The integration of genetics and the social sciences will lead to a more complex understanding of the articulation between social and biological processes, although the empirical difficulties inherent in this integration are large.

One key challenge is the implications of moving “outside the lab” and away from the experimental tools available for research with model organisms. Social science research methods used to examine human behavior in nonexperimental, real-world settings to date have not been fully taken advantage of during this disciplinary integration, especially in the form of gene–environment interaction research.

This article outlines and provides examples of several prominent research designs that should be used in gene–environment research and highlights a key benefit to geneticists of working with social scientists.

SINCE THE PUBLICATION IN Science of empirical evidence of gene–environment (G×E) interaction, there has been growing interesting in integrating biological and social science approaches, data, and models. The original results by Caspi et al.1 suggested an important, genetic source of heterogeneity in responses to early life insults, attempting to partially answer the question of why some individuals are resilient to stressors, whereas others experience deleterious psychological sequelae. Although these studies created substantial interest in potential gene-by-environment interactions, they also needed to be replicated and extended by other researchers using alternative data. There are now competing meta-analyses suggesting either that the original results linking differential response to stress by the serotonin transporter gene (5-HTT) is robust2 or lacks consistent supporting replication.3

The discussion generated by this line of research in the biological and social science communities has been valuable in highlighting the shortcoming of the research design by Caspi et al. A key concern that has been the subject of much debate is whether the study (and studies like it) is adequately powered.4,5 We point to another concern that is the subject of less inquiry. Even with highly powered studies (many current collaborative groups have amassed data sets that include tens of thousands of individuals), an important conceptual (and statistical) issue is the likelihood that the measured environments may be correlated with unmeasured genetic variation, and thus, may be acting as proxy for a gene-by-gene interaction rather than a G×E interaction. As sample sizes continue to get larger, a shift in focus should be from the statistical issue of power to the conceptual issue of modeling interactions between variables that are not themselves correlated (gene–environment correlation [rGE]). Although for studies aiming to detect main effects of genotypes, approaches that try to control for population stratification—such as genomic control,6 principal components,7 or family-based analysis8—may be adequate to account for rGE, when trying to model G×E interaction effects, the added burden of obtaining exogenous environmental variation is present, lest models become misspecified.

In light of this uncertainty, many researchers have turned to examinations of model organisms to be able to control—through random assignment—the environment as well as the genotype of animal subjects. Because human research focusing on genetic and environmental interactions will be unable to use truly experimental research designs in the near future, this leaves G×E research in a precarious position. On the one hand, results from animal models, where both the genetic and environmental contributions of phenotype can be experimentally altered, will no doubt continue to be used to suggest likely mechanisms involved in similar human phenotypes. However, it is often difficult for social scientists, and others, to fully disregard the difficulty in translating results from animal models to human populations. On the other hand, research on humans often has little leverage in experimentally altering genotype and social environment (outside of laboratories) to facilitate causal inference in G×E research. Although there is active involvement in enrolling individuals in randomized controlled trials and examining genetic heterogeneity of causal effects, this is only a small area of, typically pharmacological, research and likely does not have the capacity to answer many important G×E questions of broader relevance to public health. Because many public health interventions occur on a large scale, such as state soda taxation, federal alcohol access policies (e.g., the minimum legal drinking age of 21 years), and federal guidelines for clinical care, only large epidemiological and social science data and methods, combined with genetic and biomarker measures, will be able to examine issues related to broad public health questions.

In this essay, we suggest a way forward in G×E research in humans, which is for social scientists to utilize their training in methods of causal inference using nonexperimental data and collaborate with biological and genetic scientists to leverage the large advances in social science data that now contain biomarkers and genotype measures. Such an approach represents a path forward that is truly interdisciplinary, where both sides bring important expertise to the table. Social scientists generally lack knowledge of biological functionality important in selecting credible gene targets for examination, whereas geneticists have not been trained in advanced econometric methods. The bread and butter of large sections of modern empirical economics, political science, and sociology is leveraging so-called “natural experiments” and institutional quirks that, under reasonable assumptions, can allow causal inference using observational data. We outline specific methods and examples in this essay and also suggest new approaches.

There are multiple research designs aimed at causal inference that form the core empirical instruction for economists as well as many sociologists and epidemiologists. Rutter9 outlines a large list, including growth curve analysis, propensity score matching, and a broad set of natural experiments, where he includes within-family (e.g., sibling or twin) comparisons, the use of genetic instrumental variables (leveraging “Mendelian randomization”), “special situations” such as the Dutch Famine, and regression discontinuity. Although many research designs are plausible candidates for use in G×E research, we suggest a hierarchy in their usefulness. From our perspective, the key focus when choosing appropriate methods should be isolating variation in an environmental exposure that is plausibly unconfounded by other characteristics (including and especially genotype). In this way, we seek to mimic the laboratory experiment used in model organisms, but instead use “quasi-natural experiments” on humans outside the laboratory. We suggest that a basic dichotomy is whether the analyst can pinpoint the source of variation in the environmental exposure. Here, we view growth curve analysis, propensity score matching, Mendelian randomization, and within-family analysis as typically unable to isolate environmental variation or experiencing other problems of inference. (Note that Mendelian randomization attempts to isolate variation in the genotype rather than variation in environmental exposure—also addressing an important concern. However, although such an approach, if used in sibling-fixed effects or other family-based analysis, can yield a reduced form causal estimate of allelic effects, the use of genes as instrumental variables can be flawed in that they may not satisfy the exclusion restriction in the presence of pleiotropic effects.10–14 See Ding et al.15 for a seminal example of the approach in economics.)

By contrast, we suggest that G×E research could be advanced in its ability to estimate causal effects though the broad incorporation of research designs that aim to leverage exogenous environmental variation: difference-in-differences, regression discontinuity designs, and a narrow set of natural experiments that focus on macro-level policy variation or institutional rules and quirks. We see opportunities both in adding a genetic component to a large set of already published investigations and incorporating these research designs into typical G×E studies.

The use of these designs would then be able to overcome some of the most pervasive empirical issues in G×E research: the issue of rGE, where the danger is that statistical interactions between genetic variants and environmental exposures may reflect differential risk of exposure (e.g., “genes selecting environments”) rather than the genetic moderation of exogenous environmental exposures. Relatedly, there is the issue of environments acting as proxy for unmeasured genetic differences; therefore, G×E can sometimes reflect gene-by-gene interaction. Population stratification, when genotype and phenotype are both related to stratifying characteristics, such as race/ethnicity, may also confound analysis.

Example 1

A natural experiment design examining G×E interactions between stress and genotype in predicting depressive phenotypes. The question of whether genotype moderates the effects of stressful events in the determination of depressive phenotypes is an initial motivating question in this literature. The investigation by Caspi et al. linking the 5-HTT genotype as a moderator of life stress in predicting adult depression published in Science is one of the seminal works in the G×E area.1 However, this study has also been the subject of wide debate. Although much of the debate focuses on measurement issues, a potentially severe limitation is the lack of experimental variation in the exposure. Without this type of variation, an alternative hypothesis that fits these data is that of rGE, where (parental) genotype selects exposure rather than moderates exposure. To reduce issues of rGE, Fletcher16 leverages the sample design of the wave III collection of the National Longitudinal Survey of Adolescent Health (Add Health) data, when respondents had been surveyed throughout the latter half of 2001. Building off of work by Ford et al.,17 who showed increased reports of depressive symptoms for those who were interviewed following the terrorist attack of September 11, 2001 (9/11) in comparison with those interviewed before the attack, Fletcher16 shows that the effects of this external stressor are moderated by genotype—a G×E interaction. In this study, the environmental exposure—the date of the survey response—is plausibly uncorrelated with genotype. Of course, the limitation of this approach is that the environmental exposure used in the analysis may be of less theoretical interest in some circumstances. It could be the case that researchers and policymakers are more interested in the effects of specific life stressors, such as divorce, than the external stressor of a terrorist attack. More generally, a principal tradeoff is often between estimating an unbiased effect of an exposure that may be of less theoretical interest (terrorist attack) versus estimating a biased effect of an exposure of more theoretical interest (life stress checklist item). In related work, Lee et al.18 have examined whether the impacts of the Great Recession on harsh parenting are moderated by the DRD2 genotype in a similar study design using maternal data from the Fragile Families Study. These methods have in common that they potentially sacrifice some external validity in the cause of internal validity. However, it remains plausible that stressors such as 9/11 and the Great Recession do generalize to other more quotidian forms of psychological hardship.

Example 2

Enhanced family designs are used to address concerns about population stratification in the G×E literature. Conley and Rauscher19 provide an example of this by analyzing the same locus as Caspi et al.1—the long versus short promoter region variation in 5-HTT. They try to reproduce the Caspi findings on depression, but deploy a different environmental condition—prenatal nutrition as proxied by birth weight. These authors use both MZ (monozygotic or identical) and DZ (dizygotic or fraternal) twin differences in birth weight to predict depression (and other outcomes) using national data from Add Health. By deploying MZ twins, they are able to hold constant genetic differences that may influence both birth weight and the outcome of interest. They then examine whether the treatment effect of low birth weight varies across twin sets that are divergent on measured alleles.

However, because the strategy of using MZ twins leaves open the possibility that genetic loci across which they are stratifying are correlated with other unmeasured environmental or genetic differences because of population stratification, they also present estimates using intra-sibship comparisons among dizygotic twins. In this way, with the exception of other genes that may be linked to the gene in question through linkage disequilibrium during recombination, all other genes are orthogonal to the measured genetic difference. Linkage disequilibrium occurs when alleles at different loci do not recombine independently of each other and are not randomly distributed. In cases of high linkage disequilibrium, some allele combinations occur more frequently than others. However, in the DZ models, one cannot say that birth weight differences are exogenous to unmeasured genetic characteristics that vary between the twins. Thus, Conley and Rauscher19 present findings that are robust to both these estimation strategies. They also show, in auxiliary analysis, that birth weight seems unaffected by these genes.

Using this cautious approach, Conley and Rauscher19 find evidence of a G×E interaction effect that works in the opposite direction predicted by general understandings in the literature. They find that birth weight interacts in both DZ and MZ twin models, such that decreased birth weight (previously considered a risk factor) results in lower risk of depression, but only for those who have the “risky” serotonin transporter promoter region allele. Despite having focused, out of necessity, on a very different environmental stressor, these results should fuel debate and future research in social genetics and provide a methodological framework for moving that debate forward. Although such results lack a high degree of external validity (specifically with respect to previous work), they fit with the central methodological point that G×E interaction research must address both genetic and environmental endogeneity.

Example 3

Policy variation is another source of environmental variation that has been used in the social sciences to examine the determinants of important health phenotypes. This variation will have important uses in G×E interaction research, but is now in its infancy. In a seminal example of this type of research, Guo et al.20 explore changes in risky behaviors, such as tobacco, alcohol and illegal drug use, as a function of individual genotype and of the changing legal status of the behavior as individuals age. For example, at age 21 years, the legal status of alcohol abruptly changes because of policy, which changes the total costs of consuming alcohol for individuals (total costs = health costs + legal costs). Although the health consequences remain the same when individuals turn 21 years old, the legal consequences are largely eliminated. Here, the source of environmental variation (i.e., consequences of use) comes from specific policies that vary by age. The authors find a protective effect from the 9R/9R genotype in the VNTR (variable number of tandem repeats) of the dopamine transporter gene (DAT1) on a range of risky behaviors that varies based on the policy environment. The effect is most important at ages when a risky behavior is illegal and largely vanishes when the environment (i.e., legal status) of the behavior changes. The key innovation in this design is the use of exogenous policy variation that should be unrelated to individual genotype, and thus, the results are less likely to be spurious because of rGE. In a related example, Fletcher21 examines the interaction between state-level tobacco tax rates and genotype in predicting tobacco use patterns. In this case, the policy variation occurs based on state of residence.

Finally, we highlight opportunities that should be the subject of future research. One only needs to read a few articles in empirical social science journals to realize the vast potential to add a genetic component to an existing research question, as Fletcher16 did with the Ford et al.17 article. For example, the regression discontinuity design has been used to examine the effects of access to alcohol, using the sharp discontinuity in the minimum legal drinking age laws, on other health outcomes, such as mortality, and risky behaviors, such as arrests and tobacco use.22,23 This design combined with the availability of genetic data in the Add Health sample could be used to examine whether there is genetic modification of the effects of access and to extend the findings in Guo et al.20 More generally, because of the growing access to social science data sets that have collected DNA (Health and Retirement Study, Add Health, Fragile Families), one only needs to search for “Health and Retirement Study” and “natural experiment” or “Difference-in-Differences” to find studies that might be relevant for this added G×E direction. Except for the previously mentioned study by Lee et al.,18 all the examples come from Add Health data; this is to be expected, given that this is the social survey that has had genetic data for the longest time. Now that genetic markers are coming online for other important surveys, we do not expect this Add Health quasi-monopoly on G×E research in public health and behavioral science to continue.

This plethora of opportunities may be overwhelming, and they raise a new set of issues for both disciplinary and interdisciplinary research. First, for disciplinary social scientists, there are limited training opportunities to gain expertise in this new area of research, and there are few individuals with joint expertise in the biological and social sciences. This general lack of experience with the genetics and biology literatures likely will lead to predictable problems regarding the selection of candidate polymorphisms and associated misspecified and implausible models and findings. A longer term solution likely will require greater opportunities of interdisciplinary training programs. A short-term solution would be to team up with geneticists and biologists, although few may be interested in examining social science and public health problems. However, this lack of interest may also be changing. We think that a considerable nudge could be made through the demonstration of these research designs (for which biologists and genetics have limited expertise) to address G×E questions.

This research direction has the promise of increasing our understanding of both why certain environments affect some people but not others and the functioning of specific genes in determining important health phenotypes.


J. M. Fletcher thanks the Robert Wood Johnson Foundation Health & Society Scholars program for its financial support. D. Conley thanks the John Simon Guggenheim Foundation Fellowship program for its financial support.

Human Participant Protection

Human participant protection was not required because this study did not involve human participants.


1. Caspi A, Sugden K, Moffitt TE, et al. Influence of life stress on depression: moderation by a polymorphism in the 5-HTT gene. Science. 2003;301(5631):386389. Crossref, MedlineGoogle Scholar
2. Karg K, Burmeister M, Shedden K, Sen S. The serotonin transporter promoter variant (5-HTTLPR), stress, and depression meta-analysis revisited: evidence of genetic moderation. Arch Gen Psychiatry. 2011;68(5):444454. Crossref, MedlineGoogle Scholar
3. Risch N, Herrell R, Lehner T, et al. Interaction between the serotonin transporter gene (5-HTTLPR), stressful life events, and risk of depression: a meta-analysis. JAMA. 2009;301(23):24622471. Crossref, MedlineGoogle Scholar
4. Eichler EE, Flint J, Gibson G, et al. Missing heritability and strategies for finding the underlying causes of complex diseases. Nat Rev Genet. 2010;11(6):446450. Crossref, MedlineGoogle Scholar
5. Munafò MR, Durrant C, Lewis G, Flint J. Gene × environment interactions at the serotonin transporter locus. Biol Psychiatry. 2009;65(3):211219. Crossref, MedlineGoogle Scholar
6. Devlin B, Roeder K. Genomic control for associational studies. Biometrics. 1999;55(4):9971004. Crossref, MedlineGoogle Scholar
7. Price AL, Patterson NJ, Plenge RM, Weinblatt ME, Shadick NA, Reich D. Principal components analysis corrects for stratification in genome-wide association studies. Nat Genet. 2006;38(8):904909. Crossref, MedlineGoogle Scholar
8. Price AL, Zaitlen NA, Reich D, Patterson N. New approaches to population stratification in genome-wide association studies. Nat Rev Genet. 2010;11(7):459463. Crossref, MedlineGoogle Scholar
9. Rutter M. Epidemiological methods to tackle causal questions. Int J Epidemiol. 2009;38(1):36. Crossref, MedlineGoogle Scholar
10. Conley D. The promise and challenges of incorporating genetic data into longitudinal social science surveys and research. Biodemography Soc Biol. 2009;55(2):238251. Crossref, MedlineGoogle Scholar
11. Conley D. Learning to love animal (models) (or) how (not) to study genes as a social scientist. In: Pescosolido B, Martin JK, McLeod JD, and Rogers A, eds. Handbook of the Sociology of Health, Illness, and Healing: A Blueprint for the 21st Century. New York, NY: Springer; 2011:527542. CrossrefGoogle Scholar
12. Fletcher JM. The promise and pitfalls of combining genetic and economic research. Health Econ. 2011;20(8):889892. Crossref, MedlineGoogle Scholar
13. Fletcher JM, Lehrer SF. The effects of adolescent health on educational outcomes: causal evidence using ‘genetic lotteries’ between siblings. Forum Health Econ Policy. 2009;12(2):1180. CrossrefGoogle Scholar
14. Fletcher JM, Lehrer SF. Genetic lotteries within families. J Health Econ. 2011;30(4):647659. Crossref, MedlineGoogle Scholar
15. Ding W, Lehrer SF, Rosenquist JN, Audrain-McGovern J. The impact of poor health on education: new evidence using genetic markers. Working Paper No. 12304. Cambridge, MA: National Bureau of Economic Research; 2006. CrossrefGoogle Scholar
16. Fletcher JM. Enhancing the gene-environment interaction framework: evidence from differential responses to September 11th. Paper presented at: The Population Association of America Annual Meeting; March 31–April 2, 2011; Washington, DC. Google Scholar
17. Ford CA, Udry JR, Gleiter K, Chantala K. Reactions of young adults to September 11, 2001. Arch Pediatr Adolesc Med. 2003;157(6):572578. Crossref, MedlineGoogle Scholar
18. Lee D, McLanahan S, Brooks-Gunn J, Garfinkel I, Notterman D. The effect of the great recession and dopamine receptor gene DRD2 on maternal harsh parenting. Paper presented at: 106th American Sociological Association Annual Meeting; August 20–23, 2011; Las Vegas, NV. Google Scholar
19. Conley D, Rauscher E. Genetic interactions with prenatal social environment: effects on academic and behavioral outcomes. Working Paper No. 16026. Cambridge, MA: National Bureau of Economic Research; 2010. CrossrefGoogle Scholar
20. Guo G, Cai T, Guo R, Wang H, Harris KM. The dopamine transporter gene, a spectrum of most common risky behaviors, and the legal status of the behaviors. PLoS ONE. 2010;5(2):e9352. Crossref, MedlineGoogle Scholar
21. Fletcher JM. Why have tobacco control policies stalled? Using genetic moderation to examine policy impacts. PLoS ONE. 2012;7(12):e50576. Crossref, MedlineGoogle Scholar
22. Carpenter C, Dobkin C. The minimum legal drinking age and public health. J Econ Perspect. 2011;25(2):133156. Crossref, MedlineGoogle Scholar
23. Fletcher JM. The effects of alcohol access and use on risky behaviors: regression discontinuity evidence. Working Paper. New Haven, CT: Yale University; 2011. Google Scholar


No related items




Jason M. Fletcher, PhD, and Dalton Conley, PhDAt the time of this study, Jason M. Fletcher was with the Department of Health Policy and Management, Yale School of Public Health, New Haven, CT, and the Robert Wood Johnson Foundation Health & Society Scholar, Columbia University, New York, NY. Dalton Conley was with the Department of Sociology, the School of Medicine, and the Wagner School of Public Service at New York University, New York. “The Challenge of Causal Inference in Gene–Environment Interaction Research: Leveraging Research Designs From the Social Sciences”, American Journal of Public Health 103, no. S1 (October 1, 2013): pp. S42-S45.

PMID: 23927518