Interval-censored log-likelihood contribution — .hzr_logl

The single place the interval-censored contribution is written. Both .hzr_logl_multiphase() and the finite-difference closure inside .hzr_gradient_multiphase() delegate here, so the optimizer cannot step by the gradient of a different objective than the one it evaluates.

Usage

.hzr_logl_interval(
  cumhaz_lower,
  cumhaz_upper,
  lower,
  upper,
  weights,
  objective = c("likelihood", "sas")
)

Arguments

cumhaz_lower: Cumulative hazard at the interval lower bounds, \(\Lambda(l)\).
cumhaz_upper: Cumulative hazard at the interval upper bounds, \(\Lambda(u)\).
lower: Interval lower bounds. Used only by objective = "sas".
upper: Interval upper bounds. Used only by objective = "sas".
weights: Case weights. In a SAS parity run these are the ICENSOR variable, which is a weight (a death count in an aggregated study), not merely an indicator.
objective: "likelihood" for the interval probability – the default, and the only statistically consistent form – or "sas" for the interval-mean-hazard density term PROC HAZARD accumulates. See inst/dev/SAS-INTERVAL-OBJECTIVE-DESIGN.md.

Value

Scalar summed contribution; -Inf for infeasible parameters.

Details

Callers pass only the interval rows – already subset by status == 2 – so this helper never sees the status mask and cannot disagree with a caller about which rows are intervals.