary (married vs. unmarried), I don’t believe there is much potential for measurement error. The only issue may be the self-reported nature of this information.
Do you think there are unmeasured confounders of the mediator-outcome association and how would that affect the results of the mediation analysis?
I’m not totally convinced that disclosure is the most important variable to consider as a mediator in this situation because since the mother is already HIV+, this would have no biological effect on the transmission of HIV to her infant. However, disclosure might be an indicator of health seeking behavior.
Do you have any critiques of the paper?
Given that all the predictors of MTCT were entered into the model and included if they had a p-value of <0.05 using the Hosmer and Lemeshow test, and they considered a wide variety of predictors, it feels like the authors were doing a bit of fishing.
Please identify a quantitative research article evaluating mediation in your field and provide the citation.
Li, R., Daniel, R., & Rachet, B. (2016). How much do tumor stage and treatment explain socioeconomic inequalities in breast cancer survival? Applying causal mediation analysis to population-based data. European journal of epidemiology, 31(6), 603-611.
What is the primary discipline of the authors?
Epidemiology and Biostatistics
What is the exposure of interest?
Deprivation
What is the outcome of interest?
Survival status (dead vs. alive)
What is the hypothesized mediator of interest and how is it measured?
1) Stage at diagnosis
2) Treatment
Stage at diagnosis was obtained from cancer registry. Each patient was assigned one of four categories based TNM (tumor size, lymph nodes affected, and distant metastasis) cancer staging.
Surgical treatment was retrieved from national hospital dataset. Treatment codes were categorized and then dichotomized into major treatment (axillary dissection or axillary nodal procedures, breast conserving surgery, mastectomy, and plastic surgery) and minor or no surgery (other surgical procedures and none).
Describe the modeling approach and briefly report the estimated total, direct, and indirect effects (if these are reported).
Author employed g-computation formula using Monte Carlo simulation to allow for interactions and other nonlinearities. They conducted three analyses to investigate the mediating role of stage and treatment. They first estimate the proportion of the effects of deprivation on survival that was mediated by difference in stage at diagnosis. Then they estimated the proportion of the effect of deprivation on death that was mediated by differences in treatment (here stage at diagnosis was considered to be a confounder). In the third analysis they estimated the proportion of the effect of deprivation on treatment that is mediated by differential treatment. They stratified the outcome (dead vs. alive) according to time since diagnosis: 6 months, 1 year (conditioning on 6-month survival), 3 years (conditioning on 1-year survival), and 5 years (conditioning on 3-year survival). Analyses were performed separately on each of these four binary survival outcomes. They used multinomial regression to model stage at diagnosis and logistic regression was used for treatment and survival status. Age at diagnosis was modeled as cubic splines. To handle missing data, the authors used single stochastic imputation within the f-computation. Confounders adjusted for include effect of region, year of diagnosis.
Total causal effect: conditioning on 6-months was 2.77(2.17, 3.53) but conditioning on 3-years 1.67(1.39, 2.00)
Indirect effect of stage: conditioning on 6-months 1.43(1.27, 1.67) and conditioning on 5-years 1.08 (1.00, 1.61). Stage accounted for 35% (23, 24%) of the total effect of deprivation at 6-monthd and 30% (5, 54%) at 1 year.
Indirect effect of treatment: The authors did not find evidence for effect mediated by treatment
If the direct effect is reported, would you describe this as a natural direct effect, a controlled direct effect, or something else?
Direct effect was not reported
Do you think there is potential measurement error in the mediator and how would that affect the results?
Authors discuss effects of misclassification of both stage at diagnosis and treatment. The authors hypothesize that deprived cancer patients may be more likely to be misdiagnosis because they are likely to be treat by non-specialized enters and by less experienced surgeons. To test this hypothesis, they tested they tested what would happen if 10%, 30%, and 50% of the most deprived patients were under-staged. They concluded that up to 30% of the most deprived patients would have to be systematically under-staged compared to 0% in the most affluent group and that this is highly improbable and unsupported by the literature.
Setting aside the authors argument, if more deprived patients were in fact misdiagnosed so that they were assigned a stage lower cancer sage, the direct effect between deprivation and survival would be attenuated.
The authors used information on surgery alone since 1) the information on radiation and chemo therapy was too poor to be used 2) surgery is generally standard care in addition to the other two treatment options. The authors hypothesized that if there was under-estimation of treatment (i.e. surgery proportion) it is likely to affect the more affluent patients. They conducted sensitivity analysis to investigate how this misclassification may affect the effect estimates. They found no evidence of differential treatment on cancer survival the top 4 deprivation groups. They did find that treatment mediated about 30-40% of the differential mortality between the most deprived and least deprived patients. They concluded that surgical information is likely to be missed completely at random and therefore was unlikely to bias their results. Also, when the authors re-categorized treatment into 4 categories (originally it was a dichotomized as major or minor) the effect estimates remained unchanged.
Do you think there are unmeasured confounders of the mediator-outcome association and how would that affect the results of the mediation analysis?
The authors acknowledge that co-morbidity may be an important confounder not controlled for. This would lead to over-estimation of the beneficial effects of major surgery on mortality.
The authors further argue that since they found no mediating effect of treatment, co-morbidity would change the overall estimate only if stage and treatment were misclassified.
Do you have any critiques of the paper?
I think this paper was relatively well written. The authors tested various plausible models and addressed whether they believed there were systematic errors and how that would impact the effect estimates.
Please identify a quantitative research article evaluating mediation in your field and provide the citation.
Air pollution and gene-specific methylation in the Normative Aging Study
What is the primary discipline of the authors?
Environmental Health, Harvard
Marie-abele Bind*,
Johanna Lepeule,
Antonella Zanobetti,
andrea Baccarelli,
Petros Koutrakis,
Joel Schwartz
Brent Coull*,
* + Biostatistics, Harvard
Antonio Gasparrini, Medical Statistics, London HTM
Letizia Tarantini, Molecular and Genetic Epidemiology, U of Milan
Pantel Vokonas, VA Boston Healthcare System
Draw a DAG representing the implicit or explicit causal model explored in this paper (you do not need to post your DAG, but we will try to discuss in class).
What is the exposure of interest?
Air pollution
What is the outcome of interest?
Cardiovascular-related biomarkers (inflammatory and coagulation markers)
What is the hypothesized mediator of interest and how is it measured?
Gene-specific methylation
Describe the modeling approach and briefly report the estimated total, direct, and indirect effects (if these are reported).
Investigators fitted two linear mixed-effects models to determine the mediated effect of air pollution on cardiovascular-related biomarkers through a change in gene-specific methylation:
The effect of mediation was reported as the standard deviation increase from the mean. It is given by the product formula .
If the direct effect is reported, would you describe this as a natural direct effect, a controlled direct effect, or something else?
I think this is a natural direct effect because we are not fixing the mediator to some value (that would be a controlled direct effect).
Do you think there is potential measurement error in the mediator and how would that affect the results?
The participant’s blood sample was collected at every visit and DNA was isolated to assess gene-specific DNA methylation using highly quantitative methods based on bisulfite PCR pyrosequencing. Measurement error in this process was probably pretty minimal.
Do you think there are unmeasured confounders of the mediator-outcome association and how would that affect the results of the mediation analysis?
An unmeasured confounder of the mediator-outcome association would have to affect gene-specific methylation and coagulation and inflammation factors. I cannot come up with other confounders, but I do think there may be some residual confounding because race, BMI, and smoking status are imperfect measures.
Do you have any critiques of the paper?
Maybe stating more specifically the type of mediation analysis (the controlled direct effect of natural direct effect)?
Please identify a quantitative research article evaluating mediation in your field and provide the citation.
Lu, Y., Hajifathalian, K., Rimm, E. B., Ezzati, M., & Danaei, G. (2015). Mediators of the effect of body mass index on coronary heart disease: decomposing direct and indirect effects. Epidemiology, 26(2), 153-162.
What is the primary discipline of the authors?
Lu, Y: doctoral student at Harvard School of Public Health. Research focuses on mediation analysis of effects of overweight/obesity on cardiovascular diseases.
Hajifathalian, K, MD: Gastroenterologist in Iran. Visiting Scientist at Harvard. His research is focused on the global distribution of CHD risks by country, evaluating the effect of antihypertensive treatment on blood pressure trend in U.S., advanced mediation analysis of effects of obesity on cardiovascular outcomes, and updating the global estimates of distribution of metabolic risk factors of cardiovascular disease.
Rimm, Eric.B., ScD: Professor at Harvard. Studies modifiable lifestyle choices (e.g. diet and physical activity) in relation to cardiovascular disease as well as the translation of these findings into public health interventions that are effective for schoolchildren, adults and the food insecure.
Ezzati, M: Professor at Imperial College London. Does population health and environmental health with focus on preventable risk factors.
Danaei, Goodarz: Professor at Harvard School of Public Health. Estimates the effect of risk factors and preventive interventions on non-communicable disease incidence and mortality at the population level.
Draw a DAG representing the implicit or explicit causal model explored in this paper (you do not need to post your DAG, but we will try to discuss in class).
From the article:
(See DAG in article)
What is the exposure of interest? Body Mass Index
What is the outcome of interest? Coronary Heart Disease
What is the hypothesized mediator of interest and how is it measured?
|
Mediator |
Measurement method |
|
Blood pressure (systolic) |
Investigators pooled data from 9 cardiovascular cohort studies. In most, systolic blood pressure was measured 2-3 times at the arm, after a standard resting period. Values were averaged into one estimate. |
Other mediators of interest included serum cholesterol, blood glucose, fibrinogen and inflammatory biomarkers. For simplicity, I focused only on blood pressure as the mediator of interest.
Describe the modeling approach and briefly report the estimated total, direct, and indirect effects (if these are reported).
Modeling approach: they used the 2-stage regression method proposed by VanderWeele to estimate direct and indirect effects (VanderWeele TJ, Epidemiology. 2011) which has a number of assumptions in order to provide valid estimates of direct and indirect effects. These include having a rare outcome, no unmeasured confounding, and no model misspecification. The first is a linear regression model for blood pressure conditional on BMI and known and measured confounder(s). The second is a Cox proportional hazards regression model of the risk of CHD on BMI, blood pressure, and BMI-blood pressure interaction, and confounders.
The natural direct effects are then estimated using the coefficients from the two regressions.
Formula for calculating % of excess relative risk mediated: (HRTE – HRNDE)/(HRTE - 1)
HRTE: total effect hazard ratio = HRTE = HRNDE x HRNIE
NDE = natural direct effect
NIE = natural indirect effect
TE = total effect
Estimated total: Those in overweight category had 1.22 times the rate (hazard) of CHD compared to normal weight (HR 1.22 [1.14-1.3]); and the obese (BMI > 30) had 1.42 times the rate (hazard) of CHD (HR 1.42 [1.25-1.6]) compared to normal weight.
Direct: HR 1.16 [1.09-1.24] among the overweight, and HR 1.28 [1.15-1.43] among the obese.
Indirect: Blood pressure was the primary mediator of the overweight-CHD association explaining 28% of the excess relative risk among overweight, and 37% for obese; indirect-effect HR of 1.06 [1.03-1.08] for overweight, and HR of 1.13 [1.07-1.19] for obese.
If the direct effect is reported, would you describe this as a natural direct effect, a controlled direct effect, or something else?
Natural Direct effect. They aimed to decompose the total effect into direct and indirect effects. This cannot be done with controlled direct effects.
Do you think there is potential measurement error in the mediator and how would that affect the results?
Blood pressure was measured as an average of multiple measurements at the baseline visit. However, measuring resting blood pressure at one visit is unlikely to provide all of the true variance for each participant. If this measurement error is independent and non-differential with respect to true values of other measurements, this is likely to bias their results towards the null.
The authors actually took this into consideration, and in their sensitivity analyses, they assessed the impact of measurement error by calibrating the regression coefficients, with the assumption that baseline blood pressure only explained 65% of true inter-individual variability. When they took into account this presumed measurement error, they found the percentage of excess relative risk increased (for all mediators considered) from 47% to 69% for overweight, and from 52% to 73% for the obese.
Do you think there are unmeasured confounders of the mediator-outcome association and how would that affect the results of the mediation analysis?
Yes. Unmeasured confounders are likely, including genetic traits. In addition, there is likely residual confounding from not measuring dietary intake or physical activity very well. It is hard to predict how the unmeasured confounders would affect the results; it could bias the results in either direction depending on the causal relationships.
Do you have any critiques of the paper?
One important critique is that exposure (BMI), mediators and confounders were all measured at baseline. Although it’s hard to imagine a scenario where we may observe reverse causation from blood pressure to BMI, it is possible that someone’s blood pressure at time t minus x may have affected BMI at baseline, as well as blood pressure at baseline. Temporality would need to be absolutely met in order to estimate a causal indirect and direct estimate of this relationship.
In addition, in a letter to the editor, Fritz et al. critiqued this paper. Specifically, they highlight that the effects of BMI on CHD decreases with age and recommended including this interaction effect in their analyses.
Please identify a quantitative research article evaluating mediation in your field and provide the citation.
Fotso, Jean-Christophe, Alex C. Ezeh, and Hildah Essendi. "Maternal health in resource-poor urban settings: how does women's autonomy influence the utilization of obstetric care services?" Reproductive Health 6.1 (2009): DOI: 10.1186/1742-4755-6-9
What is the primary discipline of the authors? Reproductive health researchers in academia and at the Population Counsel
What is the exposure of interest? women's autonomy
What is the outcome of interest? place of delivery in resource-poor urban settings
What is the hypothesized mediator of interest and how is it measured? Women's education coded as none; primary; and secondary or higher
Describe the modeling approach and briefly report the estimated total, direct, and indirect effects (if these are reported).
First, multivariate models are used to identify factors associated with place of delivery and quantify their net effects. The effect of women's overall autonomy was insignificant and counter-intuitive, with low autonomy women the least likely to deliver in appropriate health facilities. The effect of household wealth on the choice of place of delivery appeared to be strong and in the expected direction (coefficient = 0.438 for less poor households).
Second, interaction models are examined to test the extent to which the effects of women's autonomy on choice of place of delivery vary by household wealth. The interaction between wealth and autonomy was largest in the poorest with middle and high autonomy. Being in the middle and higher wealth groups and having high autonomy was slightly predictive of facility birth.
Third, we examine the potential mediating effect of women's autonomy on the link between education and place of delivery by adjusting the effect of education for autonomy and assessing the change in the coefficients. The change in coefficient when adjusting for education was significant in both crude and full models.
If the direct effect is reported, would you describe this as a natural direct effect, a controlled direct effect, or something else? I think the effect reported is the controlled direct effect because the mediator is fixed.
Do you think there is potential measurement error in the mediator and how would that affect the results? There could be measurement error because the quality of the education may not be comparable between communities and therefore the levels used in the analysis may not be meaningful.
Do you think there are unmeasured confounders of the mediator-outcome association and how would that affect the results of the mediation analysis? I think use of contraception may be an interesting addition to this question because of the role education has in that decision. It may mirror the understanding gained by including number of antenatal visits – capturing something about access to the system and history of engaging with it.
Do you have any critiques of the paper? The whole sample were women living in slums. Being able to compare to a different population would have increased the understanding of autonomy and health outcomes. I wonder about the relationship between autonomy over health decisions in the setting of little autonomy over economic position. Though autonomy was looked from two different perspectives which may take care of this concern. And I do appreciate the contribution of this study to the idea that these are the women knowledge from this research can impact so focusing on them from this stage is important.
Factors That Mediate Racial/Ethnic Disparities in US Fetal Death Rates
Scott A. Lorch, MD, MSCE, Charlan D. Kroelinger, PhD, Corinne Ahlberg, MS, and Wanda D. Barfield, MD, MPH
What is the primary discipline of the authors?
The first author is a pediatrician, the others are in the field of reproductive health.
Draw a DAG representing the implicit or explicit causal model explored in this paper (you do not need to post your DAG, but we will try to discuss in class).
See figure one in the paper ;)
What is the exposure of interest?
Race/ethnicity
What is the outcome of interest?
Risk of fetal death
What is the hypothesized mediator of interest and how is it measured?
SES: maternal insurance status, education, trimester started prenatal care, age.
Pre-existing morbid conditions: ICD-9 codes listed on discharge records
Anetpartum/Intrapartum conditions: ICD-9 codes
Fetal factors: ICD-9 codes
Delivery hospital: Only included as a mediator when the hospital accounted for more than 15% of change in effect. Otherwise was treated as confounder.
Describe the modeling approach and briefly report the estimated total, direct, and indirect effects (if these are reported).
They used Baron and Kenny’s framework of:
The measured the unadjusted association for #1 and #2 above using logistic regression and then used chi2 test to assess the association between race/ethnicity and each set of potential mediating factors. To test #3 and #4 the use sequential logistic regression models and added each group of mediating factors in the temporal order that they appear in pregnancy. For #4 they reported the percentage of the fetal death disparity in each of the mediated factors.
Unadjusted total effect: Blacks OR 2.24, Hispanics OR 1.37, Asians, 1.18 (reference, whites)
Couldn’t find the direct effect? It may be that when the added SES, AP/IP complications, and fetal factors the entire direct effect for black race disappeared (OR 1.04 95%CI 0.95-1.14). Delivery hospital did not further change the OR. So, I think they are saying that after adding these mediating factors, there is no direct effect of race/ethnicity on risk of fetal death.
Indirect effects: Blacks 49% due to fetal factors; Hispanics 35.8% due to SES factors; Asians 62% due to antepartum and intrapartum factors.
If the direct effect is reported, would you describe this as a natural direct effect, a controlled direct effect, or something else?
Not reported as a direct effect, or there appears to be no direct effect left after the mediating factors are all added.
Do you think there is potential measurement error in the mediator and how would that affect the results?
There could be the potential for measurement error whenever using birth certificate data. It may be possible that ICD-9 coding for cases of fetal death is more intensive, so more complications are listed because these cases “need explanation” compared to cases that did not have a fetal death (misclassification of the mediators). Per Vander-Weele, measurement error and misclassification of the mediator will weaken the relationship between the mediator and the outcome. If this is true, then the investigators could have underestimate the relationship between the mediators and the outcomes in this case.
Do you think there are unmeasured confounders of the mediator-outcome association and how would that affect the results of the mediation analysis?
They did not conduct a sensitivity analysis to investigate unmeasured confounding, so I do believe there could be unmeasured confounding. I am especially interested with how they treated SES as a mediator in this situation and not a confounder, given the known confounding of race with maternal insurance status, maternal education, and starting trimester of prenatal care.
Do you have any critiques of the paper?
The concern I have with this paper relates to Vander-Weele’s discussion of multiple mediators, who says that you cannot use this “summed multiple mediator” method used by these investigators when the mediators affect one another, or if there are interaction between the effects of the mediators on the outcomes. Vander-Weele says you can use regression techniques (which these authors did use) to examine an entire set of mediators. The set of mediators here seem very likely to affect each other, for instance pre-existing high blood pressure can increase the chance of intrapartum complications. I am not sure how adequately the authors dealt with mediator-to-mediator interactions.
Please identify a quantitative research article evaluating mediation in your field and provide the citation.
Nandi A, Galea S, Ahern J, Bucciarelli A, Vlahov D, Tardiff K. What explains the association between neighborhood-level income inequality and the risk of fatal overdose in New York City? Soc Sc & Med. 2006 (63);662-674.
What is the primary discipline of the authors?
Arijit Nandi: Dept of Epidemiology, Johns Hopkins Bloomberg School of Public Health, Baltimore, MD, and Center for Urban Epidemiologic Studies, The New York Academy of Medicine, New York, NY
Sandro Galea: Center for Urban Epidemiologic Studies, The New York Academy of Medicine, New York, NY, and Dept of Epi, U Michigan School of Public Health, Ann Arbor MI
Jennifer Ahern: Center for Urban Epidemiologic Studies, The New York Academy of Medicine, New York, NY, and Dept of Epi, U Michigan School of Public Health, Ann Arbor MI, and Dept of Epi, UC Berkeley School of Pub Health, Berkeley CA
Angela Bucciarelli: Center for Urban Epidemiologic Studies, The New York Academy of Medicine, New York, NY
David Vlahov: Dept of Epidemiology, Johns Hopkins Bloomberg School of Public Health, Baltimore, MD, and Center for Urban Epidemiologic Studies, The New York Academy of Medicine, New York, NY
Kenneth Tardiff: Dept of Psychiatry, Weill Medical College of Cornell University
What is the exposure of interest?
Neighborhood-level income inequality
What is the outcome of interest?
Overdose mortality
What is the hypothesized mediator of interest and how is it measured?
This study hypothesized three different factors as potential mediators between neighborhood-level income inequality and overdose death.
The authors used data from the NYC Mayor’s Management Report, the NYC Police Department, and the NYC Housing and Vacancy Survey to define constructs for the level of environmental disorder, police activity and quality of the built environment.
Describe the modeling approach and briefly report the estimated total, direct, and indirect effects (if these are reported).
The authors used multivariable models (with individual-level and neighborhood-level covariates and an individual-level outcome) to measure the odds of death due to drug overdose in neighborhoods in the top decile of income inequality compared to the most equitable neighborhoods, and adjusted for the three potential mediators. The authors also measured this potential mediation using path analysis, which was conducted at the neighborhood-level. Thirty-six percent of the association between the distribution of income and the rate of overdose death is explained by the direct effect of income inequality on the rate of drug overdose death (direct effect=0.21). Sixty-four percent of the association is indirect and mediated by levels of environmental disorder, levels of police activity and the quality of the built environment (indirect effect=(-0.55x-0.40)+(0.34x0.09)+(0.36x0.36)=0.38). After adjusting for the three mediators, the odds of overdose death in neighborhoods in the top decile of income inequality decreased from 163 to 1.23 when compared to the most equitable neighborhoods. The authors concluded that the association between income inequality and the rate of drug overdose mortality was primarily explained by an indirect effect through the level of environmental disorder and the quality of the built environment in a neighborhood.
Do you think there is potential measurement error in the mediator and how would that affect the results?
One potential measurement error relates to the way the authors chose to measure their mediators (they point this out in their limitations section). “Quality of the built environment” was based on the percent of housing units in dilapidated conditions which was based off of an assessment of occupied housing units only. This is likely to underestimate the mediating effect of the “built environment” variable on the relationship between overdose mortality and income inequality. Another concern I have is that the data sources the authors used were not collected at the same time. While neighborhoods change at a fairly slow pace, I still think it’s problematic to assess neighborhood-level characteristics in the same analysis as income inequality and overdose mortality if they are captured at different time points.
Do you think there are unmeasured confounders of the mediator-outcome association and how would that affect the results of the mediation analysis?
One potential confounder that was not included in the analysis was individual-level income (the study only uses neighborhood-level SES). We cannot assume that individual income levels are the same among all inhabitants of a specific neighborhood and it could have an important effect on both the exposure and outcome here. Another potential confounder is the type of drug involved in the fatal overdose. The authors do not account at all for the difference in drug types e.g. heroin versus prescribed opioids but I believe drug type or drug of choice could be a potential confounder in this analysis.
Do you have any critiques of the paper?
The mediators considered in this analysis are difficult to measure. The authors used reasonable proxies to measure quality of the built environment, police activity and environmental disorder. However, that may not necessarily represent the perception of these mediators from the neighborhood residents (see measurement error above). Police activity, for instance, was measured by the total number of recorded misdemeanors in the neighborhood. This, however, does not necessarily represent the total amount or type (e.g. aggressive, passive) of police activity within a neighborhood. Finally, I am concerned that the measure of overdose mortality was measured by the location of the death. Individuals at risk for overdose may be at risk of overdosing in neighborhoods where they buy drugs (which could be more likely to be neighborhoods with higher income inequality) however this may not necessarily represent their neighborhood residence. Finally, the outcome relies on medical examiner data but we know that it is difficult to capture all overdose deaths by toxicology reports and it is likely that a substantial number of overdose deaths were not necessarily captures in the medical examiner reports.
Assignment 6: Mediation example
Please identify a quantitative research article evaluating mediation in your field and provide the citation.
Roman Shrestha, Pramila Karki, Tania B. Huedo-Medina, Michael Copenhaver, 2017 “Intent to Use Pre-exposure Prophylaxis (PrEP), HIV Risk Behaviors, and Self-Report Neurocognitive Symptoms by High-Risk Drug Users: A Mediation Analysis” Journal of the Association of Nurses in AIDS Care Available from https://ucsf.idm.oclc.org/login?url=https://doi.org/10.1016/j.jana.2017.04.005
What is the primary discipline of the authors?
Roman Shrestha, MPH, is a doctoral student, Department of Community Medicine & Health Care, University of Connecticut Health Center, Farmington, Connecticut
Pramila Karki, RN, is a graduate student, Department of Allied Health Sciences, University of Connecticut, Storrs, Connecticut
Tania B. Huedo-Medina, Ph.D., is an Assistant Professor, Department of Allied Health Sciences, University of Connecticut, Storrs, Connecticut
Michael Copenhaver, Ph.D., is an Associate Professor, Department of Allied Health Sciences, University of Connecticut, Storrs, Connecticut
Draw a DAG representing the implicit or explicit causal model explored in this paper (you do not need to post your DAG, but we will try to discuss in class).
What is the exposure of interest? Neurocognitive impairment
What is the outcome of interest? Intent to use pre-exposure prophylaxis
What is the hypothesized mediator of interest and how is it measured?
Mediator: HIV risk behavior.
Measurement: HIV risk behavior was assessed using a validated HIV Risk-taking Behavior Scale (HRBS). The HRBS is a brief, 11-item questionnaire developed to measure both drug- and sex-related behaviors that put individuals at risk of either contracting or transmitting HIV. The overall HRBS score is obtained by adding responses to all items on a scale (range = 0-55), with higher scores representing higher levels of HIV risk behavior
Participants are asked to self-report drug- and sex-related HIV risk behaviors during the past 30 days.
Describe the modeling approach and briefly report the estimated total, direct, and indirect effects (if these are reported).
The authors use the 4-step approach of mediation analysis proposed by Baron and Kenny. These four steps include
1. Show that the exposure variable is correlated with the outcome – this step established that there is an effect that may be mediated (total effect B = 0.974)
2. Show that the exposure variable is correlated with the mediator – this step treats the mediator as an outcome variable (exposure-mediator B =0.706
3. Show that the mediator affects the outcome variable (mediator-outcome B = 0.194
4. Show the effect of exposure variable on outcome when controlling for the mediator (direct effect B = 0.538)
They also use the Sobel test and bootstrap method to estimate the indirect effect. B = 0.075
Lastly, the control for covariates, which include age, gender, sexual orientation, ethnicity, marital status, education, income and employment
If the direct effect is reported, would you describe this as a natural direct effect, a controlled direct effect, or something else?
The direct effect denoted as c1 is the controlled direct effect
Do you think there is potential measurement error in the mediator and how would that affect the results?
The mediator is measured via self-report on a standardized scale. I think measurement error in the mediator could occur if there is differential reporting among those with neuro-cognitive impairment. It is possible that those with neuro-cognitive impairment misreport their risky sexual behavior due to memory impairment as opposed to those who are more functional.
This error is likely to result in the effect of the mediator on the outcome being underestimated, and the effect of the exposure on the outcome (direct effect) is likely over-estimated (if the product of the coefficients of exposure-mediator and mediator-outcome is positive -which is typical). The over-estimation of the direct effect is exacerbated to the extent by which the exposure-mediator Standard error is large.
Do you think there are unmeasured confounders of the mediator-outcome association and how would that affect the results of the mediation analysis? Per my DAG, the covariate marital status confounds the mediator-outcome association and not the exposure variable and outcome association.
Do you have any critiques of the paper?
My main issue with this paper is that given the numerous similar covariates confounding exposure-outcome, mediator-outcome, and exposure-mediator pathways, this paper would have benefited from a sensitivity analysis which would have allowed them to assess how robust their direct and indirect effect estimates are to violations in the four assumptions of mediation (there is also no mention of whether the data meets the assumptions of the OLS)
I also think that the authors should have controlled for the exposure variable in establishing the effect of the mediator on the outcome. This is because it is not enough just to correlate the mediator with the outcome because the mediator and the outcome may be correlated because they are both caused by the exposure variable.
One of the key assumptions for mediation analysis that there should be no mediator-outcome confounder that is itself affected by the exposure. Although not depicted in the DAG. – It is easy to see/imagine how neurocognitive impairment (exposure) itself affects confounders of the mediation-outcome path such as education, employment status and income.
My exposure is poverty and blindness from trachoma. Poverty is a huge risk factor for trachoma infection, but allied with poverty is the idea that hand sanitation and access to clean water is a mediator between poverty and increased risk for trachoma. I think it likely that accumulation is the most appropriate model, where total time in poverty increases your risk of blindness from trachoma later in life. There may be some arguments for critical periods however, as sanitation is in part a learned behavior, so it may be that there are critical periods for taking maximum advantage of the access to sanitation that is available, or alternatively children continuing to have friends who remain in poverty/without access to clean water and continue to be high sources of transmission. I think there is little evidence for mobility having a large impact upon the risk of blindness from trachoma, as it is not obvious that changing upwards or downwards in and of itself would create increased or reduced risk from trachoma.
Ideal data- poverty data from a cohort study in a trachoma endemic area following villagers over 30+ years, with reported blindness from trachoma. I suspect that such data does exist, but I’m not sure if I personally would have access to it. I would really need longitudinal data. I would be interested on having data of poverty status during infancy (0-2), early childhood (2-5), mid-childhood (5-10), late childhood (11-15).
Saturated model:
E(blindness) = b0 + b1*P1+b2*P2 + b3*P3 + b4*P4 + ∂12*P1*P2 + ∂23*P2P3 + ∂34*P3*P4 + ∂13*P1*P3 + ∂14*P1*P4 + ∂24*P2*P4 + ∂123*P1*p2*P3 + ∂134*P1*P3*P4 + ∂124*P1*P2*P4 + ∂234*P2*P3*P4 + ∂1234*P1*P2*P3*P4
Testing critical period model – Compare the critical period model to the saturated model of poverty in which each period matters and each interaction between each. The critical period model would assume being exposed in at least one time period is significant, but non-significant in zero. This would assume that at least one of b1, b2, b3 or b4 = 0, and at least one is non-zero. It also assume that all ∂’s = 0. You would use the F-statistic for n-16 degrees of freedom to compare the two models on consistency with the data.
Testing the accumulation model: Compare again to the saturated model. Here you would assume that all ∂’s were = 0 since there was no interaction over time between poverty exposure status, and instead you would also assume that b1=b2=b3=b4. Here you can use the partial F-test to compare the saturated model with 16 parameters to the simpler accumulation model with only two parameters.
Testing social mobility model - compare again to the saturated model, but easier to compare to the saturated model specified as a change in poverty status rather than the direct model specified above.
Exposure: Postmenopausal women exposed to bisphosphonates
Outcome: Atypical femur fracture (AFF) over the course of treatment (radiographically confirmed).
Hypothesis: Greater exposure to BPs is independently associated with AFF risk, particularly among those with other risk factors.
Model: The accumulation model would be the most appropriate in this example. Longer bisphosphonate use is expected to be independently associated with increased risk of AFF. Conversely, after adjustment for other risk factors and BMD, it is expected that the risk of fracture will be reduced (either minimal or moderate) or not associated with bisphosphonate use.
Regression Model: To determine whether other factors are associated with AFF, Cox models for time to AFF will be derived using a set of candidate predictors. For the proposed analysis, duration and timing of BP use will be fixed covariates, evaluated at the time the prediction is made. The analysis will also estimate the independent effect of BP use on AFF in the population age≥50, both overall and in the subset with pre-treatment BMD measurements. Longer bisphosphonate use is expected to be independently associated with increased risk of AFF.
Data sets: It would be difficult to find a dataset that may answer the above questions due to the rare outcome (AFF). However, one could use the Kaiser California data. The information would be available via electronic medical records, with standing linkage to cause of death and fact of death, laboratory data both inpatient and outpatient readily accessible, geocoding (census blocks – race, income, education), drug utilization, and radiological testing and results, demographic and enrollment data, and hospital discharge data.
Assignment: For a specific exposure-outcome combination of interest to you, specify which lifecourse model is likely most appropriate and why you think this is the case. Describe the regression models you could use to test your hypothesis. Are there any possible data sets in which this test could be conducted, and if so, what concerns would you have about interpreting your proposed test of the lifecourse model?
Exposure: malaria in pregnancy
Outcome: adverse birth outcome, including premature birth, low birthweight, and neonatal mortality (i.e. neonatal death within the first 28 days).
Hypothesis: I think relatively little is known of the risk of malaria in pregnancy in the first trimester and how that can lead to increased risks of adverse birth outcomes. Additionally, it is not known known if there is a cumulative effect of the exposures (which I am guessing there is) and if being protected in your third trimester would completely reduce the risk of adverse birth outcomes? My hypothesis is that although there may be critical windows where infection may induce the greatest harm, protection during the 1st trimester vs. 2nd trimester vs. 3rd trimester may have different effects.
Proposed model: To study my hypothesis, I believe that the social mobility model would be an appropriate model to use to study this outcome including interaction terms for each of the trimesters. However, since little is known about this topic, I wonder if I can’t test for the different hypotheses using data?
Cohort: We would need to use a large cohort study of women at reproductive age and include all women who have conceived during follow-up and closely monitor their pregnancy until delivery and 28 days after to assess neonatal death. I think these cohorts are far and few, and I can only think of cohorts (i.e. randomized control trials) that have enrolled women usually late in their first trimester (~12 weeks). If anyone knows of an ideal cohort, let me know!
Potential concerns:
Exposure: poverty level (e.g. % below federal poverty level)
Outcome: HIV infection
Hypothesis: people exposed to higher levels of poverty are more likely to develop an HIV infection, and at a younger age.
Lifecourse model: I think the “social mobility hypothesis,” or as described by Gita et al as the “life course model of a critical period with later effect modification (where the irreversible change of the critical period can be either enhanced or diminished by a later effect)” is most appropriate, because I believe poverty during childhood/teenage years has a different effect on the likelihood of HIV infection than poverty during adulthood.
Regression model: The most appropriate regression model would be a Cox model, since I’m interested in the time to event (event being testing positive for HIV infection). This could be written as:
Log(h(t|x)) = log(h0(t)) + β2S2 + β3S3 + θ23S2S3,
Where β2 gives the slope of childhood poverty level, β3 gives the slope of the adult poverty level, and θ23 is the coefficient for the interaction term.
Dataset: I could use the Community Health Applied Research Network (CHARN) data. CHARN is a research network of community health centers and universities that was established to conduct research on patient-centered outcomes among underserved populations. Advantages of this dataset include that it comprises different geographical areas of the US and is very large (over 500,000 patients). Potential problems with using this dataset are that because it focuses on underserved populations, there may not be much variability in the exposure; because it is comprised of electronic health record data, family income is unlikely to be recorded and participants will have to be asked directly (which leads to another problem – may not accurately recall early life family income levels, especially at the level of detail that would be needed for research purposes).
I am interested in looking at
Outcome- adolescent sexual reproductive health outcomes ---binary outcome – positive vs negative
Exposure: household level food insecurity status as measured by the household food insecurity assessment scale (HFIAS)
The life course model that is likely most appropriate is the cumulative model. Although by itself adolescence is both a sensitive and critical period, I believe that sexual reproductive health outcomes among adolescents are secondary to a gradual accumulation of promotive and risk factors right from in utero, through early childhood, pre-adolescence and during the adolescent period in this case- here food insecurity. In my study, food insecurity is a measure of personalized poverty.
My research question: Does household level food insecurity exposure over the life course have a cumulative effect on adolescent SRH outcomes.
Data collection and definition
I would need to measure exposure and health outcomes at 3 time points,
Exposure: Household food insecurity (HFI) is measured using the Household food insecurity assessment score (HFIAS) which divides food insecurity into 3 tertiles of 1-3 = low HFI (0), 4-6 = moderate HFI (1) and 7-9 = severe HFI (2)
Data analysis: A combination of the 3 levels of HFI (0,1,2) at 3-time points would result in 27 different life course trajectories. Life course trajectory 1, for example, would be represented by HFIS exposure 000, implying low HFIS at all three time points in the life course. Trajectory 27 would be represented by exposure 222, implying high HFIS at all three time points.
In order to represent cumulative HFIS exposure over the life course, I would create a summary score by summing up the exposure level at the three time points for each trajectory. Potentially the cumulative scores could look as follows
Summary score 0: 000
Summary score 1: 001, 010, and 100.
Summary score 2: 002, 011, 020, 101, 110, and 200.
Summary score 3: 012, 021, 102, 111, 120, 201, and 210.
Summary score 4: 022, 112, 121, 202, 211, and 220.
Summary score 5: 122, 212, and 221.
Summary score 6: 222
It is also possible for me to examine the accumulation hypothesis for a given level of household food insecurity (HFI) by doing a trend analysis. Thinking about it this way- there are nine possible ways of getting to each of the three levels of HFI.
Low HFI- the different trajectories are: 000, 010, 100, 110, 020, 200, 120, 210, and 220
Moderate HFI at time point 3: 001, 011, 101, 111, 021, 201, 121, 211, and 221
Severe HFI at time point 3: 002, 012, 102, 112, 022, 202, 122, 212, and 222
For this trend analysis: I would do a Logistic regression, with the category having the lowest cumulative score for each level HFI as the reference group and determine the odds of negative SRH outcomes.
Data sources/ sets
Realistically, without setting up a new study, the International Epidemiology Databases to Evaluate AIDS (IeDEA) is a database that provides a rich resource for globally diverse HIV/AIDS data on infants, children, adolescents, and pregnant women. It is the lowest hanging fruit that would enable to look at this question (may need to tweak the measurement of my exposure)
This data base contains longitudinal data from pregnancy, HIV outcome for the infant. Infants who test HIV positive are followed up prospectively for life beginning from the year 2006
Infants who are HIV –exposed are followed up prospectively upto 18 months.
The database has got rich programmatic level data for individual adolescents including the very young adolescents and the older adolescents.
For the Kenya data, there is baseline household socio-economic demographic data but this is not collected beyond this.
Concerns about my tests:
These are mainly do with the nature of my data.
The limitations of this data set are that it is limited largely HIV-infectedted infants and adolescents and so the findings may not be easily generalizable to adolescents who are HIV negative.
The data – does not also look at household food insecurity but rather measures levels of malnutrition using height for weight z scores, BMI and Mid upper arm circumference. This is measured atleast once every 1,2, or 3 months. While important, may not be able to tell us the HFIs but can serve as a proxy. Also the measurement is not standardized across different facilities and so there may be errors arising from measurement due to instrumentation
Other challenges with this data is incompleteness of data, children, adolescents who are lost to follow up, high mortality rate and issues around structural validity of the data (I am thinking here – maturation)
From an analytical perspective, this database only provides observational data and so the issue of being able to infer causality from this data would have to be looked at carefully.
Exposure: obstetricians exposed to midwifery training opportunities during residency (binary variable: no midwives vs yes midwives; and a multilevel categorical variable grouped into quartiles: estimated percentage of normal deliveries supervised by midwives).
Outcome: one’s personal caesarean rate over the course of a clinical career (continuous outcome).
Regression models: I would choose a regression model for the “critical period”, with the hypothesis that residency training is a formative time for new physicians’ clinical decision-making style. Previous evidence has suggested that exposure to abortion training in residency predicts future provision of abortions. The cesarean rate varies dramatically over time and over institution. It would be important to understand whether elements of residency training predict stability of clinical decision making over time, as residency training may be an important target for interventions to reduce the cesarean rate.
Data sets: I don’t know of any data sets that could satisfy this answer. The American Board of Obstetrics and Gynecology does conduct periodic surveys of newly certified ob/gyn’s, and their email list was used in the above-mentioned abortion research. However, this doesn’t allow for reaching later-career Ob/Gyn’s to assess the course of an entire career. Additionally, it is difficult for Ob/Gyn’s to know and self-report their own personal caesarean rate. I could consider asking them to self-report the institution/practice caesarean rate. However, this would introduce a meso-level variable, and I am not sure how the “critical period” regression model mixes with multi-level models.
Assignment: For a specific exposure-outcome combination of interest to you, specify which lifecourse model is likely most appropriate and why you think this is the case. Describe the regression models you could use to test your hypothesis. Are there any possible data sets in which this test could be conducted, and if so, what concerns would you have about interpreting your proposed test of the lifecourse model?
Exposure: air pollution
Outcome: small for gestational age and preterm birth à chronic disease later in life
Lifecourse model: Critical period
Babies exposed to air pollution in utero are more likely to be born pre-term and be small for gestational size. This irreversible intermediate outcome predisposes them to chronic disease later in life, most closely aligned to the critical period hypothesis (Mishra). This model assumes Y111 = Y101 = Y110= Y100= Y1**, and Y011= Y001= Y010 = Y000= Y0**; corresponding to an average change over this critical period time = Y1** - Y0**. A linear regression model corresponding to the early critical period is: E(Y) = a + B1S1.
I am not aware of any datasets that include air pollution exposure during gestation and later life outcomes (probably ethical issues). I think we can probably conduct natural experiments in heavily polluted cities or areas that experienced periods of heavy/reduction of pollution. I think the biggest obstacle in my ‘dataset’ is exposure assessment.
Assignment: For a specific exposure-outcome combination of interest to you, specify which lifecourse model is likely most appropriate and why you think this is the case. Describe the regression models you could use to test your hypothesis. Are there any possible data sets in which this test could be conducted, and if so, what concerns would you have about interpreting your proposed test of the lifecourse model?
Exposure: adverse/traumatic event (categorical variable) measured at different time points across the life course
Outcome: development of opioid use disorder (binary variable)
The accumulation of risk model is the most compelling model to answer the question: do adverse events across the life course increase the risk of developing opioid use disorder (OUD)? Under this model, we would expect that individuals with more adverse/traumatic events will have a higher likelihood of developing OUD than individuals with a lower total number of adverse/traumatic events over the life course. This model functions irrespective of the timing of adverse events, which is distinct from the critical or sensitive models which focus on the timing of adverse events, rather the accumulation of events. I’m not sure if the outcome is more influenced by the linearity or amount of the exposure. According to Mishra et al, we can assess each of these according to two different models. If the outcome depends on the amount of exposure, then Y011 < Y001 and Y111 < Y011 (where Y is exposure and 001-111 represent the different exposure options at the three different time points. For this example there are three time points). If, however, the outcome depends on the linearity of the exposure then we can create a lifetime adverse events score and assume that every trajectory can be represented by the total number of exposed periods. For every additional adverse event, the change in mean OUD is assumed to be:
^acc = (Y111 – Y000) / 3 = Y111 – Y011 = Y111 – Y101
If we assume a direct and cumulative causal effect of adverse event on OUD then ^acc is the causal parameter of interest. The linear regression would be:
E(Y) = α + β ∑ Sj
I don’t know of any datasets that routinely track the exposure, adverse events. The outcome can be tracked by medical chart and can be extracted by a manual chart review. There are some events that would be recorded in medical charts that could be used as proxies for adverse/traumatic event (e.g. sexual assault, traumatic injury) but that would only capture a small percentage of potential adverse events in someone’s life course.
If there was an irreversible physiological or biological change based on fetal exposure to opioids in the presence of maternal opioid use during pregnancy, the the critical period model would be ideal for assessing the development of OUD based on this critical event.
One of my primary interests is how educational quality affects later life cognition, specifically whether there are critical points in which educational quality matters more, such as in elementary school regardless of status in the high school period. My hypothesis is that disparities in educational quality in early childhood would have a substantively different effect on shaping interpersonal relationships, agency, coping skills, and other behaviors that may affect cognitive aging than educational exposures in adolescence, due to differences in development (e.g. a higher influence of outside social factors in adolescence) and classroom structure/teacher interaction. Mishra et al. propose a critical period model using the assumptions that Y111 = Y101 = Y110 = Y100 = Y1** and Y011 = Y001 = Y010 = Y000 = Y0** - in my situation, because educational characteristics are available only in elementary school and high school, the assumption would be that children with exposure to high quality education in early years (Y1*) would have the same benefit in cognitive outcomes, regardless of their later status, and vice versa for those exposed to low quality education (Y0*). The change in the early critical period (Δearly crit. period) would be equal to Y1* - Y0*, which would be represented by β1 in the linear regression model of E(Y) = α + β1S1. One consideration would be the potential inclusion of actual educational attainment as a third subscript, which would lead to the assumption that the effects of early childhood educational quality on later life cognition are the same regardless of actual grade completion, though the validity of this assumption is questionable, and may be more suited as a separate accumulation model for inclusion in a saturated model.
A secondary question that I would be interested in exploring is whether moving between districts of differing quality affects later life cognition. One way that this could be addressed is using the general model of social mobility proposed by Mishra et al., considering changes in downwards mobility as Y10 – Y11 (high to low verses constantly high) and upwards mobility as Y01 – Y00 (low to high verses constantly low). The corresponding regression equation would be E(Y) = α + δ12D12 + γ12U12 = α + β1S1 + β2S2 + θ12S1S2, in which Y is a function of the first school quality exposure and the post-migration quality exposure. Mishra et al. propose an alternative model that assumes all downward changes are equally harmful and all upward changes are equally beneficial, but I do not believe that relevant for this research question. One issue that arises in addressing this research question is how to address children moving school districts at different ages. Conceptually, I would propose basing an analytic approach off the results of the previous question and looking only within one age range (e.g. elementary) if a critical point of exposure was found, and condensing all ages if no critical points of exposure were found.
Because this exposure-outcome combination involves both a late-life outcome and details of an early-life exposure, there are few datasets with adequate residential characteristics. Educational characteristics are publicly available dating back to the early/mid 20th century for the US at the state and county levels, but a cohort of older adults with information on childhood residence is required for linkage. Two potential datasets with residential history are REGARDS and PSID, though PSID may be subject to lower availability of cognitive measures. A major concern in interpreting my proposed model is the dependency between early education and other educational/socioeconomic factors, and how to define assumptions in the presence of such interaction.
I am interested in how improved leadership skill leads to better managed programs and ultimately superior health outcomes. It is possible to apply the lifecourse model, taking into account the critical period, accumulation and effect modification assumptions.
In this case, the critical period is in childhood and whether an individual was taught critical-thinking and problem solving skills through their primary education.
The application of the accumulation hypothesis in this case is the frequency and duration of exposure to good problem solvers/managers, both in school and in early jobs.
It is unlikely that an effect modifier like having a very good boss/mentor later in life would be effective in reversing the habits and transmitting leadership skills, but nonetheless possible.
I am not sure how I might develop a regression model for this analysis.
I am not aware of any data sets that could be used for this analysis. I have been in early conversations with colleagues at the University of Global Health Equity in Rwanda about doing the very first part of this analysis around the skills that their students have before entering a Master’s program, their performance in school and their placement and progress in jobs after graduation. I would be worried about how to balance the relative weighting of each experience on the outcome. It’s not clear to me whether or how much more influential the early formative experiences of how to analyze information and solve problems is compared to recent experiences with on the job training and a strong model and mentor. Certainly, the likelihood of getting a good job with a strong mentor is diminished without the initial foundational skills, so we need to model the interactive effect, but I don’t know enough through empirical examples to be sure how to consider the relative importance.
Assignment: For a specific exposure-outcome combination of interest to you, specify which lifecourse model is likely most appropriate and why you think this is the case. Describe the regression models you could use to test your hypothesis. Are there any possible data sets in which this test could be conducted, and if so, what concerns would you have about interpreting your proposed test of the lifecourse model?
I am interested in evaluating the effect of maternal and child dietary composition (refined carbohydrates and added sugars) on the development of non-alcoholic fatty liver disease during adolescence and early adulthood (<30 years). Some studies would favor the hypothesis of prenatal preprogramming in which the fetal environment (in this case maternal diet) interacts with fetal genetics and epigenetics to alter risk of chronic diseases. This hypothesis would favor the Critical Period Model in which exposure to high amounts of refined carbohydrates and sugars determines an individual’s risk of developing fatty liver later in life during adolescence and adulthood. However, multiple studies have shown that dietary modification during adolescence or early adulthood can reverse fatty liver, thus the accumulation model is likely most appropriate to test the causal effect of carbs and sugars on fatty liver.
I could use a linear regression model with a lifetime diet score that takes the value between 0 and 5. Where 0 is always diet low in refined carbs and sugars, and 5 is always a diet high in refined carbs and sugars.
0 = maternal diet before pregnancy
1 = maternal diet during fetal development
2 = maternal diet during breastfeeding period, or presence of sucrose in infant formula
3 = diet during infancy
4 = diet during adolescence
5 = diet during adulthood
E(Y) = alpha + B(Sum of Sj)
B =change in the score; for every unit increase in this score, the change in mean liver fat attenuation is expected to be constant and equal to the change in lifetime accumulation.
We could potentially use the Project Viva cohort out of Harvard. Project Viva enrolled 2,670 women during pregnancy and is following them and their children over time. The project was established to examine prenatal diet and other factors in relation to maternal and child health (Oken et al. International Journal of Epi, 2015). Diet exposures were ascertained and included pre-pregnancy, during pregnancy, and after birth. Even though liver fat attenuation was likely not measured in infancy, it could be measured during adolescence and early adulthood since data on teenagers are currently being collected.
Concerns in interpreting my proposed test would include potentially lacking heterogeneity in the exposure due to having a high education sample with health insurance within the Boston area, and who were primarily white. Also, temporal changes of decreased sugar consumption from the year 2000 to 2010 may have brought down consumption of sugars to within tolerable levels to be able to measure cumulative effects on fatty liver attenuation. This exposure-outcome relationship is also susceptible to confounding, including the fact that a diet high in refined carbohydrates and sugars may be a marker of a general poor diet or a sedentary lifestyle.
Exposure: malaria prevalence in the region where you live (continuous measure, that could be dichotomized). Region must be defined. Pa is the prevalence exposed to at age a. For every individual we have a vector of ni exposure measurements Pi=(Pa1, Pa2, …, Pni).
Outcome: being a symptomatic with malaria parasite at age 40. Binary outcome Y.
Time dimension: age
1) Saturated model
2) Critical model: childhood exposure
3) Cumulative: over your entire life
4) Social mobility: changes between childhood malaria-prevalence region and adulthood malaria-prevalence region
Malaria is an infectious disease that provide some sort of short-term immunity and repeated infections can strengthen your organism response to further infections. Therefore, I think life-course exposure to malaria affects your probability of both being infected and if infected of being symptomatic today. Following Mishra et al. recommendations, I would ideally study different complementary models but I think the cumulative model and the critical model are the most appropriated, given the sensibility of children to malaria infection and the importance of childhood years as a developmental stage. Individual longitudinal cohort data exist to test this research question but may lack variability if constrained to a small region (say a village where everyone has had the same exposure history other than the cohort effect) or be susceptible to confounding. In a context where malaria is seasonal with outbreaks of different sizes from year to year and depending on environmental conditions, confounding is serious concern.
Age as the time dimension:
Article: Both Baseline and Change in Lower Limb Muscle Strength in Younger Women Are Independent Predictors of Balance in Middle-age: A 12-yr Population-based Prospective Study.
Article link: http://onlinelibrary.wiley.com/doi/10.1002/jbmr.3103/epdf
Research Question: To examine whether lower limb muscle strength (LMS) in young women and changes in LMS are independent predictors of balance in middle-age.
Significance: Among young women, greater LMS at baseline and slower decline over time are both associated with better balance in midlife. Analogous to the contributions of peak bone mass and bone loss to fracture risk in older adults, this suggests that both improvement of muscle strength in younger age and prevention of age-related loss of muscle strength could be potentially useful strategies to improve balance and reduce falls in later life.
Study sample: This was an observational 10-yr follow-up of 470 women aged 25-44 years at baseline who had previously participated in a 2-yr population-based randomized controlled trial of osteoporosis education interventions.
Longitudinal study design: Women aged 25-44 years were randomly selected from the 2000 Tasmanian Electoral Roll. Women were recruited if they were free of the following conditions: previous measurement of bone density, thyroid disease, renal failure, malignancy, or rheumatoid arthritis, a history of hysterectomy or hormone replacement therapies, pregnancy or planning pregnancy within 2 years of study entry, or lactating. At baseline, 470 women were randomly assigned to one of two osteoporosis educational interventions: group education using the Osteoporosis Prevention and Self-management course or an information leaflet.
Analysis approach: Bone mineral density was measured at the spine and hip at baseline and mean spine and hip T-score used to provide feedback of relative fracture risk as part of the education intervention (higher risk (mean T-score < 0) vs. normal risk (mean T-score ≥ 0)). Linear regression was used to examine the association between baseline LMS (by dynamometer) and change in LMS over 12 years with balance at 12 years (timed up and go test (TUG), step test (ST), functional reach test (FRT) and lateral reach test (LRT)).
Time since study enrollment as the time dimension:
Article: Association of Trabecular Bone Score (TBS) with Incident Clinical and Radiographic Vertebral Fractures Adjusted for Lumbar Spine BMD in Older Men: A Prospective Cohort Study.
Article link: http://onlinelibrary.wiley.com/doi/10.1002/jbmr.3130/epdf
Research question: To determine the associations of TBS with incident radiographic and clinical vertebral fractures in a cohort of community-dwelling older men. Among men in the Manitoba Bone Density Cohort, TBS was not associated with incident clinical vertebral fracture ascertained using administrative claims data, but the accuracy of claims data for incident vertebral fracture may be suboptimal. Additionally, among older men no prior study has examined the association of TBS with incident vertebral fractures identified on the basis of radiographs alone.
Significance: Trabecular Bone Score (TBS) is associated with incident hip and major osteoporotic fractures in older men, but its association with incident vertebral fracture is uncertain.
Study sample: From 2000 to 2002, 5994 community-dwelling men >65 years old were enrolled into the prospective MrOS study in six regions of the United States, described in previous publications.
Longitudinal study design: Participants have been followed since enrollment for incident fractures, falls, prostate cancers, and death.
Analysis approach: TBS was estimated from baseline spine DXA scans for 5,831 older men enrolled in the Osteoporotic Fractures in Men (MrOS) study. Cox proportional hazard models were used to determine the association of TBS (per 1 SD decrease) with incident clinical vertebral fractures. Logistic regression was used to determine the association between TBS (per 1 SD decrease) and incident radiographic vertebral fracture among the subset of 4,309 men with baseline and follow-up lateral spine radiographs. The authors also examined whether any associations varied by body mass index (BMI) category.
Another time dimension:
Article: Longitudinal assessment of urinary PCA3 for predicting prostate cancer grade reclassification in favorable-risk men during active surveillance.
Article link: https://www.nature.com/pcan/journal/vaop/ncurrent/full/pcan201716a.html
Research question: To assess the utility of urinary prostate cancer antigen 3 (PCA3) as both a one-time and longitudinal measure in men on active surveillance (AS).
Significance: Although methods of monitoring and triggers for curative intervention vary among Active Surveillance programs, the majority of protocols require serial prostate biopsy, a procedure associated with patient discomfort and risk of complications. As such, alternative methods of monitoring are needed. One potential option is prostate cancer antigen 3 (PCA3), a noncoding mRNA first described in 1999 that is highly overexpressed in PCa tissue. Since its introduction to clinical use, several studies have exhibited a significant association of PCA3 with PCa. To this point, however, evaluation of PCA3 in the AS population remains quite limited. Findings suggest potential utility of PCA3 in the AS setting but remained limited by small sample size and short-term follow-up. Furthermore, the utility of repeated PCA3 measures has yet to be demonstrated.
Study sample: Since 1995, the institutional review board-approved Johns Hopkins AS program has enrolled 1511 men with favorable-risk (low-risk or very-low-risk) PCa with informed consent.
Longitudinal study: Monitoring included semiannual PSA and digital rectal exam (DRE) as well as annual prostate biopsy and/or more recently prostate magnetic resonance imaging (MRI) for most men. Since 2007, urine samples were obtained at clinic visits after standard DRE. In order to assess for longitudinal changes in PCA3, the study cohort was limited to subjects with at least two urine samples obtained over greater than or equal to 3 years of follow-up (that is, at least 3 years apart) and a prostate biopsy within 6 months of each PCA3 assessment (n=294). Further, owing to variable effect of 5-alpha reductase inhibitor (5-ARI) medications on PCA3, 34 men who were on 5-ARI at the time of PCA3 assessment were excluded. The sample size is similar to a prior analysis of the AS cohort. The chosen outcome of interest was grade reclassification (GR) defined as any Gleason score >6 cancer detected on follow-up biopsy.
Analysis approach: Patient demographics as well as first (fPCA3) and subsequent (sPCA3) PCA3 scores were compared between men who did and did not undergo GR using t-test, Mann–Whitney test or chi-squared test, as appropriate. PCA3 scores were transformed into logarithmic scale to correct for skew and stabilize variance. A linear mixed-effects model with random effects was used to assess the longitudinal changes in PCA3 over time and also to evaluate its association with the outcome of interest. Additionally, the utility of a single PCA3 value for independently predicting high-grade disease was assessed using a multivariable logistic regression model adjusting for age, disease volume (very-low-risk or low-risk status) and PSAD. Model accuracy was assessed by measuring the area under the receiver operating characteristic curve (AUC), and the goodness-of-fit of the multivariable model was evaluated using Hosmer–Lemeshow goodness-of-fit test. One-sided tests were used for a priori hypotheses that were directional, and statistical significance was set at P<0.05.
Age
Longitudinal physical activity and sedentary behavior in preschool-aged children with cerebral palsy across all functional levels
This study had three aims: to describe habitual physical activity and sedentary behavior in young children with cerebral palsy (CP) from 1 year 6 months to 5 years of age, to compare habitual physical activity between time points, and to examine the rate of change in habitual physical activity and sedentary behavior across all gross motor functional capabilities.
The study sample was drawn from two population-based cohort studies conducted in Queensland, Australia. Eligible children were born between 2006 and 2009 and diagnosed with CP by a physician; and for this study, children had to have completed 3-day physical activity monitoring at any of the follow-up time points in the cohort studies (at 18 months – 2 years, 2 years 6 months – 3 years, 4 years, and 5 years of age).
Study participants were classified for gross motor function at each follow-up visit. Physical activity was measured using an accelerometer worn for 3 days, and with an activity diary kept by parents. The authors used a linear mixed-effects model to assess the association between physical activity and gross motor functional capabilities. This model was chosen to account for non-independence of repeated measures on the same participants.
http://onlinelibrary.wiley.com.ucsf.idm.oclc.org/doi/10.1111/dmcn.13439/abstract
Time Since Study Enrollment
Substance use patterns and factors associated with changes over time in a cohort of heterosexual women at risk for HIV acquisition in the United States
The aim of this study was to analyze substance use patterns in a large, geographically diverse cohort of women at increased risk for HIV infection.
The study sample is a multisite, longitudinal cohort of women at elevated risk of HIV infection. Eligible women were enrolled in 2009 and 2010 from 10 urban and peri-urban communities in six geographic regions of the US. Women were recruited during randomly selected venue-time intervals within each community (e.g. women entering the venue during a recruitment interval were approached for screening), and included if they reported unprotected sex with a man in the previous six months and at least one other personal or partner HIV risk characteristic.
Women meeting the inclusion criteria (for this analysis, all women reporting substance use at baseline were included) were followed at 6-month intervals for either 6 or 12 months. Self-reported substance use (drug use and/or binge drinking) in the past six months was assessed by type of substance and frequency. The association between substance use (specifically decreased substance use between baseline and the follow-up visit) and baseline predictors of substance use was assessed by bivariate and multivariate log regression (is this the same as logistic regression?).
Other Time Dimension
Radiation effects on cognitive function among atomic bomb survivors exposed at or after adolescence
The aim of this study was to investigate the effects of radiation on pre-dementia cognitive decline among participants who did and did not develop dementia during follow-up.
The study population consisted of participants in the original Adult Health Study cohort consisting of atomic bomb survivors who were within a set distance of the Hiroshima atomic bomb, and age- and sex-matched subjects who were beyond a set distance to the same, and who were >= 60 years of age in 1992. Participants from the cohort were included if they did not meet a set of exclusion criteria based on completeness of data, age, and timing of dementia onset.
Participants were followed every two years, and cognitive function was assessed at baseline using the Cognitive Abilities Screening Instrument (CASI) and thereafter using a short version of the CASI at each follow-up visit. Dementia was assessed for using criteria in the DSM, 4th edition. Radiation dose was estimated using the Dosimetry System 2002. The authors used a random coefficient regression model to estimate the association between radiation exposure and cognitive decline and cognition level in both demented and cognitively intact individuals.
http://www.sciencedirect.com.ucsf.idm.oclc.org/science/article/pii/S0002934315009109
Longitudinal analysis based on age
Longitudinal analysis based on time since study enrollment
1. Age as the time dimension:
Title: Longitudinal Trends in Sexual Behaviors with Advancing Age and Menopause Among Women With and Without HIV-1 Infection available from AIDS Behav. 19(5): 931–940. doi: 10.1007/s10461-014-0901-1 https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4370800/
Title: Long-Term Impact of Malaria Chemoprophylaxis on Cognitive Abilities and Educational Attainment: Follow-Up of a Controlled Trial PLOS clinical trials available from http://journals.plos.org/plosclinicaltrials/article?id=10.1371/journal.pctr.0010019
Title: Longitudinal changes in engagement in care and viral suppression for HIV-infected injection drug users AIDS. 2013; 27(16): 2559–2566. doi: 10.1097/QAD.0b013e328363bff2 Available from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3795966/
The investigators looked at
1. temporal trends in engagement in care across the entire cohort: Here, they calculated the proportion of participants reporting HIV care visits in each calendar year. They then used a linear trend time-series model with a first-order auto-regressive covariance. They determined whether there were significant improvements from 1998 to 2011 in the annual proportion of the cohort that was fully engaged in care (in care all at both ALIVE visits during the year), was partially engaged in care (in care at 1 of 2 study visits) and achieved an undetectable HIV RNA level.
2. Time-varying factors associated with the two main negative outcomes, lapses in HIV care and virologic failure. The outcomes were assessed at every follow-up study visit and thus could be experienced multiple times during the study. A lapse in care was defined as reporting that no HIV care visits were attended in the prior 6 months after being in care at the previous study visit. Virologic failure was evaluated using the same framework: study visits at which a participant was noted to have viral suppression were analyzed to determine whether the viral load remained suppressed at the subsequent visit (success) or had increased above the limit of detection (failure). Analysis: To identify significant predictors of the outcomes while accounting for intra-subject correlation resulting from repeated measures per participant, they used logistic regression models with generalized estimating equations (GEE) with robust variance estimates. An alpha level of 0.10 and 0.05 were used for model entry and retention, respectively. To account for the potential confounding effects of secular trends favoring improved engagement in care over time and differential loss to follow-up among higher-risk IDUs, they forced into the adjusted models variables for calendar year and total follow-up time, respectively.
Age
Biro, F. M., Greenspan, L. C., Galvez, M. P., Pinney, S. M., Teitelbaum, S., Windham, G. C., ... & Kushi, L. H. (2013). Onset of breast development in a longitudinal cohort. Pediatrics, 132(6), 1019-1027.
Research question: Does age of the onset of breast development (Thelarche) vary by race/ethnicity and BMI?
Study sample: 1200 girls 6 to 8 years of age living in San Francisco Bay Area, Cincinnati, and New York City from the Breast Cancer and the Environment Research Program (BCERP).
Longitudinal design: Semi-annual or annual visits from date of enrollment (between 2004 and 2008) through March 2012. Mean follow-up was 4.3 years.
Analysis: Kaplan-Meier analyses were used to describe age at onset of breast maturation
Time since enrollment
Browner, W. S., Pressman, A. R., Nevitt, M. C., & Cummings, S. R. (1996). Mortality following fractures in older women: the study of osteoporotic fractures. Archives of internal medicine, 156(14), 1521-1525.
Research question(s): Does mortality increase after fractures?
Study sample: Women 65 years of age or older who had bilateral hip replacement in the Study of Osteoporotic Fractures (SOF).
Longitudinal design: Baseline measurements were taken and participants or their proxy returned a postcard every 4 months reporting the occurrence of fractures. Death certificates and hospital discharge summaries were reviewed for women who died in order to determine if any fractures had occurred since the last follow-up.
Analysis: Multivariable proportional hazards models were used to determine association between fractures and age-adjusted mortality.
Time since diagnosis
Tzeng, H. E., Muo, C. H., Chen, H. T., Hwang, W. L., Hsu, H. C., & Tsai, C. H. (2015). Tamoxifen use reduces the risk of osteoporotic fractures in women with breast cancer in Asia: a nationwide population-based cohort study. BMC musculoskeletal disorders, 16(1), 123.
Research question: Is Tamoxifen use associated with Osteoporotic fractures?
Study sample: 75488 women with breast cancer with no prior history of fractures from the Longitudinal Health Insurance Database for Catastrophic Illness Patients in 2000-2011.
Longitudinal design: Women were followed from date of diagnosis of breast cancer (index date) to the date a hip, vertebrate or wrist fracture occurred.
Analysis: Cox proportional hazards model with time-dependent covariates to estimate risk of a fracture.
Age:
Cammack AL, Hogue CJ. Retrospectively self-reported age of childhood abuse onset in a United States nationally representative sample. Inj Epidemiol. 2017 Dec; 4:7.
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5346510/pdf/40621_2017_Article_103.pdf
Research question – What is the prevalence and onset of different types of retrospectively self-reported child abuse in the united states (sexual, physical and emotional), using a nationally representative sample?
Study sample – Wave IV from the National Longitudinal Study of Adolescent to Adult Health who were initially interviewed in grades 7-12 during 1994-1995 in schools located in 80 communities throughout the United States. Data on childhood abuse comes from a follow-up questionnaire performed when the subjects were 24-32 in 2008-2009. All participants had to have a Wave IV sampling weight to enable the construction of a representative sample.
Longitudinal design – The National Longitudinal Study of Adolescent to Adult Health is a study that followed adolescents overtime in four in-home interviews, the last happening in 2008, with a scheduled Wave V planned for 2016-2018. Childhood abuse was reported retrospectively, with age of first abuse and frequency also being reported.
Analysis approach – age at abuse onset was examined as a continuous variable (they didn’t do Kaplan-meier or other time-based analysis) in a simple linear regression to assess the mean and percentiles for individual abuse subtypes. They then performed a multi-variate linear regression on all individuals how had reported data on all the determinants of abuse.
Time since study enrollment:
Pikalov A, Tsai J, Mao Y, Silva R, Cucchiaro J, Loebel A. Long-term use of lurasidone in patients with bipolar disorder: safety and effectiveness over 2 years of treatment. Int J Bipolar Disord. 2017. 5:9.
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5332323/pdf/40345_2017_Article_75.pdf
Research question – To determine the long-term safety and effectiveness of lurasidone in patients who initially presented with a major depressive episode associated with bipolar disorder.
Study sample – Patients who had a DSM-IV-TR criteria for chronic schizophrenia or bipolar I depression were were enrolled over three 6-week, double-blind, placebo-controlled trials and followed for two years. Patients were initially participating in a 6-month, open label safety trial, and those finishing the initial six-months were and were eligible for an 18-month, open-label, continuation of the initial trial.
Longitudinal design – Patients were followed first over a six month period, and then, following successful completion of the initial six-months were followed for a further 18-months for a total of 2-years of drug safety and effectiveness. Patients were followed for a total of two years, with safety and effectiveness evaluations performed every 3-months following study enrollment, with additional visits scheduled if indicated due to adverse events or clinical worsening. Outcomes included adverse events or clinical worsening and movement disorders, body weight and vital signs.
Analysis approach – They performed simple descriptive statistics on the prevalence of various adverse outcomes, and treatment effectiveness at five different time-points. Additionally, they performed a Kaplan-meier survival analysis to estimate the probability of replapse during the 18-months of lurasidone continuation.
Other: Time since ICU admission
Martin-Loeches I, Muriel-Bombín A, Ferrer R, Artigas A, Sole-Violan J, Lorente L, Andaluz-Ojeda D, Prina-Mello A, Herrán-Monge R, Suberviola B, Rodriguez-Fernandez A, Merino P, Loza AM, Garcia-Olivares P, Anton E, Tamayo E, Trapiello W, Blanco J, Bermejo-Martin JF; GRECIA group. The protective association of endogenous immunoglobulins against sepsis mortality is restricted to patients with moderate organ failure. Ann Intensive Care. 2017 Dec; 7(1):44.
http://annalsofintensivecare.springeropen.com/articles/10.1186/s13613-017-0268-3
Research question –Evaluate the impact of endogenous immunoglobulins on the mortality risk in sepsis following ICU admission as it depends on disease severity.
Study sample – retrospective observational study of 278 patients admitted to the ICU with sepsis fulfilling the SEPSIS-3 criteria, who come from the Spanish GRECIA and ABISS-EDUSEPSIS cohorts. Patients were separated into two groups by their score on organ failure, with SOFA <8, SOFA≥8.
Longitudinal design – Patients were followed after their admission into the ICU (days) until either death, discharge or censored at 28 days after ICU admission. Immunoglobulin levels were assessed on blood samples collected in the first 12 hours following ICU admission.
Analysis approach – Association between immunoglobulin levels at ICU admission with mortality was studied in each group by Kaplan-Meier and multivariate logistic regression analysis.
Age as the time dimension:
Developmental and behavioural associations of burns and scalds in children: a prospective population-based study
Alan Emond1, Clare Sheahan2, Julie Mytton3, Linda Hollén1
Arch Dis Child. 2017 May;102(5):428-483. doi: 10.1136/archdischild-2016-311644. Epub 2016 Nov 13
Research Question. Investigators wanted to understand child developmental and behavioral characteristics and risk of burns and scalds.
Study sample. 12,966 participants up to age 11 from the Avon Longitudinal Study of Parents and Children.
Longitudinal design. Preinjury profiles of the children derived from material questionnaires completed in pregnancy, and at 6, 18, 42, 47, and 54 months. Injury data collected by questionnaire at 6,15, and 24 months, and 3.5, 4.5, 5.5, 6.5, 8.5, and 11 years of age.
Analysis approach. Authors calculated a period prevalence and an incidence rate, and used a multivariable logistic regression to control for confounders.
Time since study enrollment as the time dimension:
Diabetes incidence and influencing factors in women with and without gestational diabetes mellitus: A 15 year population-based follow-up cohort study
Sonia Minooeea, Fahimeh Ramezani Tehrania, Maryam Rahmatib, a, Mohammad Ali Mansourniab, Fereidoun Azizic
Diabetes Res Clin Pract. 2017 Apr 8;128:24-31. doi: 10.1016/j.diabres.2017.04.003. [Epub ahead of print]
Research Question. Investigators wanted to determine if the incidence of diabetes was greater in women with gestational diabetes compared to those without gestational diabetes.
Study sample. 476 women with gestational diabetes and 1982 women without diabetes were selected from participants of the Tehran Lipid and Glucose study.
Longitudinal design. Participants undergo a follow-up at 3-year intervals, and the study began in 1998 and has continued for 15 years.
Analysis approach. The age-adjusted overall time to development of type 2 diabetes was calculated using a Kaplan-Meier curve. Pooled logistic regression was used to assess the odds ratio between time-dependent covariates age, BMU, and baseline family history of diabetes and GDM and type 2 diabetes.
Neither Age nor Time since study enrollment as the time dimension:
Motivational and contextual determinants of HPV-vaccination uptake: A longitudinal study among mothers of girls invited for the HPV-vaccination
Mirjam Pota, b, , , Hilde M. van Keulenb, Robert A.C. Ruitera, Iris Eekhoutb, Liesbeth Mollemac, Theo W.G.M. Paulussenb
Prev Med. 2017 Apr 4;100:41-49. doi: 10.1016/j.ypmed.2017.04.005. [Epub ahead of print]
Research Question. Investigators would like to know whether HPV-vaccination intention, as well as other social psychological determinants, are good predictors of future HPV-vaccination uptake in a longitudinal design.
Study Population. Random sample of mothers of girls invited for vaccination in 2015 from Praeventis (Dutch vaccination register) and from three online panels.
Longitudinal design. Baseline was a questionnaire given Jan 2015, before the HPV round. After completing the vaccination round, data about girls’ actual HPV-vaccination uptake were derived from Praeventis. Questionnaires were then given to determine what attitudes led to HPV-vaccination. The time element is time from HPV-vaccination round.
Analysis approach. Authors used a multivariate linear regression analysis to assess the total amount of variance explained in HPV-vaccination intention.
I agree with Emily--this one was a tough assignment. Happy to hear some discussions about my examples, because I'm not 100% confident in my answers.
Assignment:
Defining the time dimension is a fundamental challenge in longitudinal data analysis. The most common choice is age or time since study enrollment, however, for many questions other time dimensions are relevant, for example grade in school, or time before or after stroke incidence, or time since release from prison. Identify an article in the applied literature using a longitudinal analysis based on age as the time dimension, time since study enrollment as the time dimension, and one other possible time dimension (ie, not age or time since study enrollment). For each study, briefly describe the research question, the study sample, the longitudinal design, and the analysis approach. Please post links to the studies.
Article: Age at menarche, total mortality and mortality from ischaemic heart disease and stroke: the Adventist Health Study, 1976–88
Article link: https://academic.oup.com/ije/article/38/1/245/698081/Age-at-menarche-total-mortality-and-mortality-from
Research question: Does earlier age at menarche predict mortality, particularly death from ischaemic heart disease and stroke?
Study sample: Female participants of the Californian Seventh-Day Adventist Cohort followed from 1976 to 1988.
Longitudinal study design: A total of 19,462 women were followed throughout the course of their life. Age at menarche was self-reported.
Analysis approach: Relationship between age at menarche and mortality were assessed using Cox proportional hazards regression using attained age (I’m guessing this is age from menarche to end of follow-up?) as the time variable. Models adjusted for possible confounders and an interaction between age at menopause and attained age.
Article: Effect of paracetamol on parasite clearance time in Plasmodium falciparum malaria.
Article link: http://www.sciencedirect.com/science/article/pii/S0140673697022551
Significance: Routine antipyretic therapy in children with infectious diseases has long been the source of controversy. Each year, in addition to antimalarial medication, millions of children with Plasmodium falciparum malaria receive paracetamol to reduce fever. However, the usefulness of this practice has not been proven.
Study sample: Children between the ages of 2-7 with asexual P. falciparum parasite loads between 25,000 and 20, 0000 parasites/μL blood treated with intravenous quinine at the Albert Schweitzer Hospital in Lambaréné, Gabon between February and June 1996.
Longitudinal study design: Randomized control trial of 50 children with P. falciparum malaria treated with intravenous quinine, and randomized to receive either mechanical antipyresis alone, or in combination with paracetamol. Rectal body temperature and parasitemia were recorded every 6 hours for 4 days upon study enrollment. Primary endpoints were parasite and fever clearance time.
Dimension of time: Authors used time since enrollment (i.e. time since treatment/admission) to measure the primary outcomes.
Analysis approach: Differences between the groups were assessed by the Mann-Whitney test and χ2 test. Differences between time points within a group were assessed by the Wilcoxon signed-rank test.
3. Trimester when exposed to Dutch Famine (from November 1944 to May 1945)
Article: Decreased birthweights in infants after maternal in utero exposure to the Dutch famine of 1944-1945.
Article link: http://onlinelibrary.wiley.com/doi/10.1111/j.1365-3016.1992.tb00764.x/epdf
Research question: Did in utero exposure to Dutch famine cause decreases in birthweight?
Significance: During the last months of WWII from October 1944 until the surrender of the German forces on May 7, 1945, the Western Netherlands was affected by an acute famine. Circumstances of the Dutch famine provides a unique opportunity to study the long-term effects of this environmental exposure.
Study sample: 1808 first-born singleton offspring of mothers born between January 1, 1944 to June 20, 1946 living in The Netherlands.Mothers were variably exposed to Dutch Famine (first, second, third trimester, or not exposed).
Longitudinal study: Maternal records from the University of Amsterdam teaching hospitals were extracted for births occurring between 1960-1984. Exposure was measured as birth cohorts that represented the trimester the mother was first exposed to the Dutch Famine. Outcomes assessed were birthweight (g), crown-heel length, Quetelet index (kg/m2) and head circumference.
Analysis approach: Differences in birth outcomes were compared with ANOVA. Multivariate linear regression was used to assess differences in offspring birthweights between maternal birth cohorts, testing interactions between place of birth (rural where food supply was more abundant) and maternal birth cohort.
1. Time dimension: Age “The ripples of adolescent motherhood: social, educational, and medical outcomes for children of teen and prior teen mothers.” https://www.ncbi.nlm.nih.gov/pubmed/20674531
Research Question: This study, in contrast, presents evidence suggesting that the problems associated with teen motherhood have been substantially underestimated. We expand current knowledge on the effect of teen motherhood in several important ways. First, we examined the outcomes experienced by these children across multiple domains (medical, educational, and social). Second, we reviewed long-term consequences throughout childhood and into early adulthood. Third, we evaluated outcomes experienced by the later children of teen mothers—identifying those born to older mothers with a history of teen birth as children of prior teen mothers. Finally, by using rich population-wide data, we explored the societal-level consequences of teen motherhood.
Study Sample: This was a retrospective cohort study examining all children residing in the city of Winnipeg (population 620 000) at age 17 years who were born and raised in Manitoba. Six cohorts covering all children born 1979–1982 and 1984– 1985 were analyzed.
Longitudinal design: Retrospective cohort study
Analysis approach: Child health status was measured by: 1) mortality, 2) hospitalization (excluding birth hospitalization and admissions for pregnancy/childbirth), and 3) being among the top 10% of hospital day users (total days). These health measures were examined over 3 time periods: infancy (infancy (<1yr), toddler/preschool (1-5 years), and school age/adolescence 6-17 years). Markers of educational achievement included: 1) mean course marks in ninth grade, and 2) successful high school graduation within 6 years of entering ninth grade. Social outcomes included: 1) whether or not the child was taken into foster care at any point from age 8–17 years (ages for which these data were available), 2) the child’s family was monitored by Child and Family Services (typically at the request of the physician or another agency) at some point from age 8–17 years, 3) the child received income assistance at some time as a young adult (age 18 to 25 years), and 4) the female child became a teen mother herself. The education, foster care, and Child and Family Services outcomes were available for only the 1984 and 1985 birth cohorts. Multivariate regression analyses were used to adjust for likely confounders in several categories.
2. Time dimension: time since study enrollment “Continuous Combined Estrogen Plus Progestin and Endometrial Cancer: The Women’s Health Initiative Randomized Trial”
https://www.ncbi.nlm.nih.gov/pubmed/26668177
Research Question: In the current report, we extend that finding by providing analyses on estrogen plus progestin effects on endometrial cancer type, findings in relevant clinical subgroups, and information on endometrial cancer–related mortality
Study Sample: Briefly, postmenopausal women age 50 to 79 years with an intact uterus were enrolled at 40 US clinical centers. Exclusions were prior breast cancer, anticipated survival of less than 3 years, and previous invasive cancer within 10 years. A three-month wash-out period was required for hormone therapy users at screening.
Longitudinal design: Prospective cohort. “After 5.6 years of median follow-up, the intervention was stopped when overall risks exceeded benefits and participants were instructed to discontinue study drug on July 8, 2002 (16). Protocol-defined follow-up continued through March 31, 2005, the prospectively determined trial completion date. Follow-up after that date required reconsent, obtained from 12 788 (83%) of 15 408 surviving participants.”
Analysis approach: Results for endometrial cancer incidence and deaths directly attributed to endometrial cancer were assessed with time-to-event methods based on the intent-to-treat principle, which included all 16 608 randomly assigned participants. Cancer incidence rate comparisons, presented as hazard ratios (HRs) and 95% confidence intervals (CIs) from Cox proportional hazard models, were stratified by age and randomization groups in the WHI dietary trial, and proportionality was verified with the Grambsch and Therneau’s test (19).
3. Another time dimension: duration of adult overweight “Duration of Adulthood Overweight, Obesity, and Cancer Risk in the Women’s Health Initiative: A Longitudinal Study from the United States”
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4987008/pdf/pmed.1002081.pdf
Research Question: While recent studies have suggested that the risk of cancer related to obesity is mediated by time, insights into the dose-response relationship and the cumulative impact of overweight and obesity during the life course on cancer risk remain scarce. To our knowledge, this study is the first to assess the impact of adulthood overweight and obesity duration on the risk of cancer in a large cohort of postmenopausal women.
Study Sample: The WHI is a large, multi-center prospective cohort study of postmenopausal women. In total, 93,676 women were enrolled in the observational study, and 68,132 were enrolled in the clinical trial arm (n = 161,808) [16]. For this study, we included all participants from the observational study cohort, except those who reported cancer prior to or at baseline or without data on cancer history (n = 12,827) and women with incomplete follow-up information (n = 411) (Fig 1).
Longitudinal design: Prospective cohort. Information on BMI for was obtained from retrospective self-reports at baseline for ages 18, 35, and 50 y, from weight and height measurements at baseline and at 3-y follow-up, and from self-reports at follow-up years 4–8. BMI was calculated by dividing weight in kilograms by height in meters squared. For inclusion in the study, women were required to have valid body weight information from at least three occasions and a valid baseline measurement of body weight and height.
Analysis approach: The analysis was carried out in two steps. In the first step, BMI was modeled across all ages using a quadratic growth model with random intercept and random slope, incorporating all Fig 1. Flowchart of participant inclusion. *Including data at baseline and excluding data from the year preceding cancer diagnosis; including BMI from self-reported as well as measured height and weight; BMI values lower 70 kg/m2 were excluded. doi:10.1371/journal.pmed.1002081.g001 Adulthood Overweight, Obesity, and Cancer Risk PLOS Medicine | DOI:10.1371/journal.pmed.1002081 August 16, 2016 5 / 16 available BMI information from all included participants [19]. No random coefficient was included for the quadratic term. Using this approach, we allowed individuals to have their own BMI trajectory. Using the full model, BMI was predicted from age 18 y until the age at study exit for every cohort member. In the second step of the analysis, Cox proportional hazard models with time since enrollment as the underlying time metric were fitted to estimate hazard ratios (HRs) and 95% confidence intervals (CIs) for the relationship between BMI overweight/obesity duration, OWY/ OBY, and the risk of developing specific cancers. Overweight/obesity duration and OWY/OBY were treated as continuous, time-varying covariates.
Age as the time dimension: Pearson JD, Morrell CH, Brant LJ, Landis PK, Fleg JL. Age-associated changes in blood pressure in a longitudinal study of healthy men and women. J Gerontol A Biol Sci Med Sci. 1997;52(3):M177-83.
a. Research question: to assess an association between aging and increases in blood pressure among healthy men and women.
b. Study sample: Participants from the Baltimore Longitudinal Study of Aging (BLSA). 1,307 men (age 17-97) and 333 women (age 18-93) who have been screened for health problems or medications that affect blood pressure.
c. Longitudinal design: cohort followed for 32 years
d. Analysis approach: they used longitudinal mixed-effects regression models to estimate the age-associated changes in blood pressure among the men and women in their population.
Time since study enrollment as the time dimension: Riley E, Evans J, Hahn J, Davidson P, Lum P, Page K. A Longitudinal Study of Multiple Drug Use and Overdose among Young People Who Inject Drugs. Am J Pub Health. 2016; 106(5):915-7.
a. Research question: is there an association between multidrug use and nonfatal overdose among young people in San Francisco?
b. Study sample: 173 injection drug users younger than 30 years old living in SF
c. Longitudinal design: potential participants were recruited between April 2012 and February 2014 to participate in a longitudinal cohort study of young injectors, UFO (“U Find Out”). Participants completed a visit every 3 months in which they self-reported overdose events and a set of other factors including polysubstance use through an ACASI (computerized interview) and HCV test.
d. Analysis approach: The research team used longitudinal logistic regression models fir by generalized estimating equations to estimate the effects of polysubstance use on overdose rates.
One other possible time dimension (time since retirement): Wu C, Odden M, Fisher G, Stawski R. Association of retirement age with mortality a population-based longitudinal study among older adults in the USA. BMJ. 2015. https://ucsf.idm.oclc.org/login?url=http://dx.doi.org/10.1136/jech-2015-207097
a. Research question: Is there an association between retirement age and mortality among healthy and unhealthy retirees and is this relationship modified by SES.
b. Study sample: 2,956 retirees from the Health and Retirement Study
c. Longitudinal design: tracked retirement age and reason for retirement from the Health and Retirement Study. Participants were followed from 1992-2010 – during which time 2956 participants retired. They were tracked over the course of the study and death was recorded.
d. Analysis approach: association of retirement age with all-cause mortality was analyzed using the cox model, adjusting for sociodemographics.
1. Age as the time dimension:
Albanese, E., Matthews, K.A., Zhang, J., Jacobs, D.R., Whitmer, R.A., Wadley, V.G., Yaffe, K., Sidney, S. and Launer, L.J., 2016. Hostile attitudes and effortful coping in young adulthood predict cognition 25 years later. Neurology, 86(13), pp.1227-1234.
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4818565/pdf/NEUROLOGY2015672964.pdf
i. Research Question: Are high levels of hostility and effortful coping in early adult life (under the age of 30 years) associated with lower cognitive function in early/midlife?
ii. Study Sample: The Coronary Artery Risk Development in Young Adults Study (CARDIA)
iii. Longitudinal Design: CARDIA sampled 5,115 participants between the ages of 18 and 30 from 4 US centers in 1985-1986, and re-examined the cohort at years 2, 5, 7, 10, 15, 20, and 25. For this study, the analysis was restricted to the 3,126 participants with data for all covariates at year 25 (89% of the sample).
iv. Analysis Approach: Multivariable linear regressions were used to assess the association of midlife (year 25) cognitive function to quartiles of psychological characteristics, adjusting for age, race, and sex.
2. Time since study enrollment as the time dimension:
Khondoker, M., Rafnsson, S.B., Morris, S., Orrell, M. and Steptoe, A., 2017. Positive and Negative Experiences of Social Support and Risk of Dementia in Later Life: An Investigation Using the English Longitudinal Study of Ageing. Journal of Alzheimer's disease: JAD.
i. Research Question: Do close social relationships in later life affect risk of developing dementia?
ii. Study Sample: The English Longitudinal Study of Ageing (ELSA), a panel study of community dwelling individuals over the age of 50 in England
iii. Longitudinal Design: ELSA participants were surveyed every two years using computer-assisted personal interviews for up to six waves covering a period of 10 years. The outcome was defined as time-to-dementia from the start date of the ELSA study.
iv. Analysis Approach: Because data collection occurred in two year intervals, neither time-to-dementia for incident cases nor censoring time were known. Both were treated as interval censored in proportional hazards regression models to estimate the risk of dementia by the end of the follow-up period.
3. Time since (and prior to) installation of Pokémon GO as the time dimension:
Howe, K.B., Suharlim, C., Ueda, P., Howe, D., Kawachi, I. and Rimm, E.B., 2016. Gotta catch’em all! Pokémon GO and physical activity among young adults: difference in differences study. BMJ, 355, p.i6270.
https://www-ncbi-nlm-nih-gov.ucsf.idm.oclc.org/pmc/articles/PMC5154977/
i. Research Question: Does playing Pokémon GO affect the number of steps taken daily?
ii. Study Sample: Survey participants of the Amazon Mechanical Turk (MTurk), a recruitment of online workers who receive a small compensation for completing tasks, between the ages of 18 to 35 residing in the US with an iPhone 6 series smartphone.
iii. Longitudinal Design/Analysis Approach: Participants were asked to upload screenshots of their number of steps daily while carrying their iPhone for each day, which are recorded by default in the Health application of the iPhone 6 series, as well as screenshots of their Pokemon GO application, which shows the installation date and was independent of survey recruitment. The authors conducted a difference in difference analysis by comparing the average number of steps taken each day for each of the four weeks prior to installation compared to the steps during each of the six weeks after installation in players of Pokémon GO and non-players, respectively. For non-players, the median date of installation among players was used as the cutoff for the difference-in-differences analysis. Multivariate regression was used, and the estimate of change in number of steps was obtained using an interaction indicator for playing status and week after installation of the game.
Assignment:
Defining the time dimension is a fundamental challenge in longitudinal data analysis. The most common choice is age or time since study enrollment, however, for many questions other time dimensions are relevant, for example grade in school, or time before or after stroke incidence, or time since release from prison. Identify an article in the applied literature using a longitudinal analysis based on age as the time dimension, time since study enrollment as the time dimension, and one other possible time dimension (ie, not age or time since study enrollment). For each study, briefly describe the research question, the study sample, the longitudinal design, and the analysis approach. Please post links to the studies.
Safety and efficacy of early amino acids in preterm <28 weeks gestation: prospective observational comparison.
Research question: Kotsopoulos et al evaluated the safety and efficacy of early administration of amino acids among preterm infants < 28 weeks gestational age.
Study Sample: the study sample included two groups of preterm infants (each of sample size 54), born at Mount Sinai Hospital in Toronto who were admitted to the Neonatal Intensive Care Unite. One group of children received amino acids parenterally (via intravenous catheter) using standard practice (at about 12-30 hours after birth), and the second group was recruited a year later, after practice changed, and amino acids were administered < 6 hours after birth.
Longitudinal Design: Prospective, before-after (standard change in practice) comparative study. Children were followed over time (days of life).
Analysis Approach: X2-test comparing mean differences of endpoints; t-test or Wilcoxon sum test depending on distribution of continuous variables.
http://www.nature.com/jp/journal/v26/n12/pdf/7211611a.pdf
Cardiovascular Disease Mortality Among Breast Cancer Survivors.
Research question: Bradshaw et al. estimated the relative burden of death due to cardiovascular disease among breast cancer survivors compared to women without breast cancer.
Study Sample: Study included women from the population-based Long Island Breast Cancer Study Project (case-control sample). 1,411 women without breast cancer, and 1,413 breast cancer survivors were included in this study.
Longitudinal Design: women were followed prospectively over time until 2009 (to assess mortality and cause of mortality) since study enrollment in 1996-1997.
Analysis Approach: Hazard ratios were calculated using Cox regression for overall mortality, and a competing risks analysis for cardiovascular disease-specific mortality; sub-distribution hazard ratios for cardiovascular disease mortality was also estimated using the Fine-Gray model.
Beverage purchases from stores in Mexico under the excise tax on sugar sweetened beverages: observational study.
Research question: Colchero et al. investigated the effect of purchases of beverages from stores in Mexico after the sugar-sweetened beverage tax was implemented in 2014.
Study Sample: Sample size was 6,253 households which participated in Nielsen Mexico’s Consumer Panel Services between January 2012 and December 2014.
Longitudinal Design: Purchasing data was obtained from the 6,253 households and provided 205,112 household months between January 2012 and December 2014 (tax started January 1, 2014).
Analysis Approach: Authors used a difference in difference fixed effects model, comparing predicted volumes of mL/capita/day of different beverages to the counterfactual consumption, using pretax trends, had the tax not passed.
Article with age as the time dimension: “Identifying children with excess malaria episodes after adjusting for variation in exposure: identification from a longitudinal study using statistical count models”
https://bmcmedicine.biomedcentral.com/articles/10.1186/s12916-015-0422-4
In this article, the authors are interested in identifying why some children are more susceptible to symptomatic (vs asymptomatic) malaria than others. The study sample comprises 2 cohorts of about 1500 children followed over 15 years on a weekly basis in coastal Kenya. In this longitudinal study, the time dimension used is age and children where enrolled at birth and followed for 5 to 15 years. Analytically, the idea was to identify children with excessive number of symptomatic malaria and age-matched (nested case-control) them to children with “average number of symptomatic malaria” but similar exposure (environmental exposure index based on overall transmission intensity). Higher parasite densities and hemoglobin genotype were found as risk factors.
Article with time since study enrollement as the time dimension: “Malaria morbidity and pyrethroid resistance after the introduction of insecticide-treated bednets and artemisinin-based combination therapies: a longitudinal study”
https://www.ncbi.nlm.nih.gov/pubmed/21856232
In this study, the authors used longitudinal data collected daily between 2007 and 2010 in a cohort of 405 people from a village in central Senegal resulting in 17858 person-months of follow-up. The research question aimed at assessing the impact of mass distribution to the entire village of long-lasting insecticides treated bed nets (LLIN) in 2008. Time since enrollment in the study was used as the time dimensions but was sub-divided also considering time since treatment and time since 2010 to understand the long term effects of the intervention. The analysis consisted in comparing incidence rates before intervention, after intervention and 2 years after intervention to assess the short-term and long-term effect of LLINs.
Article with time since study enrollement as the time dimension: “Intervals to Plasmodium falciparum recurrence after anti-malarial treatment in pregnancy: a longitudinal prospective cohort”
https://malariajournal.biomedcentral.com/articles/10.1186/s12936-015-0745-9
In this study, the authors were investigating the effect of different malaria treatment combinations on recurrences of malaria episode among pregnant women. In this longitudinal study, a cohort of 700 pregnant women was followed between 1994 and 2009. 481 novel episodes of malaria and 428 recrudescent episodes were observed. Time since first malaria episodes was used as the time dimension. Analytically, the geometric mean number of days to recurrence was compared by treatment groups using log-linear regression.
This was more difficulty than I thought it would be because it seems like sometimes the design can be both time since enrollment and something else? Looking forward to seeing others examples.
Murray, L. (1992). The impact of post-natal depression on infant development. The Journal of Child Psychology and Psychiatry. 33(3)
http://onlinelibrary.wiley.com/doi/10.1111/j.1469-7610.1992.tb00890.x/abstract
The research question: The principal aim of the investigation was to compare the cognitive, social and emotional development of infants of mothers with unipolar, non-psychotic postnatal depression with that of infants of non-depressed mothers.
The study sample: 702 women presenting on the postnatal wards of the Cambridge maternity hospital during the period February 1986-February 1988 who met the inclusion criteria relating to health history and birth outcome.
The longitudinal design: The exposure was maternal depression and the outcome was cognitive and language development of the infant. While the time component was age of the infant, the exposure was measured in post-partum months and not age of the mothers. Infants were assessed at 9 and 18 months. Mothers were assessed at 6, 12, and 18 months. A subsample of mother-infant diads were assessed every 2-3 months.
The analysis approach: Mothers whose depression developed later in the infants first year were excluded. Comparisons between infants were made based on exposure to maternal postpartum depression, style of interpersonal contact associated with depression, history of depression or postpartum depression. Covariates included maternal education, employment, social class, paternal psychiatric history, marital friction, infant gender. Regression analysis was used to examine the effects of the duration of depression. All two-way interactions were tested.
Becoming married and mental health: a longitudinal study of a cohort of young adults
https://www.jstor.org/stable/353978?seq=1#page_scan_tab_contents
The research question: Do becoming and staying married enhance mental health after controls for premarital rates of disorder? Do becoming and staying married benefit the mental health of women as well as women when male as well and female related outcome variables are considered? How are becoming and staying married linked to mental health?
The study sample: Subjects were identified through a telephone survey using random digit dialing, between 1979-1981 from counties in New Jersey. Quota sampling guided initial phase of obtaining subjects and targets were 450 subjects each year evenly divided by age (three groups) and gender.
The longitudinal design – Time since enrollment. Surveys were administered at the time of enrollment (T1), 3 years later (T2), and three years after that (T3). 91% of the sample completed the final assessment 7 years after T3 (T4).
The analysis approach: Scales for depression and alcohol use were administered at every time. Marriage was coded as a binary variable. Analysis included comparison of gender and marital status in depression and alcohol use scales. Regression models to predict marital status and depression at T4, marital status and depression at T3 on depression at T4, and marital status, depression at T3, and gender on depression at T4 were all fitted.
Filippi et al. (2007). Health of women after severe obstetric complications in Burkino Faso: a longitudinal study. Lancet. (370). 1329-37.
https://www.ncbi.nlm.nih.gov/pubmed/17933647
The research question: How do severe obstetric complications affect a range of health and other outcomes in the year after the end of pregnancy in hospitals in Burkina Faso?
The study sample: a prospective cohort of women with severe obstetrical complications recruited in hospitals when their pregnancy ended with a live birth, perinatal death, and a lost pregnancy. For all cases of obstetric complications, two unmatched controls were sampled from among normal births concurrently as the cases with sever obstetric complications were recruited.
The longitudinal design – Time since end of pregnancy. Baseline data was collected at 3 days post-partum, followup interviews at 3, 6, and 12 months after pregnancy ended.
The analysis approach: The sample size was determined with the goal of detecting an approximate 10% risk difference between complicated and uncomplicated cases for outcomes with prevalence between 25% and 75%, and the power between 80% and 90%. Determining differences in mortality were not the objective. Women were stratified into groups by birth outcome and unadjusted and adjusted odds ratio were calculated. Fisher's exact test to compare death rates between women with severe complications and those with uncomplicated births. Cox regression was used to calculate hazard ratios.
Choose at least 3 distinct data sources (e.g., ARIC, HRS, death certificate data, NHS, etc), and give an example of a research question (e.g., a hypothesis about the effect of a specific exposure on a specific outcome) you consider the study exceptionally strong to address. For each, provide an example of a research question you consider the design very weak to address. Explain why the data source is strong or weak for each question. Do not just discuss the questions addressed in the readings, think of new questions, preferably things you might be interested in. This is not supposed to be a commentary related to the substantive questions in the readings: the goal is to focus on the pros and cons of various data sources. For hypotheses each study would not be well equipped to address, if possible describe another study that could address the hypothesis.
California Health Interview Survey (CHIS): This is the state’s data for public health surveillance and tracking changes in health insurance coverage and eligibility for healthcare programs. It asks questions about a broad range of health conditions and behaviors, mental health, health insurance, healthcare use and access.
Strong question: How has the prevalence of asthma/asthma symptoms changed over time in different areas of the state as local air quality changes?
Alternative strong: Looking forward into the future, how would the passing of laws affecting health care coverage and uptake affect health (looking at disparate outcomes of diabetes, cancer, asthma, mental health, etc.)?
Weak question: Investigating disparities in SES status and presence and control of diabetes.
There is a very low survey response rate to the random digit landline and cell phone dialing. Tracking changes over time is more likely to be successful than trying to ascertain absolute prevalence measures of diseases. There is a strong selection bias of selection into the survey, and those who feel they have something to say, or who have more time on their hands are more likely to participate, and these are a fairly selective subsample of the population. Those who are very sick are unlikely to participate, but those who are well and busy working and with families are also less likely to participate. Elderly and unemployed individuals have higher participation rates. Furthermore, these surveys do not include individuals in nursing homes, a population with a high risk for diabetes. As elderly individuals are also some of the poorer segments of the population, this can lead to an underestimate of the effect.
Better dataset- Kaiser Permanente patient data would not be subject to the same concerns of selection bias.
NHAMCS (National Ambulatory Health Care Data):
Strong question: Are the number of sports related concussions/other injuries increasing or decreasing in youth (also in men and women)? – this information tends to be coded well and is relatively consistent over time
Weak question: What is the prevalence of asthma?
NHAMCS does not measure individuals, only occurrences, this leaves people open to being double counted. Thus, if a person visits an emergency multiple times, their data would be counted more than others. Thus this dataset is extremely powerful in measuring the causes of uses of ambulatory care and changes in episodic disease occurences, but is completely inappropriate for chronic disease measurements in the general population.
Better dataset - NHIS
National Vital Statistics System: mortality data that is collected is comparable across a wide-swath of demographic and geographic regions. They also go back a long time in history collecting information on those who pass, age, illness, etc.
Strong question: Understanding the distribution of the burden of breast cancer mortality, trying to understand if certain populations or states or regions of the country have a higher than average fatalities from the disease. This could help to identify areas where studies should be directed to investigate disparities in breast cancer management and inform areas for further intervention.
Weak question: Finding out if there is excess burden of breast cancer in certain populations or if there are disparities in the successful management of breast cancer. Only deaths are recorded, so those who did not die from the disease will not be recorded in this database. Thus, since it is unknown what the full denominator was for the full population of breast cancer patients, it is not possible to determine whether disparities are present in management or in incidence of breast cancer.
Better dataset: statewide cancer registries, which include breast cancer patients. This would allow for investigation into disparities in incidence and trying to determine causes.
HRS-Health and Retirement Survey
Strength: Does spousal support influence cancer prognosis? HRS collects information on relationships (spousal and other family members) and specifically asks about whether a participant’s cancer has gotten better, worse, or stayed the same.
Weakness: Does social media use influence cancer screening?
Better dataset: Health Information National Trends Survey (HINTS) is a nationally representative survey that provides information on women’s attitudes and behaviors towards cancer screenings. HINTS collects information on most sources of health information and screening practices/behaviors.
PSID: Panel Study of Income Dynamics
Strength: Does parental SES influence cardiovascular risk in daughters?
Panel study of income dynamics collects information on a multitude of social factors across generations (grandparents, parents, and children) so SES for both parents can be constructed. PSID also collects information on a variety of health outcomes including cardiovascular disease.
Weakness: Does BPA exposure in adolescence increase breast cancer in post-menopausal women?
Better dataset: Breast cancer and the environment study has collected information on environmental exposures for pre-adolescent and adolescent girls. Though it is too early to collect information for cancer (as the average onset of breast cancer is roughly 62 in the United States) the dataset is useful for future analyses.
SEER-Surveillance, Epidemiology, and End Results
Strength: Are Black-White cancer disparities worse in southern states?
SEER collects cancer data from across the United States. Though they do not collect data from all 50 states, they do collect data from southern states including Georgia, New Orleans, and Louisiana. Black-White disparities in these states may be compared with states such as California, Connecticut, Iowa, and Hawaii.
Weakness: Does neighborhood SES influence cancer risk? Though SEER provides Zip code data, it does not provide census tract or block group information for cancer cases. Previous studies have shown that census-tract provides stronger effect estimates when assessing neighborhood effects and cancer incidence.
Better dataset: The Cancer Prevention Institute of California (CPIC) which houses the Northern California cancer registry has created a neighborhood SES index (Yang index) for the state of California which can be paired with cancer cases.
HealthCore
HealthCore Integrated Research Database (HIRD) has one of the largest commercially insured population databases in the nation. Information is available on nearly 60 million individuals from multiple health plans across the U.S and over 175,000 physicians. 44+ million private U.S. commercial lives with medical and pharmacy claims, spanning 14 states from Anthem’s affiliated health plans dating back to 2006. This would include 14 WellPoint affiliated Blue Cross and/or Blue Shield licensed plans and 2 non-WellPoint affiliated Blue Cross and/or Blue Shield licensed plans. Regions represented include: Northeast, Southeast, Mid-Atlantic, Midwest, Central, and West. Lab results are also available for 13+ million lives integrated with claims data as well as clinical oncology data. Other data captured includes demographic data, enrollment data, outpatient prescription data, outpatient diagnosis data and procedures, hospital discharge data, deaths – captured using National Death Index or Social Security Administration’s Death Master File (deaths in the hospital may also be captured), and facility information.
Linkages available include standing linkage with SSA Master Death files and National Death Index, state cancer registries, state immunization registries, patient and provider surveys.
Strong RQ – What are the treatment patterns and cost of erythropoiesis stimulating agents in patients with cancer receiving myelosuppressive chemotherapy?
This sample includes medical and pharmacy claims data, clinical oncology data, lab results covering 13+ million individuals, and also a wide geographical representation from the Northeast, Southeast, Mid-Atlantic, Midwest, Central, and West. Therefore, it is advantageous to use this data source due to broad geographic representation for assessing outcomes in specific populations and comparison of regional practice pattern variations. Also, there would be the ability to validate claims versus EMR data for a subset of the population without going to paper charts and ability to validate claims with EMRs in other regions and provider/hospital systems (where EMRs are available). Also, this data source would be good to study broad population representative of the commercially insured in the United States.
Weak RQ – What are the inpatient treatment patterns and cost of erythropoiesis stimulating agent treatment in elderly patients ≥65 years of age with cancer receiving myelosuppressive chemotherapy?
This would not be a great research question to answer using this data source because complete capture of elderly care is also not possible, since Medicare is the primary payer, however it does capture those on Medicare Advantage Plans. Moreover, there is no linkage available to Medicare FFS data, so capture of elderly patients (≥ 65 years old) most likely will be incomplete. Also, formularies can be specific to the network and may not be representative of the population treatment patterns.
The Surveillance, Epidemiology, and End Results (SEER) – Medicare linked Database
The SEER-Medicare data reflect the linkage of two large population-based sources of data that provide detailed information about Medicare beneficiaries with cancer. The data come from the Surveillance, Epidemiology and End Results (SEER) External Web Site Policy program of cancer registries that collect clinical, demographic and cause of death information for persons with cancer and the Medicare claims for covered health care services from the time of a person's Medicare eligibility until death. The linkage of these two data sources results in a unique population-based source of information that can be used for an array of epidemiological and health services research.
The Surveillance, Epidemiology, and End Results (SEER) Program of the National Cancer Institute (NCI) collects data on cancer incidence and survival from population-based cancer registries throughout the United States. Data collection began in 1973 with a limited amount of registries and has expanded over time to include registries that cover 28% of the United States population (updated annually and data is available through 2011). The SEER Program registries routinely collect data on patient demographics, primary tumor site, tumor morphology and stage at diagnosis, first course of treatment, and follow-up for vital status. The SEER data do not capture information about surgery and radiation provided past four months of diagnosis, nor is there information about recurrence or metastasis that is detected subsequent to the initial diagnosis.
The Medicare data available as part of SEER-Medicare include claims from hospital, outpatient, physician, home health, and hospice providers. Each Medicare file varies in the data elements included and the types of procedure and diagnostic codes used, either International Classification of Diseases (ICD-9) codes for procedures and diagnoses or HCFA Common Procedure Coding System (HCPCS) codes for procedures. HCPCS are the AMA's Common Procedure Terminology codes (CPT-4) with additional codes used exclusively by CMS. In general, all Medicare files have fields for race, sex, and date of birth or age, the date(s) of service, diagnostic codes (for many files), and procedure codes in addition to the amounts for charges and reimbursement. In addition, every Medicare file contains a provider identification number for the hospital or physician.
Strong RQ – What is the comparative effectiveness of carboplatin and paclitaxel (with and without) bevacizumab in older patients with advanced non-small cell lung cancer?
It is advantageous to use this data source to answer this research question because the linkage to SEER data provides the identification of incident cases (which is not possible from the Medicare claims) with detailed site and stage reporting and information about the cause of death. The Medicare data also offers a longitudinal perspective, making it possible to look at medical services before, during, and after diagnosis. Claims before diagnosis can also be used to measure preexisting co-morbidities that might influence treatment decisions and cancer screening, to a limited extent. Also, this data source is appropriate as Medicare would include health insurance information for the elderly (65+), as well as those with end-stage renal disease and some disabled and is pertinent to this population.
Weak RQ – Identification of cancer recurrence in colorectal and breast cancer patients (in the general population).
Neither SEER nor Medicare collect information about recurrence. Medicare claims can be used to identify recurrences indirectly only if the patient receives treatment for the recurrence. It is not possible to identify recurrence through diagnosis codes only. Investigators who have used claims to identify treated recurrence have used an approach that involves reviewing claims longitudinally for cancer related treatment (cancer-related surgery, chemotherapy, RT) after the initial care period. The later surgeries should be selected carefully (ex. hepatic resection for a colon cancer patient likely shows recurrence while a hemicolectomy may be for disease recurrence or adhesions).
Using a treatment-based approach is dependent on the patient receiving additional treatment in the event of a recurrence. However, many elderly patients are not offered/decline additional treatment if their cancer recurs. While it is possible to use SEER-Medicare to identify patients with "treated recurrence", this approach can miss a large number of cases and the cases identified are a biased sample of the elderly.
Also, Medicare data limited to those over the age of 64 and the disabled and oral medications are not covered (prior to Part D data). Moreover, it is also important to note that SEER registry areas may not be totally representative of patterns of care and there is also a long-time lag to obtain data.
Hunger in America (HIA 2014)
The Hunger in America survey, the largest of its kind, is a series of quadrennial studies providing comprehensive demographic profiles of people receiving food assistance through the charitable sector, and offers in-depth analyses of the partner agencies in the Feeding America Network. Feeding America is a nationwide network consisting of 200-member food banks serving all 50 states, the District of Columbia, and Puerto Rico. The Feeding America network of food banks provides food assistance to an estimated 46.5 million Americans in need each year, including 12 million children. Data collected from these surveys, last completed in 2014 (HIA 2014), help guide the development of programs and solutions that improve FI for individuals and their households and inform public awareness and policy development for addressing hunger in the United States. Of note, this survey was available to be administered in English, Spanish, Vietnamese, Russian, and Mandarin, and could be completed either by the client independently or with assistance from a proxy.
The HIA 2014 Client Survey was fielded from April, 2013 through August, 2013, and was implemented through a force of data collectors recruited by each participating food bank. Based on pre-testing, the survey was revised, finalized, and then programmed into a computerized version of the survey to be implemented using a touchscreen tablet device and Audio Computer-Assisted Self-Interview (ACASI) technology. A second pretest was then performed on the digital/audio survey, and client responses were used to make additional improvements before the final survey was created. HIA 2014 represented the first HIA client survey utilizing this computerized technology.
HIA 2014 aimed to collect information directly from Feeding America clients, and to describe the numbers and characteristics of clients using the networks for charitable food assistance. Because conducting interviews with every client served by every program over an extended period of time was not feasible, probability sampling was used to select a subset of programs at which data collection should occur, the days on which data collection should occur at those programs, and the clients who should be asked to complete the survey. As it applies to HIA 2014, probability sampling is an approach in which each client has a known, positive chance of being selected to complete the survey. This technique makes it possible to use the sample to estimate population-level information. As the full population of Feeding America clients in the US is unknown, it was not possible to select from a known list of clients, as is sometimes possible in probability sampling. Consequently, the HIA 2014 was designed with a multistage design to facilitate selection of the probability sample.
At least 6,000 data collectors were trained and registered to carry out client data collection. Data collectors followed a prescribed study plan in order to select a random sample of clients at nearly 12,500 assistance programs across the Feeding America network.
Strong RQ – Are individuals seeking assistance at food pantries who have a personal or household history of diabetes mellitus able to obtain diabetes appropriate foods in 2014 (cross-sectional study)? Is the prevalence of individuals with a personal or household history of DM seeking assistance at food pantries requesting and unable to obtain fresh fruits, vegetables, proteins, grains, and dairy the same as the prevalence of individuals without a personal or household history of DM seeking assistance at food pantries?
This is an appropriate database for this research question because the target population would be adults seeking assistance at food pantries with diabetes or at risk for developing diabetes. Moreover, the presence of food insecurity (FI) makes diabetes mellitus (DM) management more challenging; a common coping strategy is shifting dietary intake towards cheaper and more obesogenic foods. There is clearly a role for dietary education in the management of DM, and having knowledge and availability of proper nutrition is important to improving glycemic control and preventing the development of T2D in at-risk individuals.
This survey is also advantageous because participants surveyed included grocery (i.e. food-bank operated pantry programs, food panties, community gardens, school pantries); meal (i.e. community kitchen, group home, shelter, transitional housing); food-related benefits (i.e. WIC outreach, SNAP); and non-food programs (i.e. financial assistance, GED programs, health clinics, job training programs).
Weak RQ – Are the types of food requested and unable to be obtained from assistance programs in the US between those with a personal or household history of DM and those with no personal or household history of DM associated with increased mortality rates due to DM? Is the lack of access to proper cooking and refrigeration amenities associated with increased mortality due to DM?
This would not be a great research question to answer using this data source because of the cross-sectional nature of the data source. Moreover, mortality record data was not captured and it would not be possible to address this question. One would also need to validate cause-of-death determinations (via death record information with hospital records, etc).
National Family Growth Survey
https://www.cdc.gov/nchs/nsfg/
This source gathers data from a national sample on pregnancy, birth control, men's and women's health, marriage and divorce, and infertility.
Question this data is strong to address: what factors are associated with women having received birth control counseling?
This data set reports that about 18% of women report having received counseling even though 33% report having a method. Predictors of interest would be SES, kind of insurance, kind of service provider, OB history all of which are available in this data set.
Weak to address: What was the quality of the contraceptive counseling received?
Quality data is not included in this survey. This data could be used in conjunction with tools which measure care quality.
Vital Statistics
https://www.cdc.gov/nchs/nvss/
This national system collects all marriage, death, marriage, divorce, and fetal death data.
Question this data would be strong to address: What was the impact on mortality rate related to chronic diseases (CHD, DM) among vulnerable populations after implementation of Obamacare?
This is a national data set with yearly data, so it would be straightforward to compare data before and after a policy change which happened on certain date. The coding is standardized so little bias would be present. Temporality would be clear.
Question this data would be weak to address – What is the impact on cost to patients who die due to chronic diseases among vulnerable populations after implementation of Obamacare?
There would not be a reliable measure to estimate cost related to care prior to death relayed on a death certificate. Due to the wide range of healthcare costs in the market, estimates drawn from cause of death would be unreliable. More specific data could be used to answer the question or direction of cost based on cases may be estimate-able from this data (more or less cost because of more or less cases).
CDC Wonder
Robust public health data from the CDC.
Question this data is strong to address: Is there a differences in day of the week mortality between rural and urban locations?
Multiple causes of mortality data with geographic descriptions and (interestingly) day of the week data, crude and adjusted rates are available through the system.
Question this data is weak to address – How do the common causes of death in 1978 compare to cause of death in 2008?
This data set is weak to address this question because of changes in the coding system used to describe cause of death. Therefore the findings may suffer from measurement bias. Any methods of standardizing the data may result in the loss of specificity.
1. CDC Birth certificate data: The Centers for Disease Control and Prevention’s (CDC) National Center for Health Statistics (NCHS) has been collaborating with colleagues in the State vital statistics offices to revise the certificates of live birth and death and the report of fetal death. This process is generally carried out every 10 to 15 years. Prior to 2003, the most recent revisions in effect were implemented in 1989. The 2003 revisions have recently been approved by the HHS Secretary and are going into effect in the States. Some States began the revision process in 2003, but full implementation in all States will be phased in over several years. A critical component of the recommendations for this revision focuses on fundamental changes in the way that data are collected, especially for births. Partly as a consequence of these recommendations, States are engaged in re-engineering their vital statistics systems as they implement the revisions.
https://www.cdc.gov/nchs/nvss/vital_certificate_revisions.html
Strong research question: What is the state-level prevalence of cesarean section?
The CDC birth certificate data consistently asks at the state level about several items, including mode of birth. Mode of birth is less prone to misclassification errors.
Weak research question: Does home birth increase the risk of poor neonatal outcome?
Because of inconsistencies in state reporting of home birth, most birth certificates cannot distinguish between planned and unplanned home birth, or identify home-birth transfers. These create problems of misclassification of the exposure. For instance, unplanned home births can have worse neonatal outcomes (often due to poor prenatal care), and home-birth transfers can also have worse outcomes (because they transferred for a medical reason). Thus, national studies of US home births have been limited in making causal claims. The largest cohort study to date on home birth, the UK Birth Place study, had full ascertainment of home births (better vital statistics in regards to home birth in the UK) and women there need to “book into” a home birth early in pregnancy (thus able to do an intention-to-treat analysis).
2. Diethylstilbestrol Adenosis Project (DESAD) - The DESAD began in 1974 at Baylor College of Medicine, Gundersen Clinic, Massachusetts General Hospital, the Mayo Clinic, and the University of Southern California. The DESAD, the largest DES cohort, included 4,014 DES Daughters and 1,033 unexposed women. Exposed women had documented evidence of DES exposure through review of prenatal records or by physician referral. The DESAD was assembled to conduct studies to determine if DES Daughters were at an increased risk for health problems related associated with their exposure to DES (Labarthe, 1978).
Strong research question: Does in utero DES exposure predict the development of endometrial cancer?
As the largest cohort of DES exposed girls, this is an ideal study question to examine the later-in-life outcomes of in utero DES exposure.
Weak research question: Did DES exhibit any effect-modification with environmental exposures on the outcome of clear cell adeonocarcinoma of the cervix?
Because the DES cohort was assembled only with detailed information gathered from prenatal records about DES, it would be difficult to examine DES’s interaction with any other exposures.
3. NHANES: A primary aim of NHANES is to collect and compile data on the health and nutritional status of study participants through a repeated cross-sectional design. Particular emphasis is placed on data regarding the prevalence of major diseases and risk factors for diseases. Information is derived from participant interviews as well as physical and laboratory examinations. Minority populations are oversampled to produce reliable statistics.
Strong research question: Has there been a change in the prevalence of morbid obesity in African Americans between the last two NHANES surveys? Repeated cross sectional survey designs can give important information about population trends in health.
Weak research question: How does the risk of endometrial cancer for morbidly obese African American women change as they age? Because this design requires individual level data as it changes over time, repeated cross sectional design would not work.
Birth certificate data (U.S. Standard Certificate of Live Birth). This database is an administrative record of all births, as mandated by state and federal law. It includes information certifying the circumstances of the birth, and registers the child for administrative purposes (taxes and social security). Using form: https://www.cdc.gov/nchs/data/dvs/birth11-03final-ACC.pdf
Strong research question: Is prenatal care associated with pre-term birth or low birth weight?
These data elements are both collected on birth certificates, and not likely to be misclassified. A mother would report whether or not she had received pre-natal care, and pre-term birth and low birth weight can be recorded pretty accurately by a health care professional (whose position is also recorded). Furthermore, since prenatal care and birth weight/weeks are likely to be associated, further covariates such race and education level can be adjusted from data collected on the birth certificate.
Weak research question: How does psychosocial stress contribute to pre-term birth or low birth weight?
Same outcome, but it is very unlikely to get a reliable measure of psychosocial stress from a mother soon after she has given birth from data collected on a birth certificate. There are data elements for whether or not a mother is married, principal source of payment for this delivery, previous preterm birth, which could be weak proxies for psychosocial stress. We can change the research question to better align with these data elements, or we could conduct a case-control study looking at maternal stress and depression during pregnancy.
Google trends. A database that shows how often a particular search term is entered relative to the total search-volume globally, and in various languages. Data is available starting in 2004, and while some terms translate into other languages automatically (like ‘flower’ and ‘fleur’), others do not.
Strong research question: How did the announcement of the HPV vaccine increase awareness of the vaccine?
Strengths. Good for general awareness and popular trends, such as media coverage about the HPV vaccine, and associated terms. Because of the widespread use of Google, Trends is a good measure of popular interest and news.
Weaker research question: How did the announcement of the HPV vaccine increase vaccination rates of the vaccine?
Searching for information about the HPV vaccine would need to be highly correlated with seeking the vaccine from a provider. This question could not be answered alone from Google Trends. We could use medical insurance databases to see how many more women got the vaccine after the public announcement it was available in a natural experiment/regression discontinuity design.
Medical Expenditure Panel Survey. MEPS is a set of large-scale surveys of families and individuals, their medical providers, and employers across the United States starting in 1994.
Strong research question: How often do Americans seek preventive care services?
This research question is particularly well-suited for MEPS; data is collected on what health services Americans are using, how frequently they use them, the cost of these services, how these services are paid for, and the cost, scope, and breadth of health insurance held by and available to U.S. workers.
Weak research question: What is the incidence of cardiovascular disease?
MEPS seeks to enumerate medical expenditures from households, medical providers, and employers, and therefore records are aligned to reflect costs, not necessarily diagnoses. Depending on how correlated cardiovascular disease is with seeing a cardiologist, the Medical Provider Visits (MV) Section to households does ask specialty of seen medical providers. Therefore, we could look at the cost per household for cardiovascular-related health conditions, which may be more interesting.
I have selected data sources and matched them with my context – Africa and in the fields which I am interested in – sexual reproductive adolescent and child health
Data source: Verbal autopsy
Strong research question and why: What are the medical and socioeconomic causes of maternal deaths in Kenya or in an African context
Weak research question and why: What are the clinical practices, biological factors and socio-behavioral factors related to HIV-associated morbidity and mortality in the African context
Alternative design: prospective cohort design e.g. the African Cohort Study (AFRICOS) that enrolls a cohort of HIV-infected persons and follows them. Or use existing databases such as the International Epidemiological Databases to Evaluate AIDS (IeDEA) in East Africa which harmonizes data collected by geographically disparate, but representative, cohorts of persons infected with HIV or at risk for HIV (healthcare system‐based population).
The beauty about this IeDEA cohort is that one can generate inferences about the natural or treated history of HIV, particularly regarding uncommon exposures or outcomes for which large samples would be otherwise needed.
Discussion: Verbal autopsy is problematic with diseases that have few specific symptoms. For example HIV may present in multiple myriad presentations and with varying co-morbidities and for which there may be underlying disease progression factors such as chronic malnutrition that may not be easy to tease apart. On the contrary, for conditions with well defined cause of death lists e.g. maternal mortality, it is easier to collect standardised data with a reasonable degree of reliability.
Data source: health and retirement study (HRS): Prospective cohort data using the longitudinal cohort design such as the health and retirement study – over the years, it has been able to incorporate specific age cohorts such as the children of the depression (born 1924-1930), “war babies” (born 1942-1947), “early boomers” born in 1948-1953
Strong research question and why: What is the impact of early-life adversity (e.g. early childhood malnutrition,) on adult cognitive functioning in Kenya e.g. – can take advantage of the 1983 and 1989 famines in Kenya and look at the famine survivors who were between the ages of 0 and 8 at the time of the famine. I think there is an opportunity for a natural experiment here
Weak research question and why: What are the clinical practices, biological factors and socio-behavioral factors related to cognitive impairment among adults aged >55 years in Kenya
The reason why data from a data source such as the HRS may not be sufficient to answer this question is that because it is cohort based, it may not be able to provide important information for other subpopulations in different cohorts that are not included in the cohorts they are looking at. Or like for this question in Kenya, while we can draw important understanding of etiology course and outcome for this cohort aged 0.8 years at the time of the 1983 and 1989 famines, we cannot provide inferences for other groups outside this time frame
Alternative design: The alternative design here would be to have repeated longitudinal cohort design that is complemented by repeated cross-sectional surveys such as the 10-yearly census or 5-yearly demographic health surveys or 3-yearly multiple indicator cluster surveys
Discussion: Prospective cohort studies can be used to study multiple complex diseases and risk factors simultaneously over an individual's lifetime. Such studies have proved crucial in understanding the etiology, course, and outcome of diseases in other populations and have informed the design of prevention programs.
Unfortunately, longitudinal studies of the same individuals are not sufficient to provide a useful base for analyses of change in the process of aging for other subpopulations in different cohorts.
Data source: ARIC : uses both cohort and community surveillance.
Within my context, I think a similar example to this ARIC is The KEMRI/Centers for Disease Control and Prevention (CDC) Health and Demographic Surveillance System (HDSS) is located in Rarieda, Siaya and Gem Districts (Siaya County), lying northeast of Lake Victoria in Nyanza Province, western Kenya. The KEMRI/CDC HDSS, has approximately 220 000 inhabitants
Strong research question and why:
“What is the impact of malaria transmission reduction activities on the all-cause mortality rate in the local population undergoing community-based malaria control interventions.” Or can narrow it down further to pregnant women
The method is appropriate because
Weak research question and why
What are the factors associated with lower expectancy among persons in Kenya when Siaya county when compared to the rest of Kenya
The downside of the HDSS is that unlike the ARIC it is not generalizable since it is only located within the Siaya County. While the rationale for hosting the HDSS in Siaya is because Siaya has one of the lowest life expectancies in Kenya, this information is un-informative if it cannot be compared to other regions
Alternative design: It might be necessary to set up a similar HDSS within other areas in the country with higher life expectancies and less morbidity and mortality to allow for comparability
I. Prescription Drug Monitoring Program dataset (PDMP/CURES)
CURES (Controlled Substance Utilization Review and Evaluation System) is California’s statewide prescription drug monitoring program (PDMP) and is maintained by the Department of Justice. CURES is a database of all scheduled II, III and IV controlled substances prescribed by California prescribers. The database includes records of opioid prescriptions by provider, date, location, and patient. As of 2017 prescribers in California are required by law to check CURES every 4 months for their patients receiving opioid prescriptions. PDMPs have been scaled up in an effort to monitor potentially hazardous opioid prescribing, flag patients with potential opioid use disorder, and reduce prescription opioid
Strong research question: the effects of a provider-level academic detailing intervention on opioid stewardship efforts. This study targets a randomized list of high opioid prescribers in the top 10 overdosing counties in California with one-on-one academic detailing educational trainings. Our hypothesis is that a tailored academic detailing intervention on opioid stewardship will change providers’ prescribing practices in the following ways: decrease the number of opioid prescriptions and increase the number of buprenorphine prescriptions. The CURES database is well suited to answer this question because it is an up-to-date, comprehensive data source of all opioid prescriptions written in California. All prescriptions in the database are recorded with date and linked to providers. Exposures will be measured based on our own tracking of who has received academic detailing training from a list of MediCal prescribing providers. We will give DOJ the list of providers with the date on which they were detailed. DOJ will return to us a de-identified list of those providers with the number of opioids and buprenorphine prescriptions written 6 months before and after the educational intervention so we can conduct a pre/post analysis. We will include a comparison group of providers in each county that are also high opioid prescribers but who did not receive the academic detailing intervention.
Weak research question: CURES only has the ability to answer questions related to provider-level activity – it is therefore not useful in drawing any patient-level inferences because it does not record information at the patient-level and is not currently linked to other datasets that would allow for that type of analysis. Questions this dataset cannot answer include: do changes in opioid prescribing at the provider-level reduce opioid-related ED visits for patients with long-term pain? Do mandatory PDMP checks by providers improve patient pain-related outcomes? Or, does switching from a full agonist opioid to a partial agonist like buprenorphine decrease pain and increase function among chronic pain patients? While this is useful, CURES data needs to be merged with additional datasets (like ED records or medical examiner databases) in order to fully understand the utility at the patient-level of interventions aimed at changing opioid prescribing practices.
II. United Nationals Office on Drugs and Crime (UNODC) database
The UNODC Drugs and Crime database collects data on drug use, trafficking, retail/black market opioid prices, production and diversion. The data are derived from results of annual UNODC national surveys: the Annual Report Questionnaire, the Individual Drug Seizure report, and the UN Survey on Crime Trends and Operations of Criminal Justice Systems. The UNODC maintains the database to increase cross-national data comparability and to produce regional and global estimates on drug trends.
Strong research question: This database can be useful to look at changes in country’s opioid consumption across time related to changes in national legislation. A within-country comparison manages the problems with the data mentioned below as long as countries employs consistent data reporting processes over time (E.g. uses the same operational definition for “illicit opioid use”, and the same mechanisms of collecting and reporting data). A strong research question would be related to what effects new legislation could have on opioid consumption and distribution at the national level. For example, we can compare national opioid consumption pre and post a law enforcement policy related to interdiction, or pre and post the establishment of nation-wide harm reduction programs such as the legality of opioid substitution therapy to assess the effect of the new policy. This data is aggregated at the national level so we will still not be able to assess changes at the individual level (opioid consumption is based off of national retail consumption rates from supply requests). Even with this limitation, from a policy perspective it is still useful to assess changes in opioid at the national level.
Weak research question: The opioid consumption data recorded by the UNODC does not actually represent the quantity of opioids used at an individual level, or distributed at the provider/clinic/hospital level, nor does it give insight into geographic dispersion of opioid dissemination e.g. rural vs urban opioid consumption. Additionally, data is not highly standardized across countries (e.g. different countries have different definitions of “illicit opioid use” and different mechanisms for collecting and reporting data) which makes it difficult to draw cross-national comparisons. Research questions related to comparing opioid consumption trends between countries is going to be flawed or infeasible with this datasets. Therefore, we cannot answer questions such as: Are opioid consumption rates by country associated with national statistics such as GDP, open versus closed democracy, criminalization of drug use, life expectancy, etc)?
III. Death certificate data
Death certificate data from the coroners’ reports record all pertinent death information for individual decedents (e.g. primary and secondary causes of deaths, age, race, gender, address/location, medical comorbidities etc). The difference between a strong and weak study design here could be based merely on the degree of specificity we’re looking for in our results. Using data from coroners’ reports is useful in determining broad categories death (e.g. drug poisoning) but it is not necessary useful at specificities within each category (causal opioid involved in opioid overdose).
Strong research question: Assessing changes in opioid overdose rates based on geographic and demographic criteria like gender, race, age, and resident and death location. Based on the geographic information provided, we could answer questions such as: is there an association between opioid overdose fatalities and proximity to syringe exchange programs?
Weak research question: Death certificate date is not always useful for determining specific causes (e.g. opioid type) of overdose, preventing us from accurately answering questions related to changes in the prevalence of different opioid types in overdose events. Different drug toxicology screens test for different opioids therefore if medical examiners change their drug screens it can make it appear as though causal opioids are changing when it may only be a result of measurement bias not actual changes in opioid type in overdose. One current problem with this dataset is that the SF medical examiners’ toxicology screens only recently started to test for buprenorphine. Therefore, buprenorphine will only now begin showing up as a causal agent in opioid overdoses. At the same time, we are working to encourage providers to prescribe more buprenorphine to manage patients’ pain and OUD. This will coincide with the new toxicology report that now tests for buprenorphine and could lead researchers to draw biased conclusions about the association between prescribing buprenorphine and rates of buprenorphine-involved overdoses. If the death data is taken out of context it may look like increases in buprenorphine prescribing lead to increases in buprenorphine deaths which would draw a potential causal inference where there is not one.
World Management Survey (Schools 2008-2012)
A survey to collect management practices data from firms across different countries and industries. This dataset is specific to 1,800 high schools in eight countries (Brazil, Canada, Germany, India, Italy, Sweden, UK, and US). Data was collected through telephone interviews with school principles and includes questions about management practices related to: operations, monitoring, target setting and people. In addition, school level student outcomes were collected from examination results across regions and countries.
Strong RQ – How do school management practices differ between countries and between types of schools (government, private, charter, etc.)?
This sample includes data about schools from across multiple countries and types, and so it’s possible to describe the differences and even to understand whether the classification of a school impacts the management practices that are implemented to see if there is a relationship between the type and strength of management practices.
Weak RQ – What school management practices are most important to better student outcomes?
Student outcomes are not consistently measured across countries. Similar to the comparison of HRS and ELSA data to ascertain differences in disease incidence and prevalence, the lack of consistency in data collection makes comparing results suspect.
Pew Research Center Gender and Leadership Survey (2014)
A survey of 1,835 randomly selected adults (921 women and 914 men) conducted online using a nationally representative online research panel recruited through probability sampling methods. The sample was weighted using a technique to match gender, and within gender, age, race, education, region, HH income, home ownership and metropolitan area to the parameters from the 2013 Current Population Survey.
Strong RQ – How have perceptions about equal leadership opportunities for women changed over time?
Data has been collected using this survey instrument since 2007 and so it is possible to explore trends over time in the aggregate.
Weak RQ – How do perceptions about equal leadership opportunities for women differ by respondent background (SES, geography, etc.)?
Given the relatively small sample size and reliance on weighting, it might not be as accurate to look at comparisons between sub-populations.
WHO/CDC Global Health Professional Survey (GHPS)
A survey developed in 2004 to collect data on tobacco use and cessation among health professional students because they are responsible for providing health care resulting from and education about tobacco use. The survey was conducted in 2005 among third year students pursuing advanced health science degrees in ten countries (Albania, Argentina, Bangladesh, Croatia, Egypt, Bosnia & Herzegovina, India, Philippines, Serbia, and Uganda).
Strong RQ – Is there a difference in the attitudes of smokers/non-smokers as to how health professionals should be trained about smoking cessation?
For the specific sub-populations studied in this survey, this question can be answered to provide some directional guidance.
Weak RQ – How does tobacco use vary between different types of health professionals and the general population?
The survey did not include students from each type of health profession in each country and the sample sizes were not established or weighted to be representative of the population of health professional within country, so comparisons between these categories cannot be extrapolated to represent the population of health professionals of any type in any country. Data about prevalence of adult smoking in each country is not collected in a uniform manner and so comparisons are not appropriate. Furthermore, given that this survey was done among students, it is not clear that the same results would persist among practicing clinicians.
CHIS (California Health Interview Survey): a random-dial telephone survey of over 20,000 Californians yearly that is statistically representative of the population of California with respect to age groups (adults, adolescents, and children are represented) and race/ethnicity. The survey encompasses a wide range of health-related data, including chronic disease outcomes, mental health, and access to healthcare.
Strong: Do adults reporting low levels of safety and social cohesion in their neighborhoods have higher prevalence of asthma, diabetes, and heart disease? Note that this is not, and cannot be, a causal question since we can’t definitively establish temporality for the exposure-outcome relationship from the available data. This research question would just be examining a potential association in order to allocate resources efficiently. CHIS data asked about both the exposure parameters of adults in its surveys from 2011 onward, so there is a good deal of available data. Since this question is just addressing an association, cross-sectional survey data is adequate.
Weak: After implementation of the Affordable Care Act, were more children diagnosed with asthma in 2015 (compared to 2013, before implementation of the Affordable Care Act)? This question is not possible to answer with CHIS data because, as stated above, we can’t establish when the diagnosis of asthma was made – in other words, these are prevalent and not incident cases of asthma. I could change the question to look at a different outcome, like the number of asthma-related emergency room visits in 2015 compared to 2013 (which I would hypothesize should decrease, since more people should have health insurance and should be seeing their primary care physician for better long-term management of the disease which should result in fewer ER visits). Or I could use a different data source for the original question, like Kaiser medical records (essentially an open cohort). It would still be difficult to attribute any changes to the ACA definitively, since they could be due to temporal trends unrelated to the ACA implementation.
Add Health (National Longitudinal Study of Adolescent to Adult Health): a nationally representative longitudinal study of adolescents who were in grades 7-12 during the 1994-95 school year. The cohort has been followed in adulthood with periodic in-home interviews; the most recent of which took place in 2008. The surveys have collected data on individuals’ social, economic, psychological, and physical health, including a wide range of social network data.
Strong: Do women who reported intimate partner violence in wave 2 have a higher rate of incident migraine headaches in wave 4, and is there effect modification by race/ethnicity or socioeconomic status? The longitudinal nature of the study allows us to see how an earlier exposure affects an outcome for the same group of people. The specification of the time period between exposure and outcome assessment is due to when these questions were asked of the cohort.
Weak: Do women who reported intimate partner violence in wave 2 have a higher rate of incident coronary heart disease in wave 4 (+/- effect modification as above)? While it is certainly possible to estimate an answer with this dataset, we would be answering a somewhat different question. The cohort simply hasn’t been followed long enough for CHD to manifest for most of those in the cohort, so any estimates obtained would likely be a mix of outliers (very early developers of CHD) and people with other forms of heart disease, which is not as useful an answer as one that could either be more specific to isolate CHD forms of heart disease, and one that would have longer follow-up – for example, a cohort like the Nurse’s Health Study could be used for this, if the relevant exposure question had been or could be asked.
NVSS (National Vital Statistics System): data are collected by local jurisdictions that are legally responsible for the registration of vital events (births, deaths, marriages, divorces, fetal deaths) and then reported to the National Center for Health Statistics (NCHS), which is responsible for maintaining the NVSS.
Strong: Did adolescent motor vehicle accident deaths in California decrease after implementation of the provisional driver’s license for those under 18? In order to lend more weight to the potential causal impact of the law, California’s statistics could be compared to those of other states with no such law as a “control” group. Of course, differences between the states could be strong confounders and would need to be well thought-out and adjusted for. Since this is a national database, it is possible to obtain mortality data for multiple states for comparison. Motor vehicle trauma as cause of death should be highly accurate, unlike many other causes of death later in life where there may be more ambiguity with regard to the proximate and ultimate causes of death.
Weak: Is low education associated with increased mortality rates due to cancer? While this question could potentially be answered using mortality record data (since education level and cause of death are both recorded on death certificates), cancer is generally under-reported as a cause of death on death certificates and so we may not get accurate results using this dataset. It may not be possible to address this shortcoming with a different dataset, since mortality records are ultimately the source of cause-of-death determinations. It would likely be possible and help provide more insight to attempt to corroborate death record information with hospital record and cancer registry information. Again, likely this would not completely solve the issue since often there is some subjectivity to the assessment of cause of death, but we may be able to get a better idea of how likely it is that the official cause of death is accurate.
1. US Linked Birth/Death Data: dataset from the National Vital Statistics System that links birth certificate data to information from death certificates for children under 1 year of age who dies in the US, Puerto Rico, Virgin Islands, and Guam. Linkage is used to provide comprehensive data on age, race/ethnicity of parents, birthweight, gestational period, gravidity, maternal education, maternal smoking, congenital anomalies, etc.
Strong research question: The effect of Planned Parenthood funding cuts to teen pregnancy and adverse birth outcomes, using a DiD analysis. Our hypothesis is thatcuts to Planned Parenthood funding will increase teen pregnancy, preterm deliveries, and low birthweight babies. The US Linked Birth/Infant Death Data Set will be a well-suited dataset to obtain high-quality outcome data, given the expansive information on comprehensive maternal and newborn characteristics. Exposure will be measured as the date of either when the Planned Parenthood bill was instated by President Trump (April 13, 2017) or when individual state governances issued cuts to funding. Effect modification by race/ethnicity and education will be tested. As with all DiD analyses, well thought-out comparison groups need to be identified.
Weak research question: Does maternal smoking increase the risk of congenital anomalies? Data on exposure is well defined in this data set, however, by studying an environmental exposure on an observed outcome that is conditional on whether the child survives to delivery (i.e. live birth) may induce collider bias, if there are other factors that cause fetal death and congenital anomalies. We could try to conduct a quantitative bias analysis; however, other pregnancy cohort datasets that have better follow-up data on the pregnant mother (i.e. data on stillbirths, spontaneous abortions, etc.) may be better suited to answer this question (i.e. data from Kaiser).
2. African Collaborative Center for Microbiome and Genomics Research’s (ACCME) Human Papillomavirus (HPV) and Cervical Cancer Study: Large representative cohort of non-HIV infected women >18 years old living in Abuja, Nigeria, had sexual intercourse, and no previous history of cervical abnormalities, cervical cancer or total abdominal hysterectomy. Enrollment began in February 2014. ACCME collected comprehensive demographic data (adapted questions from NHS study and biological samples including blood, mid-vaginal, ectocervical cell samples, at baseline, 6, 12, 18, and 24 months (but is planning to expand follow-up and participant recruitment).
Strong research question: Does ethnicity (measured by tribe) have an effect on persistent HPV infection among sexually active women in Abuja, Nigeria? Data on tribe, genetic data, SES is collected at baseline and possible confounding by sexual behaviors and incident HPV infection is collected at each follow-up visit
Weak research question: Does anal sex increase the risk of persistent HPV infection? This question would be relatively hard to conduct because for some individuals, reverse causality cannot be ruled out (i.e. all women were eligible regardless of the HPV status). However, if we decide to restrict our analyses to women who were HPV-naïve), we are selecting out a population of women who may have less risky sexual behaviors or may be more likely to clear the infections. Given the relatively little we know about why women are more likely to have persistent HPV vs. clear the infection, we could not draw a complete DAG that would allow us to adjust for covariates to mitigate the selection bias induced (I think).
3. ARIC dataset: prospective cohort study of 16,000 adults aged 45-64 examined twice, 3 years apart in 4 US communities (4,000 participants/community). Primary aim of study is to investigate the aetiology of atherosclerosis and clinical sequelae, and variation in CVD risk factors.
Strong research question: Does glucose-6-phosphate dehydrogenase deficiency protect against coronary heart disease?
Weak research question: Does urban/rural status modify the effect of ethnicity on diabetes? The sites were very specific in their urban and rural status (i.e. Jackson, Mississippi and Minneapolis suburbs, Minnesota were 100% urban), it would be difficult to study effect modification given the study must account for the clustering by site (addition of random effect).
i. Research Question: our aim is to evaluate the effect of usual consumption of trans fatty acids on incident coronary heart disease (CHD). We hypothesize that a diet with a greater proportion of calories coming from trans-fat will be positively associated with incident CHD.
i. Research Question: our aim is to evaluate the effect of smoking on incident type 1 diabetes. We hypothesize that smoking is positively associated with incident type 1 diabetes. Because most patients with type 1 diabetes are diagnosed during childhood, incident cases among the ARIC cohort will be few (rare disease).
SUPREME-DM (SUrveillance, PREvention, and ManageEment of Diabetes Mellitus) DataLink: SUPREME-DM includes de-identified health information from nearly 1.1 million people with diabetes in 10 states: California, Colorado, Georgia, Hawaii, Michigan, Minnesota, Oregon, Pennsylvania, Washington and Wisconsin. Participating health plans include six regions of Kaiser Permanente, Geisinger Health System, Group Health Cooperative, Health Partners, Henry Ford Health System and Marshfield Clinic.
Strong research question: How does the timing of diabetes onset affect stroke risk among diverse racial groups?
This sample encompasses urban and rural regions, includes both private insurance and Medicaid/Medicare-insured individuals, and is comprised of a diverse racial and ethnic range, making it well equipped to answer questions of clinical outcomes across a wide range of socioeconomic status.
Weak research question: What is the national prevalence of diabetes in the US?
SUPREME-DM only has the ability to answer questions about patient subgroups - population-level inferences are not possible, as this datalink only covers a fraction of the states and is not representative of the US. It also does not cover individuals who are not enrolled in insurance programs and/or who do not attend health organizations.
Monitoring the Future (MTF): MTF is a trend study, that implements cross-sectional nationally representative surveys of high school seniors, collected since 1976 (and for 8th and 10th graders since 1991).
Strong research question: Did teenage attitudes towards cigarette use shift following the implementation of minimum age cigarette purchasing restrictions?
MTF questionnaires cover a wide range of topics, including attitudes towards drug use, planned career trajectories, and drug use patterns, making it apt for measuring population average changes and trends over time.
Weak research question: Do teenage attitudes towards drug use persist into young adulthood?
MTF is unable to examine long-term outcomes, as surveys collect information only on cross-sectional points in time and don't follow individuals longitudinally.
Panel Study of Income Dynamics (PSID): PSID is a panel study, a type of longitudinal study design defined as describing cases at two points or more in time. Panel studies differ from cohort studies, which follow cohorts (a group of people selected for their inclusion in a “cohort”, often based on age/year of birth) at often infrequent intervals in that they generally sample from the entire age range, and are used extensively to monitor poverty dynamics, movements into and out of the labor market, and demographic change.
Strong research question: Are there differences in long-term downstream effects of self-earned income verses spousal earned income on mortality (e.g. does the act of earning income differ from having access to income)?
PSID includes multiple generations of families, making it well suited to answer questions regarding intergenerational characteristics and life-course exposures over time.
Weak research question: Are effects of the "healthy immigrant" generalization on mortality mediated through income?
Although the PSID is designed to be nationally representative, there is little information on immigrants. Furthermore, PSID is subject to reactivity (responses to subsequent questions being affected by prior questions) and retention as participants are repeatedly surveyed.
SUPREME-DM (SUrveillance, PREvention, and ManageEment of Diabetes Mellitus) DataLink: SUPREME-DM includes de-identified health information from nearly 1.1 million people with diabetes in 10 states: California, Colorado, Georgia, Hawaii, Michigan, Minnesota, Oregon, Pennsylvania, Washington and Wisconsin. Participating health plans include six regions of Kaiser Permanente, Geisinger Health System, Group Health Cooperative, Health Partners, Henry Ford Health System and Marshfield Clinic.
Strong research question: How does the timing of diabetes onset affect stroke risk among diverse racial groups?
This sample encompasses urban and rural regions, includes both private insurance and Medicaid/Medicare-insured individuals, and is comprised of a diverse racial and ethnic range, making it well equipped to answer questions of clinical outcomes across a wide range of socioeconomic status.
Weak research question: What is the national prevalence of diabetes in the US?
SUPREME-DM only has the ability to answer questions about patient subgroups - population-level inferences are not possible, as this datalink only covers a fraction of the states and is not representative of the US. It also does not cover individuals who are not enrolled in insurance programs and/or who do not attend health organizations.
Monitoring the Future (MTF): MTF is a trend study, that implements cross-sectional nationally representative surveys of high school seniors, collected since 1976 (and for 8th and 10th graders since 1991).
Strong research question: Did teenage attitudes towards cigarette use shift following the implementation of minimum age cigarette purchasing restrictions?
MTF questionnaires cover a wide range of topics, including attitudes towards drug use, planned career trajectories, and drug use patterns, making it apt for measuring population average changes and trends over time.
Weak research question: Do teenage attitudes towards drug use persist into young adulthood?
MTF is unable to examine long-term outcomes, as surveys collect information only on cross-sectional points in time and don't follow individuals longitudinally.
Panel Study of Income Dynamics (PSID): PSID is a panel study, a type of longitudinal study design defined as describing cases at two points or more in time. Panel studies differ from cohort studies, which follow cohorts (a group of people selected for their inclusion in a “cohort”, often based on age/year of birth) at often infrequent intervals in that they generally sample from the entire age range, and are used extensively to monitor poverty dynamics, movements into and out of the labor market, and demographic change.
Strong research question: Are there differences in long-term downstream effects of self-earned income verses spousal earned income on mortality (e.g. does the act of earning income differ from having access to income)?
PSID includes multiple generations of families, making it well suited to answer questions regarding intergenerational characteristics and life-course exposures over time.
Weak research question: Are effects of the "healthy immigrant" generalization on mortality mediated through income?
Although the PSID is designed to be nationally representative, there is little information on immigrants. Furthermore, PSID is subject to reactivity (responses to subsequent questions being affected by prior questions) and retention as participants are repeatedly surveyed.
ARIC data: The ARIC data, following 4 different communities with different racial mixing is ideal to study effect modification by race of risk factors for atherosclerotic diseases (strong RQ). On the other hand, this study design with only a baseline and one follow-up period is inappropriate for any survival analysis (weak RQ). Longitudinal data would instead be necessary.
DHS: the demographic health survey, conducts cluster randomized cross-sectional surveys to gather health data in many countries around the world. DHS data from several rounds in 1 country could be used to study Age-Period-Cohort effect on malaria indicators such as mortality in children under 5 years old (Strong RQ). On the other hand, because of the cross-sectional design, the DHS dataset can’t be used to assess risk factors for malaria incidence (weak RQ).
NHS: The nurse health study is a cohort study established in 1976 and included 121,700 female registered nurses aged 30 to 55 years. A baseline questionnaire and follow-up questionnaires (every 2 years) were self-answered by the cohort participants.
Strong RQ: impact of anorexia/eating disorders on divorce rates.
Weak RQ: Caregiving and CHD risk. I actually think the NHS data was not appropriate to answer the RQ from Lee’s article. First, if interested in caregiving as an exposure, I don’t think the study should be restrained to nurses who are professional caregivers. There are probably plenty of confounding variables associating caregiving at home and caregiving in their work and the effect on CHD risk has the potential of being heavily biases. Second, a 2 year resolution over a 4 year period (1992-1996) will only give 2 data point per individuals and the survival analysis might not be powered enough. Last, as detailed in the paper, missing data on caregiving and loss to follow-up seem to have suffered from selection bias.
Aboud, FE, Bougma K, Lemma T, Marquis GS. Evaluation of the effects of iodized salt on the mental development of preschool-aged children: a cluster randomized trial in northern Ethiopia. 2016. Maternal and Child Nutrition.13, e12322.
In this study, clustering was at village, household (focus of study), and individual levels
Intervention: Insecticide-treated bed net use versus non-bed net use
Experimental units: Clusters of Households (907/2178 registered households)
Exposure measured at household level (Bed net use)
Outcome: Incident malaria among bed net users versus non-users, determined at the individual level (children) - 1199 children tested for malaria over a 12-months period and determined.
Hypothesized effects were a reduction in incident malaria among bed net users
The statistical model used to estimate the effect: Used Poisson regression with a generalized estimating equations approach (PROC GENMOD) to account for correlation from repeated measures on the same child using exchangeable correlation structure and households from the same village.
Exposure variable and covariates could be time varying.
Confounders included baseline parasitemia, exposure to malaria transmission, household altitude, socio-economic status, stunting/wasting/nutritional status, the number of sleeping rooms and density of bed nets used per household.
Log-transformed person-time was used as an offset and an exchangeable working correlation structure was specified. Any covariate with a p-value of >0.1 in univariate analysis was included in the multivariate model.
Rate ratios (RR) and 95% CI were calculated from model parameters and model-adjusted incidence rates for covariates presented.
Protective effectiveness (PE) was calculated as 100%*(1-R1/R0) where R1 is the rate among bed net users and R0 is the rate among non-bed net users.
The attributable rate difference was calculated as R1-R0 and interpreted as the number of malaria infections prevented by ITNs annually.
Alternative analytic method: Mixed effects model which can handle more than one level of clustering, able to conduct likelihood fits and more robust to potential bias due to missing data or drop-out.
Interaction terms: none mentioned/assessed. (Net use and socio-economic status
https://malariajournal.biomedcentral.com/articles/10.1186/s12936-015-0554-1
Assignment 1
Article: Availability of Reproductive Health Care Services at Schools and Subsequent Birth Outcomes Among Adolescent Mothers.
Madkour AS, Xie Y, Harville EW
J Sch Health. 2016 Jul;86(7):488-94. doi: 10.1111/josh.12399.
The unit of clustering: The unit of clustering is high school: Public, private and feeder high schools were included
The hypothesized effects: The investigators hypothesized that young women who attended schools offering reproductive health care services on-site would evidence better birth outcomes in subsequent pregnancies compared with young women who attended schools without those services.
The level at which the exposure is measured: The exposure (on site reproductive health care services) is measured as a characteristic of the cluster(structural) i.e. presence or absence of on site reproductive health care services coded as follows
1 = provided on site,
0 = not provided on-site
Reproductive health care services were defined as diagnostic screening (including but not limited to STDs), treatment for STD, family planning counseling, and prenatal/postpartum health care
The statistical model used to estimate the effect: They used multilevel random intercept linear regression analyses, which was appropriate in this case.
Level 1: school level characteristics
level 2: individual characteristics
Analyses began by examining univariate distributions and bivariate relationships. Proportions and means of individual-level and school-level characteristics were calculated. Bivariate associations between each covariate and birth outcomes were assessed with a series of bivariate random intercept linear regression models. They then entered all variables into a multilevel model simultaneously (1 model for each outcome).
Describe whether there are any other statistical models that might be appropriate and whether they would be preferable (e.g., GEE vs mixed).
- they could have used mixed effects models as these can handle a wide variety of data structures e.g. time within adolescents within schools. These methods are also able to give additional information about correlation and are more robust than GEE when you have missing data
I am not too sure but generalized estimating equations (GEE) may be an option as well (to a lesser extent). GEE ae good when you have only one level of clustering e.g pregnant adolescents. In this case we have adolescents only clustering within schools. In general GEE would require a fairly large number of clusters (ideally > 50) and works best when the number of clusters is greater than the number of observations per cluster – in this case – there are 107 clusters of schools , each with ~11 individuals per cluster and so this method would work pretty well
-
Article: The Devon Active Villages Evaluation (DAVE) trial of a community-level physical activity intervention in rural south-west England: a stepped wedge cluster randomised controlled trial. Solomon et al. International Journal of Behavioral Nutrition and Physical Activity 2014, 11:94.
Unit of clustering: village
Hypothesized effects: increased levels of physical activity after exposure to the intervention
Exposure measurement level: the exposure was “offering people of all ages increased opportunities to experience the enjoyment of sport and physical activity.” The precise nature of the intervention varied by village, but each village had at least three options that were targeted to different age groups so that the full range of ages was covered in each village. The intervention/exposure was measured at the group/cluster level (village) – e.g. when the village was randomized to receive the intervention.
Statistical model used to estimate the effect: the researchers used multilevel, random effects linear regression models for continuous outcomes, specifying the village effect as random; and marginal logistic regression models using generalized estimating equations (GEEs) for binary outcomes. Unadjusted and adjusted (for region, gender, and age) analyses were conducted.
Other statistical models: The authors could have considered using either mixed effects models or GEE models for all outcomes rather than using mixed for continuous and GEEs for binary. Considering the authors’ interest in determining the effects for all adults, a marginal approach like GEEs might have been preferable between the two.
Does the FIFA 11+ Injury Prevention Program Reduce the Incidence of ACL Injury in Male Soccer Players?
Silvers-Granelli HJ1,2, Bizzini M3, Arundale A4, Mandelbaum BR5, Snyder-Mackler L6.Clin Orthop Relat Res. 2017 Apr 7. doi: 10.1007/s11999-017-5342-5.
Overview of the study:
Population: Men who play competitive college soccer (NCAA DI/II)
Intervention: FIFA 11+ injury prevention program
Comparison: NCAA men’s soccer teams in DI and DII who did not undergo FIFA 11+ injury prevention program
Outcome: ACL injury
Study Design: prospective cluster randomized controlled trial
Time Frame: Fall 2012 season
Unit of clustering: Authors randomized the intervention by NCAA institutions’ team.
Hypothesized effect and level at which the exposure is measured (is it a characteristic of the cluster or the observation within the cluster): Authors hypothesized that the FIFA 11+ injury prevention program can reduce the overall number of ACL injuries in men who play competitive college soccer. Sixty-five institutions were randomized using a simple computer-generated randomization. Individual player informed consent, ACL injury, mechanism of injury, and date of return to play, age, position played, leg dominance were collected at the individual level.
Statistical model used to estimate the effect: Authors used frequency counts, t-tests, chi-square tests, factorial analysis of variance, and logistic regression tests.
Other statistical models that might be appropriate/preferable: We could conduct a multilevel model because data from the participants was collected a multiple levels, and a GEE would be preferable. However, I am not sure with my sample sizes (are they big enough?)—intervention group included 27 teams with 675 players, and control group included 34 teams with 850 players.
Assignment: Find any article using clustered data and describe: the unit of clustering; the hypothesized effects and the level at which the exposure is measured (is it a characteristic of the cluster or the observation within the cluster); and the statistical model used to estimate the effect. Describe whether there are any other statistical models that might be appropriate and whether they would be preferable (e.g., GEE vs mixed).
Article: Cissé, Badara, et al. "Effectiveness of Seasonal Malaria Chemoprevention in Children under Ten Years of Age in Senegal: A Stepped-Wedge Cluster-Randomised Trial." PLoS medicine 13.11 (2016): e1002175.
Study design: stepped-wedge, cluster randomized trial
Study setting: 3 districts of high malaria transmission in Mali
Trial duration: 3 years
Intervention: Seasonal malaria chemoprophylaxis (SMC) with sulphadoxine-pyrimethamine plus amodiaquine (SP-AQ), further classified into 3 arms:
Unit of randomization: 54 health posts () were randomized to receive different combination of arms, according to the stepped-wedge design.
Hypothesized effects: SMC among children 3-59 months of age may reduce all-cause mortality
among children during the high malaria transmission season, and when extending SMC to children up to ten years of age may further prevent mortality.
Level at which exposure was measured: at the cluster level.
Level at which outcomes were measured: Primary endpoints, all-cause mortality and malaria incidence were measured at the cluster level.
Statistical model used to estimate the effect: Effectiveness of SMC was estimated by fitting a Poisson regression model to the data on the number of deaths and the number of RDT-confirmed malaria cases in each health post occurring in the period starting from the date of the first round of SMC and ending one month after the last round of SMC each year. The number of person-months at risk obtained from the DSS was included as an offset, and a gamma-distributed random effect was used to allow for correlation within clusters. Covariates included in the model were age group, calendar year, and indicator variables for the effect of SMC in children (under five years in 2008 and under ten years in 2009 and 2010) and for the indirect effect of SMC in older age groups. Interaction terms were included to compare effects in the two age groups, and combined effects were estimated where there was no evidence of interaction. The indicator variable representing SMC direct effects was set to 1 if that age group received SMC in that cluster in that year and set to 0 otherwise. The indicator variable for indirect effects was set to 1 for all non-SMC age groups if SMC was implemented in children in that cluster and set to zero otherwise. Thus the direct effect that was estimated represents the sum of direct and indirect effects, and the indirect effect estimated represents the indirect effect only.
Magadi, M., & Desta, M. (2011). A multilevel analysis of the determinants and cross-national variations of HIV seropositivity in sub-Saharan Africa: Evidence from the DHS. Health & Place, 17(5), 1067–1083.
Unit of clustering: Clustering was conducted at three levels: individual (first), regional (second), and country (third).
Hypothesized effects: This was an exploratory study designed to: (a) determine the individual and contextual socio-economic and demographic risk factors of HIV seropositivity among males and females in sub-Saharan Africa; (b) explore potential pathways of the determinants of HIV seropositivity with respect to proximate factors related to awareness, stigma, and sexual behavior; (c) explore contextual regional and country factors associated with HIV seropositivity; and (d) examine national and sub-national variations in the risk of HIV seropositivity. It emphasized the differences between males and females as well as cross-national differences.
This was a secondary analysis of existing data from the international DHIS programme and AIDS indicator surveys collected between 2003 and 2008. In includes data from 20 countries, 199 regions, and 174,592 individuals (95,759 women and 78,833 men).
Level at which exposure is measured: Exposures were measured at all three levels: individual, regional and country.
Statistical model used to estimate the effect: Four different multilevel logistic regression models were used to explore the individual and contextual regional and country level factors associated with the risk of HIV seropositivity. Model 0 has only the random region and country effects (no covariates) and subsequent models add a few country and regional level factors and several individual factors. Model 1 adds media exposure at the country, regional and individual level, as well as, socio-economic and demographic individual factors. Model 2 adds HIV stigma factors at the regional and individual levels, proportion tested for HIV at the regional level, and HIV awareness at the individual level. Model 3 adds marital information and sexual behavior factors at the individual level.
Describe whether other models might be appropriate and whether they would be preferable: Given the small number of country-level units (20), it may have been more appropriate to conduct this as a two-level analysis. Also, I’m concerned that the exposures measured at an individual level are not independent of one another, in particular those concerning awareness, stigma, and media exposure. Furthermore, it’s not clear to why these were included at multiple levels.
Ramaswamy, M. & Kelly, P. (2015). Sexual Health Risk and the Movement of Women of Women Between Disadvantaged Communities and Local Jails, Behavioral Medicine, 41:3, 115-122 https://www.ncbi.nlm.nih.gov/pubmed/26332929
The unit of clustering for this study was zip codes - the 13 most disadvantaged zip codes in Kansas City Kansas and Missouri which overlapped with areas of high incarceration compared to other zip codes with low incarceration rates.
The researchers hypothesize that fears and insecurities at the community level reflect the reality of disadvantage, serving as barriers to preventative health care which minimized risk. The exposures - fear of one's neighborhood, level of neighborhood violence, and victimization by neighborhood violence, were measured as level 1 variables - through surveys of incarcerated women.
Bivariate tests and logistic regression were used to determine relationships between the variables of interest. It seems to me that there are other methods which would provide a richer understanding of the relationships in addition to these straightforward analysis. GEE may be appropriate because of the ability to describe risk in a population instead of an individual. Not that I want it to be more complicated, but the independent variables were neighborhood-level violence, social capital and trust, and incarceration density which are complex concepts which may require a different approach?
Hospital Variation in Cesarean Delivery: A Multilevel Analysis
Andres I. Vecino-Ortiz, MEcon, PhDc1, *, David Bardey, PhD2,3 , Ramon Castano-Yepes, PhD3,4
1. The unit of clustering: Hospitals within a single insurance network in Colombia and individuals covered by this insurance network who gave birth at in-network hospitals
2. The hypothesized effects: To explore the amount of unexplained variance between hospitals in their utilization of cesarean section while accounting for reimbursement rates and individual health factors. They hypothesized that hospitals would account for the majority of the variation in cesarean section utilization.
3. Level at which the exposure is measured: Hospital level (complexity of patient base; physician reimbursement schedule; region; public hospital; teaching hospital) and individual level (mother age; mother income; users per insurance contract; education; previous births; type of admission; gender of newborn; cesarean)
4. Statistical model used to estimate the effect: Multi-level regression model and an alternative variance decomposition to explain the proportion of the variance explained by the region. They chose this method because they could generate an error term for each level of analysis, as compared to the single error term from logistic regression analysis. 5. Describe whether there are any other appropriate statistical models: Because the authors were interested in estimating an error term at each level of clustering, the multi-level model appears to be the best choice. The GEE model gives an average coefficient in the presence of clustering, so it only pertains to the entire model.
Assignment: Find any article using clustered data and describe: the unit of clustering; the hypothesized effects and the level at which the exposure is measured (is it a characteristic of the cluster or the observation within the cluster); and the statistical model used to estimate the effect. Describe whether there are any other statistical models that might be appropriate and whether they would be preferable (e.g., GEE vs mixed).
Frisvold, D. and Golberstein, E., 2011. School quality and the education–health relationship: Evidence from Blacks in segregated schools. Journal of health economics, 30(6), pp.1232-1245.
Abstract: In this paper, we estimate the effect of school quality on the relationship between schooling and health outcomes using the substantial improvements in the quality of schools attended by black students in the segregated southern states during the mid-1900s as a source of identifying variation. Using data from the National Health Interview Survey, our results suggest that improvements in school quality, measured as the pupil–teacher ratio, average teachers’ wage, and length of the school year, amplify the beneficial effects of education on several measures of health in later life, including self-rated health, smoking, obesity, and mortality.
The authors based their hypothesis on prior arguments that that changes in school quality for black students were conditionally uncorrelated with unobservable variables (e.g. parental support or decisions to move neighborhoods based on district quality) that may be correlated with health outcomes in later life.
Freedman, V. A., Grafova, I. B., & Rogowski, J. (2011). Neighborhoods and chronic disease onset in later life. American journal of public health, 101(1), 79-86.
Unit of clustering: neighborhood
Hypothesized effects: neighborhood conditions would affect chronic disease (coronary heart disease, angina, congestive heart failure or other heart problems, high blood pressure, stroke, diabetes, cancer, and arthritis or rheumatism) onset in later life.
Level of exposure measure: Exposure is observed as a characteristic of the cluster, in this case a characteristic of the Neighborhood. Neighborhood conditions were characterized by using 8 previously validated scales reflecting the economic, social, and built environments.
Statistical model: Two-level random-intercept logistic model
If the authors were interested in estimating the average population effects as opposed to individual effects, then GEE would be a better approach. For example, comparing low SES neighborhoods vs. High SES neighborhoods with regard to disease incidence.
Assignment: Find any article using clustered data and describe:
Article: Carra G, Crocamo C, Borrelli P, Tabacchi T, Bartoli F, Popa I, et al. Area-level deprivation and adverse consequences in people with substance use disorders: findings from the Psychiatric and Addictive Dual Disorder in Italy (PADDI) Study. Sub Use & Misuse, 52:4, 451-458. DOI: 10.1080/10826084.2016.1240696.
1. The unit of clustering
The authors were interested in examining the effects of geographic area-level factors on overdose, HCV, HBV, and long-term incarceration at the individual level. They explored both individual-level (level 1) and geographic area-level (level 2) data. The purpose was to explore the effects that contextualization characteristics may have on health at an individual level, even after accounting for individual level variation. The authors’ used data from a national survey (Psychiatric and Addictive Dual Disorder) and the national census track.
2. The hypothesized effect and level at which the exposure is measured (is it a characteristic of the cluster or the observation within the cluster)
Hypothesized effect: The authors hypothesized that among people with substance use disorders, area-level deprivation would be significantly associated with overdose, HCV, HBV, and long-term incarceration beyond what would be expected from individual level data. The exposure variables (measurements of geographic level deprivation based on aggregate census data) was measured as a characteristic of the area-level cluster. Deprivation was measured based on five census track markers at the census block level: low level of education, unemployment, non-home ownership, one parent family, and overcrowding – the index is calculated by summing standardized indicators, which then corresponds to four area levels of deprivation: very deprived, deprived, intermediate, affluent and very affluent. The same approach is applied to aggregate data at the municipality level. Because of a small sample size for this study, the authors combined levels into three overall area-level categories: affluent, intermediate and deprived.
3. The statistical model used to estimate the effect
Due to small sample size (small number of subjects (level 1) per area level context (level 2)), the authors did not use a multilevel analysis. Instead, they fitted a logistic regression model with cluster-robust error for each outcome, modeling individual-level and geographic area-level effects. They included identified predictors as covariates and used a likelihood ratio test to compare nested models to examine the effect that different variables had on the regression equations.
4. Describe whether there are any other statistical models that might be appropriate and whether they would be preferable (e.g., GEE vs. mixed).
It wouldn’t have been appropriate to use multilevel modeling in this instance due to the low sample size. However, due to the sample size, I’m really not sure what other statistical methods would have been appropriate for the analysis. Suggestions??
