Defining Gene–Environment Effect Modification

What Does It Mean for Genotype to Modify an Exposure Effect?

Suppose the association between an environmental exposure and a trait appears larger in one genotype group than another. The relevant question is not simply whether genotype and exposure occur together, or whether the outcome differs by genotype. It is whether the effect of the exposure differs across genotype groups, and on what scale that difference is defined.

This distinction matters because three ideas are easily conflated: gene–environment effect modification, gene–environment correlation, and biological interaction. They describe different features of data or mechanisms. A statistical contrast can establish effect modification on a specified scale, but by itself it neither explains a biological mechanism nor guarantees a causal interpretation.

Definition: Let $G$ denote genotype group, with levels $g_0$ and $g_1$, and let exposure $E$ have levels $e_0$ and $e_1$. On a specified outcome scale, gene–environment effect modification means that the effect of changing exposure from $e_0$ to $e_1$ differs between the genotype groups. For a quantitative outcome measured in its original units, the additive-scale contrast is $$ \Delta_{\mathrm{add}} = \left\{\operatorname{E}[Y(e_1)-Y(e_0)\mid G=g_1]\right\} - \left\{\operatorname{E}[Y(e_1)-Y(e_0)\mid G=g_0]\right\}. $$ There is additive-scale effect modification when $\Delta_{\mathrm{add}}\ne 0$. The definition is scale-specific: a different scale can give a different conclusion.

Here $Y(e)$ denotes the outcome that would be observed under exposure level $e$, and $\operatorname{E}$ denotes an expectation. The expression compares exposure effects within genotype groups; it does not compare the outcomes of the genotype groups at just one exposure level. Genotype is used to define strata. This contrast does not, on its own, estimate a causal effect of changing genotype.

The same idea applies to categorical genotypes with more than two levels: compare exposure effects across the genotype groups of interest. A single pairwise contrast answers a question about that pair; a broader claim that effects differ somewhere among several groups requires considering the relevant collection of contrasts.

Statistical Contrasts and Causal Effects

A statistical effect-modification statement describes how exposure–outcome associations differ by genotype on a specified scale. A causal statement is stronger: it says that the effect of intervening to change exposure differs by genotype. The potential-outcome definition above is causal, but observed data do not automatically identify its quantities.

For example, within each genotype group, the exposure groups must be comparable with respect to factors that affect the outcome, either through the study design or suitable adjustment. In a simple randomized exposure study, randomization within genotype groups can provide that comparability. In an observational study, causal interpretation generally requires conditional exchangeability: after conditioning on appropriate covariates $C$, exposure assignment is independent of the potential outcomes within genotype groups. There must also be positivity, so each exposure level is possible at relevant covariate values, and consistency, so the observed outcome under a person's received exposure agrees with the corresponding potential outcome.

When adjustment is needed, the target population and how covariates are handled matter. In particular, comparing adjusted associations is not automatically the same as comparing causal effects standardized to a common target. Without the needed design or assumptions, the contrast should be described as statistical effect modification in the observed data, not as a causal gene–environment interaction.

Worked Example: Comparing Quantitative-Trait Exposure Effects

In a hypothetical study, a beneficial exposure is recorded as $e_1$, while $e_0$ is its absence. A quantitative trait is measured in units. Assume exposure was randomized within each genotype group, all participants received their assigned exposure as defined, and both exposure levels were represented in each group. The observed group means are:

Genotype group	Mean at $e_0$	Mean at $e_1$
$g_0$	10 units	12 units
$g_1$	14 units	20 units

Under these design conditions, the within-genotype differences estimate the causal exposure effects. In group $g_0$, the effect is $12-10=2$ units. In group $g_1$, it is $20-14=6$ units. The additive-scale contrast is

\Delta_{\mathrm{add}} = (20-14)-(12-10)=6-2=4\text{ units}.

Thus, on the original trait scale, the exposure effect is 4 units larger in $g_1$ than in $g_0$. This supports additive-scale causal effect modification under the stated design and measurement conditions. It does not show that genotype itself causes the difference, or identify a biological mechanism.

Why the Scale Must Be Named

An effect is a contrast, and the contrast depends on the scale used. For a quantitative outcome, one common scale is the difference in means. For a binary outcome, effects may be compared as risk differences or risk ratios. These need not give the same judgment about whether effects differ across genotype groups. Therefore, “there is interaction” is incomplete unless the scale and effect measure are specified.

The additive-scale definition can be understood as a difference between two exposure effects. If $\mu_{ge}$ denotes the mean outcome for genotype $g$ at exposure level $e$, then

\Delta_{\mathrm{add}} = (\mu_{g_1e_1}-\mu_{g_1e_0}) -(\mu_{g_0e_1}-\mu_{g_0e_0}).

Rearranging gives the equivalent contrast

\Delta_{\mathrm{add}} = (\mu_{g_1e_1}-\mu_{g_0e_1}) -(\mu_{g_1e_0}-\mu_{g_0e_0}).

The first form compares exposure effects across genotypes; the second compares genotype differences across exposure levels. They are algebraically identical because both use the same four group means. A nonzero value says that the corresponding differences are not constant across the other factor, on this scale. This identity does not establish causality: the means must still have a valid causal interpretation for the contrast to describe causal effect modification.

Worked Example: Effect Modification Depends on the Risk Scale

Consider hypothetical risks of a binary outcome under two exposure levels, in two genotype groups. Assume for this illustration that the risks represent causal risks under the two exposure conditions, with consistency, positivity, and adequate control of confounding. The risks are:

Genotype group	Risk at $e_0$	Risk at $e_1$
$g_0$	0.10	0.20
$g_1$	0.30	0.60

On the risk-difference scale, the exposure effect in $g_0$ is $0.20-0.10=0.10$, or 10 percentage points. In $g_1$, it is $0.60-0.30=0.30$, or 30 percentage points. The difference between these effects is

$$ (0.60-0.30)-(0.20-0.10)=0.30-0.10=0.20. $$

There is additive effect modification: the risk difference is 20 percentage points larger in $g_1$. On the risk-ratio scale, however, the exposure effect in $g_0$ is $0.20/0.10=2$, and in $g_1$ it is $0.60/0.30=2$. The risk ratios are equal, so there is no effect modification on this relative-risk comparison. Neither calculation is contradictory: they answer different scale-specific questions.

Effect Modification Is Not Gene–Environment Correlation

Gene–environment correlation concerns whether exposure distribution varies by genotype: for example, whether $P(E=e_1\mid G=g_1)$ differs from $P(E=e_1\mid G=g_0)$. Effect modification concerns whether the exposure effect differs by genotype. One can occur without the other.

A genotype–exposure association can arise for many reasons, including ancestry structure, family or social processes, or selection into the analyzed sample. Its presence does not establish that the exposure has different effects across genotypes. Conversely, exposure can be distributed similarly across genotype groups even when genotype-specific exposure effects differ. It is important to examine exposure–outcome contrasts within genotype groups rather than infer effect modification from the joint distribution of genotype and exposure alone.

Worked Example: Correlated Exposure Without Effect Modification

Suppose a hypothetical population has two genotype groups. The exposure prevalence is 20% in $g_0$ and 80% in $g_1$, so genotype and exposure are associated. Within each group, the mean outcome is 10 units at $e_0$ and 12 units at $e_1$. Assume these means are comparable for the intended association description; causal interpretation would additionally require the relevant causal assumptions.

The exposure effect within $g_0$ is $12-10=2$ units. Within $g_1$, it is also $12-10=2$ units. Therefore, the additive effect-modification contrast is $2-2=0$ units, despite the different exposure prevalences. There is gene–environment correlation in this example, but no additive effect modification in the specified within-group means.

The distinction can also be seen in pooled means. If the genotype proportions are 50% each, then 80% of exposed people are in $g_1$, while only 20% of unexposed people are in $g_1$. The pooled exposed mean is $0.20(12)+0.80(12)=12$ in this particular setup, and the pooled unexposed mean is $0.80(10)+0.20(10)=10$. The pooled contrast is 2 units. It agrees with the within-group contrast here because the exposure effect is equal across genotypes, even though the genotype composition differs. With genotype-specific baseline means that differ, the composition difference can also distort the pooled association; pooled comparisons should not be mistaken for within-genotype effect contrasts.

Statistical Effect Modification Is Not a Biological Mechanism

A biological interaction is a claim about how biological processes work together: for example, whether an exposure changes a molecular process in a way that depends on a gene product. Such claims require mechanistic evidence. A statistical effect-modification contrast is a pattern in outcomes on a chosen scale; it does not reveal the molecular pathway that produced that pattern.

A nonzero statistical contrast might reflect a genuine biological difference in response, but it can also reflect confounding, measurement differences, selection, or features of the chosen scale. A biologically meaningful joint mechanism might not produce a detectable contrast on a particular statistical scale. Statistical interaction and biological interaction are therefore not interchangeable terms, and neither can be inferred solely from the other.

Common Mistake

Wrong reasoning: “The exposure is more common in one genotype group, so genotype modifies the exposure effect.” This confuses exposure distribution with an exposure-effect contrast.
Correction: Estimate or define the exposure effect within each genotype group, then compare those effects on a named scale. Separately describe any association between genotype and exposure.
Wrong reasoning: “A nonzero interaction contrast proves a gene and an environmental factor interact biologically.”
Correction: It establishes a statistical difference on the specified scale, subject to the validity of the estimates. A biological mechanism needs separate evidence.
Wrong reasoning: “There is no interaction because the relative risks are equal.”
Correction: Equal relative risks rule out effect modification on that relative-risk comparison, not necessarily on the risk-difference or another scale.

Key takeaway: Define gene–environment effect modification by comparing genotype-specific exposure effects on an explicit scale. Keep that contrast distinct from genotype–exposure correlation and from biological mechanism; causal interpretation additionally depends on design and assumptions.

Check Your Understanding

Answer each question by identifying the relevant comparison and, where appropriate, the scale.

What two quantities are compared in an additive-scale gene–environment effect-modification contrast?
Can genotype and exposure be associated when exposure effects are equal across genotype groups? Explain.
Why can a binary outcome show effect modification on the risk-difference scale but not on the risk-ratio scale?
What additional requirements are needed to interpret an observed genotype-specific exposure contrast causally?
Does statistical effect modification establish a biological mechanism? Why or why not?