Dear All, For a logistic mixed-effects model, I am trying to visualize the fitted response on top of the original data points. What confuses me is that the fitted curve seems to run much too low as compared to the (eye-balled) mean of the data points. Could anybody advice me on why this may be happening (or simply point out what I am misunderstanding)? The syntax is very simple (as implemnted in v.9.4): proc glimmix data= data1; class Site; model Success/Trials = Explanatory /solution link=logit err=binomial; random Site; run; I then extract the intercept (a) and coefficient (b) from solutions and plot the fitted curve at the original (ilink) scale as y=exp(a+b*Explanatory)/(1+(a+b*Explanatory)) Here, I expect the resultant prediction y to apply to the mean value of the random effect, and hence to – on average – hit the middle of the original data points (expressed as the ratio Success/Trials). But as said, when visually examined, the fitted function seems to run much too low at the ilink scale. I am now wondering whether I am misunderstanding something basic about the default parameterization of the random effect in glimmix? Or is there some other reason why this should not work as outlined above? If I am completely off the mark then how would you create the plot described above, with the idea of showing the original data and the fitted function at the original scale, for immediate and intuitive interpretation by the non-expert viewer? With very many thanks for any advice you may offer, Tomas
... View more