# Can we simply compare the predicted percentages of the outcome between studies?

I used the multinomial logistic regression to predict the percentages of students who voted 'acceptable', 'uncertain', and 'unacceptable' to natural ventilation use in three observed classrooms during cool and hot seasons.

The sets of significant IVs of the cool and hot season cases were different; for example, the effect of window size (i.e. small, medium, large) on the acceptance of natural ventilation was significant for cool season case but not for hot season case.

Question 1: how can I compare the percentages of 'acceptable' votes of the two seasons? Can I look at the differences in the predicted percentages between two seasons directly?

Question 2: how can I compare the percentages of 'acceptable' votes of students in the rooms with different window size?

Thank you. Please forgive me for my bad English as I am not a native English speaker.

-
Interesting topic but I wonder how well the technique you've chosen matches your research questions. It fits as a way of identifying stronger and weaker predictors of each individual's vote. But you talk mostly about the group-level results--the percentage who vote a certain way. These can be compared using a much simpler method such as a proportions test, which can yield confidence intervals for percentages and/or their differences, or a Chi-Square test. – rolando2 Aug 24 '12 at 2:38
Many thanks Rolando2. I intended to identify the main factors of students' votes on the use of natural ventilation and compare the effect of each factor. However, the pseudo-R2 values I have got from the multiple logistic regression models are very low (the models are not much better than guessing). So, the models seem to be not very useful for predicting individual's vote. Still, the models are useful for predicting students' votes as a group (if I am right). In this case, how can I compare the effect of the factors on the group's votes? – tida Aug 24 '12 at 10:26
If you want to use logistic regression, then the model for question 1 should have "season" as an independent variable, and the model for question 2 should have "size" as an independent variable. But I think @rolando2 is right - tests of proportions or chi-square seem simpler ways to answer your questions. – Peter Flom Aug 24 '12 at 10:58
Thanks Peter. I will try what you have suggested. – tida Aug 24 '12 at 17:00