I'm looking to compare linear regression models from different data subsets in r. The models are not nested. I have a model from the complete dataset with sex as a factor and then 2 separate models from 2 subsets of this data (subset1 = all male data, subset2 = all female data). I am trying to see if it is better to have one model with sex as a factor or if there is benefit having a separate model for each gender. I am looking at the MAE (mean absolute error) and RSME (root mean square error) specifically.

Can I compare the raw values or the MAE and RSME (e.g. MAE from the complete model on the complete set vs female on the female set and male on the male set)?

  • $\begingroup$ The difference is discussed in several of our regression + modeling threads (but a quick search doesn't immediately turn them up, alas). You need to estimate almost twice as many parameters unless your original (full) model interacts every variable with sex. Using two separate models allows for different error variances by sex. Shall we presume "RSME" is intended to be "RMSE" (root mean square error)? $\endgroup$ – whuber Apr 16 at 20:35
  • $\begingroup$ I had tried searching but couldn't find anything sorry so I decided to post, and yes I meant root mean square error, I'll edit that in $\endgroup$ – Moogle Apr 16 at 20:37
  • 1
    $\begingroup$ @whuber: One search which works use the keywords flag separ* $\endgroup$ – kjetil b halvorsen Apr 17 at 13:16