2

Is there a way to perform multiple logistic regression on repeated measures data using Matlab?

I have a data set containing a daily measurement recorded from 20 participants for 60 days. I am calculating 18 features for each participant for each day. I wish to determine if there is a relationship between each of these features and a 8 binary response variables, also recorded for each participant each day.

What is the best way to do this in Matlab? My question is related to a previous question

Community
  • 1
BGreene
  • 3,045
  • 4
  • 16
  • 33
  • `glmfit`, `mnrfit` or `regress`. pure speculation but I think `regress` is the closest to your objective –  Oct 10 '12 at 14:44
  • All function do regression alright but there is no obvious way to handle repeated measures using these functions – BGreene Oct 10 '12 at 15:02
  • so you cannot "average" the 60 regressions produced? Or am I misunderstanding the question here? –  Oct 10 '12 at 15:52
  • No afraid not, I want to examine the variation between days – BGreene Oct 10 '12 at 19:38

2 Answers2

2

I will just try to give you the direction. What you are looking for is described as Repeated measures design. The most common method to deal with such data is to use Repeated measures analysis of variance (rANOVA).

There are different types of measures you can do (repeat them at one day, differently distributed in time, for one patient or for few simultaneously), so be careful when choosing the package to work with.

Hope it will help.

Dmitry Laptev
  • 2,859
  • 21
  • 15
  • 1
    What package would you recommend for *logistic regression* with correlated errors? And it might be a good idea to discuss marginal (GEE) vs. conditional (GLMM) approach to the problem, and possibly how to handle multiple outcome. – chl Oct 11 '12 at 17:53
  • @chl Thanks for the key-words! I didn't know about the methods you mentioned. They seems to be [better then rANOVA in some cases](http://www.pitt.edu/~super4/33011-34001/33151-33161.ppt). If you can, that would be great if you post the short description as a new answer. – Dmitry Laptev Oct 12 '12 at 09:26
  • 1
    Unfortunately, I don't know enough of Matlab to perform such task and I would personally rely on R or Stata for that. Regarding GEE vs. GLMM, a [good summary](http://stats.stackexchange.com/a/16415/930) was provided on a related thread. – chl Oct 12 '12 at 23:01
  • I am unclear as to whether repeated measures ANOVA will work here. I have 18 features measured each day for 60 days on 20 subjects. I also have 8 binary response variables, also measured each day for each subject. A repeated measures ANOVA requires me to do 18x8 separate ANOVAs, leading to issues with multiple comparisons – BGreene Oct 16 '12 at 16:54
2

To answer my own question:

Generalized estimating equations (GEE) seem to be a suitable solution for my problem, i.e. multiple features measured longitudinally. There is a Matlab toolbox here for GEE.

In response to posters above, I did not find a Matlab repeated measures ANOVA package that satifactorily handles longitudinal data, I think repeated measures ANOVA is better suited for a relatively small number of categorical factors rather than longitudinal data.

BGreene
  • 3,045
  • 4
  • 16
  • 33