Self-Contained Gene-Set Tests and Their Null Hypothesis

What Does “Self-Contained” Ask?

A gene-set test is only interpretable once its null hypothesis is explicit. A self-contained test asks whether the genes in a specified set show any association with a defined study variable, such as treatment condition, after accounting for the covariates in the gene-level model. “Self-contained” refers to the scope of the hypothesis: the test concerns the pathway genes themselves, not whether their association is stronger than that of genes outside the pathway.

For example, suppose expression is measured for individuals in two treatment groups. For each gene, a model can estimate the treatment-associated expression difference, adjusted for specified covariates. A self-contained test for a pathway jointly asks whether all of those gene-specific differences are zero. The association parameter and the contrast must be stated: a test of a treatment main effect is not a test of a treatment-by-time interaction.

Definition: Let \(G\) be a specified gene set within the tested-gene universe. For each gene \(j\in G\), let \(\beta_j\) denote the coefficient for a specified association or contrast in a stated gene-level model. A self-contained test has global null hypothesis \(H_0:\beta_j=0\) for every \(j\in G\), and alternative hypothesis \(H_a:\beta_j\ne 0\) for at least one \(j\in G\). Thus, the global null is the intersection of the gene-level null hypotheses, and the alternative is their union.

This is a statement about the model’s chosen association parameter. If \(\beta_j\) is a log fold change, zero means no modeled log fold change for gene \(j\); if it is a regression coefficient for a continuous exposure, zero means no modeled linear association with that exposure. The definition does not itself specify a test statistic or its reference distribution. Those depend on the gene-level models and on how uncertainty across genes is handled.

Formulating the Joint Hypothesis

Write down the pathway membership and the target coefficient before interpreting a result. If the set contains \(m\) genes, collect the coefficients into \(\boldsymbol{\beta}_G=(\beta_1,\ldots,\beta_m)^\mathsf{T}\). Then the self-contained null is \(\boldsymbol{\beta}_G=\mathbf{0}\). Its alternative is \(\boldsymbol{\beta}_G\ne\mathbf{0}\), meaning that at least one component differs from zero. It does not require every component to differ from zero, nor does it require all effects to have the same sign.

A joint test might use a Wald statistic, a likelihood-based statistic, or a calibrated combination of gene-level tests. Whatever procedure is used, its null distribution must be appropriate for the fitted models and the dependence among the gene-level estimates. Testing each gene separately and reporting the smallest unadjusted p-value is not, by itself, a valid global test: the chance of at least one small p-value can exceed the intended significance level.

Interpretation: Rejection of the self-contained null is evidence, at the chosen error rate and under the test’s assumptions, against the claim that every gene in \(G\) has zero association for the specified contrast. It supports the conclusion that at least one member is associated. It does not establish that all genes are associated, that the pathway is uniformly up- or down-regulated, or that the pathway causes the observed outcome.

Worked Example: A Joint Wald Test for Three Pathway Genes

Suppose a hypothetical study estimates treatment-associated log fold changes for three genes in a prespecified pathway. The estimates are \(0.8\), \(0.1\), and \(-0.6\), with standard errors \(0.4\), \(0.5\), and \(0.3\). Assume for this illustration that the estimates are independent and normally distributed under the null, and that the reported standard errors are known. These assumptions make the joint Wald statistic a chi-square variable under the global null.

The hypotheses are \(H_0:\beta_1=\beta_2=\beta_3=0\) versus \(H_a:\) at least one of these coefficients is nonzero. With independent estimates, the Wald statistic is the sum of the squared gene-level z-statistics:

W=\sum_{j=1}^{3}\left(\frac{\hat{\beta}_j}{SE(\hat{\beta}_j)}\right)^2 =\left(\frac{0.8}{0.4}\right)^2+ \left(\frac{0.1}{0.5}\right)^2+ \left(\frac{-0.6}{0.3}\right)^2 =2^2+0.2^2+(-2)^2=8.04.

Under the stated assumptions and \(H_0\), \(W\sim\chi^2_3\), so the p-value is \(P(\chi^2_3\ge 8.04)\approx 0.045\). At \(\alpha=0.05\), this test rejects the global null. In this hypothetical example, there is evidence that at least one of the three genes has a nonzero treatment-associated log fold change. The joint result alone does not identify which gene drives the evidence; the gene-level estimates are descriptive follow-up information, not proof that all three effects are nonzero.

The independence assumption is crucial to this particular calculation. If the estimated coefficients are correlated, summing their squared marginal z-statistics generally does not have the stated chi-square reference distribution. A valid joint Wald test then uses the covariance matrix of the coefficient vector, or another procedure with a justified calibration.

Worked Examples of the Null in Different Settings

Worked Example: A Pathway Test for an Exposure Association

A hypothetical study measures expression in 120 individuals and fits a regression for each gene, adjusting for age and sex. For gene \(j\), \(\beta_j\) is the coefficient for a one-unit increase in a specified exposure. A pathway contains four genes. The self-contained hypotheses are \(H_0:\beta_1=\beta_2=\beta_3=\beta_4=0\) versus \(H_a:\) at least one \(\beta_j\ne0\). The adjustment variables are part of the gene-level model; they do not change the definition of the global null.

Suppose the hypothetical gene-level p-values are \(0.012\), \(0.20\), \(0.61\), and \(0.83\). To illustrate a valid, conservative global test that does not require independence, use the Bonferroni rule: reject the global null if any of the four gene-level p-values is at most \(\alpha/4\). At \(\alpha=0.05\), the threshold is \(0.05/4=0.0125\). Since \(0.012\le0.0125\), the global null is rejected. Equivalently, the Bonferroni-adjusted p-value is \(\min(1,4\times0.012)=0.048\).

This conclusion is evidence that at least one pathway gene is associated with the exposure under the adjusted models. It does not establish that the other three genes have no association; their p-values may reflect weak effects, limited precision, or no effects. The test is conservative in general, and its conclusion depends on the gene-level p-values being valid for their stated models.

Worked Example: Testing an Interaction Rather Than a Main Effect

In a hypothetical repeated-measures experiment, investigators compare how expression changes over time between treatment groups. For each of two pathway genes, \(\beta_j\) is the treatment-by-time interaction coefficient. The gene-level null is \(H_{0j}:\beta_j=0\); the pathway-level self-contained null is \(H_0:\beta_1=\beta_2=0\). This asks whether either gene has evidence of a difference in time trends between groups, not whether either gene has a treatment main effect.

Suppose the estimated interaction coefficients are \(0.30\) and \(-0.10\), with standard errors \(0.15\) and \(0.20\). For a simple illustration, assume independent normally distributed estimates with known standard errors. A joint Wald statistic is

W=\left(\frac{0.30}{0.15}\right)^2+ \left(\frac{-0.10}{0.20}\right)^2 =2^2+(-0.5)^2=4.25.

Under the global null, \(W\sim\chi^2_2\). For two degrees of freedom, \(P(\chi^2_2\ge4.25)=e^{-4.25/2}\approx0.119\). At \(\alpha=0.05\), the test does not reject the global null. This does not prove that both interaction coefficients are zero; it means this test has not provided sufficient evidence against their joint zero null at the chosen level. The example also shows why naming the contrast matters: a test of these interaction coefficients answers a different question from a test of treatment main effects.

Justifying the Global Null and the Test

The logical form of the self-contained hypothesis follows directly from the phrase “any genes.” Let \(H_{0j}\) be the null hypothesis that gene \(j\) has no association for the specified parameter. “No genes in \(G\) are associated” means that every \(H_{0j}\) is true, which is the intersection of the gene-level nulls. Its logical complement is “at least one gene is associated.” By De Morgan’s law, the complement of an intersection is the union of the complements:

H_0=\bigcap_{j\in G}\{\beta_j=0\}, \qquad H_a=\left(\bigcap_{j\in G}\{\beta_j=0\}\right)^c =\bigcup_{j\in G}\{\beta_j\ne0\}.

This justification concerns the hypotheses, not the distribution of a particular test statistic. For instance, a Wald test using a vector of estimates \(\hat{\boldsymbol{\beta}}_G\) and its covariance matrix \(\widehat{\Sigma}\) can use

W=\hat{\boldsymbol{\beta}}_G^\mathsf{T} \widehat{\Sigma}^{-1}\hat{\boldsymbol{\beta}}_G.

When the coefficient estimates are approximately multivariate normal under \(H_0\), the covariance estimate is suitable, and regularity conditions hold, this statistic is often compared with a \(\chi^2_m\) distribution. With estimated dispersion or small samples, a different reference distribution or calibration may be needed. The formula does not make those conditions true: they must be justified for the study design and fitted models.

Assumptions, Diagnostics, and Common Mistakes

The null hypothesis can be stated even when a particular test is unsuitable. A trustworthy p-value additionally requires that the gene-level association parameters represent the scientific question and that the global test is calibrated under the joint null. Check that the chosen contrast is correct, the gene set is defined against the tested-gene universe, and the expression measurements and sample labels are aligned. The gene-level model must be appropriate for the data and design, including its handling of covariates, repeated measurements, and distributional features.

Dependence matters for calibration. Expression measurements for different genes in the same individuals can be correlated, and so can their estimated coefficients. A test that assumes independent gene-level results may give an incorrect null distribution when that assumption is false. Use a joint procedure that accounts for the relevant covariance, or a procedure whose validity does not require independence. Also check that the pathway contains enough tested genes for the chosen method and that the estimates are not dominated by unstable fits or missing observations.

Wrong reasoning: “The pathway test is significant, so every gene in the pathway is differentially expressed.” Correction: Rejection concerns the global null that all member-gene coefficients are zero. It supports an association in at least one gene, not all genes.
Wrong reasoning: “A nonsignificant test proves that no pathway gene is associated.” Correction: Failure to reject does not prove the null. Power may be low, effects may be small, or estimates may be imprecise.
Wrong reasoning: “The pathway is enriched because its genes are more associated than genes outside it.” Correction: That is a relative, outside-gene comparison. A self-contained null asks whether the pathway genes themselves have zero association.
Wrong reasoning: “Any gene-level p-value below 0.05 makes the pathway test significant.” Correction: A global test must account for testing multiple genes, through its statistic and calibration or an appropriate adjustment.

Self-contained tests are useful when the scientific question is whether a prespecified gene set contains evidence of association at all. Their scope is also a limitation: a rejection does not measure how the set compares with the rest of the genome, establish a biological mechanism, or by itself summarize direction and magnitude. Those claims require additional analyses and careful interpretation.

Key takeaway: State the self-contained null as zero association for every member gene, for a clearly named model coefficient or contrast. Its alternative is that at least one member has nonzero association; the global test must then account for its assumptions and, where relevant, dependence among genes.

Check Your Understanding

For each question, identify the hypothesis or interpretation before checking the brief solution below.

A set contains three genes with coefficients \(\beta_1,\beta_2,\beta_3\). Write its self-contained null and alternative.
A pathway test rejects its global null. What does this establish, and what does it not establish?
A two-gene pathway has independent, normally distributed estimates with known standard errors. The estimates are \(0.4\) and \(-0.3\), with standard errors \(0.2\) and \(0.3\). Calculate the Wald statistic and state its null reference distribution.
For a four-gene pathway, the smallest valid gene-level p-value is \(0.018\). Does a Bonferroni global test reject at \(\alpha=0.05\)?

Solutions. (1) \(H_0:\beta_1=\beta_2=\beta_3=0\); \(H_a:\) at least one coefficient is nonzero. (2) It is evidence against all member coefficients being zero, supporting association for at least one gene; it does not show that all genes are associated or that the pathway causes the outcome. (3) \(W=(0.4/0.2)^2+(-0.3/0.3)^2=4+1=5\), compared with \(\chi^2_2\) under the stated assumptions. (4) The Bonferroni threshold is \(0.05/4=0.0125\). Since \(0.018>0.0125\), this test does not reject.