STATS MAKE ME CRY

Dear Jeremy,

I have learned a lot from your website and the answers to many of the questions asked here, so thank you for that.

My question is regarding the following (all in SPSS 23)
After multiple imputation, I ran different regression analyses. After running a backward regression however, I noticed the following message for my pooled coefficient results:
"For at least one model, pooled estimates could not be computed because model parameters vary by imputation."

Comparing the results from the original and the pooled results, I can see that in the original, various steps are taken and variables are removed when p > 0.1. In the pooled results, however, these steps aren't taken, as they are different between the various imputations.

How should I proceed from this? What should I report from these results? Or should I run an additional analysis with just the variables that stayed in the regression in the original data, in order to obtain more reliable coefficients (pooled) for these variables?

Hope you can help! Thanks in advance!

Frederick, Amsterdam, The Netherlands

My post disappeared after an error, but in short:
I am planning to use the following method, S2, from Wood et al 2008:
2.4.3. Separate imputations (S1, S2, and S3). Rather than choosing one of the imputed data setsfor model selection, better use may be made of the data by performing model selection separatelyin each imputed data set. This approach will typically result in models with different selectedpredictors. We propose three strategies to provide a single selected model across the imputed datasets:
S1: select predictors that appear in any model
;S2: select predictors that appear in at least half of the models;
S3: select predictors that appear in all models.

Would love to hear some feedback on my choice!

Sincerely,

Frederick