@mikemadowitz and I had a productive conversation about this data.

It turns out, the disparity in marriage employment *counts* is mostly an artifact of something else going on in the household survey entirely. /1 https://twitter.com/mikemadowitz/status/1313439549809397760

Namely, look at what's happened to the number of married men and women reported in the household survey during the pandemic: a rapid rise in February, and then a fall after June. /2

So a great deal of the declines in employment *counts* for married men and women in the household survey is explained by the declines in just the number of married people the survey shows in the first place. /3

Is this actually what's going on? Were there 1 1/2 million shotgun weddings during the early pandemic, and then a surge of divorces thereafter? Well, maybe... stranger things have happened in the last 8 months (like a penny shortage). But there's another explanation. /4

Another explanation is that response rates fell in the household survey during the pandemic, mainly because in-person interviews for the survey were cancelled, and this resulted in a bias in the population being captured by the CPS. /5

When survey responses deviate from what BLS knows to be the makeup of the population, they use survey weights to try to correct for the disparities. But BLS doesn't appear to correct for marriage rates. We can still see the surge in marriage rates even without survey weights. /6

This all makes sense by the way -- marriage is not immutable, so BLS may not feel they have a stable, reliable benchmark estimate of the married population with which to fully correct for response differences by marriage. /7

But it also means that the household survey may have been *overestimating* the number of married people throughout this pandemic. That overcount peaked in June, and has been correcting itself since then. /8

If true, that means that if you looked at a *count* at the *number* of married employed people since June, it would look like it was declining sharply, when in fact all that's happening is CPS response rates are gradually returning back to normal, with more single people. /9

One way to adjust for much (though not necessarily all) of this potential sample issue is to look at employment *rates* among married men and women -- that is, as a percent of their respective populations. /10

Here are what employment *rates* for married men and women look like in level terms (I'm using non-seasonally-adjusted data so that this isn't confounded by the potential population count problems).

Both grew by +0.3pp in September. /11

And here's how both have evolved since February 2020.

Married male EPOP is -3.3pp below February levels. For women, it's -3.9pp. The gender disparity has closed only slightly... married women were -0.9pp further behind men in April. /12

And both rates have grown modestly since June.

None of this is to say that there aren't important differences in the labor market by marriage, but there's a risk given the broader survey response issues that *changes in counts* of married people will overstate these. /13

Let me end by saying: in walking through all of this with @mikemadowitz beforehand, he was a model of patience, intellectual curiosity, and valuing rigor in the data above all. Everyone interested in economic data should follow him. /FIN

Two links to add in addendum.

One is a @BLS_gov overview of how the CPS works, including weighting. H/t @Brett_Matsumot0

The other is an excellent paper from Jason Ward & @keds_economist going deeper into CPS response issues.

https://www.bls.gov/cps/cps_htgm.pdf
https://www.rand.org/pubs/working_papers/WRA842-1.html