Why Model Variant Effects as Random?
A burden model is most effective when the variants in a set are expected to have effects pointing in a shared direction after weighting. But a set may contain effects that differ in size or direction. If their contributions partly cancel in a weighted sum, a single common-direction summary can be a poor description of the association pattern. A random-effects model offers a different target: rather than estimating one shared effect, it represents the variant effects as varying around zero and asks whether their variation contributes to the phenotype.
This is a modeling choice, not a claim that each variant has no effect on average in every biological context. The mean-zero assumption describes a distribution for the effects across the set, while a variance component describes the allowed amount and pattern of variation. The resulting phenotype model is useful only if that allowed variation can be expressed by the observed genotype patterns.
Model and Assumptions
The entries of \(\beta\) are effects per allele copy, in phenotype units per allele copy. Thus the entries of \(\tau W\) have squared effect units. The matrix \(W\) specifies relative variances and covariances among effects; \(\tau\) scales that structure. For example, a diagonal \(W\) allows independent effects with variant-specific relative variances. If a diagonal matrix is built from nonnegative variant weights \(a_j\), one possible convention is \(W_{jj}=a_j^2\). Off-diagonal entries can instead encode dependence among effects.
The normality assumptions make the marginal distribution of \(Y\) normal. The covariance calculation below requires only the stated means, covariances, and independence; normality is additionally needed for the exact marginal normal model. The residual assumption is independent, equal-variance errors conditional on the covariates, genotypes, and effects. In practice, population structure, relatedness, or other dependence may require a different residual covariance model. The genotype matrix and \(W\) are treated as fixed in this formulation, and the covariate mean model must be adequate for the phenotype.
From Random Effects to Phenotype Covariance
Because \(E(\beta)=0\), the phenotype mean after averaging over the variant effects is \(E(Y\mid C,G)=C\gamma\). The genetic contribution appears in the covariance. Applying the law of total covariance gives
In the first term, conditional on \(\beta\), the only random part of \(Y\) is \(\varepsilon\), with covariance \(\sigma^2 I_n\). In the second term, the conditional mean is \(C\gamma+G\beta\); its covariance over \(\beta\) is \(G\operatorname{Cov}(\beta)G^\mathsf{T}\). Substituting \(\operatorname{Cov}(\beta)=\tau W\) yields the result. Since \(W\) is positive semidefinite, \(GWG^\mathsf{T}\) is positive semidefinite as well.
For any fixed vector \(c\) of length \(n\), the variance of the phenotype contrast \(c^\mathsf{T}Y\), conditional on \(C,G\), is \(\sigma^2 c^\mathsf{T}c+\tau c^\mathsf{T}GWG^\mathsf{T}c\). Thus the genetic component adds variance only in contrasts for which the genotype patterns overlap a direction allowed by \(W\). The parameter \(\tau\) is a variance scale, not a signed average allele effect.
Worked Example: Computing the Induced Covariance
Suppose three hypothetical individuals have two variant dosages \[ G=\begin{pmatrix}1&0\\0&1\\1&1\end{pmatrix}, \] and let \(W=\operatorname{diag}(4,1)\), \(\tau=0.25\), and \(\sigma^2=1.5\). The dosage entries count allele copies, and suppose the phenotype is measured in a fixed unit such as a laboratory concentration unit. Find the genetic covariance and total phenotype covariance.
First, the effect covariance is \[ \operatorname{Cov}(\beta)=\tau W =0.25\begin{pmatrix}4&0\\0&1\end{pmatrix} =\begin{pmatrix}1&0\\0&0.25\end{pmatrix}. \] The first variant therefore has effect variance \(1\) squared phenotype units per allele copy, and the second has variance \(0.25\) in those squared units. Next, \[ GWG^\mathsf{T} =\begin{pmatrix}4&0&4\\0&1&1\\4&1&5\end{pmatrix}, \qquad \tau GWG^\mathsf{T} =\begin{pmatrix}1&0&1\\0&0.25&0.25\\1&0.25&1.25\end{pmatrix}. \] Adding \(1.5I_3\) gives \[ \operatorname{Cov}(Y\mid C,G) =\begin{pmatrix}2.5&0&1\\0&1.75&0.25\\1&0.25&2.75\end{pmatrix}. \] For instance, individuals 1 and 3 share the first variant dosage, producing genetic covariance \(1\); individuals 1 and 2 have no overlapping dosage under this diagonal \(W\), so their genetic covariance is zero. These are covariances in squared phenotype units.
How This Target Differs From a Burden Effect
A simple common-direction model can be written \(\beta=\theta a\), where \(a\) is a specified vector of variant weights and \(\theta\) is one fixed scalar effect. Then the genetic mean contribution is \(G a\,\theta\): each individual's weighted burden \(Ga\) is multiplied by the same signed coefficient. In the random-effects model, by contrast, \(E(\beta)=0\), and association is represented by nonzero effect variance in permitted directions. Individual effects can point in opposite directions. The two models therefore encode different alternatives, not two notations for the same assumption.
This distinction explains why a burden summary can lose information under effect heterogeneity. When some effects are positive and others negative, contributions to a common weighted sum can cancel. A variance-component model does not require all effects to share one sign, although it can still miss signal if \(W\) assigns little variance to the relevant effects or if the genotypes do not express the allowed directions.
Worked Example: Opposing Effects and a Common Burden Direction
Consider two hypothetical variants with equal burden weights \(a=(1,1)^\mathsf{T}\), and a realized effect vector \(\beta=(0.8,-0.8)^\mathsf{T}\) phenotype units per allele copy. Three individuals have rows of \(G\) equal to \((1,0)\), \((0,1)\), and \((1,1)\). Calculate their genetic contributions and compare the pattern with the common-direction model.
The realized contributions \(G\beta\) are \[ \begin{pmatrix}1&0\\0&1\\1&1\end{pmatrix} \begin{pmatrix}0.8\\-0.8\end{pmatrix} =\begin{pmatrix}0.8\\-0.8\\0\end{pmatrix}. \] By comparison, the burden values are \(Ga=(1,1,2)^\mathsf{T}\). Under \(\beta=\theta a\), the contributions would be \((\theta,\theta,2\theta)^\mathsf{T}\), so the first two individuals must have contributions in the same direction. No value of \(\theta\) reproduces \((0.8,-0.8,0)^\mathsf{T}\). The example shows the structural mismatch; it does not assert that these realized coefficients were estimated from data.
For contrast, if a random-effects model has \(W=I_2\), then for the third individual's genotype row \(g=(1,1)\), the variance of its genetic contribution is \(\tau gWg^\mathsf{T}=2\tau\). This model allows variable effects on the two variants without imposing one shared sign, but its expected contribution is zero. The distinction is between a fixed directional mean and a distribution of effects with nonzero variance.
When an Allowed Effect Direction Is Invisible
A positive-semidefinite \(W\) may be singular. It can allow effects to vary only in a lower-dimensional subspace. Even if \(W\ne0\), the genotypes may have no information about that subspace. The following example makes the qualification important: a nonzero effect covariance alone does not guarantee a phenotype covariance contribution.
Worked Example: Nonzero Effect Variation With No Genetic Contribution
Suppose two variants have effect-shape matrix \[ W=\begin{pmatrix}1&-1\\-1&1\end{pmatrix} \] and every individual has genotype row \((1,1)\), so \(G\) has identical rows \((1,1)\). Verify that \(W\) is positive semidefinite and compute \(GWG^\mathsf{T}\).
The eigenvalues of \(W\) are \(0\) and \(2\), so \(W\) is positive semidefinite and is not the zero matrix. For an individual genotype row \(g=(1,1)\), \[ gW=(1,1)\begin{pmatrix}1&-1\\-1&1\end{pmatrix}=(0,0), \quad\text{and hence}\quad gWg^\mathsf{T}=0. \] Because every row of \(G\) is \(g\), every entry of \(GWG^\mathsf{T}\) is zero. Thus \(\tau GWG^\mathsf{T}=0\) for every \(\tau\), including positive \(\tau\). The model permits effects to vary in the contrast direction \((1,-1)^\mathsf{T}\), but these genotype rows only measure the sum of the two effects. That contrast is invisible in these data.
Model Checks, Alternatives, and Limitations
Before interpreting a variance component, check the biological and statistical meaning of the inputs. Variant weights and \(W\) should be specified on a scale that makes sense for the effect units, and allele coding should be consistent: changing which allele is counted changes genotype columns and can change the interpretation of effect directions. Check that the covariates account for the intended mean structure and that residual independence and variance assumptions are plausible. In related samples or structured populations, an independent residual model may be inadequate. A nonzero estimate of \(\tau\) cannot repair confounding or an incorrect mean model.
The alternatives differ in what they assume and target:
| Model | Variant-effect structure | Primary target |
|---|---|---|
| Common-direction burden | \(\beta=\theta a\), with specified weights \(a\) | A shared signed effect along the burden direction |
| Fixed multivariate effects | Separate fixed coefficients for variants | Specific variant effects, when estimable with adequate information |
| Random effects | \(E(\beta)=0\), \(\operatorname{Cov}(\beta)=\tau W\) | Effect variation in directions represented by \(W\) |
The fixed multivariate model avoids a distributional assumption about variant effects, but estimating many coefficients can be unstable when variants are rare or correlated. A burden model reduces dimension but can be poorly aligned with opposing effects. A random-effects model represents heterogeneity compactly, but conclusions depend on the chosen \(W\). Its mean-zero formulation also does not directly estimate the direction or identity of each variant's effect.
One parameterization issue deserves attention: if both the overall scale of \(W\) and \(\tau\) are left unrestricted, only their product is identified by \(\tau W\). The scale of \(W\) should therefore be fixed by convention or construction before interpreting \(\tau\). Also, because \(\tau\geq0\), the null value lies at the boundary of the parameter space. This affects inferential calibration; a routine that assumes an unconstrained, interior variance parameter may not be justified. The model formulation alone does not specify a valid p-value calculation.
A common mistake is to say, “\(\tau>0\) means the variant set contributes genetic variance.” That conclusion is not always true. If \(W=0\), the effect covariance is zero for all \(\tau\); and even with \(W\ne0\), the observed genotypes may satisfy \(GWG^\mathsf{T}=0\). The correction is to check the induced covariance itself. A positive scale permits variation in the directions encoded by \(W\), while a contribution to the phenotype requires those directions to be visible through \(G\).
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- If \(\operatorname{Cov}(\beta)=\tau W\), what does \(W\) specify and what does \(\tau\) scale?
- For one individual with genotype row \(g=(1,0)\) and \(W=I_2\), calculate the genetic variance of that individual's phenotype contribution.
- Why does a mean-zero random-effects model not mean that every variant has a zero realized effect?
- If \(W\ne0\) and \(\tau>0\), is a genetic covariance contribution guaranteed? State the condition needed.
- What key effect assumption distinguishes \(\beta=\theta a\) from a mean-zero random-effects model?
Solutions. (1) \(W\) specifies the relative variances, covariances, and permitted directions of variant effects; \(\tau\) scales the entire covariance structure. (2) \(gWg^\mathsf{T}=(1,0)(1,0)^\mathsf{T}=1\), so the genetic variance is \(\tau\), in squared phenotype units. (3) The zero is the distributional mean across the random-effects model; a realization can have nonzero, positive, or negative entries. (4) No. The induced contribution requires \(GWG^\mathsf{T}\ne0\). (5) The burden specification constrains all effects to one weighted direction and one shared signed coefficient, while the random-effects model allows variation around zero according to \(W\).