Monday, January 26, 2015

Results of my unemployment bets with Kurt Mitman

Back in February, I made a bet with Kurt Mitman, who wrote a theoretical paper about unemployment benefits and unemployment. Based solely on my prior that "all macro models are massively misspecified," I bet that both unemployment and the employment-to-population ratio would remain flat from December 2013 through December 2014. Kurt bet on the point estimates of his model. So it was:

Noah's beta:
Unemployment: 6.7%
Employment-to-population ratio: 58.6%

Kurt's bets:
Unemployment: 5.2%
Employment-to-population ratio: 60.6%

Actual figures:
Unemployment: 5.6%
Employment-to-population ratio: 59.2%

Result: We split the bets. My miss on unemployment was 1.1% to Kurt's 0.4%, so Kurt wins that one. My miss on employment-to-population ratio was 0.6% to Kurt's 1.4%, so I win that one.

As per the terms of the bet, Kurt and I will each buy each other a pizza dinner. However, since Kurt now works in Europe, it may be awhile before we can make good!

Anyway, in retrospect, it was a little careless of me to bet on "no change" for unemployment, which was clearly trending downward as of January 2014:

And it was probably a little careless of Kurt to bet on employment-to-population ratio when his model didn't model the margin between unemployment and out-of-the-labor-force, or the margin between employment and out-of-the-labor-force.

But anyway, as a follow-up, Kurt is co-author on a new paper that gives an empirical estimate of the effects of the termination of extended unemployment benefits. Looking at differences across state borders (and between neighboring counties in different states), they find that the termination of benefits increased the total number of employed people by 1.2%. That doesn't sound huge, but it's actually 1.8 million jobs. Whether it qualifies as an employment "miracle", as the paper's title claims, is up for debate, but 1.2% is not nothing. (Of course, I'm accepting the paper's result at face value; some other studies have yielded very different results.)

So although unemployment benefits don't seem to have been the main reason for the huge decline in U.S. employment since 2008, and their effect isn't as big as in Kurt's original model, they were indeed a non-trivial factor. There were a substantial number of people out there who were being paid not to work, and are now back working again.


Mike Konczal has a blog post in which he reviews some evidence regarding UI extensions and unemployment. The other studies that examined this question looked not at states or counties but at individuals. It would take a lot of effort for me to compare the two methodologies in detail, so I won't. The other studies find that UI extensions decreased employment by about 0.1 to 0.5 percentage points, which is much smaller than the estimate of the recent paper on which Mitman is a co-author. (A 2011 paper by Makoto Nakajima, however, finds effects close to those Mitman finds).

But I think the bigger point is this. The employment-to-working-age-population ratio fell by a bit little over 5 percentage points in the Great Recession, and since then has recovered by a little over 3.5 percentage points. Mitman and his co-authors estimate that UI extensions were responsible for about 17% of the drop, or about 26% of the recovery. The other commonly cited paper estimates that UI extensions were responsible for around 5.7% of the drop, and 8.7% of the recovery.

Looking at this range, it's clear that UI extensions do hurt employment, but were not the dominant factor in the recent recession.


  1. AKAIK, people are split on if Chicago Deep Dish is really pizza.

  2. No, you didn't split the result. He bet on movement and got the direction right. To expect quantitatively correct results out of a macro model is silly. Direction is enough

    1. To expect quantitatively correct results out of a macro model is silly.

      But that's what I was betting on. I agreed with him about what the direction of the movement would be.

    2. You bet on flat. He bet on direction. This is no tie.

    3. I don't understand. How flat would it have had to have been for me to "win" in your weird view of things?

      The EPR was closer to flat than rising. It rose 0.4 percentage points. That's pretty flat!


  4. Luc Hansen4:58 PM

    It would be nice to know how many benefits were terminated to gain the 1.8m newly employed, and the personal consequences for those terminated but still out of work, and what those 1.8m think of the jobs they were forced to accept.

    1. Right. It's not at all clear that this is a net gain for the economy in welfare terms...

  5. Once again, the results validate the adage that the average of the forecasts is better than any individual forecasts.

    Sorry, why does one person on another continent mean that payment will be delayed? Money is mobile and worldwide these days, most places sell gift certificates.

  6. Well, we you do get around to settling that bet, make sure you don't do it with whatever that is in your pic for this post, because that is not pizza.

    1. It's not pizza in the sense that Superman is not a human.

    2. Right. Superman is an alien masquerading as human and deceiving everyone he has sworn to protect, as well as almost everyone he claims to love.

  7. Anonymous12:49 PM

    From the CEPR link helpfully provided by MaxSpeak, written by Dean Baker:

    "The third problem is that the LAUS data are largely model driven. There is little direct data for many counties. The Bureau of Labor Statistics (BLS) generates employment estimates for these counties from a variety of variables, including unemployment insurance claims. This makes them of questionable value in this sort of exercise."

    Red alert. GIGO alert.

    You're a blithe fool for throwing out prior studies with barely a glance at the methodology behind this study, knowing nothing about the dataset and its limitations, and general Dr. Ozzery.

    There's literally no macroeconomic issue that doesn't have a study claiming two sets of effects that directly contradict each other. If you're not adjudicating the validity of the study, but swinging back and forth like a weathervane, you're useless.