I'm using Bill Browne's MLPowSim to do some sample size estimation for a multilevel model. It creates an R program to carry out the estimation using lmer in the lme4 library. When there are predictors with more than two categories one has to modify the code generated to account for the multinomial nature of the predictor. Browne makes the following warning in his documetation based on one of his examples: "Since the probability of choosing a boys’ school is low, we may have all zeroes in the first row of the generated multinomial variable: i.e. no boys’ schools in n2 schools generated. Consequently, the whole of the third column of the design matrix for the fixed parameters, X, would then be zero. In such instances it would not be possible to estimate the parameters, and attempting to fit this model would lead to an error message in R." This was in reference to a three category predictor with probabilities of .15, .30 and .55. Browne points out that the solution is for the associated fixed parameters to be set to zero when there is an entire column of zeroes. Since his documentation is several years old, I'm wondering if the multilevel package in R will now properly set fixed effects in such cases to zero or if the problem remains?
______________________________________________ R-help@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.
