Research ToolsTime-dependent ROC

Time-dependent ROC / Time-dependent AUC(t)

⚗️ Frontier method: time-dependent AUC uses the IPCW (inverse-probability-of-censoring weighting, Uno 2007) cumulative-case/dynamic-control definition. For formal analysis, cross-check with the R packages timeROC/survivalROC; this tool does not yet output a CI.

A plain ROC has no time dimension. For a prognostic marker/model, discrimination changes over follow-up. Time-dependent AUC(t) measures "at time t, can we separate those who have had the event from those who have not", handling censoring with IPCW. Computed locally in your browser; data are not uploaded.

① Input survival data

Each row: time status marker (status 1=event, 0=censored). Space/Tab/comma-separated.

Evaluation time points (comma-separated)

How to use & methodology

How does time-dependent AUC differ from ordinary AUC?

Ordinary ROC/AUC targets a fixed binary outcome with no time dimension. In prognosis research, 'whether the event occurred' depends on which time point you look at, and there is censoring. Time-dependent AUC(t) evaluates discrimination separately at each time point and corrects censoring bias with IPCW.

Why does the AUC change over time?

A marker may discriminate high vs low risk well early on but lose discrimination later (or the reverse). The time-dependent AUC(t) curve shows this change in discrimination over follow-up.

How are cases/controls defined?

This tool uses the 'cumulative/dynamic' definition: at time t, those with event time ≤t and an event are cases, those with time >t are controls; those censored before t are handled indirectly via IPCW weighting.

Why is there no confidence interval?

The CI for time-dependent AUC needs the bootstrap or asymptotic variance and is sensitive to the censoring estimate, making it error-prone. To be safe this tool gives only point estimates; for CIs and more robust estimates, use R's timeROC/survivalROC packages.