0
$\begingroup$

I am using lmer for estimating effect sizes and I am comparing the results with estimations obtained through OLS. I first ran the LMM model with simulated data where the effect sizes are known and I get in both cases correct estimations(lmm model estimation is the same as OLS estimations). However when I run some real data through the lmm model, I get for certain effects different answers between lmm and OLS.

Both techniques use the same designmatrices, yet i fail to understand why I get different estimations between the two methods, is this normal?

Also, when I use the commands ranef on the lmm model, I obtain my random effects how do I verify that these are BLUPS?

Improve this question
$\endgroup$
6
  • $\begingroup$ Could you possibly have some oddness in the distribution of your actual data? Or perhaps the LMM estimation had an issue with convergence? The ranef function will return BLUPs as explained further here: stats.stackexchange.com/questions/111702/… $\endgroup$
    – kurtosis
    Commented Aug 25, 2020 at 0:47
  • $\begingroup$ No not really, I also scale and center the data in both situations. After that I plotted density plots of the variables and I can't reall see anything weird $\endgroup$
    – Hedayat
    Commented Aug 25, 2020 at 5:18
  • $\begingroup$ What is the random effect structure? Maybe that has an effect? (I'm grasping a bit since I could see a difference in a GLMM but not an LMM, but... maybe more information will ring a bell.) $\endgroup$
    – kurtosis
    Commented Aug 25, 2020 at 6:08
  • $\begingroup$ I am not sure if understand what you are asking, do you want to know how the distribution of the random effects look like? I do get the following warning when I run the lmm model "fixed-effect model matrix is rank deficient so dropping 1 column / coefficient" $\endgroup$
    – Hedayat
    Commented Aug 25, 2020 at 8:17
  • $\begingroup$ That error is probably the source of your problems. $\endgroup$
    – kurtosis
    Commented Aug 25, 2020 at 18:17

1 Answer 1

Reset to default
0
$\begingroup$

I went back to my simulated data and added on top of the created effects some noise(e ~N(0,1)). After that I calculated my estimators in both ways (OLS and via LMM) the estimators are no longer identical for the interaction and random effects. So the cause of the error is the noise.

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.