RunasSudo/hpstat
1
Fork 0
Code Issues Pull Requests Releases Activity

turnbull: Gracefully handle the case where ICM does not increase the log-likelihood

Browse Source
This commit is contained in:
RunasSudo 2023-12-24 22:43:10 +11:00
parent 606c1636e0
commit c67965478d
Signed by: RunasSudo
GPG Key ID: 7234E476BF21C61A
1 changed files with 7 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
src/turnbull.rs
View File

@ -513,11 +513,10 @@ fn do_icm_step(data: &TurnbullData, p: &Vec<f64>, s: &Vec<f64>, ll_tolerance: f6
step_size_exponent += 1;
// FIXME: Nasty hack!
// Limit of 10 should be good enough, but sometimes ICM step does not work well when constrained so the limit is increased
// It would be greatly preferable to simply handle this case more gracefully when constrained
if step_size_exponent > 100 {
panic!("ICM fails to increase log-likelihood");
if step_size_exponent > 10 {
// If ICM does not increase log-likelihood, then simply return the original estimate and retry EM again
// If neither increases the log-likelihood, then we will have have converged within fit_turnbull_estimator
return (p.clone(), s.clone(), ll_model);
}
}
}
@ -639,6 +638,9 @@ fn survival_prob_likelihood_ratio_ci(data: &TurnbullData, progress_bar: Progress
ci_bound_upper = ci_estimate;
}
// FIXME: Sometimes this does not converge within max_iterations - investigate why this is
// If it is not an issue with the algorithm, then we should also terminate if width of interval is narrower than a specified tolerance
ci_estimate = (ci_bound_lower + ci_bound_upper) / 2.0;
iteration += 1;