This is my first post here so apologies for the rather basic question, but after days of reading I still can't find a satisfactory answer. I want to run a cross-lagged panel model on three waves of data to look at interactions and stability effects over time between drunkenness and illicit drug use in a normal early-adolescent population. Both these behaviours were measured on an ordinal count scale, e.g. Never=0, Once last year=1, and so on up to Daily=5. In the first two waves at least (N=1300), there are approximately 95% zero scores and the positive scores are bunched heavily around the lower scores, e.g. 15 x 1s, 5 x 2s, 4 x 3s, 3 x 4s and just 1 x 5-score.

Because of the zero-inflation and under-dispersion I looked at zero-inflated poisson link function in the cross-lagged panel model, but I have two concerns/questions about this: 1) would it adequately model the extreme amount of zeros? If not, what other models could I consider? 2) presuming teh ZIP-function is the most appropriate, how to interpret the two separate coefficients for each aspect of the cross-lagged model?

Or should I use an ordinal link function? Any advice on how to proceed is very welcome!


1 Answer 1


First of all the zero-inflated poisson is not a link function, it is an error distribution. Secondly, your data are not counts but ordinal scores, the numbers assigned are purely arbitrary. As a result, count models are inappropriate for this type of data.

Ordinal logistic regression is what you are looking for.

  • $\begingroup$ Thanks for your reply @Knarpie. Would the ordinal regression still work though with the very heavy skew I have ? E.g. where 3 of the ordinal categories have less than 5 observations each and over 95% of observations are in the "0" category (total sample is 1300). $\endgroup$
    – Russell
    Mar 13, 2017 at 12:30
  • $\begingroup$ That could be a problem indeed, maybe you can pool the higher categories together, e.g. into "1" and ">1"? $\endgroup$
    – Knarpie
    Mar 13, 2017 at 19:48

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.