Why Inflation Needs Interpretation
A genome-wide association scan can produce test statistics that are larger than expected under a correctly calibrated null. This pattern is called inflation. It is a warning to investigate, not a diagnosis: genuine effects at many variants can raise statistics across the genome, while confounding or model miscalibration can do the same. Both may occur together.
The distinction matters because the remedies differ. A broad shift caused by population structure, unmodeled relatedness, or technical artifacts calls for better control of those sources of dependence or bias. Widespread genuine polygenic association is biological signal, even though it can also make many p-values small. The central task is therefore not to make every statistic look null, but to assess whether the observed pattern is compatible with the analysis and its assumptions.
Three diagnostics provide complementary evidence. A quantile-quantile (QQ) plot displays the distribution of observed test statistics against its null expectation. The genomic inflation factor, \(\lambda\), compresses that comparison to one number based on the median. LD score regression (LDSC) examines how association statistics vary with linkage disequilibrium (LD) score. Each emphasizes a different feature, and none by itself proves the presence or absence of confounding.
What the Diagnostics Measure
A QQ plot makes visible what \(\lambda\) hides. It orders the observed statistics, or equivalently their p-values, and compares them with the expected order statistics under the null. If most null tests are calibrated, the central part should track the diagonal. A departure restricted mainly to the most extreme tail can be compatible with a limited number of strong associations. A systematic departure over much of the distribution indicates broad inflation, but does not reveal its cause.
LDSC adds information about whether inflation varies with LD score, a measure of how much variation a variant tags through LD with nearby variants. At a diagnostic level, a positive association between test statistics and LD score is consistent with polygenic effects contributing to inflation: variants tagging more genetic variation tend, on average, to show larger statistics. The fitted intercept summarizes a component of average inflation not explained by that LD-related trend under the method's assumptions. It can provide evidence about confounding, but it is not a direct or universally reliable measurement of confounding.
These summaries have different roles. The QQ plot shows the shape of departure across the distribution; \(\lambda\) describes its median; LDSC asks whether inflation tracks LD score. Interpretation is strongest when the statistics are calibrated under the fitted null model, variants and samples are appropriate for the analysis, and the LDSC assumptions are sufficiently plausible. In mixed-model GWAS, the model and its calibration also matter: residual structure or a poorly specified covariance model can affect all three diagnostics.
Reading Patterns Together
| Observed pattern | What it may support | Important qualification |
|---|---|---|
| QQ plot follows the diagonal centrally, with deviation mainly in the extreme tail | A smaller set of strong associations, possibly alongside calibrated null tests | Tail deviation alone does not establish that the signals are genuine. |
| Broad QQ departure and elevated \(\lambda\) | Widespread signal, broad confounding, or a combination | Neither diagnostic separates these explanations. |
| Inflation increases with LD score and the LDSC slope is positive | Polygenic contribution is consistent with the pattern | LD-score relationships and model assumptions can be imperfect. |
| Elevated LDSC intercept with little LD-score trend | Inflation not explained by the fitted LD-related trend; confounding is one possibility | The intercept is not a definitive confounding test. |
| Elevated intercept and positive slope | A mixture of non-LD-related inflation and polygenic contribution is plausible | Other model violations can also produce misleading summaries. |
These are patterns of evidence, not a decision rule. For example, polygenic effects can be so widespread that the QQ plot departs well before its extreme tail. Conversely, confounding can have a relationship with LD score and therefore contribute to an apparent trend. The diagnostics should be read alongside study design, covariate adjustment, relatedness handling, and checks of the phenotype and genotype data.
Worked Example: Broad Inflation With Little LD-Score Trend
Suppose a hypothetical scan has \(100{,}000\) approximately one-degree-of-freedom tests. The median observed chi-square statistic is \(0.546\). Its QQ plot departs upward across much of the distribution, not only at the extreme tail. An LDSC summary reports an intercept of \(1.17\) and a small slope that is not clearly distinguishable from zero.
Calculate \(\lambda\). Using the null median \(0.4549\),
The median statistic is about 20.0% above its one-degree-of-freedom null median. This establishes a central shift; it does not identify its source.
Combine the evidence. The broad QQ departure and \(\lambda=1.200\) show that inflation is not confined to a few extreme results. The relatively elevated intercept and weak LD-score trend make non-LD-related inflation a concern. Possible explanations include residual population structure, technical artifacts, or incomplete control of relatedness. Polygenic signal is not ruled out: an imprecisely estimated or distorted LD-score relationship could obscure it.
Conclusion. The pattern warrants targeted checks for confounding and calibration, rather than a claim that confounding has been proven. Investigators should check population and batch structure, verify the mixed-model specification, and confirm that the LDSC inputs and reference information are appropriate. Any remaining evidence for polygenicity should be interpreted after those checks.
Worked Example: Inflation Consistent With Polygenic Signal
In a second hypothetical scan, the median chi-square statistic is \(0.591\). The QQ plot tracks the diagonal near its center, then curves upward progressively toward the tail. The LDSC summary has an intercept of \(1.02\), with a clearly positive LD-score trend.
Calculate \(\lambda\).
A \(\lambda\) of about \(1.299\) signals an elevated median, but it does not contradict the visual impression that much of the departure is stronger in the tail. The QQ plot retains information about where the inflation occurs; the median ratio does not.
Assess the LDSC pattern. The intercept of \(1.02\) is close to 1, while the positive LD-score trend indicates that statistics tend to be larger at variants with higher LD scores. Together with the progressive tail departure, this pattern is consistent with widespread polygenic association. It is not proof: an intercept close to 1 does not guarantee that all confounding has been removed, and the interpretation depends on the LDSC assumptions and inputs.
Conclusion. The diagnostics support polygenic signal as an important contributor to the inflation. They do not justify multiplying or dividing every statistic by \(\lambda\) to “correct” it without a scientific and statistical rationale. Such a transformation could reduce genuine signal as well as inflation from unwanted sources.
Worked Example: Evidence for Both Sources
A third hypothetical scan has median chi-square \(0.614\), a QQ plot that rises above the diagonal throughout much of its range, and increasingly strong departure in the tail. Its LDSC summary reports an intercept of \(1.08\) and a positive LD-score trend.
Calculate \(\lambda\).
The median is 35.0% above the null median. The broad departure indicates widespread inflation, while the additional tail curvature is compatible with stronger signals among a subset of variants.
Interpret the joint pattern. The positive LD-score trend is consistent with a polygenic contribution. The intercept above 1 suggests a component of inflation not captured by that trend, which could reflect confounding or other departures from the fitted model. The QQ plot agrees that the departure is not restricted to a handful of associations, but cannot partition it into causes.
Conclusion. A mixture of polygenic signal and confounding is plausible. The appropriate conclusion is not that the intercept measures exactly how much confounding exists. Rather, it motivates checking for residual structure and technical problems while retaining the possibility that the LD-related component reflects genuine widespread effects.
A Diagnostic Logic and Its Limits
A disciplined interpretation starts with the scan itself. Confirm that the test statistics and degrees of freedom are appropriate for the null calibration being used. Check that tests from different variant sets or analyses are not being pooled in a way that makes their medians incomparable. For a mixed-model scan, examine whether the model accounts adequately for fixed covariates and relatedness. Then inspect QQ shape, calculate \(\lambda\), and consider the LDSC pattern as distinct but related evidence.
Confirm the phenotype, variant filters, test-statistic definition, sample, and model are appropriate for the null comparison.
State whether departure is central, broad, tail-concentrated, or some combination; avoid describing only the most significant variants.
Report the median ratio and its scope, but do not treat it as a causal label for the inflation.
Ask whether the reported trend with LD score supports a polygenic contribution and whether the intercept raises concern about additional inflation.
Investigate structure, relatedness, batch effects, model calibration, and LDSC input limitations before assigning an explanation.
Several failure modes deserve particular attention. First, a large sample can make small, widespread effects statistically visible, so a high \(\lambda\) is not automatically evidence of a flawed analysis. Second, correcting all test statistics by \(\lambda\) can suppress genuine polygenic signal. Third, a small estimated intercept is not proof of zero confounding; the diagnostic may have limited precision or its assumptions may be violated. Fourth, QQ plots can look similar under different causes, and LDSC can be affected by LD-reference mismatch, variant selection, or other departures from its assumptions.
The main alternative is to rely on one summary alone: a QQ plot for visual assessment, \(\lambda\) for a compact central measure, or LDSC for an LD-related decomposition. Each is useful for a different question, but each loses information when used alone. The combined approach is more informative because the diagnostics emphasize different features; it remains observational and cannot replace checks grounded in study design or independent validation.
Common Mistake
Check Your Understanding
Answer each question using the diagnostic distinctions developed above. Brief solutions follow the questions.
- A scan has median chi-square \(0.500\). Calculate \(\lambda\) using a null median of \(0.4549\), and state what the result does and does not tell you.
- A QQ plot is close to the diagonal centrally but departs at its most extreme end. What is one plausible interpretation, and why is it not conclusive?
- A scan has elevated \(\lambda\), a broad QQ departure, and a positive LD-score trend. Which source of inflation is supported, and what remains uncertain?
- Why is it incorrect to conclude that an LDSC intercept near 1 proves there is no confounding?
Solutions. (1) \(\lambda=0.500/0.4549=1.099\), approximately. The median statistic is elevated by about 9.9% relative to the null median; this does not reveal whether the cause is polygenicity, confounding, or both. (2) A limited set of strong associations is plausible, but tail deviation can also arise from confounding or other model problems, so the shape alone cannot establish the source. (3) The positive LD-score trend is consistent with polygenic contribution. Broad inflation and the elevated median do not exclude confounding, and the trend itself depends on model assumptions. (4) The intercept is a diagnostic summary under assumptions, not a guarantee; residual confounding can remain undetected or be obscured by imprecise estimates or violated assumptions.