propensity score matching

[Rochelle Zhang]

Dear statalisters,

I plan to use psmatch2 for propensity score matching. In step 1, I run a logit model to get pscore

my matching variables include size, sic, and year

below is a pretend data (much fewer observation than real data which is 205,000 firm year observations)


set matsize 800
xi: logit treat size i.sic2 i.fyear

input treat size sic2 fyear
1 100 38 1990
1 120 38 1991
0 114 38 1990
0 110 36 1991
0 108 38 1990
0 138 36 1991
0 214 38 1990
0 210 36 1991
0 308 38 1990
0 391 36 1991

end

sic2 is 2 digit industry classification

my question: is my approach correct ?

log output:

stata13 log shows
i.sic2 _Isic2_1-89 (naturally coded; _Isic2_1 omitted)
i.fyear _Ifyear_1986-2013 (naturally coded; _Ifyear_1986 omitted)

note: _Isic2_84 != 0 predicts success perfectly
_Isic2_84 dropped and 346 obs not used

note: _Ifyear_1987 != 0 predicts failure perfectly
_Ifyear_1987 dropped and 3957 obs not used

note: _Ifyear_1988 != 0 predicts failure perfectly
_Ifyear_1988 dropped and 3890 obs not used

note: _Ifyear_1989 != 0 predicts failure perfectly
_Ifyear_1989 dropped and 3781 obs not used

note: _Ifyear_1990 != 0 predicts failure perfectly
_Ifyear_1990 dropped and 3782 obs not used

note: _Ifyear_1991 != 0 predicts failure perfectly
_Ifyear_1991 dropped and 3844 obs not used

note: _Ifyear_1992 != 0 predicts failure perfectly
_Ifyear_1992 dropped and 4027 obs not used

note: _Ifyear_2013 != 0 predicts failure perfectly
_Ifyear_2013 dropped and 7 obs not used

Iteration 0: log likelihood = -138441.15
Iteration 1: log likelihood = -69367.458
Iteration 2: log likelihood = -64832.816
Iteration 3: log likelihood = -59084.301
Iteration 4: log likelihood = -57394.527
Iteration 5: log likelihood = -56590.074
Iteration 6: log likelihood = -56558.129
Iteration 7: log likelihood = -55412.541 (not concave)
Iteration 8: log likelihood = -55412.541 (not concave)
Iteration 9: log likelihood = -55412.541 (not concave)
Iteration 10: log likelihood = -55412.541 (not concave)
Iteration 11: log likelihood = -55412.541 (not concave)
Iteration 12: log likelihood = -55412.541 (not concave)
Iteration 13: log likelihood = -55412.541 (not concave)
Iteration 14: log likelihood = -55412.541 (not concave)
Iteration 15: log likelihood = -55412.541 (not concave)
Iteration 16: log likelihood = -55412.541 (not concave)
Iteration 17: log likelihood = -55412.541 (not concave)
Iteration 18: log likelihood = -55412.541 (not concave)
Iteration 19: log likelihood = -55412.541 (not concave)
Iteration 20: log likelihood = -55412.541 (not concave)
Iteration 21: log likelihood = -55412.541 (not concave)
Iteration 22: log likelihood = -55412.541 (not concave)
Iteration 23: log likelihood = -55412.541 (not concave)
Iteration 24: log likelihood = -55412.541 (not concave)
Iteration 25: log likelihood = -55412.541 (not concave)
Iteration 26: log likelihood = -55412.541 (not concave)
Iteration 27: log likelihood = -55412.541 (not concave)


I do not know if it will converge, if not what is my options?

thanks,
Rochelle

[Melissa Garrido]

Hi Rochelle,
To create a valid propensity score, you cannot include any covariates that predict treatment status perfectly. The goal of a propensity score is to balance observed covariates across the treatment and comparison groups; you want the groups as comparable as possible in order to isolate the treatment effect. If you have a covariate value that always indicates inclusion in one of the groups, it cannot be balanced across the two groups. Before including any variables in your propensity score, it is often worthwhile to do a crosstab of the treatment variable with each covariate to look for cases of perfect prediction.
Hope this helps,
Melissa

[Rochelle Zhang]

Thank you, I will do a crosstab.

Best,
Rochelle

[Rochelle Zhang]

Thank you, I will do a crosstab.

Best,
Rochelle