Reading Response Discussions

ARIC – How do blood glucose trajectories vary over time vary by sex, geographic location, and SES in people with high normal blood glucose at baseline compared to people with abnormally high blood glucose? ARIC is prospective longitudinal with good biomarker measurements. It would be possible to see what happens to populations where a continuous variable falls just below and just above a disease diagnosis cut-point. Could also look to see if mortality outcomes are different. ARIC would probably be weaker at assessing early life risk factor exposures.

US Linked Infant Birth/Death data set – Among mothers with a first infant death are death rates among subsequent infants higher, lower or the same as mothers without a first infant death? The reasons the dataset may be able to answer this question is because of completeness of birth and death records (>98%) and because the information collected should enable a total birth record for each woman to be constructed. This data set would probably not be useful for answering questions about predictors of infant thriving or about variables not collected on the birth or death certificate.

HRS – Was there evidence of functional (i.e. motor) or behavioural difficulties among people who have normal cognitive screens at baseline and subsequently develop moderate/severe cognitive impairment (measured at follow-up in 2018)? HRS prospective and collects detailed information in multiple domains and at several time points. Again, weaker if trying to explore early life factors unless a sub-sample from a birth cohort where there is detailed information. Also weaker if outcome is associated with differential follow-up (if attrition is a plausible concern.)

Linked Birth and Infant Death:

Are there differences in race and prenatal usage that can be associated to differences all cause infant death? The data set collects multiple causes of infant death within the 1^st year and using the birth certificate has parental race information and prenatal usage.

But because the data set does not include measure of quality of care or previous information on prenatal services. The same question but framed to address quality of care would be difficult to address in this design. If there are racial differences in quality of prenatal care usage and if so are they associated to any difference in all cause infant mortality?

ARIC:

For high risk (high cholesterol) participants, do review sessions low risk of long term complications? ARIC review session goes over test finding and even provides assistance with proper referral for diagnosis treatment. With 3 year tests and annual surveys and the community surveillance the participants are properly tracked and any worsening condition for high risk participants can be followed same if their risk is reduced. The exposure and outcome can be modified to fit one of the many tests they do and different possible complication the community surveillance tracks stroke, heart attack or simply more build-up of cholesterol.

A similar question but looking at If review sessions after testing improves participant knowledge on tests and results related to atherosclerosis and their own health. While the data set tracks and does regular interviews with participants it does not survey knowledge of subject matter. So questions around patient knowledge or lifestyle based on test outcomes would not be easy to address. 

ISGS:

Are there differences in rates of polymorphisms of genes associated with ischemic stroke among individuals from varying ethnic origins? ISGS tracks a panel of genes associated with ischemic stroke and the polymorphism of these genes. The data set also has ethnic origin.

The data set does would be weak at addressing question about patients’ lives before the stroke and changes in life as it is self-reports after the incident. So examining possible dietary association with isxhemic stroke would be difficult.

Assignment: 

There are a lot of readings here, but most are quite short. Throughout, the goal is to understand the research design, not the specific content of the study.  Prior to class, please post on the course website answers to the following questions:

>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 ona  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.

The ARIC study is composed of 16,000 persons aged 45-64 in 4 distinct communities, examined 3 years apart with annual follow-up.  A potential research question to address from this basis could be: How does reported family history of cardiovascular disease impact retention in the cohort?  The interviews detailed in the paper explain how home interviews were conducted for potential cohort members.  These included family medical history.  A limitation to this question would be related to limitations in known family disorders, however by limiting the question to reported cases of known family histories of cardiovascular disease in family members a case-control study could be performed using incidence density sampling for controls.  A weaker research question could involve the same posed in the study or the above, but in a different target population, for example the native American population.  North Carolina and Mississippi bother have only one protected native American reservation within their states, Maryland has none, but Minnesota boasts 13.  The publication does not explicitly mention addressing this distinct ethnic/cultural group, yet they represent a considerable at-risk population.  An explicit question could be: Is there a significant distribution in the areas of atherosclerotic involvement in native American populations?  Addressing this question (which is a bit of an easy-answer, I admit) would require accessing the target population, which seems particularly difficult given the socioeconomic and cultural barriers needed to surmount.

Pickett et al. describe a data set of linked birth and death certificates for all infants born in the United States including those who died in the first year.  This data could address the question: In children born to young mothers, what is the effect of nativity (born in the US vs. outside the US) on rate of twins compared to children born to older mothers?  While the data contains mortality data this question would not require analyzing mortality at all.  A weaker question that could be addressed would be Does the rate of mixed-race parents (ie. mixed race babies) risk of morality in the first year?  This question requires distinction between specific races and does not appear to be specifically asked in the questionnaire used in the US Linked Birth/Infant Death data set.  Additionally many individuals, especially in the United States, represent mix of multiple different distinct ethnicities and may be considerably difficult to track and quantify.

Pagidipati et al. describe estimations of mortality from cardiovascular disease using death certificate data.  Their data would also be effective at answering the question: How has the rate of suicide in non-minorities changed over time in America?  A potentially weak question to address could be What has been the cause of death to increase in greatest proportion in Pakistan?  Pakistan is one country where verbal questionnaires have been used in the past.  Answering this questions would be not be reliable as if the same question were posed but “in America?”.  This relates mostly to the practice of applying a questionnaire to a family member determined to be most familiar with the circumstances surrounding the decedent’s death with the results of this questionnaire used to determine likely cause of death.  Some disease entities can be very subtle in presentation or overlap with other disease entities.  For example, septicemia is listed as the 11^th most common cause of mortality in this study group, however a number of specific infections that may lead to septicemia or share a common cause with septicemia are also listed.

Data source: Global Health Observatory Data

Strong Research Question: What effect did roll out of the Millennium Development Goals in 2000 have on under-five mortality rates across the 6 WHO regions?

Why is this a strong data source for this question? The Global Health Observatory Dataset is the WHO’s official data repository which includes global disease prevalence, substance abuse, vaccination coverage, health coverage, water and sanitation, mortality, etc. It is considered a “Metadata Registry” encompassing almost every country in the world with a standard set of 100 indicators selected by global leaders to provide accurate data on health situations and trends. As the largest global health organization in the world, the WHO has significant influence over how indicators are defined and measured, thus improving indicator homogeneity. Primary data sources are from individual countries using a combination of civil registration, census, and survey data. Though there are issues with mortality data worldwide, especially in resource-limited settings that often lack a formal vital statistics registry and where many deaths occur in the home without contact with the medical system, mortality data is still far more complete and easier to capture in these settings compared to morbidity and cause of death data. This dataset is particularly good at answering questions regarding global trends.

Weak Research Question: Were there differential effects on under-five deaths due to pneumonia by region due to the introduction of the pneumococcal vaccine in 2000?

 Why is this a weak data source for this question? As stated above, cause-specific mortality is challenging to measure even when a vital registry linked to death certificates exists. Additionally, mortality due to pneumonia requires a clinician’s diagnosis, thus excluding all pneumonia deaths that occur in the home, which is not an insignificant number. Most likely, the total number of child deaths due to pneumonia is grossly underestimated.

 Other study that could address the hypothesis: As described in Pagidipati et al., one could develop a sample vital registry system in low-income countries that requires longitudinally following a sample of nationally representative households to determine the vital status and cause of death in the representative sample. National mortality estimates could then be extrapolated from these sample data. Additionally, the use of verbal autopsy methods to determine the cause of death, especially when the individual dies at home, could be very helpful, though verbal autopsy methods are neither standardized nor validated.

Data source: Global Burden of Disease

Strong Research Question: What effect did a large-scale mass distribution of bednet campaign in 2012 have on the prevalence of malaria and malaria-related death in children under five in Malawi?

Why is this a strong data source for this question? Unlike the WHO’s Global Health Observatory Data, the Global Burden of Disease group pulls data from thousands of global sources of data and uses epidemiological modeling to handle missing data. This method is particularly helpful for resource-limited settings where limited, or no, mortality data are available. GBD is currently the most comprehensive dataset that exists estimating global trends in causes of morbidity and mortality, though it should be noted that it is only as reliable and accurate as the underlying data. However, given the multiple data sources available that are cross-referenced to create the dataset, I have more confidence in the GBD’s estimates than in the WHOs. 

Weak Research Question: How did adoption of the WHO’s Framework Convention on Tobacco control in 2003 affect tobacco-related deaths in low-income countries?

Why is this a weak data source for this question? It is very difficult to ascertain if a death is due to tobacco based on cause of death information alone; we really need individual-level tobacco use data linked with mortality data to begin to answer this question. GBD aggregates available data and models predictions based on the data sources available; it does not have the granular data required to answer this question on a global or national level and it would require significant extrapolation of existing data to estimate this.

Other study that could address the hypothesis: A prospective cohort study involving a representative sample across randomly selected LICs would better address this question. A repeated survey tool could measure risk factors for mortality, including tobacco use, which would need to be measured at least once pre-Framework Convention and confirmed on follow-up surveys. Given the limitations/lack of vital registries, cause of death information, and autopsies in these countries, use of verbal autopsy methods, despite their limitations, to determine cause of death are likely the best option for determining tobacco-related mortality.

Data source: UNICEF Child Health Coverage Database

Strong Research Question: In Bangladesh, did the 2003 mass oral rehydration and zinc administration campaign decrease socioeconomic inequalities in parental health seeking behavior for children under five with diarrhea?

Why is this a strong data source for this question? This database compiles demographic and health survey data from most countries around the world. Most countries have multiple timepoints available, some beginning in the 1990s, though the year of data collection varies by country. Bangladesh has 7 surveys available from 2000-2017.

Weak Research Question: What was the impact, in terms of caretaker acceptance and household zinc administration, across different socioeconomic strata of the first national campaign to scale up zinc administration in the treatment of childhood diarrhea in Bangladesh?

Why is this a weak data source for this question? As stated above, this is a database that compiles demographic and health survey data; it does not have household-level data or detailed survey or interview data regarding acceptance.

Other study that could address the hypothesis: An alternative study to answer this question could involve repeated household surveys administered to a representative sample of households with children with active or recent diarrhea representing five socio-economic strata. The survey would confirm socioeconomic status and ask questions regarding caretaker knowledge, acceptance and zinc administration.   

Linked infant birth/death data

Strong: These data could be used to explore whether there are racial disparities in survival of congenital abnormalities, and whether these vary by region. It is useful because it is nationally representative and records demographic and geographic data.

Weak: At first, I thought it might be interesting to look at whether day/time of birth affected infant survival – often hospitals have less experienced staff immediately available on evenings and weekends. However, I realized that this would be confounded if almost all elective c-sections and inductions happen during the daytime on weekdays. Given that elective c-sections are often done because pregnancies are higher risk, this would bias the effect estimates. A multi-center hospital dataset linked to death data would likely be better to address this question, with some measure of whether procedures were planned.

Verbal autopsy data

Strong: Verbal autopsy data could be used to decompose the beneficial effects of a public health intervention shown to reduce child mortality in rural areas without much access to medical care. For example, if a nutrition intervention was found to decrease child mortality, it could have worked by decreasing deaths from starvation, infectious diseases, or other causes. Verbal autopsy data could improve understanding of what causes of death were prevented, and therefore, what mediated the treatment effect. Verbal autopsy would be ideal for this if there were not access to formal medical diagnoses or autopsy; it would be cost-effective.

Weak: Verbal autopsy data is most useful to assess broad categories of diseases linked to symptoms – for example, acute diarrhea linked to an acute GI illness. It is not useful to identify things that cannot be observed by a caretaker, such as specific pathogens or diagnoses like leukemia. To evaluate questions with outcomes like these, one needs medical records with laboratory, physician, or other diagnostic data.

 Nurses’ Health Study

Strong: If the caregiving questions had continued beyond the 1992 survey, it would’ve been interesting to look at the 1996 Family Medical Leave Act’s effect on caregiving hours and stress and its negative health effects as a natural experiment. If the caregiving stress questions had been asked repeatedly along with the health data collected frequently by the Nurses' Health Study, it would be an excellent data source to analyze the effect of the policy as a natural experiment in the cohort.

Weak: The Nurses’ Health Study is not useful to look at the effect of education on health outcomes – everyone who is a nurse has at least an Associate’s degree. Thus, lower levels of education are underrepresented. Such questions would be better to explore in a sample that is more representative of the full range of education levels in the United States.

Data source 1: Linked birth/infant death dataset

Question this data source would be well equipped for: How did the Great Recession impact infant mortality? This data set is well equipped to answer this question as it includes data on all births/deaths over time. An interrupted time series analysis could be performed to observe the impact of the Great Recession over time (including time points before and after the Recession). This analysis could be stratified by geographic region and SES in order to determine whether the impacts differed across population subgroups.

Question this data source would not be well equipped for: How does gestational diabetes affect growth/thriving after birth? This dataset only gives information on births and deaths, not on the quality of life of the child after birth. The only children that have a second observation are those who died.

Data source 2: HRS

Question this data source would be well equipped for: How does having emotional support impact all-cause mortality? This data contains information on emotional support and social networks among the elderly. With all-cause mortality as an outcome, there is no reason to be worried about temporal order or reverse causality.

Question this data source would not be well equipped for: What is the impact of exercise on incident cancer? This dataset may not work well with outcomes that are difficult to detect and may go undiagnosed for long periods of time, such as cancers. This is because establishing temporal order may be challenging, as cancer may have gone undiagnosed for some period of time before exposure ascertainment for time invariant exposures such as exercise.

Data source 3: Uganda Malaria Surveillance Program Database (not in the readings, but pertains to my interests more…)

This is a large malaria surveillance database in 35 districts in Uganda. Individual-level data on all outpatients are collected from sentinel sites in public health facilities. The database is updated monthly (2006-present). All individuals who are suspected of having malaria due to clinical symptoms receive a laboratory test and are treated accordingly.

Question this data source would be well equipped for: How have control interventions impacted malaria incidence over time? This dataset includes health facilities in districts that did and did not receive large campaigns of indoor residual spraying of insecticides. It could lend itself well to a difference-in-difference analysis.

Question this data source would not be well equipped for: What is the probability that an individual would seek care if they were ill with malaria in Uganda? Although this dataset has information on individuals who attended the health facilities, it does not include data on any individuals who did not make it to the facility. A better dataset to answer this question may be the Malaria Indicator Survey since it asks questions on the likelihood of attending a health facility for fever episodes.

1. The Nurses’ Health Study was established in 1976 when 121,700 female registered nurses, aged 30 to 55 years, completed a mailed questionnaire providing information about risk factors for cardiovascular disease, cancer, and other major health conditions. Since then, follow-up questionnaires have been mailed to the cohort every 2 years to update information on exposures and the occurrence of major illnesses. Nurses Health study

This study design would be considered strong to address my hypothesis about the effect of taking blood pressure medications on the outcome of cardiovascular events of stroke, MI. This is not particularly interesting to me but I think the data collected would be able to provide information about this question and have the ability to adjust for known confounders.

This study design would be considered weak to address my hypothesis about the effect of exposure to cigarette smoking on the incidence of pulmonary sarcoidosis. This would be something of interest to me but based on the design, this is a targeted sample, so you could not conclude anything about the population incidence, but the incidence in female RN's. There is one study from Detroit 1997 using insurance data that found an incidence in white women of ~11 per 100K per year. I don't know if this cohort would be big enough to answer this question in this population of women.

2. Nhanes III

This study design would be considered strong to address my hypothesis about the effect of race on the outcome of spirometry measurements. Lung volumes are proportional to body size and absolute volumes must be normalized for factors that influence total volume size in order to know what is abnormal. Racial background influences body size (e.g. height). Therefore, you need to have a representative sampling of lung volumes from “normal” people in order to generate regression equations that can determine what a person’s expected lung volume should be for a given gender, race, height.  

This study design would be considered weak to address my hypothesis about the effect of nutritional problems in immigrants. Although this data set represents a nationwide probability sample of 27,801 persons with oversampling from people at or below poverty, I would be concerned that immigrants might shy away from participating in a study run by the “government” for fear of deportation or other concerns related to responding to a request to participate.

3. Insurance claims: IC9/10 codes from a large catchment area.

This study design would be considered strong to address my hypothesis that geographical exposures effect the incidence of pulmonary sarcoidosis. Provided the insurance provider was large and covered a wide area, you could look at the diagnosis of sarcoidosis by ICD codes plotted by zip code to see if there appeared to be clusters of cases relative to density of population.

This study design would be considered weak to address my hypothesis about the types of environmental exposures (dust, hobbies, chemical fumes) that cases reported. The database would not contain this type of information in a standardized or complete manner.

Wisconsin Longitudinal Study: The Wisconsin Longitudinal Study is a cohort of 10,317 men and women who graduated from Wisconsin high schools in 1957, and their randomly selected siblings. Survey data includes basic information about the life course from late adolescence through early/mid-60s. Strong question: What is the effect of adolescent political participation on lifetime risk of depression and anxiety? Since this longitudinal study started in adolescence, I think it would be an ideal data source to address questions related to early adolescent exposures. Weak question: What is the probability of intergenerational social and economic mobility among adolescents experiencing poverty? This question can definitely be answered by the WLS, but because it’s HIGHLY contextual question, it may not be of any relevance in the current political and economic climate. I do think it may be interesting to pursue for the purposes of understanding potential mechanisms that cause adolescents experiencing poverty to attain economic mobility (i.e. educational attainment, ‘white collar’ employment, etc.)

Nurse’s Health Study: The Nurse’s Health Study is a prospective cohort of female RNs residing in 11 U.S. states who were healthy (free of CHD, stroke and cancer) at baseline. Strong question: Do racial/ethnic disparities in cancer persist in a similarly educated cohort of nurses? I think this would be an interesting question that could be answered using the Nurse’s Health Study – since all of the members of the cohort are highly educated and working in the same occupation, any observed racial differences in cancer outcomes would provide evidence that education does not completely attenuate racial disparities. Weak question: What is the effect of long-term depression on mortality? While I do think that this question could be answered using the Nurse’s Health Study data, I think the relationship between depression and mortality is more socially (rather than biologically) driven. So, I wouldn’t be comfortable generalizing any results from this type of study to the larger U.S. population.

Ischemic Stroke Genetics Study (ISGC): This data comprises a multicenter cohort of 673 subjects with first-ever ischemic stroke who were enrolled between December 2002 and November 2007. Participants were assessed at baseline (when consent was signed) and approximately 3 months thereafter. Good question: Are stroke treatments/medications equally effective at improving the likelihood of physical function post-ischemic stroke across genetic differences? Since this study collected genetic data among adults with first-ever ischemic stroke and followed them over time, it’s possible there are enough patients who received the same medication/treatment, as well as adequate genetic variation per treatment. Weak question: How does post-ischemic stroke physical functional status differ across medical centers? This could be a difficult question to ask because, although this is a multicenter study, all the patients were evaluated at academic medical centers (which typically provide similar levels of care/have similar standards of care, I’d imagine, compared to safety net hospitals). So, any variation in functional status could potentially be attributed to the provider and not the center.  

Nurses health study

Strong question: what is the effect of physical activity on miscarriage and stillbirth? The study includes data on physical activity exposure (measured in MET unit quintiles) and on miscarriage or stillbirth among other reproductive life events over decades. It is longitudinal so establishing temporality would be possible regarding physical activity exposure prior to miscarriage/stillbirth.  The NHS has also been used to answer questions about the effect of oral contraception on risk of cancer (breast and cervical vs colorectal and endometrial cancer). It is well suited to these questions because it’s longitudinal and includes women of reproductive age.

Weaker question: What is the effect of education on risk of stroke in women? Are there racial disparities in the effect of oral contraception on cancer? These questions are hard to answer with this dataset because the sample includes well educated and primarily white American women. In the third round of NHS there is an effort to include more diversity. A sample such as HRS would be better equipped to answer this question because it has a much more racially/ethnically diverse sample and an important goal of HRS is to support questions on racial and ethnic disparities.

HRS

Strong question: Are there racial/ethnic disparities in the effect of depression on risk of dementia? HRS oversampled black and Hispanic populations in most of the recruitment cohorts. HRS has also been successful at maintaining high baseline and followup response rates for black and Hispanic sample members.

Weaker: what is the effect of ACEs on risk of stroke? This would be hard to answer given that the sample includes ages 50 and up and all reporting (if done at all) on ACEs would be retrospective self-reported data. 

US Linked Birth/Infant Death

Strong question: what is the effect of birthweight on the risk of SIDS death and any death? Does the effect differ before/after the campaign? Missingness is very low for birthweight (0.2%) as compared to the other variables collected in the study. Data on birth and death are linked which enables us to look at this question for both periods.

Weaker question: it would be difficult to look at the effect of maternal tobacco use during pregnancy and infant death due to a large amount of missingness (25.8%) for maternal tobacco use in the records. I would instead likely use a national behavior surveillance survey to look at this question. A survey like the Nurses Health study could also potentially work given that it collects biomarkers as well as self-reported survey data. 

Health and Retirement Study

The HRS would be exceptionally strong to determine whether those with elevated CRP levels (biomarker) are at an increased risk of death. This dataset is strong to answer this question because although it is cross-sectional, we know that elevated CRP levels would precede death. Half of the subjects in HRS provided blood samples in 2006, 2010 and 2014. The other half provided blood samples in 2008, 2012 and 2016. The mortality data in HRS is obtained in two ways: when organization attempts to interview respondents for the next survey wave and through matching with the National Death Index. Given the advanced age groups in this dataset, there will likely be many subjects who have an elevated CRP and also experience death.  

The HRS would be weak to address whether elevated CRP levels result in an increased risk of cancer. Due to the cross-sectional nature of the data, and often long detection period of many cancers, this dataset would not be able to establish temporality. For instance, malignancy may cause CRP to be elevated.

Ischemic Stroke Genetics Study

The ISGS is a prospective, multicenter study in adults with recent first-ever ischemic stroke confirmed with computed tomography or magnetic resonance imaging, compared with sex- and age- matched controls. This study would be ideal to answer the question whether polymorphisms of the β-fibrinogen gene result in an increased risk of ischemic stroke. This cohort is ideal for examining an exposure that is fixed and does not change over time. This is because the patients (cases) are recruited are included in the study after they had a stroke.

This cohort would not be ideal for answering a question about an environmental exposure such as air pollution. This information would be collected through an interview after the patients experienced a stroke, and the exposure would be subject to recall bias. This would be better answered by a cohort study that is assembled and followed longitudinally. Participants would be asked about environmental exposures such as air pollution before they had a stroke.

Death certificate data

Death certificate data would be strong to answer the research question of whether cadaveric renal transplantation is associated with a higher mortality compared to living donor renal transplantation. Death certificate data however would be weak to answer the question of whether cadaveric renal transplantation is associated with a higher cardiovascular specific mortality compared to living donor renal transplantation. This is due to inaccuracies of death certificate diagnoses. The latter question would be better answered using autopsy data or data where the death certificate diagnoses are adjudicated by a panel of physicians.

Nurses’ Health Study:

This is a prospective study of women with no cancer at enrollment (incident cases can be assessed) and participants are followed every two years. This study could be strong for answering what is the effect of breastfeeding on the risk of breast cancer, as exposure is collected before the outcome, and outcome can be ascertained with cancer registry data. On the other hand, this study could be weak for evaluating the effect of SES on the risk of breast cancer. Because of the main inclusion criteria (participants had to be registered nurses), study population is likely to be homogeneous with regards to this exposure making it difficult to find a statistically significant relationship with the outcome.  

Health and Retirement Study:

As described by Banks et al, HRS is “known to have high-quality measurements of several dimensions of socioeconomic status—education, income, and wealth” and because it is a nationally representative survey, information of SES should be diverse enough to evaluate its effect on the risk of breast cancer (as opposed to NHS). This study design is weak for assessing early life exposures, as tobacco while in utero and its association with COPD. As the study was not in place at the time the exposure occurred and by obvious reasons participants cannot self-report this, it would be hard to have this information in medical records for this period before EMR implementation. A birth cohort with several years of follow up is an alternative for this question.      

 Death certificate data:

A study evaluating overall mortality rates by race/ethnicity for the past 5 years in California could be adequately addressed with this data. However, this source of information could be problematic for evaluating the number and causes of deaths after a natural disaster. For example, months after hurricane Maria made landfall on Puerto Rico the official death toll was only 64. Because of public concern after many independent studies have estimated thousands (and also because of what people had witnessed in their communities), the government of Puerto Rico commissioned a study to the Milken Institute School of Public Health of George Washington University, and they estimated it was almost 3,000. Among the reasons for the discrepancy authors of the report mentioned physicians were not aware of guidelines for assigning causes of death after an event like this.  A second phase of the study would be to contact and interview relatives of deceased people to confirm events (example of verbal autopsy).

 https://www.nytimes.com/2018/08/28/us/puerto-rico-hurricane-maria-deaths.html  

https://publichealth.gwu.edu/content/gw-report-delivers-recommendations-aimed-preparing-puerto-rico-hurricane-season

Data Source: ARIC. The ARIC cohort was expanded to prospectively collect data on cancer incidence, recurrence, and progression starting in 2013 and moving forward.

 Research question data is well suited to answer & why: Is insulin resistance (with or without diabetes) a risk factor for pancreatic cancer? Individuals in the ARIC cohort have regularly collected insulin measures starting in 1989, allowing us to understand how changes in insulin resistance well before diagnosis may influence disease onset. This is important, as pancreatic cancer is often diagnosed late and therefore insulin measures even two years before diagnosis might be a result of disease rather than a cause.

 Research question data is NOT well-suited to answer & why: This data set would not be well-suited to study the role of early-life insulin resistance on the risk of pancreatic cancer development due to the age at enrollment.

Data Source:  HRS.

 Research question data is well suited to answer & why: Does diabetes increase risk of cognitive decline? The dataset collects data on the year of diabetes diagnosis, allowing us to ensure it is diagnosed prior to cognitive decline. HRS also has a strong (from what I understand) measure of cognitive decline/AD, compared to other cohorts.

 Research question data is NOT well-suited to answer & why: Does hypertension increase risk of developing Alzheimer’s Disease? The year of hypertension diagnosis is not provided, introducing issues of temporality.

Data Source: Nurse’s Health Study

 Research question data is well suited to answer & why:  Does eating late at night increase risk of developing diabetes?  Timing of consumption is not regularly assessed in diet questionnaires, making NHS a strong candidate to study this association. Because they also ask for anthropometric data (e.g., waist and hip circumferences) we can also look to see if this association varies between by obesity status. Timing of diabetes diagnosis is also importantly captured in this dataset.  

 Research question data is NOT well-suited to answer & why: Does eating sugary snacks late at night increase risk of developing diabetes?  NHS does not routinely assess consumption of specific foods (last assessed in 1998), making it a poor choice to study relationships with specific foods or nutrients.

NHS: 

Strong question in using NHS data in 1976: The original mailed questionnaire for the Nurses Health Study was started in 1976. In the questionnaire, it asked about a variety of health outcomes and habits including a fair amount of detailed questions about their smoking habits which were likely of great interest at the time. I think a strong question at the time would have been “What is the effect of cigarette smoking on fibrocystic breast disease?” The great thing about this particular longitudinal dataset is that it allows us to take into account the timing of our outcome of interest as well as the timing, magnitude, and other key metrics of the exposure. Another advantage to this dataset is that performing this type of a study in an RCT really isn’t possible in today’s setting since it would unethical to assign people to smoke. Weak question using NHS starting in 2000: What is the effect of cigarette smoking on fibrocystic breast disease? Back in 1976 I imagine cigarettes were not as well recognized to be as unhealthy as they are today. Hence, you likely could have gotten a fair amount of cigarette use representation in the nurses that may have likely matched that of the general population. However, nowadays I think it's likely uncommon to find many nurses that smoke and I’m not sure if those who do would be representative of the general population.

HRS:

The Health and Retirement Study is a nationally representative longitudinal survey that includes several thousand participants over age 50 in the US and data is reported every 2 years since 1992. I think this particular dataset has been great for addressing research topics geared towards individual elderly health as well as for addressing population/societal issues at large (i.e. how to properly address a rapidly aging population). I think a strong research question for this dataset would be the risk of death or poor outcomes in elderly patients using antidepressants. A weak research question would be a study that attempts to assess the risk of death or poor outcomes in all patients who use antidepressants. In this case, using the HRS dataset would be very limited in answering this research question since it would not be representative of the younger population that is commonly prescribed antidepressants. Additionally, those who had a poor outcome may not even live until 50 hence, using HRS to address this question is not ideal.

Nhanes:

   National Health and Nutrition Examination Survey (NHANES) is a program that was intended to assess the nutrition and health of children and adults in the US and it has been performed every year since 1999. In this particular program, their data would be very useful for conducting studies about the nationwide prevalence of diseases, dietary habits, living conditions, etc. I.e. what is the prevalence of asthma, lead poisoning, stunted growth. However, a major limitation to the NHANES study is that it actually isn’t a great representation of most places in the united states. The way they collect data is by going to 16 cities a year where they collect their data, thus its not geographically representative of the entire US. Hence a weak research question would include trying to answer the same questions I mentioned earlier but in more rural and local settings.

HRS: It’s well designed to address the effect of stigma on cognitive decline in the USA because of the representative, longitudinal data capture and high-quality assessment of cognition. It’s poorly designed to address the effect of stigma on stroke incidence in the USA because although there is longitudinal data capture, the study doesn’t assess clinical outcomes such as stroke, and it focuses on survey response with sub-samples of home-based visits rather than health facility-based data collection. 

REGARDS: It’s well designed to address the effect of incident stroke on cognitive decline in the USA because of the representative, longitudinal data capture by telephone with retrieval of medical records from healthcare facilities in the USA. It’s poorly designed to address the effect of stigma on cognitive decline in the USA because the survey design lacked a strong multidisciplinary perspective such that psychologists were either not part of the process or were not given enough space in the questionnaire to include those questions. 

DHS: It’s well designed to address the effect of stigma on HIV testing in Sierra Leone because of the representative, cross-sectional data capture and self-reported health assessments such as stigma and HIV testing. It’s poorly designed to address the effect of stigma reduction on HIV testing over time in Sierra Leone because the survey is not longitudinal in nature, though there are repeated cross-sectional surveys over time, which are not of the same individuals but may be able to give context of community-level indicators such as stigma. 

Healthcare Cost and Utilization Project (HCUP), Nationwide Emergency Department Sample (NEDS):

Are there state-level variations in racial disparities/inequities for psychiatric hospital admission rates? This is a multi-State database, which it makes it possible to use state-level indicators in data analysis at given time points. This database would not be suited for questions about individual change over time (aka readmissions over months/years) because it does not track individuals.

English Longitudinal Survey of Aging (ELSA):

Is there an interaction between depression and social participation on cognitive decline over time (measured by (verbal learning and verbal fluency)? This study asks about mental health and social participation as well as cognitive function. Verbal fluency and verbal learning tests were give in each wave of the study. This cohort is not as useful for addressing questions using a life-course perspective (i.e. early life experiences or childhood SES).

ARIC cohort: Does baseline depression or anxiety predict pulmonary function 10 years after baseline visit? This database tracks pulmonary function with the goal of determining predictors of change. It would not be well suited to study long term trends in groups with differential attrition/loss to follow up. Again, it would also not be suited to answer questions based on a lifecourse theoretical framework. It also might be difficult to study behavioral risk factors outside of diet and exercise or structural vulnerabilities using this database.

ARIC:

Strong: In ARIC cohort study we could investigate the effect of race on the change of total cholesterol, LDL, HDL over time (since this study measures in two occasions these biomarkers). We could also explore a potential interaction of race with the income of study participants. The aim with this research question is to evaluate if race and income have an effect on cardiovascular disease and other chronic conditions by changes in lipids with time (age). So, if we found that there is an effect of race and income on changes in lipids, we could test if that change in lipids has an effect on cardiovascular disease.

Weak: This dataset does not give information about early life exposures that could influence cardiovascular conditions later on such as parental SES, early life residency, education quality, abnormal lipid values during childhood.

The REasons for Geographic And Racial Differences in Stroke (REGARDS) Study:

Strong: In REGARDS cohort we could investigate the effect of glomerular function (EGFR) in age >45 on the risk of cognitive impairment and Alzheimer’s disease later in life. This research question would require linking REGARDS with Medicare data.

Weak: My understanding is that REGARDS have only baseline measurements of kidney function and cognitive function. Investigating the effect of kidney function on cognitive impairment when these two variables are measured at baseline would not be a strong study design.  

Health and Retirement Study and English Longitudinal Study on Aging:

Strong: These databases have similar study designs since the ELSA is was developed as the English version of the HRS. We could investigate in these databases (separate or together) the effect of the number of children (biological and adopted) in the household on cognitive decline. The HRS and ELSA have detailed information about family structure, measures of SES and other early life confounders. They also have longitudinal assessments of cognitive function.

Weak: A research question that HRS and ELSA would not answer well will be the effect of medications (e.g. blood pressure medications) that participants are taking on cognitive function since data on medication is self-reported and could lead to measurement bias.

ARIC study

Strong Question. To evaluate the risk of chronic kidney disease (measured with glomerular filtration rate) in patients with impaired glucose (glucose >= 100 mg/dl), and evaluate if race, SES and education modify the effect of prediabetes on the incidence of chronic kidney disease. 

It is a strong question since has a big sample size (16,000), that will allow us to evaluate interaction. Furthermore,  it is a representative sample, we will expect to have a wide range of SES, education, race and other sociodemographic characteristics. 

Weak question.  Is impaired glucose a predictor of cardiovascular mortality?

IT is a weak question using the community surveillance design since the mortality data will be obtained from the discharge listings from the hospitals and death certificates. According to the Pagidapti paper, cardiovascular mortality obtained from death certificates has many limitations, such misclassification of the cause of death and errors on the coding. 

 Nurse’s health study (NHS)

The NHS is a cohort study of 121,000 female nurses from 11 states of the US, enrolled since 1976. Every two years the nurses responded a follow-up questionnaire that includes diet information, with follow up rates of around 90%. Blood and urine samples were also collected.

Weak question. 

Does the sugared sweetened consumption on the adolescence increase the risk of diabetes in adulthood?

IT is a weak question using NHS since the question about the SSB consumption on adolescence is a retrospective question that could be subject to recall bias. It is possible that those with a higher risk of diabetes respond differently to the question than those that do not have diabetes risk.

 Strong Question. Do changes in sugar-sweetened beverages and water intake is associated with long term changes on kidney function on women 30-55 years? 

It is a strong question because the questionnaires are validated, and there is no risk of recall bias since the diet question is about the current consumption of beverages. It could be some measurement error on diet, but I would expect that it is random, and thus we would only observed a weaker, but a true association.

Atherosclerosis Risk in Communities (ARIC) study

Research question the study is strong to answer: A good research question would be to look at racial differences in etiology of atherosclerosis. More specifically to investigate if the association between atherosclerosis and lipid profile is modified by race, using a cross-sectional study design.

Research question the study is weak to answer: To investigate the effect of lipid profile on the onset of diabetes among different racial groups. While this question is clinically important, it is not appropriately answered using ARIC because ICD codes for diabetes are only available in the subset of participants who were hospitalized and that could bias the results.

Nurses’ Health Study (NHS)

Research question the study is strong to answer: A good question would be to investigate if the effect of self-reported depression or anxiety on cardiovascular (CV) disease is modified by exercise.

Research question the study is weak to answer: In Lee’s paper investigators looked at the association between caregiving and CV disease incidence with the hypothesis that caregiving induced stress could be related to CV incidence. NHS is a longitudinal database that by design is appropriate for answering a causal question but I think the investigator’s choice of question is not particularly a good one due to contamination.

Ischemic Stroke Genetics Study (ISGS)

Research question the study is strong to answer: ISGS is a multicenter inception cohort study of subjects with first ever ischemic stroke. A good question would be to investigate the effect of comorbidities at the time of stroke on functional status after stroke. An alternative question would be to investigate the effect of stroke characteristics (infarct size, location, etc.) on functional status at three months post stroke.

Research question the study is weak to answer: I think the investigators choice of exposure (physical activity) was not a good one due to potential for differential misclassification. Physical activity was assessed using a questionnaire (after stroke) to determine patients’ activity over the preceding year.