bug in normalization of the truncated normal?

[richarddmorey]

I believe the current version of the code in trunc_normal_lpdf (in src/zcurveEM.cpp) is incorrect in that in order to compute normalize the log density, it subtracts the difference in log probabilities instead of taking the log of the difference in probabilities. This will lead to an incorrect normalization factor.

See difference here: 05cd624

[richarddmorey]

Another issue here seems to be that a truncated normal is used, but what is needed is a truncated folded normal, right? The function seems to be missing the (small probability) other tail because z statistics are assumed to be asolute values. I can fix this and submit another pull request if you like.

[FBartos]

That is not actually correct. The model fitting dispatches to control$type == 1 in zcurve_EM.R. This calls the .zcurve_EM_fit_fast_RCpp function which relies on zdist_lpdf via compute_u_log_lik function.

The trunc_normal_lpdf is accessed only if you use the developmental settings for estimating both mu and weights (I never got to implement the EM for truncated folded -- it only had the truncated distribution with your correctly pointed out error in standardization)

[richarddmorey]

Ah, I see that zdist_lpdf has the correct folded code.