Spencer: (warning: highly biased, personal opinions)
My $.02 > Looking now at your output, I notice that "Corr" between > "(Intercept)" and "trust.cz1" for the "Random Effects" "commid" is > 1.000. This says that the structure of your data are not adequate to > allow you to distinguish between random effects for "(Intercept)" and > "trust.cz1" for "commid", while simultaneously estimating all > the fixed > effects you have in the model. Quite right. Design is the cause; overfitting/identifiability is the symptom. > > If I were you, I'd start be deleting all the terms > from the model > that don't have a "Signif. code" beside it in the table of "Fixed > effects" and then refit the smaller model, preferably also using > 'method="AGQ"'. Well, this might work, but it's also a prescription for overfitting a highly biased model. What he really needs to do is carefully rethink. What is a parsimonious model given the data at hand? Unfortunately, this is far from a trivial issue. Model choice for nonlinear model fitting is conceptually and statistically difficult. IMHO, the tendency to overfit mechanistically motivated models with insufficient, poorly designed data is a ubiquitous scientific practice, rarely understood by scientists (due to the complexity). As a result, there are a lot of questionable results published in peer-reviewed literature. Eventually it gets sorted out, but it can take a while. See Kuhn and Feyerabend, for example. Always enjoy your comments. Keep 'em coming. -- Bert ______________________________________________ [email protected] mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html
