I'm trying to predict weight change with an intervention from baseline variables. Literature search yields suggests several predictors. Univariate analyses with weight change as dependent and baseline variables give several more predictors.

I've entered both theoretical and GLM univariate predictors into a regression model using forced entry (in blocks according to significance for the GLM univariates).

The problem I'm having is that my models vary hugely depending on how I treat missing variables. My study has 329 cases. I thought that listwise was the best method of exclusion (textbook) however this leaves me with insignificant models and only 196 cases. Pairwise and replacing values leave me with much higher cases naturally and give two models with the same significant predictors.

My issue is that the predictors remaining are ones which make sense in context with the literature. Is it ok to present these? Could somebody please tell me why pairwise and replacing with means are frowned upon?

I know there are other methods of missing value imputation but I have no idea of how to do that in SPSS, so if you're feeling generous and you know, I'd be grateful for any tips!


There's a body of literature on treating missing data. In your case it sounds as if you had the worst kind of missingness, i.e. missing not at random, where you have to somehow model the missingness. It's tough.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.