İleriye Dönük Rakip Riskler Analizi

The Kaplan-Meier estimator treats all other events as independent censoring, which assumes they carry no information about the primary event probability. When competing events exist, this assumption is violated and 1 minus KM overestimates the absolute probability of the primary event. Cumulative incidence functions from competing risks analysis correctly partition total risk and always sum to no more than 1 across event types.

The choice depends on your research question. Fine-Gray subdistribution hazard regression directly models the covariate effect on absolute cumulative incidence — it is the right choice when the goal is risk prediction or estimating the probability a patient experiences the event in practice. Cause-specific hazard models condition on being event-free and are preferred when the goal is understanding the biological mechanism generating each event type. Reporting both can provide a more complete picture.

Prospective follow-up ensures that event cause, timing, and clinical context are recorded at the point of occurrence, minimizing misclassification of competing events — a major source of bias in retrospective record review. It also allows pre-specified endpoint adjudication, standardized data collection procedures, and complete covariate measurement at baseline, all of which strengthen the validity of the resulting incidence estimates.

Participants lost to follow-up are censored at their last known contact. The key assumption is that censoring is non-informative — that the probability of loss is independent of the risk of all event types. If loss is related to disease severity or treatment toxicity, the resulting estimates may be biased. Sensitivity analyses under different missing data assumptions (e.g., multiple imputation, inverse probability weighting) should be considered when loss to follow-up is substantial.

The R packages cmprsk (Gray's test and Fine-Gray regression), survival (cause-specific Cox models), and mstate (multi-state extensions) are the primary tools. The tidycmprsk package provides tidy-friendly wrappers. Stata's stcompet and stcrreg commands, and SAS PROC LIFETEST with the CIF option, also support competing risks estimation. All major platforms can produce cumulative incidence curves with confidence intervals.