

Zeroinflated mediator: choosing prop... 

Message/Author 

Maxim K posted on Friday, November 14, 2014  2:02 am



Hello, My task is to estimate the impact of job change and its characteristics (e.g. voluntariness) on wellbeing. Wellbeing is measured at two time points, T1 and T2. Some respondents have changed jobs between T1 and T2, others have not. For those who have, additional information, such as job change voluntariness is available. It can be hypothesized that both job change itself and the degree of its voluntariness have an effect on wellbeing. I wonder how to configure a model for this research question. If it were a mere multiple regression of the wellbeing score at T2, I'd use a dummy for having changed jobs, and fill in a reference value (e.g. zero) of voluntariness for nonchangers (as described here: http://stats.stackexchange.com/questions/56306/timespentinanactivityasanindependentvariable), using both variables as predictors. The difficulty is that both the change dummy and the voluntariness variable are influenced themselves by wellbeing at T1. People change jobs where they are not satisfied, after all. What would be the best strategy to model this situation? Many thanks in advance for you help. Kind regards, Maxim 


You can make the binary job change variable a DV in the model, influenced by previous wellbeing and influencing later wellbeing. Also handling the voluntariness might be more tricky. A question for SEMNET perhaps. 

Maxim K posted on Saturday, November 15, 2014  9:37 am



Thank you for the tip, Dr. Muthen. 

Noah Emery posted on Thursday, October 22, 2015  4:58 pm



I have a complex SEM model where series of latent variables are predicting two zinb outcomes at two time points. We have a series of hypotheses regarding relationships between the zinb outcomes at T1 and between T1 and T2 zinb outcomes. However, we are unsure if it appropriate to use the zinb outcomes as mediators, given the produce two estimates as an outcome (i.e., logit & count) and as predictors they are treated as a single construct. Is this appropriate and if so, how can we go about conducting this analysis? 


Counts as mediators is not a welldefined case. You can specify that the variable is count in the M ON X relation, but in the Y ON M relation, it is not clear how M should be handled. You can treat M as continuous in this relation, but that doesn't seem quite right given the piling up at zero. And counts don't have a natural underlying continuous latent response variable that can be used in both relations. I guess for the Y ON M relation you can break up the counts into categories which have different probabilities and treat it as a knownclass mixture but I haven't seen that done and it will therefore take some convincing to get it through review. 

Back to top 

