I am having a multiple categorical dependent variable and continuous independent variables. Via a multinomial logistic regression in SPSS I want to test whether the training sample makes good estimations for my validation sample. However, I do not understand how this can be tested in SPSS. I split the data into two parts (70%-30%), but I do not understand how to test the data output of the 70% sample on the 30% sample. I also tried a Bootstrap in SPSS but I cannot find exactly how this should be done in SPSS.

  1. Generate a random 70% of your data using Data > Select Cases > Random Sample of Cases. This will create a variable named FILTER_\$ indicating which cases are selected

  2. Run your mlogistic model specifying on the Save dialog to export the model to XML.

  3. Compute FILTER_\$ = 1-FILTER_\$.

  4. Using Utilities > Scoring Wizard, select the model file exported and other specifications.

You will then have the predicted value or other output statistics from the model as you select in step 4. You can then run whatever summary statistics you want on that sample or turn off the filter and use Data > Split File to generate the same statistics for both samples.

As for bootstrap, if you have the bootstrap option, you can specify bootstrapping on the logistic dialog.

  • $\begingroup$ Thank you for your answer. I have collected the predicted value and other output statistics, but when I want to test the output on the other 30% of the data, what do I have to use as DV and IV? Are my DV still the same, but do I have to use the predicted values as IV now to get a confusion table with predicted and observed classifications? $\endgroup$ – user151973 Mar 7 '17 at 9:43
  • $\begingroup$ It depends on what sort of measure you want. If you saved the predicted categories, you can just run a crosstab of predicted against actual categories using the holdout sample - or use split files and get both the within and holdout results at once. $\endgroup$ – JKP Mar 8 '17 at 14:34

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.