2024-07-11

Simple approach to setting a consistent sample in Stata.

New Ph.D. students (including me when I was) tend to hoard data, using all the possible data for each model. This isn’t wrong on it’s face, but it’s also useful to establish a sample which the reader can follow through the paper with minimal deviations. This also makes your job as a writer easier as you can address the sample once and move on. I end up explaining this once to each student that I work with, so I’m writing this up here. As always, feel free to email with questions!

Consider a dataset of random numbers, some of them missing.

. // inspect the data
. list in 1/5

     +--------------------------------------------------------------------------+
     | id   year   quarter           y   x1       x2   x3         x4         x5 |
     |--------------------------------------------------------------------------|
  1. |  1   2001         1    1.38e+11    1   668659    .   1.66e+10   56843.59 |
  2. |  1   2001         2   -5.01e+11    1   668659    .   1.66e+10   64507.01 |
  3. |  1   2001         3   -2.68e+11    1   668659    .   1.66e+10    50003.9 |
  4. |  1   2001         4    2.67e+12    1   668659    .   1.66e+10   57500.11 |
  5. |  1   2002         1    3.84e+12    1   668659    .   7.18e+10   52113.82 |
     +--------------------------------------------------------------------------+

If we summarize the data we’ll see that there are 1,600 observations but three of the variables have missing data.

. // summarize
. su y x*

    Variable |        Obs        Mean    Std. dev.       Min        Max
-------------+---------------------------------------------------------
           y |      1,600   -5.95e+10    2.48e+12  -1.79e+13   1.31e+13
          x1 |      1,200           2     .816837          1          3
          x2 |      1,600    432333.3    199654.9   118209.1     668659
          x3 |      1,120    9.535714    5.987668          1         20
          x4 |        816    3.69e+10    2.74e+10   2.22e+08   1.52e+11
-------------+---------------------------------------------------------
          x5 |      1,600    53703.42    5514.862   35966.54   71861.57

Now we can fit a series of models with different data requirements and we’ll see that this generates several samples.

. // fit a series of models
. quietly:eststo:regress y x2 x4
. quietly:eststo:regress y x2 x4 i.x1
. quietly:eststo:regress y x2 x4 i.x1 x3

. // tabulate and note sample sizes
. esttab

------------------------------------------------------------
                      (1)             (2)             (3)
                        y               y               y
------------------------------------------------------------
x2               423051.8      -6999882.8      -1833948.3
                   (0.52)         (-0.66)         (-0.17)

x4                  2.853           3.389          -6.843
                   (0.89)          (0.87)         (-1.59)

x1                           -8.33482e+11    -1.88427e+11
                                  (-0.72)         (-0.16)

x3                                           -1.47517e+10
                                                  (-0.81)

_cons        -1.50561e+11     3.36708e+12     1.23826e+12
                  (-0.71)          (0.67)          (0.24)
------------------------------------------------------------
N                     780             552             436
------------------------------------------------------------
t statistics in parentheses
* p<0.05, ** p<0.01, *** p<0.001

So far so good, but what if we want to get one sample to use in the summary and model tables? The answer is in ereturn list. After an estimation command runs is saves the details in ereturn list and return list, this is where commands like esttab get the information that they tabulate. One of these details is a vector of ones and zeros that indicates which observations were included in the estimation. This is called e(sample) and we can assign it to a variable and use that to define our sample as follows:

. // fit the most restrictive model
. quietly: regress y x2 x4 i.x1 x3
. generate mainsample = e(sample)

Now we can use mainsample to run all the models in the same sample:

. quietly:eststo:regress y x2 x4       if mainsample == 1
. quietly:eststo:regress y x2 x4 x1    if mainsample == 1
. quietly:eststo:regress y x2 x4 x1 x3 if mainsample == 1

. // tabulate and note sample sizes
. esttab

------------------------------------------------------------
                      (1)             (2)             (3)
                        y               y               y
------------------------------------------------------------
x2             -1941846.0      -3240948.1*     -3131194.0
                  (-1.63)         (-2.01)         (-1.91)

x4                 -7.289          -8.035*         -7.841*
                  (-1.90)         (-2.06)         (-2.00)

x1                           -2.23118e+11    -2.14745e+11
                                  (-1.20)         (-1.14)

x3                                            7.39664e+09
                                                   (0.39)

_cons         1.22904e+12     2.40732e+12*    2.25026e+12
                   (1.89)          (2.04)          (1.80)
------------------------------------------------------------
N                     424             424             424
------------------------------------------------------------
t statistics in parentheses
* p<0.05, ** p<0.01, *** p<0.001