Reading A GWAS Quantile-Quantile Plot

What a GWAS Quantile-Quantile Plot Shows

A GWAS quantile-quantile plot compares the observed distribution of association-test p-values with the distribution expected under a reference null model. Its value is not that it labels particular variants as associated; it gives a compact view of whether the collection of test results looks broadly compatible with that reference, and where departures occur. Reading the center and the extreme tail separately is essential: a few very small p-values can produce a striking tail departure without moving most of the plot.

For a scan of \(m\) variants, write the p-values in descending order as \(p_{(1)} \ge p_{(2)} \ge \cdots \ge p_{(m)}\). One common plotting convention assigns the \(i\)th ordered p-value the expected null quantile

The horizontal coordinate is \(x_i=-\log_{10}(q_i)\), and the vertical coordinate is \(y_i=-\log_{10}(p_{(i)})\). With this ordering, both coordinates increase toward the most extreme, smallest p-values. Some software uses slightly different plotting positions, so exact coordinates can vary a little between implementations; the broad pattern is the same.

The interpretation assumes that each null p-value is correctly calibrated: under its null hypothesis it is uniformly distributed on \([0,1]\) in the continuous case. It also assumes that the reference distribution used for the plot is appropriate for the tests being displayed. A p-value that is too small too often because of a model or calibration problem can produce a plot that departs from the diagonal, just as genuine associations can. The plot shows a pattern; it does not, by itself, establish which explanation is responsible.

Why the Reference Line Is Diagonal

Under a complete null with independent, continuous, correctly calibrated p-values, the ordered p-values have known expected ranks. To see why, let \(P_1,\ldots,P_m\) be independent Uniform\((0,1)\) random variables and let \(P_{[i]}\) denote the \(i\)th smallest. Its density is

Integrating \(u f_{P_{[i]}}(u)\) from 0 to 1 gives \(E[P_{[i]}]=i/(m+1)\). The \(i\)th largest p-value is \(P_{[m+1-i]}\), so its expected value is \((m+1-i)/(m+1)=q_i\). Thus, if observed p-values follow this reference distribution, their ranked values should generally track the expected quantiles. Applying the same transformation, \(-\log_{10}(\cdot)\), to both gives the diagonal reference line.

This argument explains the reference, not a guarantee that every point will sit on the line. Ordered values vary from scan to scan, and the ranks are not independent of one another. In GWAS, variants can also be correlated through linkage disequilibrium. That dependence affects the variability and clustering of points, even when individual null p-values are calibrated. A visual comparison with a diagonal is therefore a diagnostic, not a formal test that the whole scan is null.

Reading the Center and the Tail

The central region contains the bulk of the ranked tests. If it tracks the diagonal reasonably well, the middle of the p-value distribution is broadly consistent with the reference. This is reassuring about overall calibration, but it does not prove that every model assumption holds or that there are no associations. The center can also shift if many variants have genuine effects, if p-values are systematically miscalibrated, or if the reference assumptions do not match the data.

The extreme tail corresponds to the smallest p-values and the largest values of \(-\log_{10}(p)\), near the upper right of the plot. A pronounced upward bend or a few points well above the diagonal means that the most extreme results are stronger than the reference predicts. If only a small number of variants are strongly associated, they can separate from the diagonal at the far end while the much larger set of other variants keeps the center close to it. This is a reason to read the tail independently of the center, not to regard an extreme point as proof of a particular biological or technical explanation.

A broad upward departure, extending through much of the plot rather than being confined to a few extreme points, indicates that more of the observed tests are stronger than expected under the reference. That pattern may arise for multiple reasons, including genuine widespread association or problems with calibration. The shape and extent of a departure help describe what is happening in the distribution; they do not identify its cause.

Diagnostics, Alternatives, and Limitations

When reading a plot, first check the axes and ordering. With \(-\log_{10}(p)\) on both axes, larger values mean smaller p-values, and points above the diagonal are more extreme than expected. A reversed axis convention or a different statistic can change the visual direction, so confirm how the plot was made before interpreting it. Also check whether the displayed tests use comparable null distributions and whether the p-values are valid for their stated tests.

Look at the center and tail separately, and describe the shape before proposing an explanation. A localized tail departure can be driven by a small number of strong results. A broad displacement indicates a more widespread distributional difference. A plot that appears close to the line can still hide individual errors, and a few extreme points can be visually dominant even when most tests follow the reference.

A reference line alone does not show how much random variation is expected around it. If uncertainty bands are added, their construction matters: bands based on independent uniform p-values may not reflect dependence among variants. A calibrated simulation or other reference that accounts for the relevant dependence can provide a more suitable comparison. Even then, a band is a diagnostic for deviation from a specified reference, not a procedure that assigns a cause.

A QQ plot preserves the ranked pattern across the scan, unlike a single summary number that compresses the distribution. Conversely, it does not identify which genomic locations produce a departure; it summarizes ranks rather than genomic position. Reviewing individual association results addresses a different question. No one view replaces checking test assumptions, data quality, and the context of the analysis.

The main limitations are therefore interpretive. The plot depends on valid p-values and an appropriate reference; dependence can alter the expected scatter; and visual departures are not causal diagnoses. A careful reading states what part of the distribution departs, how broadly it departs, and what the figure cannot establish.