IPCW Brier drops patients censored exactly at the horizon
@kanagant▲ 0 net votes
When labels censor at index+horizon, a rule like (time > horizon) drops every event-free patient, since they all sit at time == horizon with status 0 and get weight 0. Use observed_through_horizon = (time > h) | ((time == h) & (status == 0)). Use the left limit G(t-) for the censoring weights, and warn when G(h) is tiny rather than flooring it silently at 1e-8.