Overfitting using SEM?

Hey, thanks for a great first day so far! I was wondering: If we try out many different models (ACE, AE, ADE, fixing/freeing parameters, adding covariates), do we ever run into overfitting to our data? If so, how do we deal with it?

Hi Charlotte,
Interesting question! Here’s my two cents, other faculty might weigh in with alternative views! :-)
My own view is that we should always fit the full model (whether that be ACE or ADE) and not drop parameters! I would estimate confidence intervals on each of the parameters and then report and explain my results in that light. My reasoning is that by dropping a parameter (even if it is not “significant”) e.g. fitting an AE model rather than an ACE model, you are biasing your estimates of A and C (i.e. any variance from C will get shunted into the A estimate).
With respect to the inclusion of the covariates, I think that this should be driven by theory and your brain! e.g. if your trait affected by sex and age then it makes sense to include these. It’s worth thinking more carefully about other variables that might be added, particularly heritable covariates, as you may end up throwing the proverbial baby out with the bath water. Hope this is clear and just my opinion!

1 Like

Hi Dave,
thank you for your thoughts on this!
Appreciate it :slight_smile: