Improving Experiments by Optimal Blocking: Minimizing the Maximum Within-block Distance

Download Presentation Slides

Want the code? Contact Jas Sekhon

Authors: Michael J. Higgins and Jasjeet S. Sekhon

Experimental designs guarantee balance between treatment and control groups “in expectation” but in practice it is possible to end up with all sorts of imbalances. Blocking at the time of assignment to treatment is a way to fix this. At EGAP 11, Berkeley’s Jas Sekhon presented his and Michael Higgins’ new method for blocking called “Minimizing the Maximum Within Block Distance” (MMWBD).

Here are a few highlights from their presentation:

To start the presentation, Sekhon reviews what blocking is:

Matched pairs is another way to block that is quite common. But using matched pairs can make it impossible to generate randomization-based estimates of uncertainty. What is pair matching and can you do better?

Now what is MMWBD?

Blocking has a similar function to covariate adjustment. It can be used to reduce variance in estimates, but also to deal with “conditional bias.” With large datasets the two approaches yield very similar results, though blocking protects you from concerns about manipulation. With smaller datasets blocking does a better job at reducing variance.

Here is the intuition behind the MMWBD approach.

And now into the weeds: here is how the algorithm works.

Read more here!