I have an experimental treatment at 15 sites and would like to build models to associate usage rates of 4 species with a set of 6 explanatory variables. In the end, I would like to rank the importance of these variable for each species. Ordinarily, I would simply fit the models and then standardize the beta coefficient by dividing by the standard error and ordering by absolute value. However, in this case, I need to model a random effect of site and my response variable must be modeled as zero-inflated Poisson, which takes me outside of the realm in which I've seen this comparative analysis performed.

Is there an analogous procedure to ranking standardized beta coefficients for generalized linear mixed models?

Much thanks in advance!

Is it possible to rank standardized beta coefficients in generalized linear mixed models?

  • 1
    $\begingroup$ I think this is probably a better fit for stats.SE. The nature of the variables won't matter to them there, only the structure of the data and models. $\endgroup$
    – kmm
    Nov 8 '16 at 13:37
  • $\begingroup$ @kmm , that is a good point. I thought I would ask here first since this is a fairly common issue in ecological modelling (as judging from the conversations I've had with folks in my university cohort) that I suspect either someone has already dealt with, or others in the biological community, would benefit from. $\endgroup$
    – et is
    Nov 8 '16 at 13:48
  • $\begingroup$ I'm voting to close this question as off-topic because it belongs to stat.SE. $\endgroup$
    – Remi.b
    Nov 8 '16 at 16:16
  • $\begingroup$ If I understand you correctly, you should be able to standardize the explanatory variables (zero mean, unit sd) before the analysis (at least if the explanatory variables are continuous). Then the effect sizes can be compared between predictors that have different scales. $\endgroup$ Nov 10 '16 at 12:43
  • $\begingroup$ Also, the "importance" of variables is a bit ambiguous. A variable can have a large effect size (top-ranked coefficient), but without a fair amount of variation in that variable it might not have a large effect in practice. On the other hand, a variable with a smaller effect size but a larger variation between sites can explain more of the variance in the response. $\endgroup$ Nov 10 '16 at 12:47