Dear all,
I doubt the way to determine effect sizes in a ZINB GENMOD model. We have analyzed several dependent variables with zero frequencies ranging from 37 to 71%. In an initial step, we tested Poisson or negative binomial distributions. However, only the zero-inflated negative binomial (ZINB) distribution met the no over-dispersion criterion tested by Scaled Pearson’s X2, being accordingly used. We were simultaneously testing a three-level factor and also one two-level fixed factor, and its interaction. We added these terms to both parts of the model (continuous and zero-inflated binomial). We then performed a backward stepwise procedure to simplify the model as much as possible (we had a low sample size). We now aim to calculate effect sizes. For pairwise comparisons in the continuous part of the model, we have used lsmeans with tukey adjust. We could, perhaps, calculate Cohen's d from lsmeans +- se of each factor level. Is this correct? On the other hand, in the zero-inflated part of the model, any procedure would surely be more difficult. Since the zero-inflated part is a sort of binomial distribution, perhaps the effect sizes can be inferred from odd ratios, but I am not sure what script to use. Any help is welcome. Thanks in advance. Carlos
.
