Communicating Uncertainty With Sparse Competing Causes

Why uncertainty needs careful reporting

A competing-risk analysis can produce a precise-looking percentage or hazard ratio even when only a few events occurred for the cause of interest. The number of decimal places does not tell readers how stable the estimate is, and a p-value alone does not show the range of effect sizes compatible with the data. Clear reporting therefore puts the estimate, its uncertainty, the event counts, and the analysis context together.

This is particularly important when the cause of interest is uncommon. A wide interval can include both a small effect and a clinically important one; the data may not distinguish between them. That is different from showing that the effect is negligible. The goal is not to make a sparse result sound definitive, but to describe what was estimated and how limited its precision is.

For competing-risk reports, name the estimand and its time horizon before interpreting uncertainty. For example, a cumulative incidence \(F_k(t)\) is the probability of experiencing cause \(k\) by time \(t\), in the presence of the other competing causes. A cause-specific hazard ratio and a subdistribution hazard ratio are different quantities, so the interval must be described on the scale actually estimated. The earlier material on choosing and naming the reported competing-risk estimand provides the relevant distinction.

What to report when a cause is sparse

Give the number of events for each cause, preferably alongside the number at risk or the cohort denominator and the follow-up horizon. A cause-specific event count is not the total number of observed events: events from other causes remain important because they affect the risk of experiencing the cause of interest. Also report the estimator or model, the uncertainty method, the confidence level, and the units or scale.

There is no universal event-count threshold at which every competing-risk estimate becomes unreliable. Stability depends on the number and distribution of events, the number of model parameters, censoring, covariate patterns, and the estimation method. A small number of events relative to model complexity can lead to unstable coefficients, large standard errors, sensitivity to a few observations, or convergence problems. Report the count and relevant diagnostics rather than declaring an estimate stable or unstable from a single cutoff.

How to interpret model estimates without a significance-only conclusion

For a model coefficient, report the estimate and interval on its natural interpretation scale. For example, a cause-specific hazard ratio compares cause-specific hazards among individuals still free of any event at each time, under the model and covariate conditions. Its interval describes uncertainty about that ratio, not a risk ratio or a difference in cumulative incidence. A p-value can summarize incompatibility with a specified null under the model, but it does not measure the size, clinical importance, or precision of an effect.

Absolute-risk predictions and limited precision

Adjusted absolute-risk predictions are often easier to interpret clinically than relative model coefficients, but they are not automatically precise. State the prediction horizon, the population or covariate profile to which the prediction applies, and the method used to obtain its interval. When comparing predictions, express an absolute difference in percentage points, not percent, unless a relative contrast is explicitly intended. Ensure both predictions use the same horizon and target population.

Choosing and presenting uncertainty measures

Different interval procedures make different approximations. A Wald interval uses an estimate and its standard error and is convenient, but can perform poorly with sparse data, skewed estimates, or estimates near a parameter boundary. Likelihood-based intervals can better reflect asymmetry when the likelihood is informative, while bootstrap intervals can capture some features of the estimator’s sampling distribution if the resampling scheme and sample size are appropriate. Neither approach guarantees reliable inference when the data contain very little information. State the procedure used; do not imply that a particular method cures sparsity.

A practical report should make it possible to judge both the magnitude and the support for the estimate. Include:

• the estimand, cause, comparison, and time horizon;
• the point estimate and confidence interval, with units and the interval method;
• cause-specific event counts and relevant denominators;
• key model or estimation conditions that materially affect interpretation; and
• an interpretation of the range of plausible values, including whether it allows clinically important effects in more than one direction.

Keep reported precision proportional to the information in the analysis. For a risk near 3%, reporting 3.2% may be reasonable, while reporting 3.2047% suggests a degree of precision the interval does not support. Avoid rounding interval endpoints so aggressively that they become equal to the point estimate or conceal a meaningful range.

Common pitfalls and limitations

• Equating a nonsignificant result with no effect. A p-value above a chosen threshold does not show that the null is true. Describe the estimate and interval, and say when the evidence is inconclusive.
• Reporting only the point estimate. A percentage or hazard ratio without an interval conceals how uncertain it is. Pair the estimate with its interval and the cause-specific event count.
• Using the event count as a universal quality rule. A low count raises concern, especially for a complex model, but does not alone prove invalidity. Consider model complexity, data configuration, and diagnostics.
• Calling a wide interval evidence of equivalence. Equivalence requires a suitable design and prespecified margin, with an analysis that addresses that margin. A wide interval that crosses zero or one is usually inconclusive, not proof of similarity.
• Confusing scales or estimands. A cause-specific hazard ratio is not a cumulative-incidence difference. Name the quantity being estimated before interpreting its interval.

Uncertainty intervals typically quantify sampling uncertainty conditional on the chosen estimator or model. They may not capture uncertainty from misclassified causes, omitted predictors, incorrect model structure, or choices made during analysis. Sensitivity checks can help reveal instability, but they do not remove these limitations. Where sparse causes make results fragile, say so directly and avoid turning exploratory estimates into definitive claims.