WEEK 5 READING RESPONSES

WEEK 5 READING RESPONSES

by Teresa Kortz -

Setting: Resource-limited settings

Exposure: Repeated gastrointestinal (GI) infections throughout childhood

Outcome: Adult educational attainment

Hypothesis: Increased number of GI infections in early childhood (<2 years of age) is associated with lower overall educational attainment in adulthood.

Lifecourse model: Accumulation of risk. While early childhood could be thought of as a “critical period”, based on the conceptual model of childhood GI infections, the potential harm is thought to increase with additional infections and there is a self-perpetuating feedback loop that encourages repeat infections (see attached figure for conceptual model). The potential harm due to one GI infection in early childhood is also not thought to be irreversible, and thus contributing to future health or socioeconomic outcomes, which would be consistent with a “critical period” model. My interest is in early childhood exposures to GI infections, infections that children are particularly vulnerable to; therefore, an adult mobility/infection model would also not be appropriate.

Regression model: As described by Mishra, et al., I could estimate the direct an cumulative causal effect GI infection by fitting a linear regression model where:

Y=a + B lifetime GI infection score + B covariates

Where B  = the change in the cumulative effect divided by the number of measurements, and the model is adjusted for confounders such as demographic (sex, village), socioeconomic factors (parents’ educational attainment, number of children <5 in the home, access to clean water, type of toilet in the home, parents’ occupations, etc.), and child comorbidities (prematurity, chronic GI disease, HIV, vaccination status, etc.). GI infection would need to be a binary exposure in each time interval. For example, if measurements are made every month, GI infection =0 corresponds to no GI infection in the last month, GI infection=1 corresponds to at least one GI infection in the last month.

The ideal dataset would be from a birth study cohort. Fortunately, one exists. The Fogarty Institute conducted a five-year multi-site birth cohort study to investigate associations between malnutrition and GI infections and their effects on children in resource-limited settings called the ‘Etiology, Risk Factors and Interactions of Enteric Infections and Malnutrition and the Consequences for Child Health and Development (MAL-ED)’ study. While this study collected detailed data over the first 5 years of life in these children, ongoing data collection throughout childhood and adulthood are possible and could potentially (eventually) answer my research question.

Concerns and limitations: Given the multisite nature of this study and that it was conduct in a resource-limited setting, measurement error between sites and missing data are potential issues. Unmeasured/residual confounding is also a concern; the relationship between early childhood GI infection and adult educational attainment is highly complex, not completely understood, and influenced by social and biological factors, meaning that other events not measured between time points may be contributing but not accounted for in this model. There could also be a survival bias; diarrheal disease is a major cause of mortality in children <5 years in resource-limited settings and in studying adult educational attainment, I will only be evaluating the outcome in subjects that survived, potentially those with less severe disease. The net result of these limitations is a potentially biased estimate.


Re: WEEK 5 READING RESPONSES

by Scott Lu -

(I imagine attaching my assignment here may make it easier to access).


Exposure: Kaposi sarcoma

Outcome: Quality of life

 

Lifecourse model: Social mobility

This is following the social mobility model analogous to the life course model of a critical period with later effect modification, as I see aspects both that could be applied to this exposure-outcome relationship.  In this case the critical period describes factors associated with HIV and HHV-8 infection which increase risk for developing Kaposi sarcoma (the exposure).  The effect of exposure on the outcome (quality of life) can be diminished later in life through a later factor (ART). 

Regression model:

A model to estimate the effect of Kaposi sarcoma on quality of life would include an interaction term for ART and another for time:

E(Y) =α + β1S1 + β2S2 + θ12S1S2

Where the outcome of quality of life depends on KS and ART status.  KS would be a binary exposure indicating presence or absence of disease and ART would also be a simplified binary indicator expressing current, theoretically continued therapy (though there may be an effect from any # of ART, I’m not completely sure how to incorporate that into this model).  Thus the effect of having no ART then starting ART (perhaps as a result of KS) can be combined to determine QOL (the model could also adjust for confounders such as location, SES, viral load, CD4 count, etc.).  Datasets to explore this would have to follow some period of time where individuals diagnosed with KS start ART and have QOL measurements taken at multiple time points.  A significant limitation to the above model is the inability to distinguish between duration of KS in individuals.  Additionally, the status of KS carries an increase in predicted mortality even on optimal therapy (compared to KS-, HIV+ counterparts).  Further, the use of 2 binary indicators limits the model to 4 possible trajectories to calculate QOL from.  The complexity of the issue makes this seem like an overly simplified approach (at the very least multiple time points and inclusion of some measure of vital status would be incorporated).


Re: WEEK 5 READING RESPONSES

by Matthew -

Population: Preterm infants

Exposure: Shortening time to reach full oral feeds

Outcome: Long term growth and cognitive development such as anthropometric growth %tiles, early childhood developmental scores (i.e. Bayley scores), academic achievement, etc.

Lifecourse model based around developmental and social trajectories

One field of interest that my team has been researching is the use of different therapies such as specialized pulsatile pacifier for preterm infants to suck on, which is meant to help enhance suck/swallow neuromuscular development. The purpose of this is the improve the infant’s ability to orally feed (as opposed to tube feeding) during this very critical period which important for the overall growth and development of the infant. Even after an infant has recovered from their acute illness, converting to oral feeds is often the bottle-neck accomplishment necessary to discharge most infants home from a NICU. I think a life course study would be of great interest to thoroughly evaluate the short- and long-term outcomes of reaching full-oral feeds earlier during this critical period. I think this touches on some important life course theory aspects such as “timing and environment.” 1) Timing- since it’s a critical developmental period important for further development not only for neuromuscular function but for gut and flora development as well. 2) Environment- since earlier oral feeds leads to earlier home discharge and the home environment is better for infant development and infant-maternal bonding which also has several benefits that are far-reaching in the mother and child’s life, Unfortunately, given the novelty of these types of interventions, there is not a database already available that has this type of information. Various regression models could be used to evaluate common short-term outcomes such as rates of necrotizing enterocolitis, chronic lung disease, transfusions, discharge home on oxygen, length of stay, etc. Regression and survival analyses could also be used to evaluate long term outcomes such as anthropometric measurements (weight, height, head circumference growth curves), mental/functional developmental scores/milestones, school grades, etc. My biggest concern about evaluating any of these hypotheses would the lack of already existing databases. Although my team and others have begun to focus on gathering long term early childhood data it will take time to accumulate enough to sufficiently investigate these ideas. Another concern would also be generalizability to other populations outside of the few sites with this information.

 


Re: WEEK 5 READING RESPONSES

by Crystal Langlais -

Exposure: obesity (≥30kg/m2)

Outcome: prostate cancer tumor size at diagnosis

I found the reading very interesting this week. In general, I don’t think I can subscribe to the critical period hypothesis under the strict definition presented by Ben-Shlomo et al. Specifically, I take issue with the idea that exposures outside the critical period have no effect on disease. Though they note that in practice there is often no distinction between critical and sensitive period, I found their formal definition of sensitive period more palatable (a sensitive period where exposure has great impact, outside of which the exposure still acts but may have diminished impact).

For my particular exposure-outcome pair, I could agree with either a sensitive period model or the chains of risk/additive effect model.  If I understand correctly, the two models vary in that in the chains of risk model, each exposure is proposed to have similar impact and the exposures may act over longer period of the life. With this understanding, the chains of risk model seems most plausible to me. Although obesity in childhood may be more harmful to prostate cancer development later in life, I could argue that this may be a result of prolonged exposure. In response to obesity, the body may develop adaptive processes to provide some semblance of temporary resilience.  As these exposures continue throughout life, the adaptive mechanisms may break down, exposing the prolonged and iterative damage, resulting in disease. 

Assuming a linear relationship between tumor size and amount of obesity (i.e., presence at different life stages), I could use a regression model like:         

E(Y) =  

where  is the presence of obesity across j time points (i.e., the lifetime obesity score).

 

 [see word doc attachment if formula doesn't come through]

Although there are datasets that have tumor size available, it is not a common characteristics. Further, it is very difficult to get a measure of obesity across the lifespan. In other countries where the medical record follows the individual across health care providers and facilities, this question would likely be feasible.

Limitation: This approach includes strong assumptions of the relationship between tumor size and lifetime obesity.  For example, it might be equally plausible that an intermittent period of non-obesity may negate some of the effects of a period of obesity. Or, the degree of obesity might increase the effect.


Re: WEEK 5 READING RESPONSES

by Maria Glymour -

Crystal

Nice discussion. I would describe the model you propose as an accumulation model rather than chain of risk.  Regardless, one key point about the equation you propose, in which a count of periods of exposure is used as the independent variable, is that even under a critical period model, this independent variable would predict the outcome.  Thus, if you estimate this model and find that the count of exposures predicts tumor size, you cannot infer that there are cumulative effects - it is consistent with a critical period model because exposure during a critical period would increase the count of exposed lifecourse periods by 1.  For example the never exposed would have a count of 0, and the exposed during only the critical period would have a count of 1.  Even though the count would also increase with exposure in other periods, you would still expect the count to predict the outcome, even under critical period.  

Also re critical periods: some typical examples are specific developmental episodes, which are biologically programmed to occur at a specific age. I agree it's not as common as sensitive periods, but for example the capacity to visually perceive depth or the capacity to speak a language fluently without an accent are both very difficult to develop outside specific developmental windows. 

maria


Re: WEEK 5 READING RESPONSES

by Maria Glymour -

Thanks Scott.  Your model specification seems to assume that duration of KS or duration of ART has no effect on QoL.  It's just whether or not you currently have KS and whether or not you are currently on ART.  If that doesn't seem plausible, you could incorporate variables such as duration of ART into the model. 

Maria


Re: WEEK 5 READING RESPONSES

by Adrienne Epstein -

Exposure: Malaria in pregnancy

Outcome: Cognitive function in childhood/adolescence

Effect modifier: Years of schooling

A recent study in a mouse model found that malaria in pregnancy harmed offspring cognition. In addition, malaria in pregnancy has been linked to children being born low birth weight, which is also associated with lower cognitive function in these children. However, studies have not examined the relationship between malaria in utero and cognitive development in children.

The question I would pose using a life-course approach is: does malaria in utero impact child cognition, and is this effect modified by years of schooling? The life-course model this invokes is the critical period model with later life effect modifiers. I would be curious to see whether children who were born to mothers with malaria in pregnancy & had lower cognition earlier in life were able to “catch up” with schooling.

I would estimate the following regression model:

E[cog_age10] = B0 + B1*MIP + B2*yrschoolage10 + B3*MIP* yrschoolage10 + ε

This model uses only information measured at age 10. MIP is a binary if indicator of whether the child was born to a mother with malaria in pregnancy. The main effect B1 would allow us to determine whether malaria in pregnancy had an impact on cognition at age 10 holding covariates constant (covariates not shown here), and the interaction term B3 would allow us to determine whether this association depends on the number of years of schooling a child has had.

An alternative regression model would be:

E[cog_age10] = B0 + B1*MIP + B2*yrschoolage10 + B3*MIP* yrschoolage10 + B4*cog_age1 + ε

The difference between this model and the one above is that we are now adjusting for cognition at age 1. It is my understanding that this model is actually evaluating the change in cognition score from age 1 to age 10 as outcome, and asking whether the rate of change differs based on years of schooling. Any feedback on which model makes more sense in this context is welcome!

To answer this question, a birth cohort in a malaria endemic country would be required. There are many birth cohorts in these regions, but I am unsure how long follow-up has been. There would also need to be adequate variation of schooling at age 10; if the setting is one in which all children go to school until at least age 10, the follow-up cognitive evaluation could occur at a later time.

Re: WEEK 5 READING RESPONSES

by Sarah Raifman -

Exposure: childhood sexual abuse (CSA) during <18years age

Sexual abuse during childhood has been associated with psychiatric disorders including PTSD, depression, suicidal behavior, and substance abuse. CSA in women is linked to increased odds of early age at menarche, adolescent pregnancy, physical/sexual/emotional abuse during pregnancy, and lifestyle risk behaviors such as smoking.

Outcome: preterm birth (<37 weeks gestation) during adulthood

Pre term birth (birth prior to 37 weeks gestation) is the leading cause of neonatal mortality and preterm infants are at an increased risk for immediate and long-term health problems. Maternal risk factors associated with preterm birth include cigarette smoking, previous preterm birth, infection, preeclampsia, obesity, psychiatric disorders, psychotropic medication use, and exposure to intimate partner violence.

Hypothesis: exposure to childhood sexual abuse (any type during age <18 years) is associated with preterm birth during adulthood.

Approach: I would hypothesize that the potential harm would increase with repeated abuse and therefore the outcome may depend on the amount of exposure (ie. accumulation model). So I think the accumulation theory may apply best here.

Alternatively, I’d also be interested to know whether the effect of childhood sexual abuse in a defined period (let’s say <12 years age) on preterm birth is modified by pregnancy intent (unintended vs. intended) – for which we could use a critical period model.

Model:

Accumulation: I could estimate the direct cumulative causal effect by fitting a linear regression model:

E(preterm) = β0 + βcumulative lifetime sexual abuse+ βcovariates. 

Beta1 is the change in cumulative effect/number of binary measurements

Bcovariates could include SES, education, race/ethnicity, age, age at first intercourse, and age at pregnancy

 For a critical period model with interaction between childhood sexual abuse (CSA) and pregnancy intention – both binary variables for simplicity:

 E[preterm] = B0 + B1*CSA + B2*preg_unintended + B3*CSA* preg_unintended + e  

The ideal data source would be a prospective cohort study that collects information on ACEs and reproductive health outcomes. Since I’m not familiar with any (yet) – I could use the CDC-Kaiser ACE Study or Behavioral Risk Factor Surveillance System (BRFSS) ACE Data. The ACE Study is a survey mailed to Kaiser members who visited a clinic in San Diego CA between Aug and Nov 1995 and Jan-March 1996.

Limitations would include self-reported nature of CSA making the exposure subject to misclassification (recall bias and underreporting). To improve outcome measurement/validity, perhaps clinical data on pregnancy and birth could be abstracted from Kaiser records for participants. Limiting the sample to women ages 20-50 who had their first pregnancy after 20 years would help with temporality (ensuring exposure to child abuse came prior to pregnancy outcomes), however it would limit the generalizability of our findings and ability to understand how child abuse affects preterm birth among women who have a first birth prior to age 20 (a perhaps particularly vulnerable population of interest for this research question). Also defining the "critical period" would be important here and would depend on the conceptual framework -- is abuse at <18years different than abuse at <12 years? I would argue yes - younger child sexual abuse is different than teenage sexual abuse - and that this would be important to consider when designing the study.


Re: WEEK 5 READING RESPONSES

by Erika Meza-Luman -

Exposure: Parental and maternal educational level when participant was 17 years old (PEd)

Outcome: Cognitive Decline

Effect modifier: Respondents educational attainment (REd)

Lifecourse model: Social mobility

Many studies have shown positive associations between socioeconomic status and adult health. However, more recent studies have suggested that upward socioeconomic mobility can impact the association between childhood socioeconomic disadvantage and markers of healthy aging. 

Hypothesis: Following the sensitive period framework, I hypothesize that an individual’s years of completed education will modify the effect of their parental and maternal educational level on cognitive decline. (Specifically, educational attainment lower than the highest level achieved by a parent would be harmful and educational attainment higher than the highest level achieved by a parent would be beneficial).

Regression Model: Following the mobility model in the Mishra et al. methods paper, the linear regression for this hypothesis is:

E(Y) = a+B1*PEd+B2*REd + B3PEd*REd

where the expectation of Y is a function of the inter-generational educational levels and their interaction and B3= - (B1-B2).

Therefore, our causal parameters of interest for the effects of mobility would be:

The change in outcome Y given less educational attainment: Y10 – Y11

The change in outcome Y given more educational attainment: Y01 – Y00

This would require a health data set that follows a middle-aged cohort and contains information on a broad range of parental and maternal educational levels as well as cognitive measures over time.

Challenges and limitations for this would include unmeasured residual confounding since there could be biological mechanisms that affect both an individual’s educational attainment and cognitive decline. Additionally, it is very possible that there could be a critical period (in early childhood) at which parental educational attainment would have the greatest effect and that would not be disentangled using this model. It is also possible that lower educational attainment results from ‘chains of risk’ or a cluster of ‘adverse effects’ however, now I wonder if these would be included in the social mobility model as covariates or if the question would become better suited for the “accumulation of risk” model.


Re: WEEK 5 READING RESPONSES

by Maria Glymour -

Erika

This is a really interesting example.  A lot of the social mobility literature emphasize the stress of mobility - whether upward or downward- but I think your hypothesis that upward is good often makes more sense.  Perhaps there's a non-linear association.  This is all testable.

Maria


Re: WEEK 5 READING RESPONSES

by Maria Glymour -

Sarah

Thanks for the example.  W/ unintended pregnancy - is it a potential mediator?

Maria


Re: WEEK 5 READING RESPONSES

by Maria Glymour -

Great example Adrienne and it raises two key issues: 

- Evaluating change in an outcome

- Addressing potential mediation. 

Short answer - in general adjustment for an earlier measurement of the outcome variable is problematic in an observational setting (but nearly always valuable in an RCT), especially for variables such as cognition, which have a lot of measurement error. 

You also have to worry about mediation here, since MIP would plausibly influence educational attainment (if it influences cognition).  

Maria


Re: WEEK 5 READING RESPONSES

by Ghila Andemeskel -

Population: African American men

Exposure: reported hate crimes

Outcome: changes in telomere length in African American men

I would use social biological model looking at accumulative risk. This model slows for the accumulative impact of hate crimes over a lifecourse. Previous research as displayed self-reported experience of discrimination in African American men had shortened telomeres. Here I would want to examine if reported hate crimes in city of residence are associated with shortened telomeres.

Hypothesis: African American men living in area with higher reported hate crimes will have shorter telomeres over their life time.

Model: Y(change Telomere length) = α + 𝛽+𝛽*X+𝛽 *C+𝜖

Issues that can be faced with this question are the difficulties of actually measuring the impact of the incidents. There have been different attempts at finding biological measure for measuring racism and how it impacts individuals. From stress reactivity to newer studies looking at telomeres. Here there will be no established data set to use except crime banks and news reports for reported hate crimes. Unlike other studies the focus isn’t simply on the individuals awerness of the crime but of their residence in an area it happens in and the missed impact that can result from that.


Re: WEEK 5 READING RESPONSES

by Maria Glymour -

Ghila

Interesting and important example.  In your model, when is telomere length measured? 

The use of ecological measures of racist environments seems very promising and something that there is increasing interest in.  Did you see eg the article on police violence and mental health - (cannot remember the first author - maybe Jacob Bor or Alex Tsai?).

Also much older but very good work by Pat Sharkey re exposure to violence in general (not specific to hate crimes) and kids test scores.

The question of timing with these models is very challenging b/c misspecification could give you a very misleading answer.

Maria


Re: WEEK 5 READING RESPONSES

by Monica Ospina Romero -

Exposure: Cholesterol (Total, HDL, and LDL) and Fasting glucose during childhood, early adulthood (20’s) and late adulthood (50’s).

Outcome: Cognitive function.

Research question: Cholesterol levels and fasting serum glucose are commonly measured in primary care of children and adults. Patients who are found to have abnormal values of these blood test are advised to modify their lifestyles. Having high cholesterol or glucose levels might be manifestations of social and biological conditions preceding the time that they were measured. Patients could improve their blood test with few modifications in their lifestyles, but they have already been exposed for some period to the high cholesterol and glucose levels. I would like to investigate whether the time exposed to these abnormal values (or the cause of these abnormal values) influence cognitive function later in life. I would like to evaluate whether having abnormal cholesterol or glucose levels during childhood or adulthood critical for cognitive function in older ages.

Model: Following Mishra et al. recommendation, I would start with a saturated model which would include the predictor value in each time period (childhood =1, early adulthood =2, mid-adulthood = 2), and their interactions.

E [Old Age Cognitive function] = α + β1 Cholesterol1 + β2 Cholesterol2 + β3 Cholesterol3 + θ12 Cholesterol1 *Cholesterol2 + θ13 Cholesterol1 *Cholesterol3 + θ23 Cholesterol2 *Cholesterol3 + θ123 Cholesterol1 *Cholesterol2 *Cholesterol3 + γk j covariates

A simpler model would assume that a cumulative effect of abnormal cholesterol levels during different periods on older age cognitive function:

E [Old Age Cognitive function] = α + β1 Cholesterol1 + β2 Cholesterol2 + β3 Cholesterol3

If I want to measure cognitive function (instead of cognitive decline) I could use linear regression. The data that I might have available for this research question is probably record data such as from Kaiser Permanent or national registry data from Danish or Sweden populations. In this case, I think I would have available diagnosis data instead of assessments of cognitive function; therefore, I would use Alzheimer’s disease as an outcome and a logistic regression model.

Limitations: Record data is collected for purposes other than research, so the quality of the data will be something to consider (measurement error or regression to the mean). It might be possible that using only three measurements of cholesterol (or glucose) are not enough to detect an effect on cognition/Alzheimer’s disease. Survival bias since the follow-up time for this study is >40 years.

 


Re: WEEK 5 READING RESPONSES

by Alice Guan -

Exposure: bisphenol A (BPA)

Outcome: Incident breast cancer

Life course model/mechanism: Accumulation of risk. BPA is a synthetic estrogen that is found in most plastic products. Mechanistically, it can function as a hormone disruptor and interrupt the equilibrium of body’s hormonal balance. Under the accumulation hypothesis, BPA exposure would have a direct effect on a person’s risk for cancer, regardless of when the exposure to BPA occurred. In other words, the total amount of exposure to BPA is what matters; not the specific times at which a person was exposed.  Considering the widespread use of BPA in bottles, food packaging, etc., it is plausible that humans are constantly consuming BPA throughout their lifetimes, thus accumulating the biological injury.

Regression Model: I would use the following regression model to estimate the effect of accumulated BPA on incident breast cancer: 

E [Breast Cancer] = \( \alpha + \beta1 * BPA level + \beta2 * age + \beta3 * covariates + \beta4 * BPA level * age + \epsilon \), where:

E [Breast Cancer] is a function of change in serum BPA level, age, and the interaction between them. Potential confounders in this model include socioeconomic status, education, and high BMI (both of which are known to be associated with breast cancer. It’s also very likely that people who are of higher SES and more educated are less exposed to BPA, whereas those who have higher BMIs are more exposed to BPA).

An ideal data set would include biospecimen data (specifically serum BPA level) measured at several different time intervals, data on potential confounders, and would capture information about incident breast cancer.

Challenges: The effects of BPA (specifically, endocrine related bioeffects) have been well documented, but the effects of BPA byproducts (including byproducts resulting in biological metabolism) are still relatively unclear. So, it’s possible that our measure of changes in serum BPA level over time are underestimating the true harmful effects of BPA that include BPA metabolic byproducts. Additionally, there are potentially other environmental factors that may co-occur with BPA, and we may be unaware of potential biochemical interactions that these may have with BPA within humans. Finally, although the accumulation of risk model makes the most intuitive sense when thinking through this research question, it’s also very possible that this approach ignore potential critical periods in which bodies are more sensitive to BPA exposure (plausible?).

Re: WEEK 5 READING RESPONSES

by Maria Glymour -

Great example Alice.  Are there any cancer examples with strong evidence of sensitive periods? 

Maria


Re: WEEK 5 READING RESPONSES

by Dan Kelly -

Exposure: Uveitis

Outcome: Depressive symptoms

Rationale/Approach: 

After reviewing the editorial of a life course approach to chronic disease epidemiology, I would like to consider the critical period model with later life risk factors. This model advocates that an exposure in a critical period results in permanent and irreversible damage or disease. Within this conceptual model, I would like to consider the exposure-outcome combination of uveitis-depressive symptoms as it connects to a cohort of Ebola survivors in Liberia. Here, the critical period is thought to be over a one- to two-year period after surviving Ebola virus disease. During this period, an individual may experience active uveitis, which is an inflammatory response in the eye, and in the case of Ebola survivors, is thought to be due to viral persistence. As a result, uveitis may turn from active to inactive, causing permanent and irreversible damage to vision. Visual impairment is connected to disability and may lead to issues with mental health. 

 Hypothesis: 

The hypothesis is that Ebola survivors who develop uveitis will be more likely to experience depressive symptoms than Ebola survivors who do not have uveitis. In Liberia, Ebola survivors are being studied as part of a five-year longitudinal cohort and there have been study visits every six months during which time the participants have had their eyes evaluated. Almost all of the participants who developed uveitis were identified with the first one-year of the cohort study. Since then, those with chronic uveitis have been monitored. Subsequently, in year four of the cohort, depressive symptoms will be collected, and once these data are collected, I will be able to test the hypothesis. 

Analyses: 

I will have a wave of uveitis data preceding the wave in which depressive symptoms will be collected, so I am considering a lagged analysis where I would use linear regression models. Another consideration may be to use uveitis as a time-varying exposure and then use generalizing estimating equations or marginal structural models to assess the association with depressive symptoms. 

Challenges/limitations:

The onset of visual impairment and depressive symptoms will need to be carefully evaluated because there will be potential limitations in measurement depending on when the exposure and outcomes occurred. Uveitis can occur from multiple etiologies and the post-Ebola syndrome is only one possibility. Likewise, there are other reasons that someone may develop depressive symptoms. Timing of depressive symptoms may be missed given that the development of uveitis may have occurred years earlier. Thus, the life course model may be a good conceptual framework for testing this hypothesis but the dataset may not be able to answer the research question. 


Re: WEEK 5 READING RESPONSES

by Maria Glymour -

Dan, 

This is an interesting example.  I would describe the plausible effects of uveitis on depression as basically immediate risk in that once the uveitis was resolved, no further effects on mental health would be expected.  A key aspect of the critical periods hypothesis is that the adverse exposure may no longer prevail but still the person has endured a biological consequence which will ultimately (perhaps decades later) manifest as disease.  

maria


Re: WEEK 5 READING RESPONSES

by Eduardo Santiago-Rodriguez -

Exposure- Marital status. I am interested in evaluating this exposure as having partner or not having partner.

Outcomes- Stage at colorectal cancer diagnosis (early vs late), CRC mortality five years after diagnosis

 Previous studies have found being married is associated with adherence to CRC screening guidelines (which lead to detecting polyps and diagnosing cancer at an early stage) and with better outcomes after CRC diagnosis (patients have support throughout the whole process; and it helps in attending appointments, surgery recovery, treatment completion, etc.). All cancer-related studies I found evaluating the effect of marital status on different outcomes, have looked at this exposure at time of diagnosis only. I propose analyzing different trajectories of marital status based on three time points: at age 50 (guidelines recommend CRC screening starting at this age), at time of diagnosis and at 5 years after diagnosis.   

 Lifecourse and regression models:

Stage at diagnosis- critical period; to evaluate the association of having a partner at age 50 and late stage diagnosis of CRC I will use logistic regression:  E(late stage)= α+β1∗marital status50   + β2 *covariates.

Mortality at five years - cumulative risk; here I will use a Cox proportional hazards model and based on Mishra et al paper will create a lifetime marital status score going from 0=never had a partner to 3=always had a partner and include it in the model along with potential confounders.

I also thought in the possibility of treating marital status as a time-dependent variable, but it then will not fit into the cumulative risk framework.  

 I do not know of an existing dataset that have the information I need to conduct the analysis. Because detailed information about cancer is required, SEER dataset could be an option, but they only collect marital status information at the time of diagnosis (I think this is the reason for previous studies only reporting marital status at this specific event). As the target population of this study would be older adults (men and women 50+ years), any dataset including them that also collect marital status at different timepoints might be useful. One option could be HRS, but the cancer information in this dataset is self-reported and data cancer-specific variables (stage at diagnosis, for example) needed for this study will not be there. Merging HRS data with information from Cancer Registries across United States would be ideal, but I also think the number of CRC cases present in the dataset at the end will be limited.  


Re: WEEK 5 READING RESPONSES

by Sepehr Hashemi -

- Exposure-outcome of interest: Association between frequent relocation (exposure), and Social Engagement (outcome) as an adult.

- Appropriate life-course model: I am not sure whether one model alone, or a combination of models can be used here. For example, one model may be sensitive period (e.g. childhood) with other later life risk factors (e.g. early loss of parents, personal illness, etc), where frequent childhood relocation shapes and disruption of social support shapes the malleable developing personality of a child to grow up to have different affinity for social engagement. However, other later life risk factor can significantly affect resulting social engagement nonetheless. On the other hand, I also see it possible for an Accumulation of risk model with chains of additive risk . For example, one who relocates often over time suffer increased financial stress, interpersonal stress (e.g. marital stress), predilection for specific employment types (e.g. self-employed), all of which will influence social engagement. Perhaps an integrated model (sensitive period with other life risk factors with possible accumulation of chain of additive risk factors) is most comprehensive, but I’ll choose sensitive period with other risk factors may be best.

- Regression models: Then a hypothesis might be that frequent relocation decreases social engagement, with sensitive period at 1-18 years of age. I would use a mixed effects model around this critical period:

Yij = B0 + B1tijδij + B2tij(1-δij) + b0i + b1itijδij + b2itij(1-δij) + eij

            Where tijδij and tij(1-δij) represent pre and post 18 years of age; B0,B1,B2 represent fixed effects (differences between persons) and b0i, b1i,b2i represent random effects (differences in the same person). I did not include the other risk factors or any interactions. Looking forward to reviewing this approach in more detail in class.

- Possible datasets: Although I need to confirm whether this datasource contains the validated items that measure social engagement, perhaps MRC National Survey of Health and Development may be a possible dataset.

- Concerns regarding life-course model: One of my concerns is that this model does not capture the possible interactions between the various variables involved (not in model). Also it does not point out possible triggering exposures (e.g. personal illness) whose risk may increase due to frequent relocation, and will have great effect on the outcome.


Re: WEEK 5 READING RESPONSES

by Maria Glymour -

Sepehr,

That's an interesting example.  Are you intending to evaluate changes in social engagement leading up to age 18 and then after age 18, or rather social engagement in adulthood as a function of early versus later relocations?

Maria


Re: WEEK 5 READING RESPONSES

by Eduardo Santiago-Rodriguez -

(Forgot to add the limitations.)

The proposed study did not cover some important considerations. First, it would be interesting to look at marital status trajectories according to different classifications, such as: married (including domestic partnership), never married, divorced and widowed. Also, it would be important to analyze time people are in each category, age and gender (effect modification?). Related to age, in recent years it has been reported an increase in CRC incidence among younger adults in the US; perhaps age of participants should no be limited in this study and additional analyses assessing the effect of marital status on screening (focusing only in those 50+, per screening guidelines) should be conducted.


Re: WEEK 5 READING RESPONSES

by Maria Glymour -

Thanks Teresa.  I'm a little confused by the model - is the outcome measured repeatedly or just once?  With education, we would typically think of just one final outcome, although it might make sense to evaluate "currently in school, yes/no" multiple times as the child grows up. 

This matters for whether the independent variables in your model are a count of previous infections or an indicator for recent infections.  

But if it's strictly the latter, it's hard to support your original argument, that there's a cumulative effect of exposure, because you cannot distinguish recent exposure from long-term cumulative.

Maria

Re: WEEK 5 READING RESPONSES

by Laura Koth -

The exposure-outcome that I do not currently study but has long fascinated people who do, is the racial and geographical association with idiopathic granulomatous disease (==sarcoidosis). I was struck by the US map pictorial image of stroke rates that Dr. Glymour presented during last week’s lecture because there are eerie similarities in the rate of stroke and the occurrence of sarcoidosis. The data for sarcoidosis came from an insurance database published in 2016 which describes the incidence and prevalence of sarcoidosis and the large racial disparities and geographical differences. The similarity between the stroke and sarcoidosis maps got me wondering about common exposures in these large areas of the country that may be associated with many diseases given the overlap observed between these two diseases. ***I uploaded an image of the sarcoidosis map*** This got me wondering whether it would be interesting to combine these types of US incidence/prevalence data for multiple diseases to serve as a rationale for such life course studies if there truly appears to be substantial overlap…..Seems like this simple idea would have already been done?

RE: the model, I think there is plausible justification for a critical period model wherein some time in childhood development, specific exposures are present that make an individual susceptible to exaggerated granulomatous disease as an adult. In this model, I would propose that there would be later life risk factors (another exposure/genetic modification changes in later life) that initiate expression of the disease. I think this argument could be made based on the geographical distribution of the disease. But some investigators also believe that sarcoidosis is the result of the other model, accumulation of risk, in play with underlying genetics as an explanation for the racial differences in occurrence regardless of the geographical distribution.  

RE: regression models. I think I may need some help with this. Presumably we are talking about longitudinal measurements related to exposures. Those may be repeated over time. Thus, I would anticipate using some type of linear mixed effects models like we are currently learning about in biostat 209.

I think I am interested in looking into the above ideas further after I complete this quarter. Regarding life course studies, there have been twin studies performed in sarcoidosis, but I don’t necessarily think they analyzed data back to time of birth or before. Databases from countries that have done longitudinal twin studies from birth would be an interesting idea but the expression of sarcoidosis is quite different in Europe so it may not inform our ideas about disease in the US. My concerns about doing this type of study is that the trigger for granulomatous disease could really be just about anything. An example is the 911 WTC attack. There is a large cohort of firefighters who developed sarcoidosis within the year after that attack. These people may have never gotten sarcoidosis if it wouldn’t have been for that severe exposure. So, the point is that I would worry that the breadth and depth of exposure history just would not be sufficient enough to answer the question about what exposures are associated with sarcoidosis in twin studies.

Re: WEEK 5 READING RESPONSES

by Zahra Izadi -

Exposure: Adverse childhood experience (ACE)

Outcome: Primary: Age at onset of systemic lupus erythematosus (SLE); Secondary: SLE disease severity and damage

Hypothesis: Stress has been implicated as a probable trigger of disease onset and flares as well as chronic disability. I hypothesize that ACE will be associated with younger age at onset, as well as increased damage and disease severity.

Life-course model: Accumulation model. I hypothesize that the longer the exposure to ACE during childhood and adolescence, the worse the SLE outcomes and the lower the age at onset. In an ideal study I would assess exposure at 2-5 years, 9-13 years and 14-18 years. Exposure at all three time points would be expected to have a greater effect than exposure at 2 time points. Similarly, exposure at two time points would have a greater effect than exposure at 1 time point, and so on.  

Regression model: As described by Mishra et al., assuming a direct and cumulative causal effect of ACE, I could estimate change in age at onset or disease damage/severity by fitting the linear regression model:

E(Y) = β0 + β1 X + β2 Z + ε

Where β0 corresponds to age at onset or damage among the group not exposed to ACE. β1 corresponds to the estimated parameter of interest (change in age at onset or disease damage/severity) and ACE indicators corresponding to different ages are summed to obtain a life-course ACE score (X) that takes the value between 0 (never) and 3 (throughout childhood and adolescent).  Z corresponds to a vector of covariates and ε is the error term.

Dataset and limitations: The California Lupus Epidemiology Study (CLUES) is a prospective cohort of individuals with SLE. Starting in 2015, participants were recruited from the California Lupus Surveillance Project and from earlier studies of genetic risk factors for SLE outcomes. Study procedures include in-person clinic visits, which comprised physical examinations and collection of clinical labs, and a structured interview which includes the Adverse Childhood Experiences Questionnaire (ACEQ). ACEQ is a validated 10-item questionnaire with 3 domains: household challenges (5-items), neglect (2-items) and abuse (3-items). Scores 4+ denote severe exposure to ACE. The major limitation of this dataset is that ACEQ is designed to assess ACE during the first 18 years of life, hence only one score is available per individual, instead of three scores each corresponding to the proposed age categories.


Re: WEEK 5 READING RESPONSES

by Dan Kelly -

Teresa, this is a really interesting question. It's great that you have an initial cohort from which you can re-sample. I would be interested to learn more about how you would balance the short-term need to measure frequency of GI infections by self-report (I assume) with the longer term need to measure educational attainment. When considering frequency of study visits as a research and pragmatic issue, what would your study visit interval be? In order words, how often would you re-sample the cohort?

Re: WEEK 5 READING RESPONSES

by Kirsty Bobrow -

Exposure of interest – lung function over time in middle-age (3 measures, years 1,5,8)

Outcomes of interest – incident cognitive impairment and dementia (annual for 16 years)

Most appropriate life course model – cumulative exposure. Reasons for selecting this include that although childhood events likely effect adult lung function (for example – being nutritionally well off and growing taller likely also reflected in better lung function test results), adult exposures like tobacco smoking, and occupation (for example being a miner) will also impact lung function tests in mid to older-age.

Use a multilevel model

1.      Cognitive trajectory

SCogij= b0 + m0i + (b1 + m1i) ageij + (b2 + m2i) age2ij + (b3 + m3i) age3ij + εij 

2.      Cognitive trajectory accounting for changes in lung function

SCogij= b0 + m0i + (b1 + m1i) ageij + (b2 + m2i) age2ij + (b3 + m3i) age3ij

+ γ1 LFTij + γ2 (ageij LFTij) + γ3 (age2ij LFTij)  + εij                      

 

HealthABC data – only have lung function data i.e. exposure data from older age (when enrolled in the study.) No information on early life events though it may be possible to construct an occupational trajectory from self-reported data. Caveat being that some people with cognitive impairment or dementia may not report all exposures of interest.


Re: WEEK 5 READING RESPONSES

by Maria Glymour -

Nice example Kirsty.  It's a major challenge in many studies of aging- we are often relying on retrospective self-reports to reconstruct life course experiences.  Some promising options involve linking data sets that recorded admin information across the life history of the individual, but that's not the norm.

Maria