Hi everyone,

I am using meologit to fit a model to individual binary outcome data from a cluster randomized trial with three arms and three occasions of measurement. The model fits fixed effects for arm, time and their interaction, and random intercepts for clusters and individuals in clusters.

Having fitted the model with the default number of integration points (7) I experimented with increasing them (on the basis that most of the literature suggests that more is better). The fixed effects of the model were very stable with more points but incrementally moving to 30 points there was an increase in individual variance (0.9792 to 1.2292) and a decrease estimated cluster variance (12.0595 to 9.3432) leading to a notable increase in the cluster ICC (0.0600 to 0.0887). Apart from varying intp, I’m using default settings.

The thing that concerns me is that the log likelihood for the 30-point solution was somewhat poorer than for the 7-point solution (-1719.4641 vs -1714.0902): I expected a model estimated with more points to fit better or, at least, not worse! Is this a naïve expectation? Is it valid to compare LLs like this? Is there anything else I need to consider and which solution should be preferred?

Thanks for any thoughts,

Andrew