Interpreting a Gray-Test Result

Reading a Gray-Test Result in Context

A Gray test answers a particular question: is there evidence that the cumulative incidence function (CIF) for a specified event type differs across groups over follow-up? Its p-value summarizes evidence against a null hypothesis; it does not say how large a difference is, which group has the higher estimated incidence, or whether a difference matters in practice. Those parts of the interpretation require the group-specific CIF estimates and the study context.

The distinction matters because the test statistic is constructed to assess differences across follow-up, whereas a report often needs to describe incidence at a meaningful time point as well. A careful interpretation reports both: what the test indicates about evidence of a group difference, and what the estimated CIFs show about its observed direction and magnitude.

The null concerns the CIF for the event type named in the analysis, not all events combined. A competing event prevents the event of interest from occurring first and is part of the event process; it is not simply treated as ordinary censoring. The CIF therefore describes the probability of the event of interest in the presence of competing events.

For interpretation, assume independent individuals and censoring that is independent of the event process, at least within the groups being compared, under the conditions used for the analysis. The test also relies on its stated follow-up period, target event definition, and variance calculation. If censoring differs in ways related to event risk within groups, or follow-up is inadequate for some groups, the estimates and test may not support the intended comparison. These conditions should be checked using the study design and follow-up diagnostics, not inferred from a small p-value.

What the Statistic and P-Value Tell You

For two groups, Gray’s test can be represented by a score statistic \(U\) and its estimated variance \(V\), with \(Q=U^2/V\). Under the null and the test’s regularity conditions, \(Q\) is compared with a chi-square distribution with one degree of freedom. For \(G\) groups, the corresponding quadratic form is compared with a chi-square distribution with \(G-1\) degrees of freedom. The reported p-value is the probability, under the null model, of obtaining a test statistic at least as large as the observed statistic.

A p-value is not the probability that the null hypothesis is true, nor the probability that the observed difference occurred “by chance.” It quantifies how unusual the test statistic would be under the null and the assumptions used for the test. A small p-value is evidence against equality of the group CIFs; it does not by itself establish the direction, size, clinical importance, or cause of a difference.

Why the Test Does Not Give a Direction or Effect Size

For two groups, squaring the standardized score removes its sign. If \(U\) is positive or negative, \(Q=U^2/V\) is unchanged. Therefore, \(Q\) and its p-value alone cannot tell which group has the higher estimated CIF. For more than two groups, a significant global test indicates evidence that at least one group differs, but does not identify which groups differ or in what direction.

To describe direction, inspect the group-specific CIF estimates over time and state the time point or interval being discussed. To describe size at a particular time \(t\), calculate an absolute difference, such as \(F_k(t\mid A)-F_k(t\mid B)\), and report it in percentage points. A relative measure may also be useful if appropriate, but it answers a different question and can be unstable when estimates are small. Neither a single-time contrast nor a test p-value summarizes every feature of the curves.

Uncertainty for an estimated contrast should be reported using an appropriate confidence interval that accounts for the estimates’ sampling variability and covariance. Do not treat two separate confidence intervals as if they were a confidence interval for their difference. Gray’s test is not itself a confidence interval for an effect size.

Worked Examples Across Common Reporting Situations

A Short Justification: Why a Squared Score Has No Direction

For a two-group test, let \(U\) denote the score contrast and \(V\) its estimated variance. The test statistic is \(Q=U^2/V\). Since \(V\) is nonnegative, replacing \(U\) by \(-U\) leaves the statistic unchanged:

The chi-square reference distribution and its upper-tail p-value therefore depend on the magnitude of the standardized contrast, not its sign. This algebra is why the statistic alone cannot identify the higher-CIF group. The direction must be read from the estimated CIFs, with the time period made explicit.

Diagnostics, Alternatives, and Limits

Before interpreting a result, check that the event type, group labels, time origin, and follow-up period match the scientific question. Then inspect the group-specific CIF estimates and the numbers still under observation or otherwise contributing information over time. Very sparse late follow-up can make late estimates unstable, even when a test returns a numerical p-value. Check that censoring patterns and the assumptions used for the variance and test are defensible; substantial group-specific loss to follow-up related to event risk can undermine the comparison.

The central alternative to a test-only report is not a different p-value but a complementary description of the estimated CIFs and their uncertainty. A table or plot can show how incidence develops over follow-up; a contrast at a scientifically meaningful time can summarize an absolute difference. Each answers a distinct question. The choice of a single time point should be guided by the study question, not selected after inspecting results merely because it shows the largest difference.

Gray’s test is a test of group differences in the CIF for a specified event type. It is not a direct test of a causal effect, a measure of hazard, or a summary of clinical utility. In observational comparisons, confounding may explain group differences. In any design, statistical evidence and practical importance are separate: a small difference can yield a small p-value in a large study, while an important-looking difference can remain uncertain in a small study.

A common mistake is to write, “The p-value is 0.02, so Group A has a larger risk and the effect is large.” The correction is to separate the claims: the p-value provides evidence against the null of equal CIFs; the estimated CIFs show which group is higher at the stated time; and their contrast, with uncertainty and units, describes the observed size. A p-value alone supports none of the latter claims.