Logistic mixed-effect regression example

Hello, I was trying to make sense out of NLMEFIT help in order to fit logistic mixed-effect regression and I could not. In R syntax is straight forward. How would I do it in MATLAB? Thank you, Michael

Answer 1

[Tom Lane]

In that case nlmefit would not be suitable, because it fits models with a continuous response. The glmfit function would be suitable, but it doesn't support mixed effects so you could only use that if you were willing to treat your predictors as having fixed effects. Unfortunately there's no Statistics Toolbox function that performs mixed effects logistic regression.

[Barry Greene]

Tom, Is it possible to do something similar with ANOVAN? i.e. if I use subject id as a random effect? I have a similar problem to the poster above - I have 18 predictors (features) and 8 binary response variables (Y/N). Each response variable and feature is recorded from 20 subjects for 60 days. I can't see an obvious way to do this with ANOVAN without doing 18x8 separate ANOVAs andd then running into multiple comparison issues. I was trying nlmefit until I came acorss this answer... Any advice would be gratefully received.

Thanks

[Tom Lane]

I can't think of a good way to do what you want. The anovan function isn't suitable for binary or multivariate responses. The glmfit function and other functions aren't set up for random effects.

You can of course use glmfit with dummy variables for the subjects, treating them as fixed effects. In the latest release you can use GeneralizedLinearModel.fit with categorical predictors, and not have to create dummy variables yourself. But neither of these supports random effects. Nor do they deal with multivariate responses.

Answer 2

[Tom Lane]

You could use NLMEFIT to fit a response with normally distributed errors around a curve with a logistic shape. But there is no function in the Statistics Toolbox for fitting a mixed-effect model to a logistic regression to model the probability for a binomial response variable.

Answer 3

[Michael Yudelson]

Tom thank you for a response. I know I am being dumb, but I'm not comfortable with stats at this high level. I guess I was looking for an example, anywhere. NLMEFIT is unusable for me with current help :(

[Tom Lane]

I need to understand what kind of model you're trying to fit. Is it the usual logistic regression model where the response is a set of counts or proportions? Or is it more like continuous measured data that just happens to have a logistic shape?

[Michael Yudelson]

My response is a binary variable (success/fail) and my random predictor are a student_id and a question_id. I wanted to start with just those two and an overall bias.

Answer 4

[Michael Yudelson]

Tom, my response is a binary variable (success/fail) and my random predictor are a student_id and a question_id. I wanted to start with just those two and an overall bias.

Answer 5

[Peng Li]

You could use fitglme now to fit mixed effect logistic regression models. You can specify the distribution as Binomial and this way the Link function will be made as logit as well. Then you will be fitting a mixed effect logistic regression model (of course you need to specify random effects correctly in the formula).