Select the phase whose log_mu will be solved by conservation

Chooses the phase contributing the largest share of total cumulative hazard at the current theta, matching the C HAZARD SETCOE strategy.

Usage

.hzr_select_fixmu_phase(
  theta,
  time,
  status,
  phases,
  covariate_counts,
  x_list,
  weights = NULL,
  time_lower = NULL
)

Arguments

theta

Full parameter vector (internal scale).

time

Numeric vector of follow-up times.

status

Numeric event indicator.

phases

Named list of validated hzr_phase objects.

covariate_counts

Named integer vector.

x_list

Named list of per-phase design matrices.

weights

Optional numeric vector of row weights (length n). Defaults to unit weights. Applied when summing per-phase cumhaz so that selection happens on the same scale as the (weighted) observed event count.

time_lower

Optional numeric vector of counting-process entry (start) times. When supplied, phases are ranked by entry-time cumulative hazard H(stop) - H(start), the scale on which events are conserved. NULL (the default) means no truncation, i.e. H(start) = 0.

Value

Character: name of the phase to fix.

Details

When one phase dominates by an extreme margin (>10x the median contribution) it is typically due to poorly-scaled shape parameters at the starting theta rather than a genuine signal that the phase should absorb all events. Selecting such a phase as the fixmu phase traps the optimizer: CoE keeps rescaling that phase's log_mu upward while the optimizer tries to drive it to near-zero. We therefore exclude extreme outliers and select among the remaining phases. If all phases are outliers (or only one phase exists) the largest is used as a fallback.