Testing Module Associations With Phenotypes

Why Test Module–Phenotype Associations?

A module eigengene summarizes the expression pattern of a module across samples. Once modules have been identified, a natural question is whether their eigengenes vary with a phenotype of interest. For example, does the eigengene of a hypothetical immune-related module tend to be higher in samples with higher inflammation scores? This is an association question: a statistical relationship does not, by itself, establish that the module causes a phenotype or that any gene in the module is responsible for it.

The analysis must reflect the kind of phenotype and the design. A quantitative phenotype can be related to an eigengene using correlation or regression. For a categorical phenotype, compare eigengene distributions across groups, or model the phenotype as an outcome. If other variables could explain part of an observed relationship, use an adjusted model and distinguish that estimate from the unadjusted association.

A standard analysis assumes samples are independent observational units, the phenotype and eigengene are correctly paired, and the measurements and model are appropriate for the study design. Repeated samples from one individual, related individuals, or clustered sampling may violate independence; those designs need models that account for dependence rather than an ordinary independent-sample test.

Quantitative Phenotypes: Correlation and Regression

For a quantitative phenotype \(Y_i\), the Pearson correlation summarizes the direction and strength of a linear relationship with \(ME_{ik}\). A simple linear regression expresses the same relationship in phenotype units:

Here \(\beta_1\) is the expected change in the phenotype for a one-unit increase in the eigengene, under the fitted model. The usual t-test of \(H_0:\beta_1=0\) requires independent errors, a correctly specified linear mean relationship, and constant error variance; exact small-sample inference additionally assumes normally distributed errors. A scatterplot should be examined for curvature and unusual observations. A correlation can be useful as a compact summary, but regression makes the phenotype’s units and covariate adjustment explicit.

Categorical Phenotypes: Group Comparisons and Outcome Models

For a binary phenotype, one direct approach is to compare the mean eigengene between the two groups. With independent samples and approximately normal within-group errors, a two-sample t-test is appropriate. A pooled-variance t-test also assumes equal group variances; Welch’s t-test avoids that equality assumption and is often a reasonable default. For a phenotype with more than two categories, one-way ANOVA tests whether all group means are equal, assuming independent errors, a suitable mean structure, and (for the classical pooled ANOVA test) equal within-group variances. A significant omnibus test indicates that not all means are equal; it does not identify which pairs differ without follow-up comparisons.

The direction of modeling depends on the scientific question. Modeling the eigengene as the response compares eigengene means across phenotype groups. Alternatively, for a binary phenotype \(D_i\), logistic regression models the probability of the phenotype from an eigengene:

In this model, \(\exp(\gamma_1)\) is the fitted odds ratio for a one-unit increase in eigengene, assuming a linear relationship between the eigengene and the log odds. Logistic regression does not require normally distributed eigengene values, but does require independent observations, a correctly specified logit relationship, and sufficient information to estimate the coefficients. These models describe different conditional quantities; report which one was fitted rather than treating them as interchangeable.

For a categorical phenotype with three or more groups, the same logic extends to an omnibus group comparison. For ordinal categories, treating category codes as a quantitative score imposes a particular trend model; use that coding only when equal spacing and a linear trend are scientifically defensible. Otherwise, model the categories as groups or use an appropriate ordinal outcome model.

Adjustment, Confounding, and Interpretation

A variable is a potential confounder when it is associated with both the eigengene and phenotype and may account for some of their observed relationship. A multiple linear regression can estimate the eigengene association conditional on measured covariates \(Z_{ij}\):

The coefficient \(\beta_1\) now describes the fitted phenotype difference per eigengene unit at fixed covariate values, under the model. It need not equal the marginal slope. Adjustment does not automatically remove bias: the relevant covariates must be measured and appropriately modeled, and inappropriate adjustment can also distort an association. Confounders should be chosen using subject-matter knowledge and the study design, not solely because their inclusion changes a p-value.

Diagnostics, Multiplicity, and Limitations

A test statistic and p-value are only as useful as the model and data behind them. For quantitative regression, inspect a scatterplot and residual-versus-fitted plot for curvature or changing variance; examine residuals for marked departures from normality when relying on small-sample t inference. For group comparisons, inspect group sizes, distributions, and spreads. A few samples can strongly influence a slope or correlation, especially if they have extreme eigengene values. Check leverage and influence measures, and calculate how key estimates change when an observation is omitted as a sensitivity analysis. An influential sample is not automatically erroneous; verify it and report the sensitivity rather than deleting it merely because it changes the result.

When many modules and phenotypes are tested, the collection of nominal p-values creates a multiple-testing problem: even if every null association is true, some small p-values can arise by chance. Define the testing family in advance and use a suitable multiplicity procedure, such as false-discovery-rate control, when screening many associations. Report effect estimates and uncertainty alongside p-values; statistical significance alone does not convey biological importance.

Standard model-based p-values also have limitations when modules, phenotypes, or covariates were selected after inspecting the same results. They generally do not account for such selection. Treating estimated eigengenes as fixed is a common working analysis, but it does not represent uncertainty from every stage of constructing the network. Independent validation or a design that separates exploration from confirmation can strengthen evidence. These cautions do not make association testing useless; they clarify what its estimates and tests support.

A practical comparison is: correlation gives a concise unadjusted linear summary; regression gives an effect estimate in phenotype units and accommodates covariates; t-tests and ANOVA compare eigengene means across categories; logistic regression models a binary phenotype’s odds. None substitutes for diagnostics, thoughtful covariate selection, or sensitivity checks.