책 리뷰 : 자연 실험과 신뢰도 높은 통제군 , 최저임금이 미치는 고용 효과는 적다. Arindrajit Dube 두베 .

(1) 1995년 데이비드 카드와 앨런 크루거가 최저임금의 효과에 대한 책 “신화와 측정 : Myth and Measurement”을 출판해서, 경제학계에 충격파를 던짐. “신화와 측정” 결론에 따르면,  최저 임금 상승이 ‘해고’에 미치는 영향은 지극히 적다는 것이다.

무엇을 주목해야 하는가? 카드와 크루거가 채택한 방법론이 중요하다. 정책들의 급격한 변화를 설명하는데, 그들은  ‘자연 실험’(natural experiments) 사용했다는 점이다.그러나 ‘자연 실험 NE’은 그것을 연구하는데 사용된 ‘ 통제군-대조군 control group’이 좋아야만 한다. 통제집단이 좋으면 그만큼 자연적 실험 방법도 좋은 결과를 냄. 

(*주의: But a natural experiment is only as good as the control group used to study it. 

자연적 실험 NE에 대한 평가들

<1> 옹호자 입장- 지리적 근사치 (geographical proximity) 란 통제군을 구성하는 좋은 방식이다.

(뉴욕주는 실험군-처리군이고, 뉴욕주 경계 주인 펜실베니아는 통제군-대조군이 됨: 최저임금 영향에 대한 연구자들이 이 주장을 옹호) 

<2> 비판론자: ‘반사실: counterfactuals ’을 구성하는 적절한 방식인가에 대한 의견불일치 논란은 아직도 계속되고 있다.

(2) 카드와 크루거의 이론적  공헌: 노동시장에서  수요와 공급 사이 불일치 때문에 발생하는 ‘탐색 마찰 search frictions’의 역할을 집중 조명했다는 것이다 . ‘탐색 마찰들’이  최저임금 효과를 중재할지도 모른다는 주장들이 생겨났고, <신화와 측정> 출간 이후 많은 연구논문들이 이 주제를 다뤘다. 그 결과 최저임금 효과에 대한 논란들이 줄어들고 있다. 

[결과 요약] 탐색 마찰 (이직이나 구직 때문에 발생하는 에너지나 비용) 때문에, 최저임금이 인상되더라도 이게 곧바로 ‘해고 상승’으로 귀결되지 않을 수도 있다.

(3) 자연 실험과 신뢰도 높은 ‘통제 집단(군)’ 

카드와 크루거의 ‘자연적 실험’ 기여도: 명료한 처리군-실험군(treatment)과 신뢰할만한 통제군-대조군(control group)을 사용하고 규정하는 다양한 ‘설계 design’이 나오게 되었다. 카드와 크루거 자연실험 기여도:  1990년대  정책 연구 효과를 검증하는데 있어, 그 신뢰도를 높이는 ‘유사-자연실험 연구 설계’ 를 아주 다양하게 만드는데 기여했다. 

(4) 실제 카드와 크루거 연구 대상: 1992년 뉴 저지, 1988년 캘리포니아 최저임금 인상 영향 연구.

그 이후, 좋은 뉴스란 지난 25년간 미국 내 각 주별로 최저임금 차이가 커지지 시작했다. 29개 주는 연방 지정 ‘최저임금’보다 더 많은 최저임금을 지불하고 있다. 경제학자들에게는 다양한  ‘자연 실험’ 방법들 가동할 좋은 기회인 셈이다.  

그러나 나쁜 소식은 이러한 최저임금 효과를 연구하는 ‘자연 실험’은 ‘무작위 통제/대조군’ (*?) 통제 실험과는 다르다. 왜냐하면 각 주별로 최저임금 격차들이 있기 때문이다. 지난 30년간 가장 많은 최저임금을 받는 주들은 주로 뉴 잉글랜드 지역과 서부 해안가 주들이다.만약  저임금 고용 성장율을 비교하는데 남부 텍사스 주와 동부 메사추세츠 주를 비교하는 건 잘못된 결과를 도출해낼 수도 있다. 

텍사스 주 같은 경우, 멕시코에서 온 이민자들, 에너지 가격, 기호, 여러가지 경제적인 요소들을 고려한다면, 텍사스와 메사추세츠 주에서  고용 형태와 방법들이 서로 많이 다를 것이다. 이러한 차이들 때문에 카드와 크루거가 쓴 ‘자연 실험’ 방법들이 양적으로 다양해졌지만, 아직도 최저임금이 미치는 ‘고용’ 효과에 대한 논란은 지속되고 있다.  

(5) 두베 (Dube)는 어떻게 자연실험을 혁신시켰는가? 그렇다면 어떻게 해야 ‘합의’ 수준을 높일 수 있을 것인가? 

 ‘통제군/대조군control group’을 제대로 규정하는 것이 논란을 종식시키는 중요하다. 신뢰도를 높일 수 있다. 

연구자가 발견하고 하는 고용 효과 크기를 고려할 때, ‘특정화 오류 specification error’가 굉장히 커질 수 있고, 아마도  확률론적 성분들을 포착하는 ‘표준 오차’만큼 중요할 것이다. 

그렇다면 이러한 실수들을 줄이기 위한 방법: Dube 방식이란?

통제군-대조군을 어떻게 구성하는가? 2010년, 두베는 윌리엄 레스터와 마이클 라이쉬 T.William Lester and Michael Reich 등과 공동작업을 했다. 두베 (Dube)는 카드와 크루거의 통찰에 기초해서 그 방법을 아래와 같이 혁신시켰다. (*두베 주장) 

17년간 조사 대상주 경계 선상에 있는 서로 붙어 있는 ‘카운티 (한국으로치면 군, 읍, 적은 시 같은 행정구역)’들에  식당과 소매업을 연구했다. 지리적으로 근접한 도시들을 연구함. 

주 (state) 경계선 불연속 설계의 매력이란, 서로 연속으로 붙어있는 ‘근접 카운티들’은 원거리에 있는 카운티보다 서로 어떻게 변하는가를 더 잘 파악할 수 있다.  경계선상 불연속 설계는 또한 “관찰되지 않는 교란변수confounders”를 설명해줄 수 있고, 최저 임금 정책들의 “내생성 endogeneity( 변화/변수)"들을 설명할 수 있다.

(6) 그래서 우리가 발견한 것은 무엇인가?

식당 부문, 임금 탄력성은 약 0.2, 그러나 고용 탄력성은 거의 ‘0’에 가까웠다. 10대 고용 문제에 대해서도, 우리가 이 ‘경계 불연속 설계’ 연구 방법을 썼더니, 그 결과 역시 위와 비슷하게 나왔다.

또 하나 주목할 것은, 카드와 크루거의 “신화와 측정”에 대한 공통적인 비판들 중에 하나가, 중장기보다는 너무 단기에만 연구를 치중했다는 것이다. 그런데 우리 연구에 따르면 이런 ‘단기’ 문제는 치명적인 결함이 아니었다. 4~5년 정도 더 긴 시간들을 연구했을 때도, 우리가 발견한 것은 고용 추산은 그렇게 크지 않고 ‘적었다’는 것이다.

지난 10년간 ‘처리군,실험군 treatment ’와 ‘통제군-대조군(control group)집단’ 의 비교가능성을 향상하기 위해서 다양한 전략들을 사용해왔다. 이러한 전략들 중에는, 매개변수 parametric 경향 통제, 팩터 모델, 종합적 통제, 경계 불연속 접근의 일반화 등이 있다.

[소결론] 이러한 연구를 통해 나온 결과들에 따르면, 미국에서 최저임금 인상이 ‘고용 (시장)’에 미친 영향력을 그렇게 크지 않고 적다는 것이다.

당연히 최저임금이 미치는 ‘고용 효과’라는 이 주제는 서로 대립하는 증거들을 들이대면서 여전히 논란거리로 남을 것이다.

예를들어  Meer and West (2016)연구에 따르면, 최저임금이 총 고용에 상당히 부정적인 영향을 끼친다. 

그러나 Dube and Allegretto (2017) 연구에 따르면, 추정 실직은 임금 분배에서 더 높게 발생한다는 게 밝혀졌고, Meer 와 West 의 측정의 인과적 의미에 대해 의구심을 낳게 만들었다.

Clemens 와 Wither (2014)도 주장하길, 저임금 노동자들은2007년~2009년 미연방 최저임금 인상 적용을 받는 states 주 들에서 실업자가 될 확률이 높다고 했다. 그러나 Zipperer (2016)가 발견한 것은, 대공황의 지역적 특성을 고려했을 때,  이러한 추정(측정들)은 실질적으로는 더 적다는 것이다.

저임금 고용에 미치는 최저임금 인상의 영향력은 지금까지는 상당히 적은 것으로 드러났다. 지금까지 논의된 연구들은 다른 선진국에 비해서 미국의 경우, 최저임금이 상대적으로 낮았던 시기를 연구한 것이다. 따라서 최저임금이 인상된 이후  (버럭 오바마 이후)는 사정이 달라질 수도 있다.

참고 자료: 왜 캐나다 최저 임금이 낮은가?

출처: 

Arindrajit Dube myth.pdf

The Long-Run Impact of Minimum Wage Research: A Case Study of Myth and Measurement

 Arindrajit Dube University of Massachusetts Amherst

In 1995, David Card and Alan Krueger published Myth and Measurement (hereafter M&M) and sent shockwaves throughout the economics community. In that book, the authors forcefully argued that the evidence on the disemployment effect of minimum wages was surprisingly weak. 

Their own case study—which had just been published in the American Economic Review—compared fast-food restaurants in New Jersey and Pennsylvania after a minimum wage increase in New Jersey and found that, if anything, employment rose in New Jersey following the minimum wage hike. But M&M was more than just the New Jersey and Pennsylvania case study. It provided a vast array of empirical evidence and then went on to argue that the totality of evidence pointed toward the inadequacy of the simple supply-and-demand model for understanding the low-wage labor market. Instead, they argued that employers have some power to choose wage policies: paying a little bit more 

would improve the company’s recruitment and retention of workers but would mean higher labor costs due to paying more to infra-marginal workers. Card and Krueger called this the dynamic monopsony model and argued that it better accorded with the data. It has been an eventful 21 years since the publication of that book. So what have we learned from—and since—M&M? In this review, I highlight two points.

 First, Card and Krueger made an important methodological contribution in pushing the use of “natural experiments,” or sharp changes in policies. But a natural experiment is only as good as the control group used to study it. Here, their idea that geographical proximity is a good way to construct a control group has been vindicated by many studies, including ones looking at minimum wage impacts.

 At the same time, disagreement over the proper way of constructing counterfactuals continues to be a source of controversy in the literature. Second, the book made an important theoretical contribution in highlighting the role of search frictions in the labor market. The idea that such frictions may mediate minimum wage impact has been taken up by numerous papers since M&M and has become less controversial than at the time it was proposed.

Natural Experiments and Credible Control Groups Card and Krueger were among the pioneers of the credibility revolution of the 1990s, which encouraged the use of quasi-experimental research designs (Angrist and Pischke 2010). Such a design defines a clear treatment group and a reliable control group and studies the changes in outcome following treatment in these two groups. In the U.S. minimum wage context, a quasi-experimental design typically uses policy variation across states. 

It is not accidental that the new minimum wage literature of the 1990s emerged at a time when states started raising their minimum wages, with 11 states paying above the federal standard in 1995. This scenario created the opportunity to study natural experiments such as the 1992 increase in New Jersey or the 1988 increase in California, which were both analyzed in M&M. The good news is that state-level variation has increased over the past 25 years.

 Today, 29 states have a minimum exceeding the federal minimum wage—offering economists a rich set of natural experiments to study. The bad news, however, is that these natural experiments are not like randomized control trials; that is, they are not distributed randomly across the United States. Over the past three decades, states with the highest minimum wage increases have been concentrated in New England and on the West Coast.

 This strong regional component to minimum wage variation can result in very misleading inferences if we compare low-wage employment growth across, say, Texas and Massachusetts. Given migration from Mexico, differential reliance on energy prices, climate, and many other economically relevant factors, we might expect very different patterns in employment in those states quite apart from minimum wages. 

These differences underlie why the economics literature has continued to struggle in producing a consensus on the question of employment effect even with an increasing number of natural experiments to draw from. (For an example, see the exchange between Neumark, Salas, and Wascher [2014] and Allegretto, Dube, Reich, and Zipperer [2017].) 

Properly defining a control group is critical to obtaining a reliable answer to this question. Given the size of the employment effect one is trying to detect, the “specification error” can be quite large and is probably at least as important as the standard error capturing stochastic components. For example, a state-panel regression that assumes parallel trends in teen employment across U.S. states tends to find a sizable negative employment elasticity as large as –0.2. 

We have quite a bit of evidence, however, that points to a downward bias in the estimates from such a two-way fixed-effects approach (Allegretto et al. 2017; Manning 2016). For example, much of this putative job loss occurs many years prior to the actual change in policy, stretching the credibility of a causal interpretation. Similarly, controlling for state-specific trends often sizably affects the magnitude (or even sign) of the disemployment estimate. 

So, what are some better approaches to constructing control groups? My work with  T. William Lester and Michael Reich built on the insight of Card and Krueger by comparing restaurant and retail employment in contiguous counties across state borders and pooling more than 64 border segments with minimum wage differences over a 17-year period (Dube, Lester, and Reich 2010). 

The attraction of the border discontinuity design is that contiguous counties track each other much better than counties farther away, and the design provides a way to account for unobserved confounders and the endogeneity of minimum wage policies (Dube, Lester, and Reich 2016; Slichter 2016). What did we find? For the restaurant sector, we obtained a wage elasticity of around 0.2, but an employment elasticity close to zero. In more recent work using this border discontinuity design, we have found broadly similar results for teens (Dube et al. 2016).

 Note that although one of the common criticisms of M&M was that it considered only short-run responses, this turned out not to be a fatal flaw. Even when we considered longer-term effects (e.g., four or five year out effects), we found employment estimates to be fairly small (Dube et al. 2010; Allegretto et al. 2017). Over the past decade, we have seen the emergence of a range of strategies to improve the comparability of treatment and control groups. 

These strategies include the use of parametric trend controls, factor models, synthetic controls, and generalization of the border discontinuity approach (e.g., Addison, Blackburn, and Cotti 2015; Dube and Zipperer 2015; Totty 2015; Slichter 2016). In balance, the evidence appears to confirm that employment effects from minimum wage increases in the United States have been fairly small. To be sure, the topic of employment effect of minimum wages remains controversial, with sometimes conflicting evidence. My reading of the literature is that estimates suggesting large job losses have often turned out to be fragile or driven by questionable control groups. 

For example, Meer and West (2016) estimated a large negative effect of minimum wages on aggregate employment. Dube (2013) and Allegretto et al. (2017), however, provide evidence that the putative job losses are occurring higher up in the wage distribution, raising questions about the causal import of their estimates. 

Similarly, Clemens and Wither (2014) argued that low-wage workers were more likely to lose jobs in states bound by the 2007–2009 federal minimum wage increase; but Zipperer (2016) found that these estimates are substantially smaller when accounting for the regional nature of the Great Recession. Abstracting from the strengths and weaknesses of particular studies, however, meta- analyses of the U.S. evidence also suggest that the impact on low-wage employment to date has been quite small (Belman and Wolfson 2014; Wolfson and Belman 2016). 

Some evidence still persists of publication bias in the minimum wage literature that was found in M&M: studies suggesting negative employment effects are more likely to be published than would be expected based on chance alone. At the same time, the bias seems to have diminished over time, and M&M probably deserves some of the credit for that trend. Finally, keep in mind that most of the existing evidence comes from a period when U.S. minimum wages have been low by historical and international standards (Dube 2014). More recently, a growing number of cities and states have pushed minimum wages substantially higher. The effect of the policy at these higher levels remains an open question that is the subject of ongoing research.

Inadequacy of the Simple Supply-and-Demand Model of the Labor Market 

M&M argued that the simple supply-and-demand model of the low-wage labor market was inadequate for understanding a small or positive employment effect of minimum wages.

 They put forward a dynamic monopsony model of the labor market in which employers have some wage-setting power. Recruitment and retention both respond positively to wages, leading to a positive but finite labor supply elasticity facing the firm. They argued that empirically plausible magnitudes of this labor supply elasticity were consistent with a small positive or null effect of minimum wages on employment at firms in which the labor supply was the binding constraint. 

Of course, for other firms, labor demand may be binding, and these firms may reduce employment or go out of business. And eventually, if the wage is raised high enough, the latter scenario becomes increasingly likely. But over a range, the labor supply effect may largely offset the traditional labor demand effect, muting the overall impact on employment. Since 1995, thinking about the labor market in terms of search frictions has become much more standard; for example, the 2011 Nobel Prize in Economics went to Diamond, 

821 BOOK REVIEWS

Mortensen, and Pissarides “for their analysis of markets with search frictions.” Burdett and Mortensen (1998) formalized the dynamic monopsony model in an equilibrium context with search frictions and wage posting. Such a model can help us understand a variety of facts about the low-wage labor market: why similar workers are paid differently, why so much jobto-job mobility occurs, and why employment effects of minimum wages may be small. Indeed, in some cases, by compressing the wage distribution, minimum wage increases may actually improve the functioning of the labor market.

 My recent work with Michael Reich and William Lester (2016) also provides relevant evidence. We estimated the effect of minimum wages on separations and new hires, along with the effects on employment and wages. We found a striking pattern when we considered either a high-impact demographic group (teens) or a high-impact sector (restaurants): while the effect of minimum wages on employment is close to zero, both separations and new hires fell sharply in response to a minimum wage hike. 

This trifecta of results—strong positive wage effect, small employment effect, and strong negative turnover effect—is a signature of a model with search frictions and on-the-job search such as Burdett and Mortensen (1998). More generally, a growing number of recent papers have pointed to the importance of using richer models of labor market competition with search friction to understand the impact of minimum wages (Flinn 2006; Brochu and Green 2013; Giuliano 2013; Gittings and Schmutte 2016). 

Additionally, recent firm-level studies have found labor supply elasticities consistent with substantial wage-setting power, typically under 2 (Falch 2010; Staiger, Spetz, and Phibbs 2010; Dal Bó, Finan, and Rossi 2013; Dube, Giuliano, and Leonard 2015; Naidu, Nyarko, and Wang 2016). Ashenfelter, Farber, and Ransom (2010) edited a volume of the Journal of Labor Economics devoted to monopsony and concluded that “[t]he evidence surveyed from a fairly broad range of labor markets suggests that monopsony may be far more pervasive than is sometimes suggested.” Similarly, the fourth volume of the Handbook of Labor Economics devoted an entire chapter to “Imperfect Competition in the Labor Market” (Manning 2011).

 In many ways, M&M pointed to the shape of things to come. Standing the test of time is a challenge for any scholarly work, but especially so for a book as controversial as M&M. In his review of the book, the economist Finis Welch wrote: “I question David Card and Alan Krueger’s models and how they do empirical research. Although the notoriety surrounding [M&M] suggests important conclusions that challenge economists’ fundamental assumptions, I am convinced that the book’s long-run impact will instead be to spur, by negative example, a much-needed consideration of standards we should institute for the collection, analysis, and release of primary data” (Welch 1995). 

Twenty-one years later, I think it is safe to say that the book’s long-run impact has not been mainly on the standards for collection, analysis, and release of primary data. Rather, the book helped spur important innovations on both empirical and theoretical fronts and has furthered our understanding of how labor markets really work.

References