nvm

lifelines/fitters/kaplan_meier_fitter.py

@@ -125,10 +125,10 @@ def fit_interval_censoring(
         label=None,
         alpha=None,
         ci_labels=None,
-        show_progress=False,
         entry=None,
         weights=None,
-        tol=1e-7,
+        tol: float = 1e-5,
+        show_progress: bool = False,
     ) -> "KaplanMeierFitter":
         """
         Fit the model to a interval-censored dataset using non-parametric MLE. This estimator is
@@ -168,6 +168,10 @@ def fit_interval_censoring(
               if providing a weighted dataset. For example, instead
               of providing every subject as a single element of `durations` and `event_observed`, one could
               weigh subject differently.
+          tol: float, optional
+            minimum difference in log likelihood changes for iterative algorithm.
+          show_progress: bool, optional
+            display information during fitting.
 
         Returns
         -------
@@ -203,7 +207,7 @@ def fit_interval_censoring(
 
         self._label = coalesce(label, self._label, "NPMLE_estimate")
 
-        results = npmle(self.lower_bound, self.upper_bound, verbose=show_progress)
+        results = npmle(self.lower_bound, self.upper_bound, verbose=show_progress, tol=tol)
         self.survival_function_ = reconstruct_survival_function(*results, self.timeline, label=self._label).loc[self.timeline]
         self.cumulative_density_ = 1 - self.survival_function_
 

lifelines/fitters/npmle.py

@@ -100,10 +100,10 @@ def create_turnbull_lookup(turnbull_intervals, observation_intervals):
     return {o: list(s) for o, s in turnbull_lookup.items()}
 
 
-def check_convergence(p_new, p_old, turnbull_lookup, weights, tol, i, verbose=False):
+def check_convergence(p_new, p_old, tol, i, verbose=False):
     if verbose:
         print("Iteration %d: delta: %.6f" % (i, norm(p_new - p_old)))
-    if log_likelihood(p_new, turnbull_lookup, weights) - log_likelihood(p_old, turnbull_lookup, weights) < tol:
+    if norm(p_new - p_old) < tol:
         return True
     return False
 
@@ -120,7 +120,7 @@ def probs(o):
     return o / (o + 1)
 
 
-def npmle(left, right, tol=1e-10, weights=None, verbose=False, max_iter=1e5):
+def npmle(left, right, tol=1e-5, weights=None, verbose=False, max_iter=1e5):
     """
     left and right are closed intervals.
     TODO: extend this to open-closed intervals.
@@ -150,17 +150,16 @@ def npmle(left, right, tol=1e-10, weights=None, verbose=False, max_iter=1e5):
 
     while (not converged) and (i < max_iter):
         p_new = E_step_M_step(observation_intervals, p, turnbull_lookup, weights)
-        converged = check_convergence(p_new, p, turnbull_lookup, weights, tol, i, verbose=verbose)
+        converged = check_convergence(p_new, p, tol, i, verbose=verbose)
 
         # find alpha that maximizes ll using a line search
-        best_alpha, best_p, best_ll = None, None, -np.inf
+        best_p, best_ll = None, -np.inf
         delta = odds(p_new) - odds(p)
         for alpha in np.array([1.0, 1.25, 1.75, 2.5]):
             p_temp = probs(odds(p) + alpha * delta)
             ll_temp = log_likelihood(p_temp, turnbull_lookup, weights)
             if best_ll < ll_temp:
                 best_ll = ll_temp
-                best_alpha = alpha
                 best_p = p_temp
 
         p = best_p