STATA-FORUM.DE

Hi all,

For my Dissertation I am calculating a multilevel regression with data from a cross-national survey. Apart from the national differences I have got a special focus on differences between postcommunist and capitalist societies. For the latter I engage a matching-procedure additionally.
However, I am not sure if it is statistically feasible to include country-level variables in my propensity score calculation. My outcome-variable is life satisfaction.

These are my results with the country-level variables:
Variable Sample Treated Controls || Difference S.E. T-stat
stflife Unmatched 6.52477649 7.30727985 || -.782503361 .010444271 -74.92
ATT 6.5630657 6.1991459 || .363919805 .047303311 7.69
Note: S.E. does not take into account that the propensity score is estimated.

psmatch2: psmatch2: Common
Treatment support
assignment Off suppo || On suppor Total
Untreated 0 || 154,234 154,234
Treated 2,333 || 52,921 55,254
Total 2,333 || 207,155 209,488

And here without country-level variables:

Variable Sample Treated Controls || Difference S.E. T-stat
stflife Unmatched 6.42518045 7.31894856 || -.893768108 .009584001 -93.26
ATT 6.42515767 6.51730024 || -.092142568 .015272386 -6.03
Note: S.E. does not take into account that the propensity score is estimated.

psmatch2: psmatch2: Common
Treatment support
assignment Off suppo || On suppor Total
Untreated 0 || 171,802 171,802
Treated 1 || 69,132 69,133
Total 1 || 240,934 240,935
The code I used for these outputs:

Code: Alles auswählen

psmatch2 postcom fem agea agesquared income attend edu_local subhealth domicil unemployed SocialTrust actandinst regulatory supranational gdppc socspend domicil_C Gini, out(stflife) common
est store pps1
psmatch2 postcom fem agea agesquared income attend edu_local subhealth domicil unemployed SocialTrust actandinst regulatory supranational, out(stflife) common
est store pps2


Nevertheless, I guess both results are quite strong.

So eventuallly my question is if it is feasible and worth including country-level variables?? And why ?


Fritz

First of all, I strongly suggest using kmatch instead of psmatch2 since it receives updates and has many useful features. For your initial question, you can either include country dummies, use exact matching within countries or cluster standard errors by country. However, I cannot tell you which option is best and I suggest to compare different results.