Read NONMEM modelfit results, with bug fixes for SIR covariance models

Wraps pharmr::read_modelfit_results() with patches for Pharmpy bugs that arise when a SIR covariance step is used ($COVARIANCE SIRSAMPLE=...). NONMEM writes a second ext-file table (method "Importance Sampling of Variance-Covariance (SIR)") that Pharmpy mishandles in three ways:

Usage

read_modelfit_results(path, ...)

Arguments

path

Path to NONMEM model file (e.g. "run1/run.mod").

...

Additional arguments passed to pharmr::read_modelfit_results().

Value

A Pharmpy ModelfitResults object (same as pharmr::read_modelfit_results()).

Details

Bug 1 (_parse_parameter_estimates — NaN final estimates): The SIR table contains ITERATION 0 (FOCE starting point) and iterations 1…N (SIR samples). Because the SIR modal OFV at -1000000000 differs from the last sample OFV, _get_iter_df appends a NaN sentinel row. This NaN row is the last row of the combined pe DataFrame, so pe.iloc[-1].isnull().all() is True and final_pe is set to all-NaN instead of the FOCE estimates.

Bug 2 (_parse_table_numbers — wrong covstatus): Including the SIR table in table_numbers makes covstatus_table_number point to the SIR table (number 2). The SIR block in the lst-file has "Elapsed estimation time" but no "Elapsed covariance time", so covariance_step_ok is None and the .cov/.cor files are never read, even though the standard covariance step ran successfully under table 1.

Bug 3 (_parse_standard_errors — cov_abort with missing -1000000005): If the SIR ext table has row -1000000001 (SEs) but not -1000000005 (omega/sigma SEs in stdcorr form), the original code overrides the valid SEs with NaN and sets cov_abort = True, blocking the .cov file read.

The fixes skip the SIR table in _parse_table_numbers, _parse_ofv, and _parse_parameter_estimates (so only the FOCE table is used for estimates and covariance-step lookup), while _parse_standard_errors still uses the last (SIR) table to extract SIR-based standard errors.