Why Intergene Correlation Matters
A competitive pathway test asks whether genes in a specified set show stronger evidence of association with a phenotype than genes outside that set. Its null hypothesis concerns a comparison between the set and the rest of the tested-gene universe; it does not say that every gene has no effect. The earlier material on competitive and self-contained enrichment distinguishes these questions.
A separate issue is whether the gene-level statistics being combined can be treated as independent. Genes in a pathway may respond together because they participate in related processes, share regulatory inputs, or reflect common sources of biological and technical variation. If their test statistics are positively correlated, they provide less independent information than the same number of uncorrelated statistics. Treating them as independent can therefore make a pathway appear more significant than the data warrant.
Correlation adjustment concerns the uncertainty of a pathway-level statistic. It does not establish that coordinated genes are causally related, nor does it alter the competitive question being asked.
Variance of a Mean of Correlated Gene Statistics
Consider a pathway containing \(m\) genes. Let \(Z_i\) be a standardized gene-level statistic for gene \(i\), oriented so that larger values indicate stronger evidence in the direction being tested. For a simple derivation, suppose under the competitive null that each \(Z_i\) has mean zero and variance one, and every pair of distinct genes in the pathway has the same correlation \(\rho\). This equal-correlation model is an approximation, not a claim that real pathways have identical pairwise correlations.
The mean pathway score is \(\bar{Z}_G=m^{-1}\sum_{i=1}^{m}Z_i\). Its variance follows by expanding the variance of a sum: there are \(m\) variance terms and \(m(m-1)\) ordered covariance terms. Since each variance is one and each covariance is \(\rho\), the result is:
If the genes were independent, \(\rho=0\), and the variance would be \(1/m\). Thus the variance inflation factor is \(1+(m-1)\rho\), and the standard error is multiplied by its square root. For positive \(\rho\), the variance is larger than under independence; a statistic divided by the correct standard error is consequently smaller in magnitude.
A competitive statistic often compares the set’s average score with a reference formed from genes outside the set or from the tested-gene universe. The derivation above isolates the set-mean component: uncertainty in the reference and any dependence between the set and reference must also be handled by the specific test. CAMERA-style procedures use a competitive statistic and adjust its variance for estimated intergene correlation; the displayed equal-correlation calculation explains the central adjustment, rather than specifying every detail of every implementation.
Worked Examples
Worked Example: Positive Correlation Inflates the Variance
Suppose a pathway has \(m=20\) gene-level statistics, each with variance one, and suppose their average pairwise correlation is \(\bar{\rho}=0.10\). Under the equal-correlation approximation, the variance inflation factor is:
\[ 1+(m-1)\bar{\rho}=1+19(0.10)=2.90. \]
The independence standard error for their mean would be \(1/\sqrt{20}=0.2236\). After allowing for correlation, it is:
\[ SE=\sqrt{\frac{1+(m-1)\bar{\rho}}{m}} =\sqrt{\frac{2.90}{20}} =\sqrt{0.145} =0.3808. \]
If the observed mean score is \(0.80\), its adjusted standardized statistic is \(0.80/0.3808\approx2.101\). Using a standard normal reference distribution gives an approximate two-sided p-value of \(2[1-\Phi(2.101)]\approx0.0357\). By contrast, using the independence standard error gives \(0.80/0.2236\approx3.578\), with a much smaller two-sided normal p-value, about \(0.00035\). In this hypothetical calculation, ignoring positive correlation makes the evidence look substantially stronger than the correlation-adjusted calculation does.
Worked Example: Correcting a Competitive-Test Standard Error
Suppose a hypothetical competitive test summarizes a pathway’s advantage over its reference genes as an observed score of \(0.80\). The independence-based standard error is \(0.5000\). There are \(m=10\) pathway genes, and the estimated average correlation is \(0.2074\). The variance inflation factor is:
\[ 1+(10-1)(0.2074)=2.8666. \]
The adjusted standard error is \(0.5000\sqrt{2.8666}\approx0.8466\). The adjusted statistic is therefore \(0.80/0.8466\approx0.9450\), giving an approximate two-sided normal p-value of \(2[1-\Phi(0.9450)]\approx0.345\). At a conventional significance level of \(0.05\), this result is not statistically significant. The p-value is the probability, under the specified competitive null and reference approximation, of a statistic at least this extreme; it is not the probability that the null hypothesis is true.
The estimate indicates that the observed pathway score is not large relative to its correlation-adjusted uncertainty. It does not show that the pathway has no biological role, and it does not test whether all pathway genes have zero effects.
Worked Example: Negative Correlation in a Small Set
For illustration, suppose \(m=5\) standardized statistics have common correlation \(\rho=-0.10\), which is mathematically admissible because an equal-correlation matrix of size five requires \(\rho\geq-1/(5-1)=-0.25\). The variance inflation factor is \(1+4(-0.10)=0.60\). The adjusted standard error of the mean is:
\[ SE=\sqrt{\frac{0.60}{5}}=\sqrt{0.12}=0.3464. \]
If the observed mean is \(0.40\), the standardized statistic is \(0.40/0.3464\approx1.155\), and its approximate two-sided normal p-value is \(2[1-\Phi(1.155)]\approx0.248\). Here negative correlation reduces the mean’s variance relative to the independence calculation. The example illustrates the mathematics; in real data, pairwise correlations vary, estimates are uncertain, and a simple equal-correlation model may be inadequate.
How CAMERA-Style Adjustment Works
CAMERA (Correlation Adjusted MEan-RAnk gene set test) is a competitive gene-set test that accounts for intergene correlation when assessing a set-level statistic. In broad terms, it compares the gene-level evidence in a set with evidence in the tested-gene universe and adjusts the variance of the set statistic using an estimate of within-set correlation. The practical consequence is often summarized by the variance inflation factor \(1+(m-1)\bar{\rho}\): positive correlation increases uncertainty, reducing the magnitude of the standardized statistic and often increasing its p-value.
For expression data analyzed with a design matrix, gene-level effects are evaluated after accounting for the modeled design. Correlation estimation should likewise reflect the model: residual expression patterns after fitting the design are more relevant to shared variation not explained by the modeled effects than raw expression correlations are. A CAMERA-style result is therefore tied to the specified design, tested-gene universe, gene-level statistic, and correlation estimate.
Define the phenotype contrast, the competitive gene set, and the tested-gene universe. The competitive null concerns relative evidence in the set versus outside it.
Fit the chosen gene-level model and form statistics with a consistent direction, so that a larger value has the intended interpretation.
Estimate average correlation among the set’s genes in a way compatible with the fitted design, recognizing that this estimate is itself uncertain.
Inflate the variance to reflect dependence, calculate the test statistic and p-value using the procedure’s reference distribution, and report the set, universe, contrast, and adjustment.
The equal-correlation derivation justifies why an adjustment is needed: covariances add to the variance of a sum. In an actual analysis, pairwise correlations need not be equal, and the estimation and reference-distribution details depend on the implemented method and model. The simple variance inflation factor is an informative description of the dependence correction, not a substitute for specifying how the software calculated the test.
Assumptions, Diagnostics, and Limitations
A variance adjustment is meaningful only relative to the model and statistic used. The gene-level statistics should have comparable scales for averaging to be interpretable; their signs must correspond to the same contrast direction; and the gene set and tested-gene universe must be defined consistently. The standard-error calculation also relies on an appropriate estimate of correlation and a suitable reference distribution for the adjusted statistic. The equal-correlation formula assumes common variances and a common pairwise correlation; real pathways can depart from both assumptions.
Useful checks include examining whether the design adequately captures important sample structure, whether the gene-level statistics are formed for the intended contrast, and whether the estimated within-set correlation is plausible and stable. A very small or very large estimated correlation can have a large effect when the set is large, because the multiplier includes \(m-1\). Compare the correlation-adjusted result with the unadjusted result as a diagnostic of sensitivity, not as a choice between the “right” and “wrong” p-value after seeing which is more favorable. Also check whether a few unusual genes dominate the pathway score; variance adjustment does not make a fragile summary robust to influential observations.
Correlation adjustment has limitations. It cannot repair confounding, incorrect gene-level models, poor gene-set definitions, or a mismatched tested-gene universe. A single average correlation compresses a complex dependence structure; negative and positive correlations can coexist, and their contributions may not be captured well by one number. Correlation estimates can also be noisy, particularly for small sample sizes or small gene sets. Finally, a significant competitive result is evidence of relative enrichment under the specified test, not proof that every member is active or that coordinated expression is caused by pathway membership.
The main alternative is an independence-based competitive test, which can be reasonable when dependence is negligible but may be anti-conservative under positive correlation. Other competitive procedures may use different statistics or reference distributions; their assumptions and uncertainty calculations should be checked rather than presumed equivalent to CAMERA. Self-contained tests ask a different null question, so changing to one is not merely another way to correct the same competitive test.
Common Mistake
Check Your Understanding
Answer each question, then check the solutions below.
- For \(m=12\) genes with average correlation \(0.05\), calculate the variance inflation factor for their mean.
- In the equal-correlation model, what happens to the variance inflation factor as \(\rho\) increases above zero, with \(m\) fixed?
- A pathway has a positive adjusted competitive result. Does this establish that every gene in the pathway is differentially expressed? Explain.
- Why are residual correlations after fitting the design generally more relevant to correlation adjustment than raw expression correlations?
Solutions. (1) \(1+(12-1)(0.05)=1+0.55=1.55\). The mean’s variance is 1.55 times its independence variance. (2) The factor \(1+(m-1)\rho\) increases linearly with \(\rho\), so the standard error increases by the square root of that factor. (3) No. A competitive result indicates stronger evidence for the set relative to the tested-gene reference under the specified test; it does not imply that every member has an effect. (4) Raw correlations can reflect the modeled phenotype or covariates as well as shared residual variation. Correlations after fitting the design more directly represent dependence not explained by the modeled effects, which is the dependence relevant to the uncertainty calculation.