mnrfit : how to include interaction terms ?

Hello,

I would like to compute a logistic regression on accuracy data (accuracy is 0 or 1).

There are 3 predictors (X1, X2, X3) of accuracy, but i would like to include the interaction terms between those predictors (X1*X2,..., X1*X2*X3) in the model.

How can I achieve this ? I can't figure out how to set up B or parameter values to achieve this. Any ideas ? Should I switch to another function (for example, compute log-likelihood ratios separately then run regstats/regress or something else on them) ? Thanks a lot.

[dpb]

I've not used it, but the doc seems to indicate that all you needs must do is use the 'interactions', 'on' optional input and it handles the interaction terms internally automagically. I'd at least give that a go w/ your data and see if it doesn't satisfy first...

[laurie]

that is what i thought at first but actually the doc seems to be saying that the "interactions" parameter refers to the interaction between outcomes and coefficients, not to the interaction between predictors. :-/

[laurie]

up

[laurie]

similarly, how to set this up so that one predictor is taken in as a random effect ? :-/

Answer 1

[the cyclist]

I think you may need to use nlinfit() or nlmefit() to do what you need.

[laurie]

thanks, i'll give it a try... it looks awfully complicated though, i really feel out of my depth here :-(

[the cyclist]

I have posted some simple example of nlinfit in this forum. Search on "nlinfit" and "cyclist" to find them. I don't have much experience with nlmefit, but you are right that it is somewhat complicated (partly because of its flexibility). There are examples, of course, in the documentation.

[laurie]

thanks ! just found the example you wrote on the forum.

this might be a stupid idea, but what if i was to compute the log likelihood of a correct answer for each predictor combination for each subject, then run a simple stepwise regression ? would that be appropriate ? i gather the only non linear part in logistic regression is in WHAT is fitted (that has to be transformed using the logit formula), not in how it is fitted (the relationship between the predictors and the logit data is actually linear) right ?