Those have to do with the underlying form of the equation. Sometimes you know a cubic functional form is unlikely. So if you have 5 time periods, while you can technically fit a cubic it will be difficult for real world data to be that curvy to justify a cubic over fewer terms.

With longer time series, it is easier to fit higher level polynomials. But that does not mean that a linear function is sufficient. In my work on crime trends and micro places, https://andrewpwheeler.wordpress.com/2016/11/15/paper-replicating-group-based-trajectory-models-of-crime-at-micro-places-in-albany-ny-published/, it was 13 years and the trajectories were clearly linear.

You can plot the original data as I have shown in this post to give a visual check. But you can also do all of the same fit statistics comparing different polynomial functions as you can for comparing more/less groups.

]]>OK, thanks so much!!

]]>I would check out Stata maybe to see if you can use the suest command like I show with the ordered logit. Not 100% sure — main thing I am not sure about is how the dispersion term is handled (but you can use it with Poisson, which should result in very similar inferences).

Another approach is to stack the data and estimate a single equation, but there again with negative binomial you have the issue of whether to let the dispersion term vary across the dependent variables or constrain them to be the same.

]]>Here is a direct link to the macro code, https://github.com/apwheele/Blog_Code/blob/master/SPSS/MacroMosaic.sps. You can uncomment out the end portion to show how to use the macro.

If you go to https://code.google.com/archive/p/andrewpwheeler-wordpress/downloads#makechanges you can download the MosaicPlot.zip file that has data as well.

]]>