PS matching

[Moomal Khan]

Hi everyone,
I am facing some troubles in understanding which STATA command is best at estimating propensity score matching
Firstly, when i applied psmatch2 command, results shows insignificant relationship.

psmatch2 t_anc $xlist, out( healthy)

Probit regression                               Number of obs     =      1,888
                                                LR chi2(3)        =       9.47
                                                Prob > chi2       =     0.0237
Log likelihood = -200.40082                     Pseudo R2         =     0.0231

------------------------------------------------------------------------------
       t_anc |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
         AGE |   .0052712   .0122269     0.43   0.666    -.0186931    .0292355
        EDU |  -.0149966   .0427118    -0.35   0.726    -.0987102     .068717
        b_od |   .0955766   .0341968     2.79   0.005      .028552    .1626012
       _cons |    1.68904   .3680223     4.59   0.000     .9677291     2.41035
------------------------------------------------------------------------------
----------------------------------------------------------------------------------------
        Variable     Sample |    Treated     Controls   Difference         S.E.   T-stat
----------------------------+-----------------------------------------------------------
         healthy  Unmatched | .725745257   .534883721   .190861537   .069055841     2.76
                        ATT | .725745257    .47100271   .254742547    .10443035     2.44
----------------------------+-----------------------------------------------------------
Note: S.E. does not take into account that the propensity score is estimated.

           | psmatch2:
 psmatch2: |   Common
 Treatment |  support
assignment | On suppor |     Total
-----------+-----------+----------
 Untreated |        43 |        43
   Treated |     1,845 |     1,845
-----------+-----------+----------
     Total |     1,888 |     1,888

However when i use following teffects command results are significant.

 teffects psmatch ( healthy ) ( t_anc $xlist , logit), atet

Treatment-effects estimation                   Number of obs      =      1,888
Estimator      : propensity-score matching     Matches: requested =          1
Outcome model  : matching                                     min =          1
Treatment model: logit                                        max =         11
------------------------------------------------------------------------------
             |              AI Robust
     healthy |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
ATET         |
       t_anc |
   (1 vs 0)  |   .2623306    .096744     2.71   0.007     .0727158    .4519454
------------------------------------------------------------------------------

Next, when I applied kmatch on my model results are significant which contradicts with the results of psmatch2. Moreover i'm facing difficulty for how to read matched and controls results.

kmatch ps t_anc $xlist (healthy)
(computing bandwidth for ATT ... done)
(computing bandwidth for ATC ... done)

Propensity-score kernel matching                Number of obs     =      1,888
                                                Kernel            =       epan
Treatment   : t_anc = 1
Covariates  : age edu b_od
PS model    : logit (pr)

Matching statistics
------------------------------------------------------------------------------------------
           |             Matched             |             Controls            | Bandwidth
           |       Yes         No      Total |      Used     Unused      Total |          
-----------+---------------------------------+---------------------------------+----------
   Treated |      1767         78       1845 |        43          0         43 |  .0016262
 Untreated |        43          0         43 |       819       1026       1845 |  .0002598
  Combined |      1810         78       1888 |       862       1026       1888 |          
------------------------------------------------------------------------------------------

Treatment-effects estimation
------------------------------------------------------------------------------
     healthy |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
         ATE |   .2216293   .0912343     2.43   0.015     .0426985    .4005601
------------------------------------------------------------------------------

similarly, significant results are also obtained from mahalanobis, for this i have applied following command

  kmatch md t_anc $xlist (healthy), ate att atc
(computing bandwidth for ATT ... done)
(computing bandwidth for ATC ... done)

Multivariate-distance kernel matching           Number of obs     =      1,888
                                                Kernel            =       epan
Treatment   : t_anc = 1
Metric      : mahalanobis
Covariates  : age edu b_od

Matching statistics
------------------------------------------------------------------------------------------
           |             Matched             |             Controls            | Bandwidth
           |       Yes         No      Total |      Used     Unused      Total |          
-----------+---------------------------------+---------------------------------+----------
   Treated |      1788         57       1845 |        43          0         43 |  1.708896
 Untreated |        43          0         43 |       944        901       1845 |  .5696319
  Combined |      1831         57       1888 |       987        901       1888 |          
------------------------------------------------------------------------------------------

Treatment-effects estimation
------------------------------------------------------------------------------
     healthy |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
         ATE |   .2013704   .0814573     2.47   0.014     .0416145    .3611263
         ATT |   .2014745   .0816728     2.47   0.014     .0412961    .3616529
         ATC |   .1970415   .0792827     2.49   0.013     .0415505    .3525326
------------------------------------------------------------------------------

Questions:
1- Problem is why psmatch2 is giving me insignificant results however treatment effects estimation, kernel matching and mahalanobis gives significant results.
2- Among psmatch2 and kmatch commands which one should be preferred?
3- How to read kernel matching results for both matched and controls?

[David Radwin]

Questions:
1- Problem is why psmatch2 is giving me insignificant results . . .

Why do you think psmatch2 is giving insignificant results? The t-statistic for ATT (average treatment effect on the treated) is 2.44, which is statistically significant at the conventional p < .05 level and is similar to the z-scores of 2.71 reported by teffects psmatch and 2.47 reported by kmatch. If you want Abadie-Imbens adjusted standard errors in psmatch2, use the ai() option.

(psmatch2 was written by Leuven and Sianesi, and kmatch was written by Jann. Both are available from SSC.)

I'm sorry that I can't help with questions 2 or 3.