Hello everyone,
I need to pick your brains! I am running a mixed effects regressions logit model (glmer) on some binomial data, where the dependent variable is coded as either 0 (failure) or 1 (success). I have two variables A and B, each with 2 levels. I have applied a contr.sum(2) coding to both variables because I want the output to be interpreted as an ANOVA - so the intercept should represent the grand mean (in logits) and the main effects should represent the deviations from the grand mean (correct me if I am wrong). To double check that I am using the correct contrast coding, along with glmer models, I run lrm models on the same data - because the lrm model does not have random effects, the intercept it estimates should correspond to the actual grand mean, am I right? I have 768 successes and 300 failures, so total number of answers is 1068, which should make a grand mean in logits of .94. When I run the lrm model: lrm(answer~A*B), I do NOT get the grand mean of. 94 in the intercept, but something much higher : 1.23. When I run lrm models with only one predictor, I get an intercept of .94 (the grand mean) when I include only B, and one of 1.23 when I include only A (the same intercept when I include both predictors and their interaction). I do not understand why. Below I give the data breakdown by condition - the numerator gives the number of successes and the denominator the total number of answers in the condition A level 1 A level 2 B level 1 135/263 250/272 B level 2 139/266 244/267 First I thought that it might have to do with the fact that the data set is not fully balanced. However, I get the same thing when I use new variables A and B that have been 'scaled as numeric' (I believe this scaling addresses the imbalance). Here is the distribution of the number of observations across condition: A level 1 A level 2 B level 1 263 272 B level 2 266 267 Does anyone have any clue as to why this is happening? I need to understand what I am doing, because later I need to analyse (actually re-analyse) some more complex data. I have only started to use contrasts recently (partly as a result of reading Parsimonius mixed models by Bates et al) and I have found that using (appropriate) contrasts makes a big difference to the ease of convergence of a model, at least on my data. Thanks! Maria Nella Carminati
