I am looking at some RNA count data where the count data is modelled as a negative binomial distribution. The problem however is that the logp of the model is -inf. I read an earlier thread where another user had the same issue, and the problem were nan values in the input. However, in my case this is not the problem. Essentially, this line is enough to reproduce the problem:
Y = pm.NegativeBinomial("Y", mu=mu, alpha=alpha, observed=Y_OBS)
Regardless of mu and alpha (which can be replaced by constants), Y will have a logp of -inf. If i replace Y_OBS with Y_OBS / 100 however it works. So I guess the count data is simple too large? The max value in Y_OBS however is only 80 000, which to me doesn’t seem like it is too big?
Anyway, is there a method for “normalizing” the count data in my case or apply any other kind of transformation? I guess simply dividing the count data by a constant won’t work since Y_OBS is supposed to be discrete.