Overall fir of binomial general linear model

Hello,

I want to test a binomial model predicting binary variable with a number of physiological variables. Physiological variables are stored in the matrix ECGRClog. The dependent variable takes values 0- for participant did not detect the stimulus, and 1- participant detected the stimulus.

I used the following command: [b,dev,stats]=glmfit(ECGRClog,RC01,'binomial','logit')

It gave me number of statistics for the coefficients, but I don't know how to assess overall significance and predictive power of the model. Can I deduct it from the stats that I've received or do I need to add additional analysis? How can I get something like overall p-value or adjusted r-squared?

Many thanks for any tips

Answer 1

[Jeff Miller]

People don't normally test the overall significance of these models. Instead, they look at the significance of the individual predictors, which is reported in the vector of 'p' values in the 'stats' output. If any of the predictors are significant, then the model is.

I suppose you could use a sign test if you wanted an overall p value for the model as a whole, counting correct predictions as positive and incorrect predictions as negatives. But the null hypothesis would be that the overall model predicts no better than chance, which is a pretty low bar.

[Sylwia Kaduk]

Thank you very much for this response. What about predictive power? Is there something simmilar to the adjusted r-squared that I could use in glm?

[Jeff Miller]

I don't know about that. You could simply report the percentage of correct predictions which is conceptually similar to an unadjusted r-squared. I suppose you could get an adjusted PC with a leave-out-one technique.

[Sylwia Kaduk]

I really appreciate the advice. Thanks a lot.