I am trying to run a zero-inflated poisson model for a count variable (cigarette use), using 35 points of assessment. I attempted to estimate a model using example 6.7 in the manual, however the model failed to estimate.
My questions are the following: 1) Since I am only estimating an intercept model (no slope), is the example 6.7 applicable for me? I deleted "s" and "si" as well as "s@0" and "si@0" in the Model statement. Is there anything else I need to adjust if I want to estimate an intercept only model?
2) I was also considering estimating this model using example 6.16 of the manual. However, it appears very similar to the model in the example 6.7. What are the differences between these two models? And which one would you recommend I use for my case (i.e., an intercept only model for a cigarette use variable, assessed weekly over a course of a semester; 35 points of assessment)?
1) The way you are doing it is correct. You might want to first try a non-inflated Poisson model. If you have problems, send your materials and license number to firstname.lastname@example.org.
2)ex 61.6 is a two-part model. This is a little different than ZIP. ZIP is a 2-class model where there are 2 classes that can produce a zero value, while two-part model is a single-class model. With two-part you would have to treat the positive number of cigarettes as continuous-lognormal; the count option should not be used because Mplus does not provide a truncated (at zero) Poisson. If you have most people at zero, you may want to use ZIP.
Qiana Brown posted on Wednesday, June 25, 2014 - 7:04 am
1. Can I use the growth model for parallel processes if my outcomes are not continuos?
I have one outcome (past month smoking) that is binary, and another outcome that is continuous. I would like to model them using the parallel process growth model, but section 6.13 in the User's guide only mentions the parallel process for continuous outcomes.
2. My binary smoking variable has several zeros at each time point. About 82% of the participants are zeros (did not smoke in the past month). Should I used a zero-inflated poisson growth model in this case, or will a growth model for binary outcomes suffice? Also, can the zero-inflated poisson growth model be modeled in the parallel processes framework?
1. Yes. A binary and a cont's process can be handled.
2. You don't want to use Poisson if you have only 2 response categories. Only if you smoking variable really consists of counts. Yes, you can combine a zIP model for one of the parallel processes with a cont's variable process.
Hello, I used a parallel processes growth model with a MLR estimator to model a binary and continuous outcome. Should the estimates be interpreted as if both outcomes were continuous? Also, are there any specific model fit indices that I should consider in this case? Would you please recommend a paper that might help me?
The analysis and interpretation of binary growth models is a large topic that we cover in Topic 3 of our short courses on our website. See the video and handout, slides 185-212. This also gives references.
See also the paper on our website under Papers, Growth Modeling:
Masyn, K., Petras, H. and Liu, W. (2013). Growth Curve Models with Categorical Outcomes. In Encyclopedia of Criminology and Criminal Justice (pp. 2013-2025). Springer.
JLuk posted on Wednesday, November 11, 2015 - 12:01 pm
I'm running a ZIP latent growth model for a count outcome, number of alcohol problems, in a 7-wave data. The proportion of zeros is about 45-50% across all waves and the BIC is lower for the ZIP vs. Poisson latent growth model. As a next step, I'm interested in looking at time-invariant (baseline) covariate effects. I regressed i, s, ii and si on the covariates. The majority of the results make sense except that I found two unexpected effects on the slope of the zero-inflated part, and so I wanted to make sure I'm interpreting the results correctly.
(1) From what I understand, a covariate with a positive coefficient on the zero-inflated intercept means that this covariate is associated with greater log odds of having zero alcohol problem at baseline. Is this correct?
(2) By extension, does a covariate with a positive coefficient on the zero-inflated slope means that this covariate is associated with greater log odds of remaining to have zero alcohol problems over time?
(3) If so, then I have results that are seemingly contradictory. Specifically, one (same) covariate is positively associated with the slope of alcohol problems in the count part (i.e., growth of alcohol problems over time), but is also positively associated with the slope of alcohol problems in the zero-inflation part (i.e., greater log odds of remaining to have zero alcohol problems over time?). Is this possible?