ignoring rather than omitting NA covariates
On 23/01/2009, at 1:40 AM, Daniel Ezra Johnson wrote:
Dear all, This is not primarily a mixed models question, so I'll ask it in the framework of glm(). But I have the same question w/r/t glmer(). In my field, sociolinguistics, researchers have used a software tool for some thirty years that performs logistic regression assuming categorical predictors. This software is usually called VARBRUL (the current version of it is called GoldVarb). Assume a data file like this: response pred1 pred2 0 a x 1 a y 1 a x 0 a y 0 b x 1 b y 0 b x 0 b y 1 a / 0 b / My question is about the behavior of the slash (/) used in the last two lines. Assume sum contrasts. The software estimates the values of a and b (which sum to zero) and of x and y (which sum to zero). The interesting part is that for the last two data points, the predicted values are calculated on the basis of pred1 only, and pred2 is ignored. Looking at the various options of na.action, I do not see anything that would correspond to this. Basically we have NA in a certain predictor column and we want this predictor ignored for the row in question - we don't want the whole row omitted. Any way to accomplish this in R?
There are ways to deal with missing data through multiple imputation, mi package is one. There are other ways the GoldVarb package may use, possibly it builds a model for the missing data. This could be done in R it just has to be programmed. :-) Possibly they aren't actually covariates but part of multivariate responses in which case missing data is much easier to deal with. Ken