Dealing with Within-Family Correlation

CE question our use of mixed-effect models to cope with multiple observations from the same family. The key focus of this critique is that “‘families’ in the Add Health dataset segregate into different ‘types’ whose members differ systematically” (9) and that this likely biases our result. Instead, they propose randomly choosing one sibling from each family to create an independent sample.

Although their recommendation is likely far too conservative (mixed-effect models are used throughout the social sciences), in the interest of testing whether our replication is sensitive to the choice of modeling framework we created a randomly chosen, fully independent sample with 8,744 unrelated individuals (in other words, we took only one subject from each family) and used a generalized linear model to estimate the association between 5HTT and the Add Health measure of self-reported turnout in the 2000 general election. Table S1 (Tables S1–S7 are in the Online Appendix) shows that this more conservative approach also yielded a significant association.Footnote ¹ Importantly, none of the 8,744 subjects in this model were included in the original study because we did not have genotype information for them (they were genotyped in Wave IV but not in Wave III). Moreover, none came from households with oversampled twins or siblings, suggesting that CE's conjecture that oversampling would bias the results is incorrect.

In Table S2, we combined the previously genotyped and newly genotyped samples to generate an estimate across both samples; however we continued to use only one individual per family (N = 9,821). The results show that the estimated effect increased slightly, and the p value on the association was 5 × 10⁻³. Note that the 95% confidence interval for the beta coefficient on the 5HTT interaction in the original study was [0.13 to 0.79] and in the replication study it was [0.06 to 0.32]. Although this suggests that the original sample was underpowered to detect an effect of the size estimated in the replication sample (0.45 > 0.32), it does not invalidate the replication: In both samples, the association was positive and significantly different from zero, and a test of the difference in the coefficients produced by both samples was not quite significant (p = 0.07). Given the much larger sample size in the replication study, however, we suspect that the true coefficient is closer to that we reported in the replication, which is confirmed by the model reporting results on the combined independent sample, suggesting a confidence interval of [0.07 to 0.33].

Genotype Specification

CE assert that the results we originally presented are sensitive to specification of the 5HTT variable. They claim that rather than coding the genotype as a 1 if a person had at least one long allele (combining the LL and Ls genotypes into one group), we should have coded it as 1 if a person had two long alleles (including only those with the LL genotype). In Table S3 we present results for alternative genotype coding criteria by including separate variables for both the LL and Ls genotypes. We found that the effect size of the Ls genotype is exactly half that of the LL genotype, suggesting an additive process that increases with the number of L alleles. In fact, an additive model of 5HTT produces a better fit (as evidenced by a lower AIC) than a model that includes the Ls and LL genotypes separately (and both models have a lower AIC than a fully saturated model that omits the genotype variables).

Splitting the Sample

Throughout their article CE argue against our findings by analyzing subsets of the overall sample and demonstrating a lack of statistical significance. However, it is important to consider that the failure to maintain a significant association found based on the total sample may simply be the result of a dramatic reduction in sample size. More concretely, any random culling of the data by X% will increase the standard error by a factor of approximately $\sqrt{100/(100-X)}$ and, as noted earlier and demonstrated in Figure 1, will cause a large reduction in power. A more appropriate way to determine whether pooling by subsample is appropriate is to test whether the effect size varies by group. An easy way to do this is to interact the effect size with an indicator variable for the subsample in question.

In their first argument that the sample should be split, CE assert that individuals voting in their first election (ages 18–20) or their second election (ages 20–22) should not be pooled with older individuals. We tested the assertion that the association varies by age group by interacting it with the 5HTT-Attendance interaction. Table S4 shows that we could reject the hypothesis that the interaction effect is different for 18–20 year olds (p = 0.17) and for 20–22 year olds (p = 0.21).

CE also assert that nonwhite individuals should not be pooled with whites because the effect for the whole population might be driven by an effect that only exists for a particular subgroup. We tested the assertion that the effect size varies by race or ethnicity by interacting it with the 5HTT-Attendance interaction. Table S5 shows that we could reject the hypothesis that the interaction effect is different for blacks (p = 0.49), Hispanics (p = 0.72), and Asians (p = 0.45). The interaction with Native Americans is significant (p = 0.02), but is positive, and the main interaction continues to be strongly significant (p = 2 × 10⁻³), suggesting that Native Americans are not driving the overall association.

Alternative Markers for 5HTT

In addition to studying the original 5HTT measure, we used a newly available classification from the recently released Add Health data. In the previous release, 5HTT was classified according to whether individuals had a “short” or a “long” version of the 5HTTPLR allele in the promoter region. In the most recent release, a single nucleotide polymorphism (rs25531) was genotyped that has been shown to interact with 5HTTPLR to affect gene expression levels (Hu et al. Reference Ioannidis2006). Specifically, only those with the “long” genotype and adenine (A) at rs25531 evinced higher expression of 5HTT. We used this “triallelic” genotype to test the association among 5HTT, church attendance, and voting. In Table S6 we show that this alternative classification also yielded a significant interaction with church attendance (p = 1 × 10⁻³).

Alternative Phenotypes

CE assert that voting in a single election is not sufficient to measure voting behavior. We have more to say about the broader issues related to this critique in the next section, but we were able to take advantage of new data to test CE's assertion. Although Add Health does not have additional measures covering voting behavior in other elections, Wave IV does contain a question tapping one's overall propensity to vote: “How often do you usually vote in local or statewide elections?” Four responses were permitted for this question: never, sometimes, often, and always. This question is not strictly comparable to the question we previously used as a measure of voting because the Wave III question is about a national election held in 2000 and the Wave IV question is about state and local elections held through 2008. Nonetheless, the polychoric correlation between these two measured variables was positive and highly significant (0.55, 95% CI 0.53 to 0.57).

To test whether the association with 5HTT is sensitive to phenotype specification, we regressed the Wave IV voting measure on the same independent variables shown in our other models. Table S7 shows that the interaction between the genotype and church attendance remained significant (p = 0.02). This model assumes that the categorical variable is continuous, so to be sure this assumption is not driving the result, we also estimated an ordered logit specification. The interaction between the genotype and church attendance was also significant in this alternative model (p = 0.03).

There are important limitations to note regarding our replication exercise. Although our analysis is based on a much larger sample than previously employed and is therefore more likely to be adequately powered, the previous track record of candidate gene-environment interaction studies suggests that our results should be evaluated cautiously. In addition, we still run the risk of finding false positives due to population stratification. Even though we compared estimates across groups and restricted our analysis to whites, we lacked additional genetic information to control for unobserved population structure. This information could be provided using genome-wide association study (GWAS) samples once they become available. Another alternative is to conduct a replication based on a non-U.S. sample in which the potential confounding relationship between allele frequency and the environmental determinants of voting behavior is unlikely to follow the same pattern. Finally, it should be pointed out that a previously reported association between depression and an interaction of 5HTT and stressful life has not been reproduced by follow-up studies (Duncan and Keller Reference Duncan and Keller2011; Risch et al. Reference Sallach, Babchuk and Booth2009). Although we conducted a direct replication (Duncan and Keller Reference Duncan and Keller2011) of a relationship among 5HTT, church attendance, and voter turnout, the hypothesis that the effect of 5HTT is moderated by stressful events has clearly been undermined by failed replications.

Findings from Medical Genetics

CE argue that “the search for genes that could predict prevalent and devastating behavioral phenotypes such as schizophrenia and autism, not to mention global killers such as diabetes and hypertension, has to date been unsuccessful” (2). They go on to identify just three single-nucleotide polymorphisms (SNPs) as “the few reliably reproduced associations between an [sic] SNP and a given phenotype of such a magnitude that the SNP can be called a risk factor” (13), but this argument is highly misleading. For example, a recent review of the literature shows that since 2007, 40 genetic loci have been discovered for type I diabetes and 50 have been discovered for type II diabetes (Visscher et al. Reference Weber and Johnson2012). These loci are among 384 SNPs associated with autoimmune and metabolic diseases with great enough significance in large samples that they are considered to be reliable associations. In fact, enough progress has been made that “it should be possible to identify groups of individuals who are at a substantially greater-than-average risk for diabetes” (Visscher et al. Reference Weber and Johnson2012, 17). Genetic susceptibility scores have already been created and are being applied for other complex phenotypes such as body mass index (Speliotes et al. Reference Speliotes2010). We do not claim that medical genetics has a complete understanding of complex diseases or that prediction at the individual level is a viable reality yet; however, neither is it accurate to summarily dismiss what has been accomplished so far.

The Current “Paradigm” in Genetics

CE argue that “CGA studies of complex behavioral traits (such as political behaviors) rely on an outdated genetic paradigm” (2). However, the journal Behavior Genetics recently updated its guidelines on the publication of candidate gene studies (Hewitt Reference Hu2012) to reflect the latest research in genetics, and it continues to publish CGA studies. Moreover, these studies are still viewed as beneficial, even by critics. The authors of a recent highly critical review of candidate gene studies write that they “may well prove to be important or even central for understanding the etiology of psychiatric disorders. At issue is how to separate the wheat from the chaff” (Duncan and Keller Reference Duncan and Keller2011,1047–48) Thus the “paradigm” is not at issue—it is the methodology that proves challenging. The solution is not to abandon candidate gene studies, but rather to implement guidelines for their publication that reduce the likelihood of false positives.

Gene Transcription and Stability

It is important to understand that what takes place at the molecular level is complex. However, CE's description of this complexity is highly misleading. For example, they write, “Once considered the paragon of stability, DNA is subject to all manner of transformation. For example, retrotransposons or “jumping genes” comprise 45% of the human genome, move about the genome by a copy-and-paste mechanism changing DNA content and structure” (10–11). Yet consider that the rate of mutation per generation per site in human DNA is estimated to be about 1 in 100 million (Lynch Reference Mackay2010). This means that all but a few dozen of the 3 billion base pairs in an individual's DNA will be exactly the same throughout our reproductive lifetimes. Thus, by and large, the genes we are born with are the genes we will die with. It is true that the environment plays a critical role in gene expression, myelination, and other biological processes that influence the way genes affect us, but it is extremely disingenuous to argue that these measures of our biological inheritance are unstable.

CE further assert, “Genes do not regulate the extent to which they are capable of being transcribed in any obvious, unidirectional manner” (10). This claim is directly contradicted by evidence from more than 5,000 genes that shows that transcription explains nearly 40% of the variation in levels of expression in mammals (Schwanhäusser et al. Reference Settle, Dawes and Christakis2011). Although this evidence also suggests an important role for the environment, it shows that heritable DNA sequence variation does have an important impact on expression. In particular, a recent review of the literature on the 5HTTLPR polymorphism shows that the transcriptional efficiency of the “long” version of this genotype is associated with significantly higher expression than the shorter allele in both mice and humans (Murphy and Lesch Reference Norton and Han2008). Causal evidence of the gene's effect on expression is especially strong in mouse studies, because many of them “knock out” (delete) or modify the gene in question (Murphy and Lesch Reference Norton and Han2008).

This does not mean that gene-gene and gene-environment effects on transcription are absent, and recent research is exploring a variety of factors that modify gene expression (Cheung and Spielman Reference Cheung and Spielman2009), including cell type (Dimas et al. Reference Dimas2009). However, it is incorrect to imply that it is impossible to detect a relationship between genotypes and transcription levels. Given that we do not know what many genes do, the best place to start is to study genotypes that we know have a measurable effect on the molecules they transcribe.

Pleiotropy

CE list many phenotypes that have been associated with variation in 5HTT and MAOA and ask, “How is it possible that the same polymorphisms of the same gene could simultaneously predict (or be risk factors for) so many different phenotypes?” (13). Genes playing a role in important hormonal systems that regulate social behavior are likely to influence a wide variety of outcomes, a phenomenon known as “pleiotropy” (Lande Reference Littvay1980). Depression and personality factors are two possible mediating mechanisms, but there may be many others.

CE question the idea of an underlying factor that may cause several of the phenotypic outcomes (an “endophenotype”). However, they neglect to point out that systems such as the serotonin system influence a wide variety of behaviors via different mechanisms (Murphy and Lesch Reference Norton and Han2008). This means that a single genotype may influence several endophenotypes, each of which is associated with particular phenotypes; there is strong evidence that this is true for 5HTT. Knock-out studies in mice where the 5HTT gene is completely removed show that a large number of phenotypes are affected, and the results are so well accepted that pleiotropy in 5HTT has been proposed as a reason for evolutionary conservation of diversity in this gene (Murphy and Lesch Reference Norton and Han2008).

Consider a genotype we have already been studying for decades: sex. A wide range of political phenomena are influenced by a person's sex (Hatemi et al. Reference Hatemi, Medland and Eaves2012), and although we do not claim that these effects are all purely genetic, neither do we doubt a reported association between sex and one political behavior merely because it also predicts other political behaviors. Sex can influence specific hormonal responses, but it can also induce gender-specific roles in the self and others (Hatemi et al. Reference Hatemi, Medland and Eaves2012). Each of these could be considered an endophenotype with many phenotypic consequences.

Another important consideration is that many of the phenotypes listed by CE have not been replicated, so it may be premature to ask what such different phenotypes have in common with one another. For example, our replication efforts here suggest that voting should be removed from the list of those phenotypes associated with MAOA. At the same time, however, our confidence that 5HTTPLR plays a role is now increased by successful replication, so (for now) it belongs on the list.

CE continue: “Consider that the polymorphisms of MAOA-μVNTR and 5-HTTLPR are the only polymorphisms for these two genes for which there are data in a large data set such as Add Health and that MAOA and 5-HTT are two out of the eight genes in total (emphasis in original) for which there are genotypic data. What is the likelihood that these same polymorphisms on these same genes will conveniently turn out to be the genetic key to so much human behavior?” (13). This question is misleading in two ways. First, we never called these genes “the key to human behavior.” Consider what we actually wrote:

Second, the investigators of the National Longitudinal Study of Adolescent Health did not choose their genes randomly—they focused on genes that had already shown promise in explaining social behavior and that were linked to behavioral and psychiatric outcomes. According to the study researchers,

It should thus not be surprising that biomarkers for the genes they chose have an effect on many social behaviors, and we would have a better chance of finding more than one association among them.

Stochastic vs. Deterministic Outcomes

CE write, “The problem is that a large set of genes (and if large enough, then we are simply talking about the human genome and hence the human organism), the transcriptional activity of which is influenced by the environment and each others’ functional products, is incompatible with the expectation that two genes could predict voter turnout” (30). This is a disingenuous critique. We clearly acknowledged that there are multiple genetic and environmental causal factors that underlie turnout: “There is some (likely large) set of genes whose expression, in combination with environmental factors, influences political participation” (Fowler and Dawes Reference Fowler and Dawes2008, 590).

Our usage of the word “predict” in our original article was meant to convey that there is a nonrandom, systematic relationship between at least two (and probably very many more) genes and voter turnout. In this article we show that one of these relationships replicates and one does not. We therefore conclude that one of these genes “predicts” turnout, even though we in no way believe and have never argued that it is the only such gene. We were very careful in 2008 to make this clear: “[I]t is important to emphasize that there is likely no single ‘voting gene’” (Fowler and Dawes Reference Fowler and Dawes2008, 590) In fact, we deliberately (emphasis added) chose to publish associations with two genes in one paper (rather than two separate papers) to emphasize this point.

CE go on to argue that “CGA studies depend on the assumption that the presence of a particular allele entails that it is turned on. . .. We can no longer assume, however, that the presence of a particular allele entails that it is capable of being transcribed in the manner associated with that allele, because it may be epigenetically silenced” (10). Again, this argument demonstrates a profound misunderstanding of scientific inference. Although it is true that gene expression is variable and it can be sensitive to the environment and other genotypes, this does not mean that the gene plays no role in influencing a particular behavior. Nor does it preclude studies that take this stochasticity into account by testing the hypothesis with many observations and under many conditions.

Polygenic Effects

CE emphasize the multifactorial nature of genetic effects, writing that “most human traits with a genetic component are influenced by a vast number of genes of small effect” (11) and that “[r]esearchers who conduct CGA studies err when they suppose that multifactorial traits exhibit a genotype-phenotype relationship analogous to that of monogenic disorders” (12). They write this in such a way that it implies that we are not aware of these complexities. Yet, note what we wrote four years ago: “[T]he vast majority of phenotypes are ‘polygenic,’ meaning they are influenced by multiple genes (Mackay Reference Martin2001; Plomin Reference Price2008) and are shaped by a multitude of environmental forces” (Fowler and Dawes Reference Fowler and Dawes2008, 581).

CE conclude that the existence of polygenic effects “calls into question the underlying assumption of CGA studies” (11) and makes genetic studies “particularly problematic” (11). In stark contrast, we argued as follows:

At the heart of this disagreement is a fundamental difference in the way we and CE approach scientific inquiry. CE claim that because the system we are studying is complex, it is pointless to examine any of its parts. But this philosophy would preclude nearly all of social science. Social systems are inherently complex, and observational and experimental studies attempt to reduce complexity by isolating parts of these systems to determine which factors and processes are most important for a given outcome. The hope is that these factors can then be integrated into a more holistic theory at the system level.

To be clear, we are sympathetic to systems-level approaches. In other work, we advocate for a better understanding of interdependence and its effect on a number of political phenomena (e.g., Christakis and Fowler Reference Christakis and Fowler2009). However, we would never make the argument that macro-level approaches are a substitute for micro-level research. The truth is that we need both.