Dear Jan, just some thoughts on your problem. The general location model distinguishes between categorical variables W and continuous variables Z. It is not clear to me how you specified your model. Perhaps the rocket phenomenon might be caused by sparseness in the data, e.g. if you use age in years as a categorical variable. I'm not that familiar with the mix software, so I can be of little help there.
I would take the following approach to solve your imputation problem. In the case where there is missing data in only one variable (e.g. astma), iteration is not necessary at all. I would fit a logistic regression model to predict astma given age, gender and response for all cases that are completely observed. Next assuming MAR, I would use that model to multiply impute the missing astma scores given the three predictors. Some care needs to be taken to include the appropriate amount of noise, but the basic tricks are known (see ch 5 of Rubin 1987). If two variables have missing scores (say astma and smoking), a similar strategy can be followed. Specify two models A and S: 1) Astma given all others, and 2) Smoking given all others. Presumably, both model A and S are logistic regression models. There is now a new problem however: the predictors in models A and S are incomplete, i.e. in model A the score for smoking is unknown for some cases, and in model S the score for astma is incomplete. The idea of now impute some starting values for A and S, then fit the imputation model A, multiply impute missing Astma scores, fit imputation model S using the last imputed values, impute missing smoking scores, refit model A using the last imputed values, re-impute missing astma scores, and so on. Eventually, such a sequence convergences to a stationary distribution specified by the conditionals. In my experience, five to ten such iterations are often adequate, especially if the number of missing values is not large. This approach is implemented in S-Plus software called MICE, which can be downloaded from www.multiple-imputation.com. Considering your mishap, perhaps you could give this a try. Lots of succes, Stef van Buuren S. van Buuren PhD Department of Statistics TNO Prevention and Health P.O. Box 2215 2301 CE Leiden ********************************************************************** This email and any files transmitted with it are confidential and intended solely for the use of the individual or entity to whom they are addressed. If you have received this email in error please notify the system manager. This footnote also confirms that this email message has been swept by MIMEsweeper for the presence of computer viruses. **********************************************************************
