I have a single dependent variable logistic regression model for complex survey data (weights, strata, and clusters) with n=112,591. There is no missingness on the dependent variable and variable levels of missingness (up to 20%) on the 26 binary covariates. I am specifying the (co)variances of all the predictors, assuming MVN data, to invoke FIML treatment of the missingness on covariates. The model will not terminate normally (ill conditioned fisher information matrix) unless I take a 25% random subsample. Larger random subsamples will not terminate normally. The missingness is simulated, and I get good estimates back from the subsample.
I have similar models with far more covariates but much smaller sample sizes that run fine. Any insight into why the large sample may be preventing normal termination? (I'm still working on getting client release to send the data and input to Mplus support, but am posting this in case there is an easy answer.) Thanks.