SPSSX Discussion

beta estimates

Classic

List

Threaded

2 messages Options

Jill Adelson

beta estimates

Hello,

I have two fairly large data sets with lots of cases and lots of variables. I am using SPSS 17.0 with Missing Values. I used multiple imputations on my data, so now I have 11 times the number of cases (original data + 1 imputed data set). I finally was able to get a stepwise logistic regression with a lot of variables to run. However, SPSS keeps running out of working memory to process the output. I need the beta estimates for the next step of my analysis as I need to predict probabilities for half my cases (they could not experience the outcome, so it was not appropriate to use them in the original logistic regression). Unfortunately, the beta estimates are not pooled estimates as the different imputations resulted in different models at some step - meaning I need to get each beta estimate for a large number of variables for all 11 data sets. I cannot get the output to copy or to extract to Excel due to the SPSS working memory issue. Is there a way to just save the beta estimates when I run the logistic regression? Any other suggestions?

Thanks for the help!

Jill

Jill L. Adelson

Doctoral Candidate and Research Associate

Measurement, Evaluation, & Assessment and Gifted Education

Department of Educational Psychology

University of Connecticut

SPSS Support

Re: beta estimates

Hi Jill,

Both of the SPSS procedures that will perform stepwise logistic regression modeling (LOGISTIC REGRESSION, which is Binary Logistic Regression in the menus, and NOMREG, which is Multinomial Logistic Regression in the menus) have the ability to write parameter estimates (optionally with covariance matrices) to XML files, where you could at least view them and have access to the numbers as text. However, that's probably not the best option.

If you're using LOGISTIC REGRESSION, you can cut down the size of the output by specifying to print results only from the final model rather than at each step, if you're not already doing that. NOMREG only produces results for the final model. In either case, in terms of cutting down the size of the output, since you're not using pooled results, you could run analyses separately for each imputation and save the results to separate .spv files. Not optimal, but something to keep in mind.

The OMS (Output Management System) approach would let you save the estimates to a file of various types. An SPSS .sav file is often the easiest if you want to compute with results.

Since your goal is to predict probabilities for (or to "score") new cases, perhaps the best option is to let the procedure do the predictions for you. In either procedure, when you save predicted probabilities, predicted groups, etc., if there are cases in the file that are complete on the predictors but not on the dependent, these cases will not be used in the analysis, but will have predictions saved. (In LOGISTIC REGRESSION, you can also apply results from a subset of cases to the entire data set even if the cases you don't want to include in the original analysis are complete on all variables, using the SELECT subcommand. In the main Binary Logistic Regression dialog box you can specify a selection variable and rule.) Thus a direct solution to getting probabilities saved for the cases that couldn't experience the outcome event is to include them in your data but have their values for the dependent be missing, and save predicted probabilities. If these cases are not already in the imputed data, some repetitive file merging would be required, but this would likely still be less work than computing predictions from saved coefficients.

David Nichols

From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Jill Adelson
Sent: Tuesday, March 24, 2009 10:52 PM
To: [hidden email]
Subject: [SPSSX-L] beta estimates