Why do results of teffects versus psmatch2 differ so much?

[Tunga Kantarci]

I use teffects and psmatch2 to check if they produce same results. They do not. Below is the code. What would be the reason for that they produce results that are very different from each other? I checked the matched samples each package produce before they estimate the ATET, and in fact find that the observations matched are quite different across the two packages. How is it possible that the two packages use the same matching method and end up with different matched samples (to estimate the treatment effect)?

use https://www.stata-press.com/data/r16/cattaneo2
teffects nnmatch (bweight mage prenatal1 mmarried fbaby) (mbsmoke), atet nneighbor(1)
psmatch2 mbsmoke, mahalanobis(mage prenatal1 mmarried fbaby) outcome(bweight) neighbor(1) ties

[FernandoRios]

Hi Tunga,
I think the answers to your question are.
1. you are not using the same method in both.
you are using nnmatch for teffects, and Mahalanobis matching for psmatch2.
2. What do you mean large differences in effects?

teffects:
-----------------------+----------------------------------------------------------------
ATET                   |
               mbsmoke |
(smoker vs nonsmoker)  |  -232.4632    24.1189    -9.64   0.000    -279.7354   -185.1911
----------------------------------------------------------------------------------------
Psmatch:

        Variable     Sample |    Treated     Controls   Difference         S.E.   T-stat
----------------------------+-----------------------------------------------------------
         bweight  Unmatched | 3137.65972   3412.91159  -275.251871   21.4528037   -12.83
                        ATT | 3137.65972   3367.56366  -229.903935    71.942121    -3.20
----------------------------+-----------------------------------------------------------

I only see a 3 points difference in both treatment effects.

[Tunga Kantarci]

Hi Fernando, thanks for your quick response. You must be right that the methods are not the same. When I execute the code in my OP, I get the results below. I wonder how you obtain what you pasted.

. use https://www.stata-press.com/data/r16/cattaneo2
(Excerpt from Cattaneo (2010) Journal of Econometrics 155: 138-154)

. teffects nnmatch (bweight mage prenatal1 mmarried fbaby) (mbsmoke), atet nneighbor(1)

Treatment-effects estimation                   Number of obs      =      4,642
Estimator      : nearest-neighbor matching     Matches: requested =          1
Outcome model  : matching                                     min =          1
Distance metric: Mahalanobis                                  max =        139
----------------------------------------------------------------------------------------
                       |              AI Robust
               bweight |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-----------------------+----------------------------------------------------------------
ATET                   |
               mbsmoke |
(smoker vs nonsmoker)  |  -232.4632    24.1189    -9.64   0.000    -279.7354   -185.1911
----------------------------------------------------------------------------------------

. psmatch2 mbsmoke, mahalanobis(mage prenatal1 mmarried fbaby) outcome(bweight) neighbor(1) ties
Ties is not implemented for Mahalanobis matching. This option will be ignored.
----------------------------------------------------------------------------------------
        Variable     Sample |    Treated     Controls   Difference         S.E.   T-stat
----------------------------+-----------------------------------------------------------
         bweight  Unmatched | 3137.65972   3412.91159  -275.251871   21.4528037   -12.83
                        ATT | 3137.65972   3298.94329  -161.283565   66.5694099    -2.42
----------------------------+-----------------------------------------------------------
Note: S.E. does not take into account that the propensity score is estimated.