A new estimate of the importance of unemployment benefits

by on October 15, 2013 at 2:52 pm in Uncategorized | Permalink

This is a new NBER paper from Marcus Hagedorn, Fatih Karahan, Iourii Manovskii, Kurt Mitman, the abstract is here:

We exploit a policy discontinuity at U.S. state borders to identify the effects of unemployment insurance policies on unemployment. Our estimates imply that most of the persistent increase in unemployment during the Great Recession can be accounted for by the unprecedented extensions of unemployment benefit eligibility. In contrast to the existing recent literature that mainly focused on estimating the effects of benefit duration on job search and acceptance strategies of the unemployed — the micro effect — we focus on measuring the general equilibrium macro effect that operates primarily through the response of job creation to unemployment benefit extensions. We find that it is the latter effect that is very important quantitatively.

There is an ungated version of the paper here (pdf).

1 dreamshade October 15, 2013 at 3:51 pm

Is this measuring against the reported unemployment rate, or against the employment/population ratio?

2 Mark_H October 15, 2013 at 5:30 pm

From the paper:
“Locations separated by a state border are expected to have similar labor markets due to the same geography, climate, access to transportation, agglomeration benefits, access to specialized labor and supplies, etc”

So their results depend on the 2 bordering counties being the same in every way except for unemployment policy. But what about all of the other state policies that have effects?

I like the discontinuity approach, but it doesn’t really seem to apply here.

3 Ethan October 15, 2013 at 6:38 pm

If you read the section 4 of the paper you’ll see that the two counties don’t have to be identical. Further, they go on to control for a whole host of other policies that could be different across states (e.g. food stamps, taxes, federal stimulus, foreclosure laws, etc). Seems like they covered all the bases.

4 Marie October 15, 2013 at 7:18 pm

State and municipal and county regulation of small business?

5 Mark_H October 16, 2013 at 11:50 am

This fascinating paper shows that controlling for variables does not even come close to guaranteeing bias elimination:

http://karlan.yale.edu/fieldexperiments/papers/00256.pdf

And controlling for the factors in states would do little to control for selection bias. Consider this example: a selection of people are working in Lake Havasu, say, a city on the border of California and Arizona. Some workers might choose to live in Havasu, Arizona because they do not feel that the taxes they’d pay in California are worth the services (like social insurances) they would get in California. Or maybe they prefer the gun laws or some other factor in Arizona. Others live in Needles, California because they feel that the property taxes they’d pay are worth the social insurances they’d get.

Fundamentally, what’s wrong with saying “these people are alike because they live close to eachother?” Well, if you have fundamentally different people choosing to live in the two places, then they might have fundamentally different employment preferences, *irrespective of the one policy in question.* That’s why this methodology is not perfect in this scenario–without randomization, it is next to impossible to cover all of the bases.

6 Dustin October 15, 2013 at 9:32 pm

Golly, isn’t 1100 pairs of adjacent counties, in addition to the controls stated, sufficient?

7 dan1111 October 16, 2013 at 3:39 am

If there is some other unmeasured factor at work that correlates with state unemployment policy, then sample size would not help overcome the problem. This is certainly plausible, since unemployment policy and many other things will be related to the political climate of the state.

And every study like this controls for lots of things. But did they control for the right things? That is always open for criticism.

Also, most of us probably are not qualified to evaluate whether their modeling approach is legitimate (I certainly am not). That is another possible area of criticism.

I am quite sympathetic to their claims (mood affiliation, obviously), but would be reluctant to put too much stock in this study without knowing more.

8 Mark_H October 16, 2013 at 12:40 pm

Please see the reply to Ethan above. Also, as dan1111 stated, sample size does nothing to counter some biases: if there is selection bias present, then it doesn’t matter how big your sample is.

Fundamentally, discontinuity works in a scenario like: somebody who is 20 years and 360 days old is not going to be very different than somebody 21 years and 5 days old, except for the fact that the latter is allowed to drink and buy handguns. So seeing how they differ on, say, likelihood to be involved in armed robbery or to wind up in the hospital with alcohol poisoning is almost certain to be due to the regulation discontinuity.

In the paper, they’re comparing people who live very close together but in two different states. Can you say that they are going to be the same except for their UI opportunities? Certainly not. And if there are fundamental unobservable (or observable but unobserved) differences between the type of person who chooses to live in one state vs. another, then the controls do not totally account for the differences. This paper is suggestive but certainly not definitive.

9 Dustin October 16, 2013 at 8:04 pm

Sure. My point was that it is incumbent upon any well-intentioned critique to highlight hypothesized absent variables so to test them, because it is certainly not possible to prove they don’t exist – the mere possibility of such variables does not indicate there existence.

Additionally, how many combinations state-adjacencies are there in the CONUS?

I would rather, and I believe appropriately, describe this research as ‘highly suggestive’.

10 Mark_H October 17, 2013 at 10:08 am

“it is incumbent upon any well-intentioned critique to highlight hypothesized absent variables so to test them”

My whole point is that you can’t know what variables are important and you can’t test for all of the variables (unobservables, say). This study just says “the types of people who choose to work in states that have high UI benefits are the types of people who tend to use use UI benefits, controlling for a handful of other policies.” What does this answer say about whether reducing or increasing UI benefits nationally reduce or increase unemployment nationally? I believe that more generous UI probably increases unemployment, but I am skeptical about the ability of this paper to prove so. As long as the type of people who choose to live in the more generous states differ systematically from the type of people who choose to live in the less generous states, you just can’t know if you’re seeing a difference in the people or a difference in the law, not to mention the difficulty in controlling for all other relevant policies.

I was highlighting the weakness of the methodology in this context. You can absolutely expect the alcohol example above to produce unbiased results because people don’t choose to turn 21. People choose which state they live in. The whole point of that methodology is that, if used right, it goes a long way toward eliminating selection bias, which it does not in this context.

Honestly, I do not mean this as any kind of personal jab, but the idea that you can just control for variables until you get a clean study has been absolutely debunked. You have to have some compelling reason that you’ve eliminated that, which discontinuity, difference-in-difference, or randomization can plausibly do, but which I contend fail to happen in this case.

11 KLO October 16, 2013 at 12:39 pm

How do they address the fact that you are considered unemployed in the state in which you live but you file for unemployment in the state in which you work? Maybe this is too small to matter, but it does seems significant in several large job markets (New York, Washington, D.C. and Chicago) and perhaps others.

12 mark October 15, 2013 at 5:39 pm

Methodology sounds very much like the minimum wage study that liberals love to cite as evidence that minimum wage hikes don’t increase unemployment. Will be interesting to see how people react to this.

13 jeff October 15, 2013 at 10:20 pm

nothing to see here, move along now. this is not the study you are looking for

14 mark October 15, 2013 at 5:53 pm

The approach sounds very much like the study done a few years ago where, of two bordering states, one raised its minimum wage, the other did not, and it was shown that employment in the former did not suffer. Liberals like that study and cite it all the time. It will be interesting to see how people with well-identified positions on that study react to this one.

15 Dustin October 15, 2013 at 8:48 pm

I think there is a sliiiiight difference in sample size between these two studies

16 Wonks Anonymous October 15, 2013 at 6:52 pm

Someone needs to do a regression on multiple policies different across state borders.

17 gavinf October 15, 2013 at 9:32 pm

Mark – the introduction states the opposite assumption, that increases in unemployment benefits will put upward pressure on the minimum wage (though it doesn’t cite where this has been shown). Can’t have it both ways.

18 chuck martel October 15, 2013 at 11:32 pm

Practically speaking, the unemployment benefit IS the minimum wage. UI recipients, being sentient beings, don’t take a pay cut to go to work unless their benefits have run out. In many ways UI is an insurance program for mortgage, rent and automobile payments. It’s also a subsidy for seasonal businesses, road construction, for instance, that pays the bills during the winter so those workers don’t move into other fields and are available when the frost goes out of the ground. Moreover, many seasonal workers would be unlikely to buy houses or new cars if they couldn’t count on UI benefits when they know they’ll be unemployed.

19 mike October 16, 2013 at 4:10 am

Seasonal workers don’t receive unemployment.

20 dan1111 October 16, 2013 at 5:53 am

This is not completely true. It likely depends on the state. For example New York clearly does cover many types of seasonal workers: https://labor.ny.gov/ui/dande/covered1.shtm

Beyond that, lots of jobs have seasonal ups and downs without necessarily being an officially seasonal position. A construction worker isn’t necessarily a seasonal employee, but has a higher than average chance of being unemployed over the winter.

21 Brandon October 16, 2013 at 9:23 am

This is correct. Many construction workers are laid off seasonally for a month or two every winter and they collect unemployment during this time. The system is structured this way and I have to imagine that everyone is aware and that it’s built into wages and labor costs.

22 Dustin October 16, 2013 at 8:35 pm

Hmm. It is not obvious to me that a construction worker is a ‘seasonal employee’ despite their being commonly laid off during slow winter months. I understand the definition of seasonal employee to be a bit more prescriptive; something along the lines of ski, tax, holiday retail, etc…

23 Donald Pretari October 15, 2013 at 11:51 pm

I looked at the following Three Papers:
http://www.iza.org/conference_files/PolicyEval_2013/mitman_k9175.pdf
http://www.iab.de/UserFiles/File/downloads/gradab/Dokumente%20Garloff/Mortensen_Pissarides_1994_Job%20creation%20and%20job%20destruction%20in%20the%20theory%20of%20unemployment_RES_pp_397_415.pd
http://math.cims.nyu.edu/faculty/avellane/BaiNG2002.pdf

The Number of Assumptions & Methods Astonished Me. Checking Each One must be tough. I am not competent to assess their Conclusions, although there’s no doubt they are serious attempts at understanding hard topics. But there have to be Methodological, Philosophical, & Practical Issues that underly some of the disagreements among Economists. I am simply more convinced by Micro or Human Agency Approaches to many Economic Policy Issues. I hope this makes sense.

24 Wonks Anonymous October 16, 2013 at 9:41 am

You have an Odd approach to Capitalization.

25 Donald Pretari October 16, 2013 at 3:18 pm

I know. But it helps me read what I’ve written.

26 John October 16, 2013 at 8:01 am

I don’t see how the study supports the argument that extended unemployment insurance puts upward pressure on the equilibrium wage because, as far as I could discern, the authors did not disaggregate unemployed persons between those who receive unemployment insurance and those who don’t:

“Everything else equal, extending unemployment benefits exerts an upward pressure on the equilibrium wage. This lowers the profits employers receive from filled jobs, leading to a decline in vacancy creation. Lower vacancies imply a lower job finding rate for workers, which leads to an increase in unemployment.”

If the authors rest their conclusions on that upward pressure, than they need to show evidence of its existence. Yet they appear to believe that disaggregating is relevant only to determining dis-incentive effects of EUI:

“Everything else equal, extending unemployment benefits exerts an upward pressure on the equilibrium wage. This lowers the profits employers receive from filled jobs, leading to a decline in vacancy creation. Lower vacancies imply a lower job finding rate for workers, which leads to an increase in unemployment.”

27 John October 16, 2013 at 8:20 am

The second quote should read:

“[B]ut comparing ineligible to eligible would only capture the difference in behavioral response of search effort between workers, not the possibly much larger macro effect.”

28 Kurt Mitman October 18, 2013 at 3:51 pm

John – In Table 5 we provide evidence on the upward pressure on the equilibrium wage. You are correct that this is not dis-aggregated. What this means is that the effect on the equilibrium wage would have been even higher if everyone were receiving benefits – which would have led to an even larger macro effect of unemployment benefits on unemployment.

29 Steve Roth October 16, 2013 at 1:05 pm

@Noahpinion They actually find UI is NEGATIVELY associated with change in unemp, but interpret this as POSITIVE effect due to anticipation.— Arindrajit Dube (@arindube) October 16, 2013

30 Steve Roth October 16, 2013 at 1:09 pm

Arindrajit Dube (the guy whose data the paper is based on):

Tweet: “They actually find UI is NEGATIVELY associated with change in unemp, but interpret this as POSITIVE effect due to anticipation.”

31 Kurt Mitman October 16, 2013 at 3:51 pm

One of the authors here. See the discussion with Dube below on twitter:

Kurt Mitman ‏@SorryToBeKurt
@Noahpinion @arindube We don’t find that UI is negatively associated with the change in unemployment. What are you basing this claim on?

Arindrajit Dube ‏@arindube 1h
@SorryToBeKurt @Noahpinion Your outcome is forward quasi-differenced (equation 9), (U(t) – c*U(t+1)). Approx -(U(t+1)-U(t)), neg change in U

Arindrajit Dube ‏@arindube 1h
@SorryToBeKurt @Noahpinion Therefore seems that you find UI(t) correlated negatively with (approximately) change in unemp rate (u(t+1)-u(t))

Kurt Mitman ‏@SorryToBeKurt 28m
@arindube @Noahpinion That’s exactly the sign that supports our conclusions. Consider the basic Pissarides model.

Arindrajit Dube ‏@arindube 22m
@SorryToBeKurt @Noahpinion As I said, you interpret the neg. correl between ∆unemp and UI as increasing unemp from modeled anticipation.

Kurt Mitman ‏@SorryToBeKurt 20m
@arindube @Noahpinion There’s no anticipation in that example. That’s why we picked it.

Kurt Mitman ‏@SorryToBeKurt 20m
@arindube @Noahpinion With anticipation only quasi-differencing is theorectically consistent.

Arindrajit Dube ‏@arindube 19m
@SorryToBeKurt @Noahpinion Look at my original tweet, and tell me what in is incorrect.

Kurt Mitman ‏@SorryToBeKurt 3m
@arindube @Noahpinion We completely agree with your formulas. Let’s be clear on the interpretation.

Kurt Mitman ‏@SorryToBeKurt 2m
@arindube @Noahpinion A one-time change in benefits implies a negative correlation between benefits and unemployment growth.

Kurt Mitman
‏@SorryToBeKurt
@arindube @Noahpinion Precisely because raising benefits increases unemployment. This is the logic of a textbook search model.

32 PL October 16, 2013 at 9:58 pm

Mitman wins again!

33 Kurt Mitman October 18, 2013 at 4:06 pm

I realized that this ended up getting pasted in a weird order. Here’s the full discussion from Wednesday:

Dube: Your outcome is forward quasi-differenced (equation 9), (U(t) – c*U(t+1)). Approx -(U(t+1)-U(t)), neg change in U. Therefore seems that you find UI(t) correlated negatively with (approximately) change in unemp rate (u(t+1)-u(t))

Kurt: That’s exactly the sign that supports our conclusions. Consider the basic Pissarides model. If benefits are high in per 0, V0 is low and thus U0 is high. In per 1 V1 fully recovers and U1 recovers but slower. So U1-U0 is negative. All this happens precisely because benefits increase unemployment.

Dube: As I said, you interpret the neg. correl between ∆unemp and UI as increasing unemp from modeled anticipation.

Kurt: There’s no anticipation in that example. That’s why we picked it. With anticipation only quasi-differencing is theorectically consistent.

Dube: Look at my original tweet, and tell me what in is incorrect.

Kurt: We completely agree with your formulas. Let’s be clear on the interpretation. A one-time change in benefits implies a negative correlation between benefits and unemployment growth. Precisely because raising benefits increases unemployment. This is the logic of a textbook search model.

Dube: So we agree you find neg correlation bewteen unemp change* (u(t+1)-u(t)) and UI(t)? (*approx due to quasi-diff). And we agree that you interpret this as causing higher unemp from the model. Right?

Kurt: If you are saying that the response of unemp to benefits in the data is consistent w/ DMP model (Nobel 2010), I agree.

Dube: My point simply that there are other interpretations. E.g., lagged Aggregate Demand also ➾ Correl(∆Unemp,UI)<0.

Kurt: If people spend their benefits in their home county, then we completely control for any potential demand effects.

34 JonFraz October 16, 2013 at 7:10 pm

Too facts from the real world.
1. Fewer than 40% of the unemployment qualify for benefits at all (this probably varies a bit by state) so there should be a large pool on non-UI recipients available for low wage jobs and hence the effect of UI pushing up wages for these jobs should be swamped, at least in high unemployment times like these.
2. UI replaces, on the average, about 1/3 of a worker’s previous salary. Since most workers expect to be hired back somewhere around their old salary (but may be willing to settle for, say, 10% less) it’s also hard to see how such a small stipend, probably lower than most people’s reservation wage, can act to push wages up.

35 Kurt Mitman October 18, 2013 at 3:58 pm

During the great recession up to 70% of the unemployment were receiving benefits (you can construct this from Bureau of Labor Statistics data from the Current Population Survey).

The equilibrium wage depends on the outside worker. If a worker is deciding whether to accept a job, they evaluate the alternative, that is, staying unemployed and waiting for another offer. The value of that to them is the combination of the benefits they receive while waiting and the expected attractiveness of the next job offer.

36 A Greek Bearing Facts November 7, 2013 at 6:53 pm

NOTE TO READERS: That 70% UI coverage rate, mentioned by Prof. Mitman, did not last long (you can look it up). During most of the recession and ensuing recovery the UI coverage rate was far below 70% of unemployed workers. But even at the point that 70% of the unemployed were collecting UI benefits, 30% were not. A little data checking and math will show that even at that point the number of unemployed who did NOT receive UI exceeded by a sizable margin the number of job vacancies reported in the JOLTS survey. That is, job seekers NOT affected by longer UI benefit durations exceeded the total number of job vacancies on offer. Employers could have filled all their job vacancies with job seekers who were not basing their reservation wages on liberalized UI benefits. A lot of job seekers are folks just completing school, college, or training, and these folks are ineligible for UI. So are all the folks, now quite numerous, who have used up their UI eligibility.

Some other challenges for the basic theory: It relies on the threat to firm profitability caused by employers’ fear that high / long lasting UI benefits will raise reservation wages / equilibrium wages and consequently make new job offers unprofitable.

(A) Where is the evidence that equilibrium wages have risen since the onset of the recession in 2008? Real wages are flat across most industries and occupations.

(B) Where is the evidence employers are not making profits on the job vacancies they’ve actually filled? The last time I checked, both business profits and the business share of output have reached new post-World-War-II highs nearly every quarter. It would be a useful, though perhaps impossible, exercise for the authors to present readers with data showing that, even though business profitability is at a historically high level, the potential impacts of temporary UI benefit extensions on unemployed workers’ reservation wages are so large that *new* job openings would be unprofitable.

(C) Hyper-rational employers are supposed to anticipate that a state’s increase in UI benefits today will signal its future UI benefit generosity. But this expectation does not appear very rational to one who knows anything about the history of the UI program: (i) All the benefit liberalizations in the Great Recession were initiated and funded by the Federal government, not by the states. Therefore, how can the benefit extensions possibly signal state-level intentions regarding either temporary or permanent state-level benefit changes (and the profitability of current or future job offers)?; (ii) All the federal benefit liberalizations were explicitly temporary: There was an end date written into the law providing for each extension; and (iii) Temporary federal extensions of UI benefits have taken place in every U.S. recession since the late 1950s, and in every single case the UI program has been restored to its pre-recession generosity level when unemployment fell toward a more normal level. So the supposedly hyper-rational employers in the model do not appear to be acting on very extensive knowledge about the recent and historical experience of the United States UI program.

37 Jon October 18, 2013 at 7:36 am

Just at a first glance–their measure of unemployment uses data from the BLS survey. Is there a risk that one contribution is that some people stop reporting themselves as unemployed and seeking work after their benefits expire.

38 Jon October 20, 2013 at 10:03 am

While they do have additional regressions on other variables, I am not persuaded by their technique. Although it is based on a well known model of employment(the Pissarides model), they have a number of their own tweaks. Notably the difference scheme replaces expectations with realized values and some factors linearized.

Also, I don’t see quickly how much of the differences between county pairs is accounted for by the principal factors and how much by the benefit differences. If the model is “perfect” that is not relevant; however these types of models are not perfect.

Georgia has 159 counties–more than California (58 counties). Thus the sample is going to be rather skewed by states. Furthermore the median county in the survey has only “1/2” of 1% of the population of the state. Only Texas has over 200 counties! Unless I am missing something on weighting, this survey will not be reflective of the counties in which most people live. For example in Maryland and Virginia–Montgomery County Md, Prince Georges County Md, and Fairfax County VA, will get the same weight as low population counties such as St. Mary’s, MD.

39 Kurt Mitman October 21, 2013 at 7:51 pm

Hi Jon,

First, regarding county size, we were talking median, not mean. However, if you’re concerned we also run the analysis on a sub-sample that only includes counties that are in a Core-Based Statistical Area (a census term that ends up exclusing most of the smaller rural counties) and we find the results are unchanged.

The counties can be different. What’s important is that the differencing cancels out state level differences that could be driving unemployment benefits. That’s all that we require for identification.

Kurt

Comments on this entry are closed.

Previous post:

Next post: