Distinguishing Polygenic Signal from Test-Statistic Inflation

Why Inflation Has More Than One Explanation

A GWAS QQ plot compares observed association statistics with the distribution expected if all tested variants followed their null models. Upward departure can reflect widespread real genetic effects: if many variants have small effects, many test statistics may be modestly larger than their null expectations. But population structure, relatedness handled inadequately, batch effects, phenotype problems or other model misspecification can also shift statistics upward. These causes can coexist.

The distinction matters because a plot that departs from the null is not, by itself, evidence that the model is wrong, nor is a plausible genetic explanation proof that the model is adequate. Lambda summarizes one feature of the distribution; it does not identify the cause. Interpretation requires considering the whole QQ plot, sample size, study design, model checks and whether the pattern behaves as expected across relevant analyses.

The reference is appropriate only insofar as the tested statistics are being compared with the intended null distribution. For a single-variant test with one degree of freedom, a common large-sample null approximation is \(\chi^2_1\). That approximation relies on a suitable association model and standard-error calculation; for example, the relevant covariates, phenotype scale and dependence among participants must be handled adequately. Tests with different degrees of freedom require their own null references.

How Widespread Genetic Effects Shift the Distribution

Under a null hypothesis of no association, a one-degree-of-freedom statistic \(T\) is approximately central chi-square: \(T\sim\chi^2_1\). Under an alternative, a Wald statistic based on a standardized estimate is approximately \(Z\sim N(\delta,1)\), so its square has a noncentral chi-square distribution with one degree of freedom and noncentrality parameter \(\delta^2\). Here \(\delta\) is the effect measured in standard-error units.

This gives a short justification for why genuine effects can move statistics upward. Since \(E(Z^2)=\operatorname{Var}(Z)+[E(Z)]^2\), under this alternative,

If many tested variants have nonzero effects, the genome-wide collection is not a collection of pure null statistics. Even when individual effects are small, their contribution can shift many ranks of the QQ plot. Linkage disequilibrium means nearby variants can carry overlapping association signal; it also means the tests are dependent. The QQ plot can still describe the marginal distribution of the statistics, but its points should not be treated as independent observations.

For a fixed effect and a well-behaved association model, increasing the sample size usually reduces the standard error and increases \(|\delta|\). In a simple setting with standardized genotype and residual variance \(\sigma^2\), the noncentrality is approximately \(N\beta^2/\sigma^2\), where \(N\) is sample size and \(\beta\) is the effect on the modeled phenotype scale. Thus genuine polygenic signal can produce greater inflation in a larger study, even if the underlying effects have not changed.

That pattern is suggestive, not decisive. A systematic bias in an estimated effect can also yield a test statistic that grows with sample size: a persistent bias becomes more statistically visible as its standard error shrinks. Sample-size comparisons help interpret a distribution, but cannot by themselves distinguish genetic signal from confounding.

What Lambda Can and Cannot Tell You

Lambda is useful as a compact summary of the middle of the distribution. It is not a complete description of the QQ plot: two studies can have the same lambda but different tails, or similar upper tails but different medians. A median statistic can also be affected by widespread alternatives, and its value may be close to 1 when genuine effects are too weak or sparse to move the median detectably.

The plot adds information by showing where the departures occur. A broad, gradual shift over many quantiles is consistent with many small contributions, although it is not specific to polygenicity. A largely well-calibrated center with a sharp upper-tail departure is more consistent with a smaller set of stronger associations than with uniform inflation, but it does not rule out localized artifacts or model problems. Shape is evidence to weigh, not a unique signature of a cause.

A value above 1 is not automatically a defect. Conversely, a value near 1 does not guarantee that every important assumption holds or that the tail is free of bias. The statistic is a diagnostic summary rather than a pass/fail threshold, and it should be interpreted alongside the association model and data-generation process.

Checks That Help Separate Explanations

A useful assessment combines several kinds of evidence rather than searching for a single definitive diagnostic. Start with the study design and fitted model: consider ancestry structure, relatedness, recruitment, genotyping or sequencing batches, phenotype measurement, missingness, and the covariates used. Ask whether each issue could create association between genotype and phenotype beyond the biological effects of interest.

Then examine how the distribution changes under defensible sensitivity analyses. For example, compare results after reviewing ancestry adjustment or batch covariates, while recognizing that adding covariates can change precision and may remove real signal correlated with those covariates. A substantial reduction in broad inflation after addressing a plausible technical or structural problem raises concern that the original pattern was at least partly confounded. It does not quantify the contribution without further analysis.

• Compare the whole distribution. Record lambda and inspect central and upper quantiles rather than relying on the most extreme variants alone.
• Review design and model choices. Check whether ancestry, relatedness, batch, phenotype definition and analysis assumptions are addressed appropriately.
• Use sample-size comparisons carefully. Increasing inflation in larger comparable studies is consistent with accumulated polygenic signal, but persistent bias can also scale with precision.
• Check sensitivity. Determine whether plausible changes to covariates or analysis subsets alter the broad shift, the tail, or both.
• Look for coherent evidence. A genetic interpretation is more credible when the pattern fits the study design and relevant effects are not driven by an obvious technical or structural artifact.

These checks do not make the causes mutually exclusive. A study can contain real polygenic association and residual confounding at the same time. The practical aim is to determine whether the observed distribution is adequately explained by the model and whether conclusions depend on questionable assumptions.

Common Mistake: Treating Lambda as a Cause Detector

A common error is to say, “Lambda is greater than 1, so population stratification caused the inflation,” or the reverse: “The study is large and polygenic, so the inflation must be real.” Neither conclusion follows from lambda. The statistic records the ratio of two medians; different data-generating processes can produce the same ratio.

The correction is to state what the diagnostic establishes and what remains uncertain. For example: “The median statistic is 1.24 times its one-degree-of-freedom null reference. The broad QQ-plot shift is compatible with polygenic signal, but its reduction after ancestry and batch sensitivity checks suggests that residual structure may also have contributed.” This distinguishes an observed pattern from a causal explanation.