The letter is archived at "Economist Statement on the Federal Minimum Wage," Economic Policy Institute, January 14, 2014, available at: http://www.epi.org/minimum-wage-statement/.
Many economists who currently support large minimum wage hikes claim that the best research now shows that such an increase would not cause significant drops in employment. However, their conclusion relies on a dubious reading of the literature. Dozens of recent empirical studies show significant employment reductions from minimum wage hikes, with some of these analyses using the newer "case study" approach, as opposed to the traditional regression analysis. Furthermore, serious methodological criticisms have been leveled against even the best of the studies used to justify increases in the minimum wage. Finally, even on their own terms, the studies purporting to show that the minimum wage is benign can justify only modest hikes: these studies' own results are consistent with the claim that aggressive minimum wage hikes will cause many unskilled workers to lose jobs.
As late as the 1980s, among economists the orthodox view—as crystallized in the literature review of Brown et al. (1982)—was that a 10-percent increase in the minimum wage would reduce employment of teenage workers by 1 to 3 percent.2
This consensus began to collapse in the 1990s. In a recent New York Times column, Paul Krugman explained why he changed his mind:
Until the  Card-Krueger study, most economists, myself included, assumed that raising the minimum wage would have a clear negative effect on employment. But they found, if anything, a positive effect. Their result has since been confirmed using data from many episodes. There's just no evidence that raising the minimum wage costs jobs, at least when the starting point is as low as it is in modern America.3
In the online version of his article, Krugman hyperlinked his phrase "been confirmed" to a 2010 paper by Dube, Lester, and Reich, which uses a clever technique to generalize the case study approach in the famous Card-Krueger study.4 Especially with Krugman's confident statement that there is now "just no evidence" of a disemployment effect, economists who are not familiar with the recent minimum wage literature—not to mention the general public—might conclude that the empirical case has been settled.
Even if this were true, and the minimum wage really didn't have a measurable impact on employment, it might still be a poor policy. For example, employers might compensate for the higher wage rate they are forced to pay workers by reducing fringe benefits. Moreover—and as I explained in a previous Econlib article5—a higher minimum wage might attract affluent new workers (such as college students) into the labor pool, displacing the originally employed, low-income workers who are the true targets of the policy.
However, in the present article, I focus on the narrow issue of the empirical estimates of the employment effect of a minimum wage hike. I cite the recent literature showing that this is still an open question even if we consider only modest hikes. Furthermore, even if we stipulate the results in the 2010 Dube et al. paper, we still should be wary of the popular minimum proposals, because the proposed increases are so large. Ironically, it would be more accurate to say that there is no evidence to justify large increases in the minimum wage.
In a comprehensive 2007 review, Neumark and Wascher summarized more than 100 studies published since the 1990s, from both the United States and abroad, and conclude that
the preponderance of the evidence points to disemployment effects.... Of [102 studies], nearly two-thirds give a relatively consistent... indication of negative employment effects of minimum wages, while only eight give a relatively consistent indication of positive employment effects.6
However, many economists believe that the best designed studies—specifically, those that construct a "control group" to isolate the impact of the treatment—tend to support the finding that there are modest (if any) disemployment effects. (See a 2014 article by Daniel Kuehn for a thorough presentation of this perspective.7) Therefore, when some economists today argue that the minimum wage has no ill effects, they must be assuming that the scores of studies to the contrary all suffer from the same methodological flaws that plagued the original consensus.
To tackle this more sophisticated position, in the remainder of this article I rebut the two specific examples of "good" studies that Krugman cited. Thus, it's not merely that there are many dozens of studies finding negative effects, but that there are serious problems with the purported evidence for a minimum wage increase.
The celebrated Card-Krueger (1994) paper did not use the traditional approach of relying on regressions conducted on large time-series of data. Rather, it was a case study, analogous to a medical trial with a "control group" and a "treatment group." In this analogy, the "medicine" was New Jersey's 1992 minimum wage hike, which raised the state's minimum wage from the federal level of $4.25 to $5.05. In contrast, adjacent Pennsylvania's government took no action. In this setting, Card and Krueger surveyed 410 fast food restaurants in New Jersey and Pennsylvania, both before and after New Jersey's 1992 hike. They concluded:
Relative to stores in Pennsylvania, fast food restaurants in New Jersey increased employment by 13 percent. We also compare employment growth at stores in New Jersey that were initially paying high wages (and were unaffected by the new law) to employment changes at lower-wage stores. Stores that were unaffected by the minimum wage had the same employment growth as stores in Pennsylvania, while stores that had to increase their wages increased their employment. [Card and Krueger (1994), emphasis added.]
The last sentence above underscores the apparent power of the result: If New Jersey had just happened to experience an economic boom relative to Pennsylvania right around the time of its unilateral 1992 minimum wage hike, then we would expect to see superior employment growth at all wage rates. Yet Card and Krueger found that New Jersey stores exhibited stronger relative growth only if the minimum wage was binding, that is, was above some of the wage rates that had previously been paid. This suggested that their findings involved causation, not mere correlation.
Notwithstanding the apparent strength of the Card-Krueger (1994) result, a more recent paper casts significant doubt on its wider relevance. Hoffman and Trace (2009) realized that a later episode would effectively provide a mirror-image case study.8 Specifically, between 1996 and 1997, the federal minimum wage was raised 90 cents (in two steps), from $4.25 to $5.15. Because New Jersey had already raised its own minimum wage to $5.05 back in 1992, this federal hike corresponded to an increase of 90 cents in Pennsylvania but only 10 cents in New Jersey. Hoffman and Trace analyzed whether this much larger effective raise in the minimum wage in Pennsylvania affected relevant employment growth compared to that in New Jersey. Looking at employment data from the Current Population Survey (not merely fast food workers) in 1995 and then in 1998, Hoffman and Trace found a negative employment effect in Pennsylvania among those workers that economic theory would predict to suffer, while Pennsylvania workers unlikely to be affected (because they were highly skilled and already were paid more than the new minimum wage) actually saw slightly higher employment growth.
To be sure, Hoffman and Trace (2009) isn't exactly the opposite of Card and Krueger (1994); one difference is that this study relies on statewide data rather than Card and Krueger's approach of using county-level data. Nonetheless, their study should give serious pause to anyone who thinks that the 1994 Card and Krueger study was a dagger to the heart of the traditional empirical findings.
Next, I turn to a methodological objection that critics have raised with the Dube et al. (2010) paper, which, as noted above, Krugman singled out when he claimed that the Card and Krueger results had been confirmed using data from "many episodes."
As I explained in my earlier Econlib article, Dube et al. (2010) claimed that the consensus findings had been systematically biased because, by throwing all of the data into one giant regression, economists were effectively treating all areas of the country as valid control groups for each other. But if states that raised minimum wages just so happened to be located in regions suffering from local economic shocks, then the coefficient on the minimum wage variable resulting from a traditional regression would be biased to indicate a greater negative impact on employment than really exists. In other words, if higher minimum wages happened to be correlated with other factors that reduced employment, and if those other factors weren't explicitly taken into account, then the standard regressions might be erroneously attributing employment losses to the minimum wage. It would be akin to running a simple regression with only a few variables and concluding that hospitals are very dangerous buildings because so many people die in them.
In order to persuade the reader that such omissions of other relevant variables were indeed at play, Dube et al. first reproduced the traditional results with modern data, showing the familiar result that the minimum wage seemed to "explain" some of the employment losses in the data. However, Dube et al. then began introducing controls for geographical trends, which had the effect of shrinking the apparent effect of the minimum wage. This was consistent with their hypothesis that historically, on average the states that raised their minimum wages just so happened to be hit by regional economic shocks. In their preferred specification—in which they cleverly matched contiguous counties that straddled a state border and could, thus, be subject to different minimum wages—Dube et al. actually found a slightly positive, though not statistically significant, impact on employment.
However, in a 2014 article, Neumark, Salas, and Wascher argued that we can empirically test whether these new techniques actually use better control groups than the traditional regressions.9 Neumark et al. drew on a broader literature on constructing "synthetic control groups." This literature shows ways to take a sample group with various attributes and build a proper "control group" according to the characteristics the investigator wishes to study. In general the technique will show how much weight to assign in the control group to various members of the overall sample population.
Relying on this literature, Neumark et al. argued that when constructing a control group to explain employment trends for a given U.S. state, the optimal "weights" placed on states within the same Census division are often not higher than the weights placed on randomly selected states. The following example (which is mine) helps explain their point: Los Angeles is, in many respects, a better "control city" than Green Bay, if one wants to evaluate mayoral policies in Chicago, even though Green Bay is much closer geographically. Neumark et al. thus undercut the whole motivation of the modern revisionist literature, which explains away the orthodox findings as due to inadequate control groups. If, in general, we have no strong reason to prefer geographically closer regions as control groups for employment trends, then there should be no major problem with the traditional regression analyses that tended to find significant employment reductions from the minimum wage.
To be sure, there continues to be a volley of arguments in the journals over this point, but the lesson for the lay reader is that cutting-edge researchers are still hotly disputing the interpretation of the data.
So far, I've shown that the alleged superiority of the technique in Dube et al. (2010) is in doubt. However, for the sake of argument, let us stipulate their results and focus on their own "preferred specification," in which they matched contiguous counties to construct what they believed to be an ideal control group. In this best-case scenario for the minimum wage, here are their results:
In our preferred specification[,]... we find that comparing only within contiguous border county-pairs, the employment elasticity is 0.016.... Bounds for this estimate rule out elasticities more negative than -0.147 at the 90% confidence level....10
It is difficult for a non-economist to parse such technical econometric talk, so I'll translate it into plain English. If we are discussing proposals to increase the minimum wage to $10.10, then Dube et al. are telling us that they are 95-percent confident that teenage employment will fall by no more than about 6 percent.11 If, instead, we consider the more aggressive proposals to raise the minimum wage to $15 per hour, then Dube et al.'s results assure us with 95-percent confidence that the hit to teenage employment will be no worse than about 16 percent. (!) These outcomes are hardly negligible, and they are fencing in a spectrum of bad outcomes, not just an isolated (and improbable) disaster. In other words, when we translate the quotation from above into plain English, we are not saying that there is just a small probability of an awful result, but that otherwise things are fine. Rather, Dube et al. are merely placing a ceiling on how bad the employment drop will likely be.
To see the big picture, note that the results in Dube et al.'s preferred model, which show a slightly positive employment effect from a minimum wage hike, are not statistically significant. This means that in their regression, there is so much variability in the data that they cannot be confident that the "true" value on the minimum wage variable is different from zero. Thus, to get a handle on our uncertainty, we can draw a "confidence interval" around their point estimate of 0.016, with the left end being lower and the right end being higher, such that we are 90 percent confident that the true value of the minimum wage coefficient falls within that range.
Now the interesting thing is that when we perform this exercise, it turns out that the lower bound falls smack dab within the traditional consensus. In terms of coefficients, the orthodox regressions found that the minimum wage variable fell in the range of negative 0.100 through negative 0.300, whereas (in the quotation above) Dube et al., in their preferred model, report that it's probably not worse than negative 0.147. This is neither the intellectual revolution nor the green light for policymakers that Krugman and others would have us believe.
In the 1980s, there was a genuine consensus that a 10-percent hike in the minimum wage would reduce teenage employment by 1 to 3 percent. However, in the 1990s, various "case studies" began challenging this orthodox view, and more recent studies have generalized techniques to apparently find negligible employment effects. Many economists have used this new research to assure policymakers and the public to pay no heed to warnings about harmful job losses from even aggressive minimum wage hikes.
However, in reality, the employment effect of the minimum wage is still an open question even for modest hikes. Since the 1990s, scores of articles have found negative effects of minimum wage increases. These include "case studies," with one serving as the mirror image of the famous Card and Krueger study. Furthermore, critics have challenged the entire premise of the new techniques, which claim to construct better control groups than the traditional approaches.
Finally, even if we take the very best examples of the "new" results at face value, they provide little comfort that large hikes in the minimum wage—such as a doubling to $15 per hour—will have modest impacts. Policymakers and the public should be wary of the glib assurances of some prominent economists when they claim that such large hikes will not cause teenagers to lose their jobs. The odds are very high that they will.
The letter is archived at "Economist Statement on the Federal Minimum Wage," Economic Policy Institute, January 14, 2014, available at: http://www.epi.org/minimum-wage-statement/.
Brown, Charles, Curtis Gilroy, and Andrew Kohen. (1982) "The Effect of the Minimum Wage on Employment and Unemployment," Journal of Economic Literature, Vol. 20, No. 2, June 1982, pp. 487-528.
Paul Krugman, "Liberals and Wages," New York Times, July 17, 2015, available at: http://www.nytimes.com/2015/07/17/opinion/paul-krugman-liberals-and-wages.html.
The two cited articles are: Card, David and Alan B. Krueger. (1994) "Minimum Wages and Employment: A Case Study of the Fast-Food Industry in New Jersey and Pennsylvania," American Economic Review, Vol. 84, No. 4, pp. 772-793. PDF available at: http://davidcard.berkeley.edu/papers/njmin-aer.pdf. Dube, A., T. W. Lester, and M. Reich. (2010) "Minimum wage effects across state borders: Estimates using contiguous counties," Review of Economics and Statistics, 92(4), pp. 945-964. PDF available at: http://www.irle.berkeley.edu/workingpapers/157-07.pdf.
Robert P. Murphy, "Economists Debate the Minimum Wage," Library of Economics and Liberty, February 3, 2014, available at: http://www.econlib.org/library/Columns/y2014/Murphyminimumwage.html.
Neumark, David and William Wascher. (2007) Minimum Wages and Employment. Foundations and Trends in Microeconomics, Vol. 3, No. 1-2 (2007), pp. 1-182. The block quotation is from pages 163-164.
Kuehn, Daniel. (2014) "The Importance of Study Design in the Minimum-Wage Debate," Economic Policy Institute, Issue Brief No. 384, September 4, 2014, pp. 1-13.
Hoffman, S. D. and D. M. Trace. (2009) "NJ and PA Once Again: What Happened to Employment When the PA-NJ Minimum Wage Differential Disappeared?" Eastern Economic Journal, 35(1), pp. 115-128.
Neumark, D., J. M. I. Salas, and W. Wascher. (2014) "Revisiting the minimum wage-employment debate: Throwing out the baby with the bathwater?" Ind Labor Relat Rev 67(3), pp. 608-648.
Dube et al. (2010), p. 953.
The reason I have translated the Dube et al. quotation of 90 percent into a claim about employment effects based on confidence of 95 percent is that a confidence interval of 90 percent constructed around a particular point estimate will have only 5 percent probability mass to the left (i.e. more negative) of the lower bound of the interval.