Troubleshooting

Every subject's NLL shows up as exactly 1e7

Symptom: FitResult.NLL (or the value returned by a fit_func call) is suspiciously uniform at 1e7 for some or all subjects, and the EM fit doesn't seem to be learning anything.

Cause: 1e7 is the capped fallback value used whenever a fit function's objective is non-finite (e.g. NaN/inf from a log of zero, or a parameter falling outside its natural bounds — most built-in fit_funcs explicitly return 1e7 when a transformed parameter like beta or alpha is outside its valid range). This is a deliberate cap (see calc_fval in pyem.utils.math) that keeps the optimizer from crashing on invalid parameter proposals, not a real likelihood value.

Fix: Check that your param_xform list matches the bounds your fit_func actually enforces (e.g. norm2beta/norm2alpha for a (0, 20)/(0, 1) model), that your data doesn't contain values (choices, rewards) the fit function doesn't expect, and that your starting Gaussian-space values aren't so extreme that every transformed parameter falls outside bounds on the first EM iteration.

compute_lme() returns NaN or seems to drop subjects

Symptom: result.compute_lme() (or the per-subject Laplace approximation) has NaN/zero entries, or the reported subject count in diagnostics is lower than expected.

Cause: The Laplace approximation used by calc_LME requires a positive-definite Hessian (inv_h) for each subject. When np.linalg.slogdet reports a non-positive Hessian for a subject (common for subjects whose fit sits at a bound or is otherwise poorly identified), that subject's contribution is treated as unreliable: it's marked as not "good" and its value is replaced by the mean of the well-behaved subjects (or 0.0 if no subject has a valid Hessian) rather than propagating a NaN into the summed LME.

Fix: Inspect which subjects have a non-positive-definite Hessian (a large fraction usually indicates weakly identified parameters or too little data per subject) and consider a stronger/narrower prior, fewer free parameters, or compute_integrated_bic() as a Hessian-free alternative for model comparison.

ValueError: mstep_maxit must be >= 1

Symptom: Calling model.fit(mstep_maxit=0) (or any value less than 1) raises ValueError: mstep_maxit must be >= 1.

Cause: mstep_maxit controls the number of EM iterations and is validated up front; the EM loop cannot run with fewer than one iteration.

Fix: Pass mstep_maxit >= 1 (the default is 200). If you're trying to do a quick single-pass fit for debugging, use mstep_maxit=1 rather than 0.