Epistemic Status (written before I looked at the paper): This new paper confirms all of my priors.
(Which is no longer the epistemic status because things started looking weird, but one step at a time). I’m publishing part 1 now because I don’t want things to drag on forever, and the calculation I just did looks weird enough that it’s a decent stopping point.
Seattle has raised its minimum wage to $13 an hour ($12 for small employers) and will soon raise it to $15. Armed with a unique data set, a new working paper claims that even the increase so far has had a big enough effect on low-wage employment that it has backfired on low-wage employees, costing them more in hours than they have gained in pay. Many of the details I read about made the paper and its conclusions appear highly credible, but some of the numbers involved seemed to naively imply crazy things, so I decided it would be worthwhile to read the actual paper.
The paper is worth reading (technically it is a working paper, but no one seems to care about that distinction). It includes details like this one, which appears as a footnote when discussing the fact that a number of big employers are missing from the data set the study uses, to answer the anticipated objection that big employers ‘have more bargaining power when negotiating wages, and can more easily absorb the additional labor costs’:
The Seattle Minimum Wage Study surveyed over 500 Seattle business owners immediately before and a year after the Ordinance went into effect. In April 2015, multi-site employers were more likely to report intentions to reduce hours of their minimum wage employees (34% versus 24%) and more likely to report intentions to reduce employment (33% versus 26%). A one-year follow-up survey revealed that multi-location employers were more likely to report an actual reduction in full-time and part-time employees, with over half of multi-site respondents reporting a reduction in full-time employment (52%, against 45% for single-site firms).
Take a step back for a moment, and look only at those numbers, ignoring the main study. What conclusions can you draw?
In terms of large versus small businesses, this likely indicates that single-site employers cut a higher percentage of jobs and hours than multi-site employers. This is because they are measuring whether magnitude is greater than zero, rather than the average reduction. A multi-site employer will have plans to reduce employment, or actually reduce employment, if this is true of at least one of those sites. If 45% of single-site firms reduce employment, then even with some amount of correlation between sites of the same firm, having only a 52% rate of reduced employment would imply that the base rate almost has to be lower, especially since their individual sites also likely employ more people per site.
Minimum wage discussions are full of competing claims, and counter-intuitive theories and results like this. A result is presented as showing one effect, but seems to show its opposite. So we have to be careful to look at all details before drawing any conclusions…
Wait, did that say that 45% of single-site firms reduced employment? And a majority of multi-site firms (even if this might be a relatively small percentage of all their sites or jobs)? How big is the average single-site firm? How could this not imply a huge disaster, if true?
That’s the thing about the results here. They are freaking huge. The reason I read the paper in its entirety is that even with priors that raising the minimum wage will reduce employment, I had to see the numbers to wrap my head around the numbers involved because they seemed absurdly big.
Before getting into detailed analysis of the central findings, I think it’s important for me to state my priors, and for you to think about yours.
My priors are:
In favor of minimum wage raising being bad:
- When you raise the price of something above its market rate, you almost always get less of it.
- When you raise the price far above the market rate, you usually get far less of it.
- It is costly along many dimensions to fire people. It is costly along many dimensions to eliminate positions. Most employers hate firing people. Short term effects will be smaller than medium term effects, which will be smaller than long term effects.
- Proportionally more damage will be done to potential future jobs, than will be done to existing jobs. Proportionally more firms will never open, or fail to expand, than existing firms will go out of business and existing sites will close down.
- As the effect persists and expands, and is expected to persist and expand, there will be substitution of capital for labor that will shift efficiency, defaults and norms towards.
- Losses will be concentrated among those who are no longer employable, who were already the worst off.
- The broader economic effects of increasing costs must be presumed to be negative.
- There is huge incentive in most of academia to find that raising the minimum wage is good. All previous studies finding no effect seemed to rely on sketchy/incomplete data and methods, and seemed not to understand the power of the priors they were fighting against.
- Unemployment is a no good, very bad thing, as is underemployment, and the cost of such things is very high so decreases in available jobs are much worse than they appear.
- If you raise wages above their equilibrium price, you have too many workers chasing too few jobs. This gives employers tremendous power over employees. They can now fire them and threaten to fire them freely, knowing there are many new applicants waiting for those jobs. They can treat the workers like dirt, give them terrible working conditions and benefits, not listen to their complaints, scramble their schedules, give them 29-hour work weeks, demand extra work off the books, and so forth, much more so than before, so we expect a lot of the wage increase to be inefficiently reclaimed.
In favor of the minimum wage raising being not so bad:
- Efficiency wages are a thing. Many employees are underpaid in the sense that the employer would be better served to be paying them better.
- There is incentive in other parts of academia, and from other places, to find large negative effects from minimum wage increases.
- The amount of evidence against the minimum wage having negative effects has been surprising, so we should update at least somewhat on it even if it’s not perfect, unless we can explain where it is coming from (as the authors here claim they can).
- Seattle is in a boom and already well-off, so a $13 minimum wage in Seattle doesn’t seem obviously outrageous. Maybe it is fine, especially near term.
- There can be positive spill-over effects on wages for those not making the minimum (although this also means there can also be negative spill-over effects for all the same reasons).
I want to stop here and emphasize how profoundly weird it is to propose that low-skill labor, the thing we most want to raise the price of beyond its market value, would happen to be the one important exception to where raising the price gives you less of it. I mean, anything’s possible and there are a lot of factors, but think about just how profoundly weird it would be for that to be true. It would be freaking bizarre. People who try to act like this would not be freaking bizarre are either engaging in wishful thinking, have so little grounding in economics and the real world as to not understand this basic point, or are lying.
Methodology and Data Sources
Isolating the effect of the minimum wage increase is hard, and assumptions matter a lot, so it’s important to look at the details of what the study is up to here.
The big win that makes this study possible is the use of a uniquely rich data set:
We study the impact of the 2015 and 2016 minimum wage hikes in Seattle using
administrative employment data from Washington State covering the period 2005 through the third quarter of 2016. Washington’s Employment Security Department collects quarterly payroll records for all workers who received wages in Washington and are covered by Unemployment Insurance (UI). Washington is one of four states in the US that collects not only data on earnings, but also on hours worked during the quarter. Employers are required to report actual hours worked for employees whose hours are tracked (i.e. hourly workers), and report either actual hours worked or total number of hours assuming a 40 hour work week for employees whose hours are not tracked (i.e. salaried workers).
The government’s invasion of your privacy and recording of the details of your lives in order to engage in redistribution is our gain! Sweet. The paper emphasizes multiple times that reductions in hours are an important negative effect that other studies have not been able to study. This makes sense, especially given other recent changes in law regarding health care and other benefits, along with employers’ reluctance to fully fire workers. The cost of employing workers goes up, so maybe they are instead used less (demand goes down) and also used in a more efficient way for the employer (not have to pay for health care), as a way of reclaiming part of the wage increase.
This also allows the study to avoid using industries or other subsets of the economy to approximate low wage employment, and instead look directly at who is being paid how much money. Seems big.
The assumption that salaried workers put in 40 hours seems like it should get further examination. I would be surprised if the average salaried worker puts in a number that close to 40 hours of work, and my guess would be on average that they put in more than 40 rather than less, which means their effective wage rates are being overestimated here and also that they have more variance. There’s also the possibility that previously minimum wage employees are being put on salary in order to avoid the new minimum wage since they would now be able to work more than 40 hours while only being paid for 40. This could mean that hours worked are not going down as much as the data would naively suggest, and there is more low-wage employment than the data would naively suggest, but does mean that workers would be getting a worse deal rather than a better one; it’s hard to see how this effect could work in their favor.
The next problem is how to identify which workers are effected by the minimum wage increase. It is known that minimum wage increases push up wages of those that were already above the new minimum, because of the need to maintain the right relative wages between different employees and jobs – the assistant manager has to earn more than the retail clerk, so the two wages need to move together.
Their solution is to look for the cutoff where employment stops increasing during the time period being studied. They found an increase in paid work for wages up to $18/hour, so they put the cutoff at $19/hour, which seems safe at first glance. The other question is whether employers are substituting high-skill labor for low-skill labor. It seems plausible that some employers would hire a $50/hour worker to do a job that serves the same purpose as multiple previous low-wage employees, since the relative costs have changed. Robots can take your job, but so can other people, and one could in theory even tell a story where Seattle’s recent boom is being encouraged by the forced substitution of high-wage labor for low-wage labor, which combines with high rents in forcing low-skilled workers to move away. That would locally look like a boom, including in the data.
Since the cutoff at $19/hour, which is roughly double the previous minimum wage, is designed to catch the entire effect of the wage increases (no false negatives) it also means that a lot of the workers considered ‘low wage’ will see no effect (inevitable false positives), which will make the effects of the new minimum wage look much smaller than they are.
The first method they try is difference-in-differences, which has been used by many previous papers. They use two different control regions, either the nearest unaffected areas (which gets you some good similarities, but risks spillover effects) or a group of areas chosen to be farther away and as similar to Seattle as possible, neither of which are great. There are no good choices, since Seattle is unique in Washington and they only have their data set inside the state, so this is certainly a source of error although it is not obvious which direction the effect goes.
The paper notes that these samples fail a falsification test for these reasons, which certainly sounds bad. In their words:
To overcome this concern, we estimate the impact of the minimum wage using two
methods which allow for flexible pre-policy trends in control and treated regions: the synthetic control estimator (Abadie and Gardeazabal, 2003) and the interactive fixed effects estimator (Bai, 2009). Both methods have been used in the regional policy evaluation literature and applied to the minimum wage as well…
Synthetic control works by choosing a set of regions that minimizes the observed error in the time periods before Seattle raised its minimum wage.
These methods assume the effects on the minimum wage are a linear combination of factors. This is probably wrong, but over small periods of time and relatively small changes, it should be fine. The interactive fixed effects simulator they say is done following Gobillion and Magnac (2016) who ‘have developed a publicly-available program to estimate the treatment effects in the regional policy evaluation context.’
There are any number of ways these methods might be flawed. They might have implemented it incorrectly. The original methods might have serious flaws. I don’t know, and I am choosing not to investigate this further unless someone tells me these threads are worth pulling at.
The most obvious flaw (assuming the methods are fundamentally sound and correctly executed) is that they are comparing all of Seattle’s policies in the pre-raise world to all of Seattle’s policies in the post-raise world, since they choose a control that matches Seattle’s previous experiences. Any number of other policies could also have changed, which I find unlikely to be too bad, but the biggest worry is that Seattle effectively had a lower minimum wage than its controls before it enacted its new policy, because Seattle’s price level is higher. This shouldn’t be an issue because we are still measuring the effect of the shift from one minimum to another minimum, but it still seems worth worrying about a little.
On January 1, 2016, the minimum wage in Seattle was raised from $11 to $13 an hour for most places.
I wasn’t able to easily and nicely reproduce the entire chart, so I retyped key sections of table 3, starting with the number of jobs Under $13, Under $19 and Overall:
|Quarters||Under $13||Under $19||All||% Under $19|
|4/1 ($11 min)||35087||92668||311886||29.7%|
|8/5 ($13 min)||24420||88431||331927||26.6%|
The $11/hour level doesn’t have a naked-eye-visible effect. The $13/hour level does have one on the Under $13 category, for obvious reasons (the number does not go to zero, I’m assuming due to the small places that only have a $12/hour minimum, which seem to have had the majority of all $13/hour and under jobs given this data. It has a less obvious effect on the Under $19 category. The dip instead seems to be in quarter six, which is odd since an anticipatory effect seems unlikely to be that coordinated, and if the effect was seasonal we would see it in the all category
Basic heuristics say that natural wage growth and inflation over two years should add something like 5% to wages given Seattle was doing well, which should move some of the jobs under $19 to over $19, but less than 10% since one must assume that each dollar interval is smaller than the previous one. So about 9.8% of all jobs under $19 an hour seem to have vanished. Given a large majority of those jobs were already above the new minimum, that seems like a huge number, and we haven’t considered hours yet.
The issue with looking at it that way is that overall employment in Seattle is clearly doing just fine. If there are enough jobs for everyone, isn’t it a good thing that now 5% more of them pay over $19/hour? We seem to be putting a lot of faith in the idea that jobs over $19/hour are on net unaffected by the change, which is even more odd if Seattle is at full employment. This seems worth revisiting later, but for now we need to look at more of the data.
|Quarters||Hours U13||Hours U19||Hours All||U19 H%||Avg U19||Avg All|
Hours worked in Under $19/hour jobs seems to decline less relative to all jobs, than the number of jobs alone, and generally seems to maintain the pattern of about 90% of baseline hours (hours counts are in thousands of hours per quarter in the second through fourth columns, in hours per quarter per worker in the last two columns).
When looking at hours worked rather than jobs worked, we seem to see a phase shift but one that is over by quarter six; after that, we see the situation stabilizing. This is very odd, since previous studies did not look at hours and found less change, whereas this study did look at hours and found larger change, but the hours data seems to point towards less change than the job counts. Time to keep digging.
Average hourly wages:
|Quarters||Avg U13||Avg U19||Avg 13-19||Avg 19+||All|
The under 19 average is driven by the number of wages that are still under 13/hour. If you look at the Average 13-19 column, which I calculated, that number doesn’t change at all, stuck smack in the middle at $16/hour. Interesting! That means that even before the minimum wage, we didn’t see (more than a marginal amount) more people earning $13-$16 than $16-$19, and the ratio did not change when we pushed a lot of people out of the Under $13 bracket. That would not have been my prior; I would have expected this number of to start off lower (something like 15.50-15.75), and more importantly I would have expected it to decline as people from the Under $13 bracket got pushed into the $13-$19 range, whereas we have postulated that the law isn’t pushing that many net people from under $19 to over $19, which is why $19 was chosen.
But if all of that is true, why isn’t this ratio moving? I’m pausing here to think about that for a while, and suggest you do the same. I can see how this might be wonderfully good news, and I can see how this might be horribly terrible news. I’ll resume in part two.