Hello ! I’m trying to model a process using a zero-inflated exponential. I’ve written up a log-probability function for it, and giving it a test-drive in theano, I believe it gives me what I’m expecting. However, sampling with PyMC3 returns strange results – the zero-inflation rate always comes out…

The likelihood was indeed wrong. The second term in the return of the zero_inflated_exponential function should be tt.log(p), not tt.log(p + lam - (lam * p)).

The zero-inflated exponential

Questions

junpenglao January 17, 2019, 7:45am 2

I would imagine that this is pretty difficult to estimate as exponential distribution has large likelihood around 0 to begin with. I would try with much stronger prior for p.

Topic		Replies	Views
Zero Inflated Poisson Log-Lik Questions	3	1767	August 2, 2019
Zero-Inflated Poisson Example Questions	10	3367	October 28, 2019
Literature on implementing a Zero-Inflated Beta likelihood Questions	19	2338	January 18, 2019
Custom discrete distribution for zero truncated Poisson Questions	10	2608	December 19, 2017
pm.DensityDist issue (pymc3 vs pymc5) v5 bug , modeling	4	679	May 22, 2023

The zero-inflated exponential

Related topics