Dear all,

Currently, I'm estimating some multilevel linear probability models and I try to obtain (pseudo) R2 measures using the Snijders/Bosker R2 obtained by -mltrsq- from the ssc -mlt-. However, the R2 at both level 1 and 2 drops if I include additional variables, which is apparently often due to misspecification (according to Snijders and Bosker, 1994 themselves). I use multilevel modelling as I'm trying to do a 2-1-1 mediation, following the methodology of Baron and Kenny. I use data from 2015 and 2016, totaling approximately 300.000 observations (level 1), over 66 unique countries (level 2). In total there are 105 country-years, ranging from 2000 to 40000 observations per country/year. My dependent variables are both dummies. I have four continuous variables of interest at level 2, four controls at level 2 and eight controls at level 1. My main model is in the form of this:

Code:
xtmixed SUByy deasystar g_procedures g_time g_log_capital g_log_cost gender i.educ c.c_age##c.c_age i.GEMHHINC knowent suskill i.opport fearfail g_log_gdpc g_growth g_unemployment g_polstab i.year || country: deasystar, cov(uns)
I attached a screenshot of several models, formatted and shown with -estout-. The second and third set of models are subsamples of the first set of models. I used Stata 15.0 and Windows 10. All variables that are not binary and not categorical are grand mean centered (g_) or centered withing the cluster (hence c_), as this is recommended by Snijders and Bosker. As said, if the Snijders/Bosker R2 drops (both at level 1 and 2) and becomes negative for level 2 in the full model. As they are negative, I think it might be because of misspecification in some part of the model, but I don't know where and why. I tried several things (only using one year, decreasing the number of observations per country, using less controls at level 2), but none of this really solved it.

I hope someone can help me out on what this might be and/or where to look next. Thanks in advance!

Regards,

Arnoud

Note: This is my first post on Statalist, I consulted the guide to correctly format my question. If not: please let me know as well!

Reference:
Snijders, T. A., & Bosker, R. J. (1994). Modeled variance in two-level models. Sociological methods & research, 22(3), 342-363.