This is a general question about ZIP outcomes. If I am using ZIP count outcomes (with a zero inflation) but am not really interested in modeling two separate process (inflation and count), would relying on a simple "count" specification induce bias in the model estimation and if yes, what kind of bias?
It sounds like perhaps you would prefer the negative binomial which has the inflation hidden away.
But you shouldn't have to stay away from ZIP. Just don't have a separate process for the inflation - let it be a free mean parameter. Just focus on the regression slopes for the count variable.
I would not switch to regular Poisson unless it had a better BIC.
Nick Roshon posted on Sunday, October 02, 2011 - 9:43 pm
I am interested in examining risk for alcohol problems (a count variable), in which 66% of cases have a “zero.” However, I am not really interested in examining the effects of predictors on problems (continuous) and the effects of predictors on being unable to assume any value except for zero (inflation variable) as separate processes (I don't think the effects should be different for non-drinkers versus drinkers without problems). I tried using ZIP but when I model the effects of predictors on the inflation factor, there are no significant effects on the inflation factor nor the count part. I have several questions. I should mention that my sample size is pretty small (n=170).
1.Is it OK to leave this statement (alcprob#1 on gender ptsd paralc) out but still treat alcohol problems as inflated? This analysis yields significant results in the expected directions. Count is alcprob (i) MODEL: alcprob on gender ptsd paralc;
2. Is it be correct to say that #1 models the effects of the predictors on alcohol problems while taking into account the zero-inflated nature of the outcome variable?
3. Would you recommend a negative binomial model (or ZINB) instead? The NB model yields results similar to those from from #1 and a significant dispersion parameter.
4. Why are results so different when I include a statement modeling effects of predictors on the inflation factor?
1. Yes. 2. Yes. 3. Could be a good idea. 4. Can't say.
Nick Roshon posted on Monday, October 03, 2011 - 4:53 pm
Thank you very much for your relpy. Would you recommend that I make the decision to use ZIP versus negative binomial model based on comparing the BIC? Substantively, is it correct to say that both methods take into account zero inflation?
Yes, BIC would be good for this. Ys, t negative binomial also takes into account zero inflation in the sense that it does not assume that the variance is equal to the mean. It fits data with an abundance of zeros better than Poisson.