Thank you for the further input. I tried running my original code with NOBOUND added as the final argument of the PROC GLIMMIX statement, which returned the same message as before, i.e. "The initial estimates did not yield a valid objective function." The variable trigger has 61 levels, while CompLemma has 345 levels. Both follow an unbalanced (Zipfian) distribution, i.e. a small number of high-frequency levels are responsible for the lion's share of the dataset while the rest contribute much less, with many levels having single-digit sample sizes. This is linguistic data, where such skewed distributions are unfortunately the norm. Both random effects are lexical items occurring at key syntactic positions of the grammatical phenomenon under scrutiny. The model includes categorical fixed-effect parameters for shared characteristics of triggers and CompLemmas -- for example, there's a categorical fixed effect for whether the trigger is a verb, noun, adjective or conjunction, and there's a numeric fixed effect representing the number of syllables in CompLemma. In both cases, however, there's every reason to suspect that the lexical identity of the trigger/CompLemma may exert an effect over and above its fixed-effect characteristics. I am averse to introduce dummy variables for individual levels of trigger or CompLemma because there are a number of research questions regarding these variables that ideally require all levels of both variables to be treated the same way rather than estimating some levels as random effects (undergoing shrinkage) and others as fixed effects (avoiding shrinkage). For example, one question of interest is the overall magnitude of inter-level variance with trigger and CompLemma, and this cannot be properly estimated unless each random effect includes all of its categories. When you talk about 'consolidating' levels, do you mean collecting low-frequency levels within the random effects into trashcan categories? Indeed, some scholars in my field do follow such a practice. I was hoping to show them up by avoiding such an information-lossy procedure though. 🙂
... View more