Random-Effects Models for Gene-Level Association

Why Model Variant Effects as Random Effects?

A gene may contain several rare variants, and their effects on a trait need not point in the same direction. One variant may increase the trait while another decreases it. A model that insists on one shared effect direction can therefore be poorly matched to the scientific question. A random-effects model offers a different target: rather than specifying the direction of every variant effect, it asks whether effects across the gene vary around zero enough to support association.

This is a model for representing uncertainty and heterogeneity across variant effects. It does not mean that the effects are repeatedly randomized inside each person, nor that every variant is known to be causal. The observed genotype data describe which variants each person carries; the random-effects distribution describes a statistical model for the corresponding variant coefficients.

Definition: Let $n$ individuals have a quantitative outcome collected in the vector $y$, and let $m$ variants in a gene have genotype dosages collected in the $n \times m$ matrix $G$. The entry $G_{ij}$ is the dosage for individual $i$ at variant $j$, commonly 0, 1, or 2. A gene-level random-effects model is $$ y = X\gamma + G\beta + \epsilon, \qquad \beta \sim N(0,\tau W), $$ where $X$ is a covariate matrix, $\gamma$ is a vector of fixed covariate effects, $\beta$ is the vector of variant effects, $\epsilon$ is the residual vector, and $W=\operatorname{diag}(w_1,\ldots,w_m)$ is a fixed diagonal matrix of nonnegative variant weights. The parameter $\tau \geq 0$ is the gene-level variance component.

In the usual independent-effects version of this model, conditional on the chosen weights, $\beta_j$ has mean zero and variance $\tau w_j$, and different variant effects are independent. A larger $w_j$ allows a larger effect variance for variant $j$ relative to a variant with a smaller weight. When $w_j=0$, that variant contributes no random effect under this model. The weights are treated as specified for the test; how to choose them is a separate modeling decision.

For a basic quantitative-trait formulation, assume that, conditional on $\beta$ and the covariates, the residuals have mean zero and covariance $\sigma^2 I$, and are independent of $\beta$. The genotype matrix and the weights are treated as given. These assumptions make the variance-component interpretation precise. Other outcome types or residual covariance structures require a model that matches those data.

What the Variance Component Says

The mean-zero assumption does not say that all realized variant effects are zero. It says that, under the model, effects are distributed around zero, with their spread controlled by $\tau$. If $\tau=0$, the distribution of $\beta$ collapses to the zero vector: each component has variance zero. If $\tau>0$, the model allows nonzero effects, including effects with either sign. Thus a test of $\tau$ is an omnibus test of whether the gene's variant effects have nonzero modeled variation, not a test that all variants increase the outcome.

Hypotheses: For a specified gene, genotype matrix, and weight matrix, the variance-component hypotheses are $H_0:\tau=0$ versus $H_a:\tau>0$. Under the stated random-effects model, $H_0$ means that the gene's modeled variant-effect vector is zero. The alternative allows a nonzero spread of effects without fixing their individual signs.

The parameter $\tau$ has units of squared outcome units per squared dosage unit when genotypes are recorded as dosages. Its square root describes a coefficient-scale standard deviation for a variant whose weight is one. More generally, the standard deviation for variant $j$ is $\sqrt{\tau w_j}$. The weights therefore matter to the scale and interpretation of $\tau$, even though the omnibus question remains whether the variance component is zero.

Worked Example: Covariance of the Genetic Contributions

Suppose three hypothetical individuals have dosages at two variants recorded as $$ G = \begin{pmatrix} 0 & 1\\ 1 & 1\\ 2 & 0 \end{pmatrix}, \qquad W = \begin{pmatrix} 1 & 0\\ 0 & 4 \end{pmatrix}, \qquad \tau=0.09. $$ Let $u=G\beta$ denote the vector of genetic contributions to their outcomes. Under the model, $E(u\mid G)=0$, and its covariance is $\tau GWG^\mathsf{T}$. First, $$ GW = \begin{pmatrix} 0 & 4\\ 1 & 4\\ 2 & 0 \end{pmatrix}, \qquad GWG^\mathsf{T} = \begin{pmatrix} 4 & 4 & 0\\ 4 & 5 & 2\\ 0 & 2 & 4 \end{pmatrix}. $$ Therefore, $$ \operatorname{Cov}(u\mid G) =0.09 \begin{pmatrix} 4 & 4 & 0\\ 4 & 5 & 2\\ 0 & 2 & 4 \end{pmatrix} = \begin{pmatrix} 0.36 & 0.36 & 0\\ 0.36 & 0.45 & 0.18\\ 0 & 0.18 & 0.36 \end{pmatrix}. $$ For example, the modeled variance of the second individual's genetic contribution is $0.45$ outcome-units squared. The covariance between the first and second individuals' contributions is $0.36$ outcome-units squared. These values follow from their genotype patterns and the model weights; they are not estimates of the actual contributions for these individuals.

Why a Variance Test Is Omnibus

A burden model combines variant dosages into a single score and asks about the effect of that combined score. Its performance can depend on how variant effects align, as the earlier material on burden tests and opposing effects establishes. A variance-component model does not require the effects to add in one common direction. It allows one coefficient to be positive and another negative, while testing whether the collection has variation beyond the null model.

This is the sense in which the test is omnibus: the null is a single gene-level statement about the variance component, and the alternative permits many possible configurations of variant effects. It does not identify which variant contributes, estimate a reliable direction for each rare variant, or establish that every included variant is causal. Those are distinct questions.

Worked Example: Opposing Effects Are Allowed

Consider three hypothetical variant effects in a gene with weights $w_1=1$, $w_2=0.25$, and $w_3=1$, under $\tau=0.04$. The model-implied standard deviations are $$ \sqrt{0.04(1)}=0.20,\qquad \sqrt{0.04(0.25)}=0.10,\qquad \sqrt{0.04(1)}=0.20. $$ Suppose one possible realized effect vector is $\beta=(0.20,-0.10,0.15)^\mathsf{T}$, in outcome units per dosage unit. The second effect is negative while the first and third are positive, and each magnitude is compatible with the corresponding model scale.

For a hypothetical individual with dosages $(1,2,1)$, the genetic contribution is $$ 1(0.20)+2(-0.10)+1(0.15)=0.20-0.20+0.15=0.15 $$ outcome units. This calculation shows how opposing effects can offset for an individual without requiring any shared direction. It does not show that this particular coefficient vector would be observed, or that the model has been tested; it illustrates a configuration the random-effects alternative permits.

A Short Derivation of the Implied Covariance

The covariance structure explains why this model can be tested as a gene-level variance component. Conditional on the observed genotypes and covariates, write the random part of the outcome as $u+\epsilon=G\beta+\epsilon$. Since $E(\beta)=0$, $E(\epsilon)=0$, and the two are independent, its mean is zero. Its covariance is

\begin{aligned} \operatorname{Cov}(G\beta+\epsilon\mid G,X) &=G\operatorname{Cov}(\beta)G^\mathsf{T} +\operatorname{Cov}(\epsilon)\\ &=G(\tau W)G^\mathsf{T}+\sigma^2 I\\ &=\tau GWG^\mathsf{T}+\sigma^2 I. \end{aligned}

The first equality uses independence to make the cross-covariance terms zero. The second substitutes the assumed covariance matrices. Consequently, $\tau$ scales the part of the outcome covariance induced by the gene's genotypes, while $\sigma^2 I$ represents the residual covariance in this basic model. Testing whether the gene contributes modeled variability is testing whether the coefficient multiplying $GWG^\mathsf{T}$ is zero.

This derivation establishes the model's covariance implication; it is not the derivation of a particular score statistic or its null distribution. Those calculations require additional testing results. One important feature is already visible: $\tau=0$ is at the boundary of the allowed parameter space $\tau\geq 0$. Therefore, the usual reference distribution for testing an unrestricted parameter away from a boundary cannot simply be assumed without justification.

Worked Example: Interpreting a Hypothetical Test Result

Suppose a correctly calibrated variance-component procedure for a hypothetical gene reports $p=0.012$ at significance level $\alpha=0.05$, along with a model-based estimate $\hat{\tau}=0.031$. The comparison is $0.012<0.05$, so the result is statistically significant at the stated level. The p-value means that, if $H_0:\tau=0$ and the testing assumptions hold, the probability of obtaining a test result at least as extreme as the observed one is $0.012$.

The estimate corresponds to a coefficient-scale standard deviation of $\sqrt{0.031}=0.176$ outcome units per dosage unit for a variant with weight one. This estimate is on the scale defined by the model and its weights. The test supports evidence of nonzero gene-level effect variation under the model; it does not give a 1.2% probability that the null hypothesis is true, nor does it establish the direction or causal status of any individual variant.

Assumptions and Common Pitfalls

The model's interpretation depends on its assumptions. The weights and variant set define what is being tested; changing them changes the genetic covariance term. The zero-mean distribution and its variance structure describe how effect heterogeneity is represented. The covariates must appropriately account for modeled non-genetic mean differences, and the residual assumptions must be suitable for the outcome. In an analysis, population structure, relatedness, or other dependence may require additional adjustment rather than the simple residual covariance $\sigma^2 I$.

A significant variance-component test is evidence against the zero-variance gene model, conditional on the analysis choices and assumptions. It is not proof of a biological mechanism. Likewise, a nonsignificant result is not proof that every effect is zero: limited sample size, few carriers, or a poorly matched variant set can make a test uninformative.

Common mistake: “The random-effects model assumes every variant effect is zero on average, so a significant test means every variant has an effect in the same direction.” The correction is that zero is the modeled mean across effects, not a shared direction for their realized values. A positive variance component supports variation in the gene-level effect distribution; it neither requires aligned signs nor identifies which variants have effects.

The random-effects formulation matters because it changes the target from a single combined direction to the spread of effects across a gene. That target can accommodate heterogeneous variant effects while still producing one gene-level hypothesis. Its strength is also its limit: the result is an omnibus statement about modeled variation, not a variant-by-variant conclusion.

Check Your Understanding

Answer each question by referring to the model and its assumptions.

In $\beta \sim N(0,\tau W)$, what does $\tau$ measure, and what is the effect-scale standard deviation for variant $j$?
Under the stated model, what does $H_0:\tau=0$ imply about the distribution of $\beta$?
Why can the alternative allow both positive and negative variant effects?
In the covariance expression $\tau GWG^\mathsf{T}+\sigma^2I$, which term represents the modeled genetic contribution?
Why does a significant variance-component test not establish that every variant in the gene is causal?