For context:
STATA has a command called ‘splitsample’ where it effectively splits your current dataset into different partitions based on some user defined value. nsplit() defines how many groups, gen() creates the group identifiers for each.

There are additional arguments that can be passed such as: balance() and rround.

balance() forces the split groups to preserve semi-equal proportions/averages of some variable

rround helps randomly round sample sizes to satisfy the sample-size ratios

splitsample, gen(group) nsplit(2) balance(clicks networkrevenue) rround

Here’s an example output for this in STATA

Since I specified rround in STATA, the proportions are near 50/50 but not exact

tab group

  group |      Freq.     Percent        Cum.
      1 |      1,903       50.44       50.44
      2 |      1,870       49.56      100.00
  Total |      3,773      100.00

And because I also specified:

balance(clicks networkrevenue)

When I look at the summary statistics for my two groups, their averages are also alike

tab group, sum(clicks)

        |          Summary of Clicks
  group |        Mean   Std. dev.       Freq.
      1 |    30.31319   108.58514       1,903
      2 |   29.740107   98.621471       1,870
  Total |   30.029155   103.75319       3,773

 tab group, sum(networkrevenue)

        |      Summary of NetworkRevenue
  group |        Mean   Std. dev.       Freq.
      1 |   44194.721   242880.98       1,903
      2 |   44970.651   220502.77       1,870
  Total |   44579.293   232029.24       3,773

What I’m looking for is a Python equivalent to this command that has similar behavior to the balance() and rround argument.

I’ve done some cursory research into the available Python sampling libraries but they all behave differently than STATA’s splitsample.

I’ve looked at Imbalanced-learn’s over and under sampling methods + scikit’s sampling methods.

Any help is greatly appreciated!