Does unionization lower employment and workers’ earnings?
That’s the counterintuitive finding of an unpublished working paper by Brigham Young University economist Brigham Frandsen titled “The Surprising Impacts of Unionization: Evidence from Matched Employer-Employee Data” that’s popped up recently in the policy blogosphere, cited by Matt Yglesias in Vox and Lydia DePillis in the Washington Post.
Frandsen’s conclusion from analyzing over 45,000 NLRB elections is that newly-unionized U.S. firms hire less, stay in business less, and pay workers less overall — not because they slash wages, but because higher-paid workers leave.
This seems to fly in the face of received wisdom and, well, common sense — not to mention a wealth of studies that DePillips lists (and Frandsen himself cites) that show the opposite. And while, yes, at a glance it has the appearance of the kind of contrarianism-for-its-own-sake that pervades some corners of the economics discipline — what Ellora Derenoncourt described as “trolling the basic ethical instincts of the rest of humanity” — this isn’t a snarky Economist op-ed. It’s a real research study: marshaling an impressive dataset, careful methodology, statistically significant findings and a sort-of-plausible mechanism to explain them — a kind of Ayn Randian scenario in which the senior and best workers, fearing that union contracts will limit future earnings, quit rather than be dragged down by the slackers. So it’s worth taking seriously.
On closer inspection, though, Frandsen’s study is less clear and less surprising than it seems.
First, there’s the research design. Frandsen’s study is built on what’s called a regression discontinuity (RD) model. Before we get into the statistical weeds here, let’s take a step back. Any social science study that purports to find out if X causes Y is going to have to deal with the problem of endogeneity. This is where you see a correlation between X and Y but can’t be sure if X really causes Y — or if Y actually causes X; or if there’s some other variable you didn’t account for, Z, that affects both X and Y and gives them the illusion of a causal relationship.
This is the classic “umbrellas cause rain” scenario you’re taught to watch out for in stats 101. The entire field of econometrics was developed to try to solve this problem, coming up with various tools to control for unobserved factors.
Regression discontinuity is one of those tools. It’s an elegant way of comparing two groups of people with different outcomes, while controlling for all the other things that might make them different: by comparing between those who are just barely on either side of that outcome.
In this case, Frandsen wants to know the effect of a workplace unionizing on future earnings, employment and company survival. But workplaces that choose to unionize might be different in other respects than those that choose not to: they may be more heavily concentrated in one sector or industry; they may be more profitable; the workers may be more white, more male, more skilled, more mobile — all of which would affect earnings, employment and survival independent of being union or not. Any credible comparison would have to take all that into account.
So Frandsen’s model looks at a lot of union elections (and I mean a lot: in statistics, when you have over 45,000 observations, you’re probably going to get a significant result) and compares the post-vote earnings and employment between shops where the union barely won with shops where the union barely lost.
This is good research design: There shouldn’t be any reason to believe shops that voted 49% union are any different, on average, than those that voted 51%.
Except they are. Frandsen’s data show clear clustering of union elections on one side of the 50% mark: there are a lot more elections in which the union barely lost than those in which it barely won; in other words, close elections clearly favor employers.
This is no news for any labor organizer who’s worked an election. But it also violates the basic logic of RD, since RD rests on the premise of random selection: when you see a clustering of results on one side, that’s evidence the distribution of outcomes isn’t random, and that someone is gaming the system — say, employers figuring out they’re about to lose and bribing and intimidating their way to a win.
Thus elections on either side of the 50% mark aren’t really comparable, since the shops where the union barely lost likely had more aggressive, better resourced bosses than the ones where the union barely won. And those shops are likely to have different outcomes than those that unionized for a variety of reasons — the same reasons that allowed them to avoid unionization.
To his credit, Frandsen points all this out and considers all the problems it presents. And he offers a solution, an alternate route, using time series data: Rather than compare different shops where the union either won or lost, he tracks payroll changes over time in the same shops, pre- and post-unionization, using what’s called a difference-in-difference model. Notably, though, he keeps the same RD logic of restricting analysis to cases where the union barely won.
But there’s a problem with measuring the impact of unionization from close elections: Unions don’t like close elections. They invest a lot into trying to predict the outcome of an election ahead of time, and if they think they will lose, they will usually call off the campaign.
This study thus draws its conclusions from a very particular subset of elections: those that unions thought they might win but didn’t, which means the union was in a relatively weak position in terms of information and leverage over the employer. Cases in which unions pulled out of elections because they predicted defeat aren’t in the data, and cases in which the union won overwhelmingly don’t impact the results.
The weakness of unions in close elections inevitably factors into how good a contract the workers can get (if they even get one: employers have no legal obligation to bargain a contract following a successful union vote, which explains why wages don’t rise right after an election). Most of these elections almost certainly involve single-workplace organizing drives or hot-shopping, and not company-wide, multi-site elections that are the product of coordinated campaigns. And bargaining with small businesses, single shops within larger companies and companies that are in financial trouble all limit union leverage over wages.
I raised this issue with Frandsen via email. He writes,
It is correct that the regression discontinuity analysis applies to effect of unionizing drives that result in close elections, and are not directly informative about cases where the election was not close. It turns out, however, that close elections are actually quite typical.
There are indeed a significant number of elections where the union wins by a landslide, but if you look at figure 1 in the paper, the most common outcome is around 40% for the union. If I were to restrict attention to drives that were heavily targeted and resourced by larger national umbrella union organizations, it may well be that there would be fewer closely contested elections, but I’m interested in all unionizing drives, not just the heavily targeted and well-resourced ones.
In other words, the type of elections that most favor a strong contract — a union landslide — are not the norm, though there is a noticeable uptick in the number of elections where the union vote share approaches 100%. So while there are a fair number of landslides, a lot more elections are pretty close. And in most of them, the union loses.
This brings up a second point: by only looking at NLRB elections, Frandsen’s analysis leaves out card check and other neutrality union recognition deals, which are increasingly the preferred route to unionization.
Unions have long seen the NLRB election process as toothless in preventing worker intimidation by bosses, and thus either attempt to set limits on tactics like captive audience meetings or avoid the process entirely. Unions that have more clout are more likely to take the card check/neutrality route and are also more likely to bargain a good contract, which is often built into that neutrality deal.
Of course, most unionization drives don’t involve card check, because in most cases unions aren’t strong enough to get it. Frandsen replies that comparing cases in which unions did or did not get card check wouldn’t really work, since that would be “severely biased for the causal effects of unionization,” and thus impossible to isolate the impact of unionization alone — although his data show that comparing cases in which unions won or lost NLRB elections doesn’t work that well, either.
Third and most critically, the study measures wage gains/losses as the change in quarterly earnings the year immediately following the election. As DePillis notes in the Post, it takes most newly-unionized firms two years or more to get a first contract (again, assuming they get one at all).
Frandsen disputes this, noting that in his sample, between 40 and 50% of the worksites that unionized reported to the Federal Mediation and Conciliation Service that they had negotiated a contract within the first year — a percentage Frandsen suspects is low because of underreporting. Extending his sample to two years out is problematic, he writes, because by then, a significant number of firms start to disappear. And in any case, wages themselves are a moot point, since the drop in overall earnings he observes is affected by changes in the relative seniority of the workforce, not changes in wage rates.
But when workers are casting their ballots, they’re thinking how much their wages will rise, not the seniority of the workforce. And whether those raises in turn improve productivity, lower turnover and ultimately turn slackers into the kind of high-value workers Frandsen suggests are fleeing union shops in droves is absolutely dependent on having a contract. So looking at the effects of unionization when perhaps half of the shops don’t even have a contract doesn’t tell you much.
So what does Frandsen’s study tell us? A lot of what we already know: Unions tend to lose close elections, and when they barely win — when they don’t have the info to predict an election or the leverage to get neutrality, or even, in many cases, a contract at all — workers tend to get a raw deal. That deal tends to be a lot better in places where unions are strong, can run comprehensive campaigns, can get card check and quick first contracts, and in the few industries and sectors where density is above the single digits. Frandsen’s study points out, correctly, that such cases are the exception and not the rule. Why should any of this be a surprise?