Instability With Rare Competing Causes

Why Rare Causes Make Absolute-Risk Prediction Difficult

A competing cause can be uncommon in a study population yet still matter for an individual decision. The challenge is that a model has limited information about a cause observed only a few times. Its estimated absolute risk may then change substantially with small changes to the data, the covariates included, or the way participants are grouped. A small point prediction is not, by itself, evidence that the risk has been estimated precisely.

For cause \(k\), the cumulative incidence function (CIF) at time \(t\), written \(F_k(t\mid x)\), is the probability of experiencing cause \(k\) by \(t\), before any competing event, for covariate profile \(x\). In a continuous-time setting with cause-specific cumulative hazards \(\Lambda_j(t\mid x)\), its standard representation is

Thus, absolute risk for one cause depends both on its own event process and on the competing causes that determine who remains event-free. A small number of events of cause \(k\) means little direct information about how that cause varies across covariate profiles or over time. Estimating the other causes well does not automatically supply that missing information.

Event Counts and a Basic Precision Argument

A useful first check is the number of target-cause events by the prediction horizon, not just the total sample size. If a cohort is large but only a handful of participants experience the cause of interest, the information for that cause remains sparse. The relevant count can be smaller still within a subgroup or among participants with particular covariate patterns.

A simple calculation shows why relative uncertainty grows as the event count falls. Suppose, as a rough approximation, that a rare cause's count \(D\) follows a Poisson distribution with mean \(\lambda\), under comparable follow-up and a stable underlying event rate. The standard deviation of \(D\) is \(\sqrt{\lambda}\), so the relative standard deviation is

This is not a general formula for the standard error of a competing-risk prediction. It is a justification for the practical warning: with very few events, the count itself has substantial relative variation. A change of one or two events can therefore represent a large proportional change in the evidence about the target cause. Prediction uncertainty also reflects estimation of hazards, covariate effects, baseline risk, and the competing causes.

Why Model Complexity Can Amplify Instability

A prediction model divides the available information among its estimated quantities. If the target cause occurs only a few times, adding separate effects for many covariates, interactions, time patterns, or finely divided subgroups can leave very little information for each part of the model. Estimates may become sensitive to individual observations; some estimates may be extreme or poorly determined. Consequently, predictions for particular profiles can vary markedly even if the model's overall average risk appears plausible.

There is no universal events-per-parameter threshold that guarantees reliable absolute-risk prediction. The difficulty depends on the number and distribution of events, censoring, covariate structure, functional form, prediction horizon, and the intended use. Counting events per fitted coefficient is a useful warning signal, not a proof that a model is valid or invalid. In particular, the total number of events can conceal a near absence of events in an important subgroup.

Model complexity also affects interpretation. If a flexible model produces different predictions after adding or removing a small number of events, those differences should not automatically be described as meaningful heterogeneity in risk. They may instead reflect weak information about how risk varies across profiles.

Assessing Prediction Uncertainty and Sensitivity

An absolute-risk prediction should be considered together with its uncertainty. Intervals quantify sampling uncertainty under the fitted procedure and its assumptions; they do not guarantee that the model form is correct. When the target cause is rare, intervals can be wide, asymmetric, or close to a boundary. A narrow interval from a procedure that fails to reflect important estimation uncertainty can be falsely reassuring.

The earlier material on uncertainty in absolute-risk predictions and bootstrap intervals provides ways to quantify uncertainty. For rare causes, the key interpretive question is whether the estimated uncertainty reflects the limited target-cause information and the full prediction process. For example, resampling should preserve the original sampling unit and repeat the relevant estimation steps; otherwise the resulting distribution may not capture the variability of the reported prediction. A bootstrap that repeatedly encounters no events in a subgroup can itself reveal how fragile a subgroup-specific estimate is.

Useful diagnostics include:

• Tabulate target-cause events by prediction horizon and by clinically relevant subgroup, alongside the number at risk and follow-up support.
• Check whether a small number of observations have large influence on the fitted prediction, such as by comparing estimates after deleting or perturbing one event at a time.
• Compare predictions under modest, defensible simplifications of the model. Large changes are evidence of sensitivity, not a reason to select whichever estimate is preferred.
• Inspect interval width and shape, and check whether uncertainty is especially large for profiles with sparse data support.
• Check probability coherence and monotonicity as appropriate, while recognizing that satisfying these checks does not establish precision or calibration.

These checks answer different questions. Probability coherence checks whether predictions obey probability constraints; they do not show that a rare-cause estimate is stable. Likewise, a smooth risk curve can be produced by a model even when few events support its shape.

What to Do—and What Not to Conclude

The appropriate response to sparse cause-specific events is not to hide uncertainty or automatically abandon prediction. It is to match the complexity and claims to the information available. A more parsimonious specification may reduce sensitivity, but it cannot create event information that the study did not collect. Pooling groups or imposing stronger modeling assumptions may yield more stable estimates, but those assumptions should be explicit and the resulting predictions should still be checked.

Alternative modeling strategies can represent competing causes and covariate effects differently. For this tutorial's purpose, the central comparison is between a highly detailed model and a more constrained one: the detailed model can express more variation but may be poorly supported when events are rare; the constrained model can be more stable but may smooth over genuine differences. Neither strategy is automatically preferable. Compare the predictions and uncertainty under plausible choices, and explain what assumptions the choice makes.

Several limitations matter. First, intervals usually describe uncertainty conditional on assumptions such as the model form and the observed follow-up process; they may not capture misspecification, data errors, or changes in the target population. Second, prediction at a horizon with little follow-up support raises a distinct problem that cannot be solved merely by having a reasonable total event count. Third, rare events can make calibration assessment imprecise: agreement between a few observed events and predicted risks is weak evidence of reliable performance. Finally, a model's risk association is not by itself a causal effect.

Common Mistake

Wrong reasoning: “The predicted risk is only 0.5%, so it must be negligible and accurately estimated.”

Correction: A point estimate describes the fitted model's prediction, not its precision or practical importance. With few target-cause events, the uncertainty may be large relative to 0.5%; the plausible range and the decision context matter. Similarly, a large total sample does not guarantee a stable estimate if only a few participants experience the cause of interest.