Difference-in-differences: regress or teffects?

[Paolo Giovanetti]

Dear Statalist users,
I have been working on trying to find whether there has been a positive impact on the voting share of Democratic candidates for governor following the expansion of Medicaid in 2014 using the diff-in-diff technique. My control group is made up of two states that did not adopt the expansion, Florida and Wisconsin, while the treatment group consists of Colorado and Minnesota. The elections I consider were held in 2006 and 2010 (pre-treatment) and in 2014 and 2018 (post-treatment: the expansion took effect in January of 2014). My data is at the county level.

My question is whether I should be considering the results obtained using the command regress or teffects.

For clarity: the "expansion" variable stands for whether a state implemented the expansion or not, while "post" stands for whether the year is after 2014 or not. So expansion#post (in the first regression) and postexp (in the second) are my diff-in-diff terms.

reg voteshare_dem expansion#post insured_perc unemployment_rate Poverty_rate Median_income Black_perc Hisp_perc Na

> tive_perc Asian_perc Pop_dens nationalgains_dem previous_pres_dem [aweight=pop_total], cl(state)

(sum of wgt is 117,873,853)




Linear regression                               Number of obs     =      1,160

                                                F(2, 3)           =          .

                                                Prob > F          =          .

                                                R-squared         =     0.8780

                                                Root MSE          =     4.1274




                                       (Std. Err. adjusted for 4 clusters in state)

-----------------------------------------------------------------------------------

                  |               Robust

    voteshare_dem |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]

------------------+----------------------------------------------------------------

   expansion#post |

             0 1  |   3.077134   1.164534     2.64   0.077    -.6289329      6.7832

             1 0  |   2.362998   4.430529     0.53   0.631    -11.73692    16.46292

             1 1  |   8.679507   2.622158     3.31   0.045     .3346303    17.02438

                  |

     insured_perc |  -.2027824   .1108363    -1.83   0.165    -.5555131    .1499482

unemployment_rate |    .748344    .520117     1.44   0.246    -.9069005    2.403588

     Poverty_rate |   .2958827   .1199389     2.47   0.090    -.0858166    .6775819

    Median_income |   .0000794   9.91e-06     8.01   0.004     .0000478    .0001109

       Black_perc |   .0356887   .0800048     0.45   0.686    -.2189222    .2902997

        Hisp_perc |  -.0426906   .0055727    -7.66   0.005    -.0604254   -.0249559

      Native_perc |  -.0281068   .0396139    -0.71   0.529    -.1541759    .0979623

       Asian_perc |  -.0856813   .1318648    -0.65   0.562    -.5053339    .3339712

         Pop_dens |    -.00026   .0003716    -0.70   0.535    -.0014427    .0009228

nationalgains_dem |   8.594712   2.096184     4.10   0.026      1.92372     15.2657

previous_pres_dem |   .8987743   .0491679    18.28   0.000     .7423002    1.055248

            _cons |   2.827052   7.991739     0.35   0.747    -22.60623    28.26033

-----------------------------------------------------------------------------------

teffects ra (voteshare_dem insured_perc unemployment_rate Poverty_rate Median_income Black_perc Hisp_perc Native_p
> erc Asian_perc Pop_dens nationalgains_dem previous_pres_dem) (postexp), vce(cluster state)

Iteration 0:   EE criterion =  2.316e-25  
Iteration 1:   EE criterion =  2.817e-29  

Treatment-effects estimation                    Number of obs     =      1,160
Estimator      : regression adjustment
Outcome model  : linear
Treatment model: none
                                       (Std. Err. adjusted for 4 clusters in state)
-----------------------------------------------------------------------------------
                  |               Robust
    voteshare_dem |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
------------------+----------------------------------------------------------------
ATE               |
          postexp |
        (1 vs 0)  |   3.582105   1.131778     3.17   0.002     1.363861     5.80035
------------------+----------------------------------------------------------------
POmean            |
          postexp |
               0  |   42.52256   1.627895    26.12   0.000     39.33195    45.71318
-----------------------------------------------------------------------------------

As you can see, the results show a positive effect in both cases, however, I am not sure about which of the two estimation techniques is more suited to my kind of research.

Thank you very much in advance

[Mead Over]

Not sure if you still ned a response here. But for the two approaches to be at least somewhat comparable, the interaction term in your regression model should have two hash marks instead of only one. That is, assuming that the variables "expansion" and "post" are both dummy variables, you should replace the term "expansion#post" in that command with "i.expansion##i.post.

You might also want to try the command -diff.ado- published in the SJ here:
Villa JM. Diff: Simplifying the Estimation of Difference-in-differences Treatment Effects. The Stata Journal. 2016;16(1):52-71. doi:10.1177/1536867X1601600108