I'm using zero-inflated negative binomial in a complex dataset (clustering within schools). If I don't change the starting values, I get a reasonable result. But if I do increase the amount of starting values, I get a result with fixed parameters in the zero-model to avoid singularity. I was also wondering which technique is used to correct the s.e. for the complex structure.
the model below has 2 zero-inflated (poisson) dependent variables. I would like to include a correlation between these 2 variables. 1. Is this correlation specified correctly? 2. Is the standard MLR the one to use? 3. should I be suprised that the other coefficients are quite different, given a high correlation between the 2 independent variables?
Thank you very much for providing these techniques. It's a really nice model I couldn't have fitted before using Mplus. Ruben.
Usevariables are prop_del viol_del delgroup att_viol s_contr mk_tot; Missing are .; categorical are delgroup; count is prop_del viol_del(i); cluster = school; analysis: type = complex; starts 100 20; model: att_viol on MK_tot ; S_contr on att_viol MK_tot; delgroup on S_contr att_viol; prop_del on MK_tot att_viol S_contr delgroup; prop_del#1 on MK_tot att_viol S_contr delgroup; viol_del on MK_tot att_viol S_contr delgroup; viol_del#1 on MK_tot att_viol S_contr delgroup; f BY prop_del viol_del; f@1;
the model is quite stable, and replicable with other start values. However, one of the coefficients is very high. The estimate for prop_del#1 on delgroup is 14.525. The oddsratio is therefore about 2million. This is very strange to report. But it does make sense that this coefficient is very high, only not that high, and the rest of the model makes sense.
Can and should I somehow restrict the size of this parameter? Or should I just report it like it is?
in the model above, is it possible to estimate indirect effects on the zero-inflated dependent variables. Can Mplus do this? And in what scale would they be? I assume that it would be an effect measured in 2 coefficients, one for the zero-part and one for the count-part?
The interpretation of the indirect effects can be drawn from the path model in terms of positive and negative effects. Which I will report like this. I was just wondering if there would be some estimation of the size of this effect, and some confidence intervals if possible.
The guiding principle for being able to produce indirect effects is that the M ON X and the Y ON M regressions are both linear. This is more general than it sounds. For instance, M can be a latent response variable for a categorical (binary or ordered) observed variable, in which case we call it M*. In for example probit regression M* ON X is then a linear regression. For this example what is required is that Y ON M is also linear. Y can then be a latent continuous response variable for a categorical outcome, a lograte for a count, or a log hazard for a survival variable. But, continuing the example with a categorical observed measure of the mediator M, the key is that Y ON M concerns the latent continuous response variable M*, not the observed categorical measurement. So for instance, with a binary variable it is not the event itself that predicts Y but the tendency for the event to happen.
In this example, Y is a count and I think you had a categorical mediator. That's a tricky combination which Mplus doesn't yet handle. Count Y requires ML which doesn't yet work with a latent response variable M for Y ON M. Bayes can do that, but can't yet do counts.