Integrating Molecular Evidence Across Tissues

Why Integrate Molecular Evidence Across Tissues?

A genetic variant may influence the same molecular trait in several tissues, but the estimated effects need not be identical. Integrating tissue-specific estimates can improve precision when an effect is shared, while also revealing when a single summary would hide meaningful tissue variation. The analysis must account for how each estimate was obtained and, especially, whether estimates from different tissues are statistically dependent because they use overlapping participants.

The aim here is to combine estimates of the same genetic effect on the same molecular trait across tissues. Before pooling, align the effect allele and ensure that the estimates have compatible units and refer to comparable genetic contrasts. A positive effect must mean the same direction in every tissue. If estimates are on different scales or refer to different estimands, a pooled number may not have a coherent interpretation.

Definition: For \(K\) tissues, let \(\hat{\boldsymbol{\beta}}=(\hat{\beta}_1,\ldots,\hat{\beta}_K)^\mathsf{T}\) be the vector of estimated genetic effects, where \(\hat{\beta}_j\) estimates the effect \(\beta_j\) in tissue \(j\). Let \(S\) be the \(K \times K\) sampling covariance matrix of these estimates. Its diagonal entries are the estimated sampling variances; its off-diagonal entries describe sampling-error covariance between tissues.

The matrix \(S\) concerns uncertainty in the estimates, not necessarily biological similarity between the true tissue effects. For example, if the same donors contribute to expression measurements in two tissues, the estimation errors can be correlated. Treating those estimates as independent when they are not can make a pooled estimate appear more precise than it is. We assume throughout that \(S\) is known or estimated well enough for the calculations, is positive definite, and that the estimates are approximately jointly normal.

A Common Effect: Covariance-Aware Pooling

A common-effect model assumes that all tissues have one true effect, \(\beta_j=\beta\) for every \(j\), and that differences among the observed estimates arise from sampling error. In vector form, \(\hat{\boldsymbol{\beta}}=\mathbf{1}\beta+\boldsymbol{\varepsilon}\), where \(\mathbf{1}\) is a vector of \(K\) ones and \(\boldsymbol{\varepsilon}\) has mean zero and covariance \(S\). Generalized least squares gives the estimate that accounts for the entire covariance matrix.

Definition: Under the common-effect model, the generalized least-squares estimate and its standard error are

\hat{\beta}_{\mathrm{GLS}} = \frac{\mathbf{1}^\mathsf{T}S^{-1}\hat{\boldsymbol{\beta}}} {\mathbf{1}^\mathsf{T}S^{-1}\mathbf{1}}, \qquad SE(\hat{\beta}_{\mathrm{GLS}}) = \left(\mathbf{1}^\mathsf{T}S^{-1}\mathbf{1}\right)^{-1/2}.

The estimate is a covariance-adjusted weighted average. When tissue estimates are independent, \(S\) is diagonal and the weights are proportional to inverse variances. With correlated estimates, the off-diagonal entries also affect the weights and the standard error. If the common effect is being tested against zero, the statistic is \(Z=\hat{\beta}_{\mathrm{GLS}}/SE(\hat{\beta}_{\mathrm{GLS}})\); under the null and the stated normal model, it is approximately standard normal.

Worked Example: Pooling Two Correlated Tissue Estimates

Suppose a hypothetical study estimates a variant's effect on expression in two tissues. Both estimates are in expression units per allele, with \(\hat{\beta}_1=0.40\) and \(\hat{\beta}_2=0.30\). Each has standard error \(0.20\), and their estimated sampling-error correlation is \(0.45\). Thus the sampling covariance is \(0.45(0.20)(0.20)=0.018\), and

S= \begin{pmatrix} 0.040 & 0.018\\ 0.018 & 0.040 \end{pmatrix}.

The common-effect estimate is the average in this symmetric two-tissue case:

\hat{\beta}_{\mathrm{GLS}}=\frac{0.40+0.30}{2}=0.350 \quad\text{expression units per allele}.

For two equally precise, positively correlated estimates, the variance of their average is \((0.040+0.018)/2=0.029\). Therefore,

SE(\hat{\beta}_{\mathrm{GLS}})=\sqrt{0.029}=0.1703 \quad\text{expression units per allele}.

A 95% normal-approximation confidence interval is \(0.350 \pm 1.96(0.1703)\), or \((0.0162,\,0.684)\) expression units per allele. This interval describes uncertainty about the common effect under the common-effect model; it is not a 95% probability statement about the fixed parameter. For the test of \(H_0:\beta=0\), \(Z=0.350/0.1703=2.055\), giving a two-sided \(p\)-value of about \(0.0398\). Under the model, the data provide evidence that the shared effect differs from zero. This calculation uses the nonzero covariance; assuming independence would instead give a smaller standard error of \(\sqrt{0.040/2}=0.1414\).

Testing Whether Tissue Effects Are Heterogeneous

A pooled estimate alone does not show whether a common effect is plausible. A standard diagnostic compares the tissue estimates with their generalized least-squares common-effect estimate. The resulting \(Q\) statistic measures their covariance-adjusted deviations. If the estimates are jointly normal, \(S\) is treated as known, and the common-effect model holds, \(Q\) has a chi-square distribution with \(K-1\) degrees of freedom.

Definition: The generalized least-squares heterogeneity statistic is

Q=(\hat{\boldsymbol{\beta}}-\mathbf{1}\hat{\beta}_{\mathrm{GLS}})^\mathsf{T} S^{-1} (\hat{\boldsymbol{\beta}}-\mathbf{1}\hat{\beta}_{\mathrm{GLS}}).

Under the common-effect model, its reference distribution is approximately \(\chi^2_{K-1}\). A small \(p\)-value indicates that the observed tissue differences are difficult to explain by the specified sampling covariance alone. A large \(p\)-value does not prove that effects are identical; with few tissues or imprecise estimates, the test may have limited power.

Worked Example: Testing Heterogeneity With Correlated Estimates

Now suppose the same hypothetical two-tissue study estimates effects of \(0.40\) and \(0.10\) expression units per allele, retaining the covariance matrix from the preceding example. The common-effect estimate is \(0.25\), so the deviations are \(0.15\) and \(-0.15\). Their difference is \(0.30\), and its sampling variance is

\operatorname{Var}(\hat{\beta}_1-\hat{\beta}_2) =0.040+0.040-2(0.018)=0.044.

With two tissues, the GLS heterogeneity statistic equals the squared difference divided by the variance of that difference:

Q=\frac{(0.40-0.10)^2}{0.044} =\frac{0.090}{0.044}=2.045.

The degrees of freedom are \(K-1=1\), and the chi-square upper-tail \(p\)-value is approximately \(0.153\). At a 0.05 significance level, these estimates do not provide strong evidence against a common effect. This is not evidence that the true effects are exactly equal. In particular, the shared-sample covariance matters: the variance of the contrast is \(0.044\), not \(0.080\), the value that would follow from incorrectly assuming independence.

Random Effects: An Average Effect and Between-Tissue Variation

If true effects may vary, a random-effects model represents each tissue effect as a draw around an average: \(\beta_j=\mu+u_j\), with \(u_j\) having mean zero and variance \(\tau^2\). In the basic model, the tissue deviations are mutually independent and independent of the sampling errors. This assumption is distinct from independence of the observed estimates: sampling errors can still be correlated through \(S\). The model says that \(\mu\) is the average effect across the modeled tissues, while \(\tau^2\) quantifies between-tissue variation.

Definition: Under the random-effects model,

\hat{\boldsymbol{\beta}}=\mathbf{1}\mu+\mathbf{u}+\boldsymbol{\varepsilon}, \qquad \operatorname{Var}(\mathbf{u})=\tau^2 I, \qquad \operatorname{Var}(\boldsymbol{\varepsilon})=S, \qquad \operatorname{Var}(\hat{\boldsymbol{\beta}})=S+\tau^2 I.

Here \(\tau^2\geq 0\) is the between-tissue variance. If effects are measured in expression units per allele, its units are (expression units per allele)\(^2\), equivalently squared expression units per squared allele.

A simple method-of-moments estimate of \(\tau^2\) starts with \(Q\), calculated using \(S\) as in the common-effect model. Define

P=S^{-1}-S^{-1}\mathbf{1} (\mathbf{1}^\mathsf{T}S^{-1}\mathbf{1})^{-1} \mathbf{1}^\mathsf{T}S^{-1}.

Then \(Q=\hat{\boldsymbol{\beta}}^\mathsf{T}P\hat{\boldsymbol{\beta}}\). Under the random-effects model, its expectation is \(K-1+\tau^2\operatorname{tr}(P)\). Equating the observed \(Q\) to this expectation gives an estimate, truncated at zero because a variance cannot be negative. The calculation assumes that the specified sampling covariance is adequate and that independent tissue deviations are a reasonable model.

Result: The method-of-moments estimate is

\hat{\tau}^2_{\mathrm{MM}} = \max\left(0,\frac{Q-(K-1)}{\operatorname{tr}(P)}\right).

When the estimates are independent with variances \(s_j^2\), this reduces to the familiar denominator \(C=\sum_j w_j-\sum_j w_j^2/\sum_j w_j\), where \(w_j=1/s_j^2\). For correlated estimates, the matrix form uses the full \(S\).

Derivation: Why the Heterogeneity Statistic Has This Expectation

The projection-like matrix \(P\) removes the fitted common mean: \(P\mathbf{1}=0\). Under the random-effects model, the estimate vector has mean \(\mathbf{1}\mu\) and covariance \(S+\tau^2 I\). For any random vector \(Y\) with mean \(m\) and covariance \(V\), the expectation of \(Y^\mathsf{T}PY\) is \(\operatorname{tr}(PV)+m^\mathsf{T}Pm\). Here \(m^\mathsf{T}Pm=0\) because \(P\mathbf{1}=0\). Therefore,

E(Q) =\operatorname{tr}\{P(S+\tau^2 I)\} =\operatorname{tr}(PS)+\tau^2\operatorname{tr}(P) =(K-1)+\tau^2\operatorname{tr}(P).

The last equality follows because \(PS\) is the residual-forming operator for one fitted common mean and has trace \(K-1\). Solving this expectation equation for \(\tau^2\) yields the stated method-of-moments formula. Truncation at zero enforces the variance constraint; it does not mean the data establish that the true heterogeneity is exactly zero.

Worked Example: Estimating Between-Tissue Variance

Consider three hypothetical tissue effects, each measured in expression units per allele: \(0.20\), \(0.50\), and \(0.80\). Suppose their sampling errors are independent, with standard errors \(0.10\) in each tissue, so \(S=0.01I\). The common-effect estimate is their mean, \(0.50\). The deviations are \(-0.30\), \(0\), and \(0.30\), giving

Q=\frac{(-0.30)^2+0^2+(0.30)^2}{0.01} =\frac{0.18}{0.01}=18.

Here \(K-1=2\), and \(S^{-1}=100I\). The matrix \(P=100(I-\mathbf{1}\mathbf{1}^\mathsf{T}/3)\) has trace \(100(3-1)=200\). Thus,

\hat{\tau}^2_{\mathrm{MM}} = \max\left(0,\frac{18-2}{200}\right) =0.0800\;(\text{expression units per allele})^2, \qquad \hat{\tau}=\sqrt{0.0800}=0.2828\;\text{expression units per allele}.

The heterogeneity test has \(2\) degrees of freedom; its upper-tail \(p\)-value is \(e^{-18/2}=e^{-9}\approx0.0001234\). The estimates show substantial variation relative to their sampling errors under the stated model. With the estimated variance, the random-effects covariance is \(S+\hat{\tau}^2I=0.09I\). The random-effects estimate of the average is still \(0.50\), and its plug-in standard error is \(\sqrt{0.09/3}=0.1732\) expression units per allele. This standard error treats the estimated \(\tau^2\) as fixed; it does not include the additional uncertainty in estimating that variance.

Interpreting the Integrated Result

A common-effect estimate answers, “What single effect best summarizes these tissues if their true effects are equal?” A random-effects estimate answers, “What is the average effect across the modeled tissues when tissue effects may vary?” These are different scientific summaries. A statistically precise average does not imply that every tissue has an effect close to that average, and a heterogeneity test does not identify which tissue differs or explain why.

The basic random-effects model also makes a substantive assumption: the tissue deviations are independent with a shared variance. This may be a useful starting point, but tissues can have structured biological similarities. If the true deviations are correlated, the covariance model \(\tau^2 I\) may not describe them adequately. Likewise, the sampling covariance \(S\) must reflect shared participants and the estimation procedure; it is not a substitute for modeling biological heterogeneity.

Common Mistakes

Assuming estimates from different tissues are independent. Separate tissue labels do not guarantee independent sampling errors. When donors overlap, use the off-diagonal elements of \(S\).
Calling a nonsignificant heterogeneity test proof of a shared effect. A large \(p\)-value means the data did not detect heterogeneity at the chosen threshold; it does not establish exact equality.
Confusing the average with a universal tissue effect. In a random-effects model, \(\mu\) is the average across modeled tissues. It need not equal each \(\beta_j\).
Reporting \(\tau^2\) in the units of the effect. Variance has squared units. If an effect is in expression units per allele, \(\tau^2\) is in (expression units per allele)\(^2\), while \(\tau\) is in expression units per allele.
Pooling estimates that do not share a coherent scale or direction. Check allele alignment, phenotype definition, and effect units before combining results.

Key takeaway: Integrating effects across tissues requires separating sampling covariance from true tissue variation. Use the full sampling covariance matrix for common-effect pooling and heterogeneity assessment; use a random-effects model when the scientific target is an average across effects that may genuinely differ.

Check Your Understanding

Try these questions without looking back.

What does an off-diagonal element of \(S\) represent, and why might it be nonzero for two tissue estimates?
Under what model assumption does the GLS pooled estimate represent one common true effect?
What does a small \(p\)-value for \(Q\) indicate, and what does it not identify?
In a random-effects model, distinguish the meanings and units of \(\mu\), \(\tau\), and \(\tau^2\).
Why does a nonsignificant heterogeneity test not prove that the tissue effects are identical?