Equalizing Wage Differentials

Equalizing wage differentials MIT 14.662 Spring 2015 Outline (1) Preliminaries (2) Theory of equalizing differences (Rosen, 1974) (3) Estimating compensating differentials (Lucas, 1977; Brown, 1980; Summers, 1989; Fishback and Kantor, 1995) (4) Application: Spatial equilibrium (Roback, 1982) 1 Preliminaries Rosen (1986) provides a (very helpful, even if slightly dated) overview of equalizing differentials. The core idea of equalizing differences dates back to Chapter X of Adam Smith’s The Wealth of Nations, where Smith noted: “The whole of the advantages and disadvantages of the different employments of labour and stock must, in the same neighbourhood, be either perfectly equal or continually tending to equality.” Why is this an important model? Rosen (1986) argues that the theory of equalizing differentials can make a legitimate claim to be the fundamental (long-run) market equilibrium construct in labor economics; equalizing differentials is also a central model in urban economics. Empirically, the model has been extremely useful in understanding and interpreting the structure of wages, and for making inferences about preferences and technology from observed wage data. One reason we are interested in estimating equalizing differentials is that we often would like to know the value of non-market goods – which by definition aren’t priced. Estimates of the value of many non-market goods are extremely policy-relevant. As one example, estimates of the equalizing differential for mortality risks associated with on-the-job risk exposure can provide evidence on how individuals trade off earnings with risk, and can be developed into estimates of the value of a statistical life. Because a variety of government policies involve balancing increased costs with risk reduction, these value of statistical life estimates are directly applicable in a variety of policy settings. As a second concrete example - and to give some historical context for this literature I want to briefly discuss what is perhaps the first rigorous analytical treatment of equalizing differentials: Income from Independent Professional Practice, a 1954 NBER text by Milton 1 Friedman and Simon Kuznets.1 This book was Friedman’s dissertation, with was published with Kuznets (his adviser). The text is a detailed study of the income structure of five professions (medicine, dentistry, law, certified public accountancy, and consulting engineering). Following on Smith’s idea stated above - that the “whole of the advantages and disadvantages” of different occupations would continually tend toward equality for persons with similar ability, persistent differences in pecuniary returns would compensate for differences in training, the attractiveness of the work, the risks involved, etc. - Friedman and Kuznets argue that actual differences in income are a combination of such ‘equalizing’ differences, temporary differences mid-adjustment to changing economic conditions, and “persistent hinderances to the free choice of occupation.” What hinderances? The authors note that hinderances could arise from the “requirement of relatively rare abilities,” or they may be introduced by society. All of that language is somewhat talking around the (very controversial) conclusion that Friedman and Kuznets arrive at in their analysis. The case study they focus on is a comparison of incomes of physicians and dentists. One version of their research question is: “What factors explain the large and seemingly persistent difference between the average incomes of physicians and dentists living in the same community and in practice the same number of years?” (p.123). The authors argue that the two professions require somewhat similar abilities and training. Yet four times as many persons applied annually for admission to American medical schools as for admission to American dental schools; the authors argue that on the basis of these figures alone one can conclude that the observed difference in incomes is apparently greater than the ‘equilibrium’ difference. But from there they ask how much of the ∼ 32 percent difference in average incomes can be explained by “factors connected with the free and moderately rational choice of profession by prospective entrants” and how much to the “greater difficulty of entry into medicine.” Specifically, they investigate whether differences in length of training, variability of income, nonpecuniary factors (such as “prestige”), or demand can explain this income gap. They conclude instead that the divergence between the observed and ‘equilibrium’ difference is primarily attributable to the greater difficulty of entry into medicine than into dentistry not because of differences in required ‘innate ability,’ but rather because of a deliberate set of decisions by the American Medical Association (AMA) to artificially limit the number of licensed physicians in order to increase doctors’ earnings. From today’s perspective, it’s hard to appreciate how controversial this assertion was at the time; the publication of the book was delayed for several years due to objections from an NBER board member. Controversies over occupational licensing continue today, for everything from doctors to hairdressers.2 While it is difficult to draw firm conclusions from tabulations of descriptive statistics like those presented by Friedman and Kuznets, the types of data they presented was incredibly influential in spurring subsequent researchers to think about whether the distribution of incomes we observed is consistent with the compensating differentials we would expect to see in a market economy, or whether other factors (such as barriers to entry) 1 This text is available on the NBER website: http://www.nber.org/books/frie54-1. NPR’s Planet Money had an interesting story covering the latter example: http://www.npr.org/blogs/ money/2012/06/22/155596305/episode-381-why-its-illegal-to-braid-hair-without-a-license. 2 2 must be at play. 2 Theory of equalizing differences (Rosen, 1974) The classic reference on the theory of equalizing differences is Rosen (1974). Here, I’m going to focus on a simplified version of the model presented in Rosen’s 1986 Handbook of Labor Economics chapter (Rosen, 1986). 2.1 Rosen (1986): Binary case As in ‘standard’ models, prices in this model adjust to achieve market equilibrium. However, whereas in standard models the identity of the traders is immaterial, here the question of “which workers work for which firms” matters because equilibrium serves a matching or sorting function that allocates specific workers to specific firms. Labor market transactions are viewed as tied sales in which the worker simultaneously sells (rents) the services of her labor and buys the attributes of her job; likewise, employers simultaneously buy the services and characteristics of workers and sell the attributes of jobs offered to the market. The wage paid in equilibrium is therefore the sum of two transactions: one for labor services, and another for job attributes. In this sense, the labor market can be viewed as an implicit market in job and worker attributes. Both the theory and applications of equalizing differences are based on the assumption of perfect information on both sides of the market. Given perfect information, firms and workers are capable of making decisions that are privately optimal, and their choices generate supply and demand for each type of job. Rosen (1986) focuses on a binary case with two types of jobs: “clean” jobs (D = 0, no airborne particulates) and “dirty” jobs (D = 1, some airborne particulates). Wages w0 and w1 are paid to workers in each type of job. Assume that workers are productively homogenous, and differ only in their preference for D. 2.1.1 Preferences, opportunities, and worker choices Workers have preferences over two types of consumption goods: market consumption C (purchased with money) and job type D. Worker preferences are represented by a utility function U = U (C, D) where UC > 0 and UD ≤ 0. For a given value of C, assume that U (C, 0) ≥ U (C, 1) - implying that D = 1 is not preferred to D = 0, other things equal. Let C0 denote market consumption when D = 0. Given C0 , let C ∗ denote the consumption level required to achieve the same utility with a D = 1 job as C0 guarantees with a D = 0 job: U (C ∗ , 1) = U (C0 , 0). Given that D = 1 is not preferred to D = 0, C ∗ ≥ C0 . Define Z = C ∗ − C0 to be the compensating differential for D = 1 compared to D = 0 - that is, the additional consumption necessary to make the worker indifferent between the two types of jobs. 3 In Figure 12.1, the disamenity D is on the x-axis while the level of market consumption C is on the y-axis. Clean jobs offer (w0 , 0) while dirty jobs offer (w1 , 1).3 Workers may also be endowed with unearned income y. The vertical distance cb represents Z - the compensating difference (relative to w0 ) paid to workers who accept jobs with D = 1. Note that this utility curve is drawn for an inframarginal worker: she would be willing to work a D = 1 job at a wage premium of Z = ab, but the market clearing wage w1 is the vertical distance cb > ab. Courtesy of Elsevier, Inc., http://www.sciencedirect.com. Used with permission. The market equalizing wage differential is ∆w = w1 − w0 . Workers choose job type to maximize utility: workers choose D = 1 if ∆w > Z and choose D = 0 if ∆w < Z (workers are indifferent if ∆w = Z, so can choose by a coin flip). 2.1.2 Market supply Given the size of the labor force choosing between D = 0 and D = 1, relative market supply conditions are completely characterized by calculating the number of workers for whom ∆w > Z and for whom ∆w < Z. Note that ∆w is the same for all workers, whereas Z is a personal taste variable that varies across individuals. Workers maximize utility given ∆w and their Z. Let Z ∼ G(Z), so that g(Z) is the density function of Z in the population of workers. Define Ls1 to be the fraction of workers applying to jobs with D = 1 (that is, the fraction of workers with ∆w > Z), and Ls0 to be the fraction of workers applying to jobs with D = 0 (that is, the fraction of workers with ∆w < Z). Then we have: 3 Note that Rosen writes this as U (C, D) even though in Figure 12.1, D is on the x-axis and C is on the y-axis. 4 Ls1 Z ∆w = Ls0 = Z0 g(Z)dZ = G(∆w) (1) g(Z)dZ = 1 − G(∆w) (2) ∞ ∆w Figure 12.2 illustrates Ls1 and Ls0 for a given value of ∆w. Left of ∆w, workers choose D = 1 jobs; right of ∆w, workers choose D = 0 jobs. Courtesy of Elsevier, Inc., http://www.sciencedirect.com. Used with permission. 2.1.3 Technology, opportunities, and firm choices Firms sell a good x (output) to the market; normalize the price of x to be 1. D is produced as a by-product of the production of x (for example, the production of steel x involves smoke D). Firms can use resources to reduce D - for example, purchasing cleaner capital equipment. Note that D must be productive for it ever to be observed in the market. Let the production technology have the following form: x = a1 L if D = 1 (3) x = a0 L if D = 0 (4) Define B = a1 − a0 . Restricting B > 0 requires that D be productive: the efficiency of labor in x production is larger when resources are not used to clean up the work environment. B represents the marginal cost per worker of producing clean worksites, denominated in terms of forgone output. The marginal labor cost per worker of providing clean jobs is ∆w. In choosing their production technology, firms thus compare B with ∆w. 2.1.4 Market demand Take firm size as exogenous. Firms differ in B, the marginal cost per worker of producing clean worksites. Let B ∼ F (B), so that f (B) is the density function of B in the population of firms. 5 Firms choose D = 1 if B > ∆w and choose D = 0 if B < ∆w. Define Ld1 to be the fraction of firms demanding workers in jobs with D = 1 (that is, the fraction of firms with B > ∆w), and Ld0 to be the fraction of firms demanding workers in jobs with D = 0 (that is, the fraction of firms with B < ∆w). Then we have: Ld1 Z = Z d L0 = ∞ f (B)dB = 1 − F (∆w) (5) f (B)dB = F (∆w) (6) ∆w ∆w 0 Figure 12.3 illustrates Ld1 and Ld0 for a given value of ∆w. Left of ∆w, firms choose D = 0 jobs; right of ∆w, firms choose D = 1 jobs. Courtesy of Elsevier, Inc., http://www.sciencedirect.com. Used with permission. 2.1.5 Market equilibrium and selection Market equilibrium is defined by equality of supply and demand for workers in each type of job: wages w0 and w1 must adjust so that the number of workers seeking positions in each type of job equals the number of positions to be filled. That is, ∆w adjusts so that Ls1 = Ld1 (or, equivalently, Ls0 = Ld0 ). The equilibrium wage differential ∆w∗ thus solves G(∆w∗ ) = 1 − F (∆w∗ ). Workers and firms are systematically matched in market equilibrium: workers in D = 1 jobs have the smallest distaste for D, and firms offering D = 1 jobs have the largest costs of cleaning up their work environment. This produces negative assortative matching in equilibrium: workers with larger Z values are systematically found in firms with smaller B values, and vice versa. This can be formally stated as: E(Z|D = 0) > E(Z) (7) E(Z|D = 1) < E(Z) (8) E(B|D = 0) < E(B) (9) E(B|D = 1) > E(B) (10) 6 Importantly, the market wage difference ∆w is only telling us about the preferences of the marginal individual. We can define the term “rent” as - for a given individual - the difference between the reservation wage and the actual wage. Using this definition of rent, we can say that the market allocation may generate significant rents in the sense of the excess return relative to what would be required to change an individual’s decision. In Figure 12-2, the average person choosing D = 1 is far from indifferent, and would keep the same job choice even if ∆w changed substantially. Writing out the definition of rent as the difference between the reservation wage and the actual wage, we can note that for the average individual choosing D = 1, the average rent is ∆w − E(Z|D = 1). 2.2 Rosen (1986): Normally distributed case As an example, we can work through the case where Z is normally distributed in the worker population. Write Z = Z¯ + ν, where Z¯ is the mean value of Z and ν ∼ N (0, σ 2 ). As above, workers choose D = 1 if ∆w ≥ Z, implying that the share n1 of workers found in D = 1 jobs is: n1 = Pr(Z ≤ ∆w) (11) = Pr(Z¯ + ν ≤ ∆w) (12) ¯ = Pr(ν ≤ ∆w − Z)   ν ∆w − Z¯ = Pr ≤ σ σ   ∆w − Z̄ = Φ σ (13) (14) (15) In words, the share of workers found in D = 1 jobs at the relative wage ∆w is the cumulative standard normal evaluated at ∆w−Z̄ . σ Differentiating this expression for n1 with respect to ∆w and converting to an elasticity by multiplying by 1 = = = ∆w n1 gives: ∆w ∂n1 n1 ∂(∆w)   ∆w ∆w − Z¯ 1  φ Z̄ σ σ Φ ∆w− σ 
 ∆w−Z ∆w  φ  σ  σ Φ ∆w−Z̄ (16) (17) (18) σ Rosen derives an expression for the average value of Z among all workers found on D = 1 jobs. To derive his equations, we need to apply some properties of the standard normal distribution along the same lines as we used in our Roy model lecture notes. Specifically, if the standardized normal is truncated from above but not below (e.g. E(z|z < a) where z ∼ N (0, 1)), then E(x) simplifies to −φ(b) Φ(b) . Using this, we can calculate the average value of Z among all workers found on D = 1 jobs as follows: 7 E(Z|D = 1) = E(Z¯ + ν|Z̄ + ν ≤ ∆w) ¯ = Z¯ + E(ν|ν ≤ ∆w − Z)   ∆w − Z̄ ν ν ¯ = Z + σE | ≤ σ σ σ    φ ∆wσ−Z̄ ¯    = Z −σ ∆w−Z̄ Φ σ (19) (20) (21) (22) Define a “selection bias” term S1 to be the difference between the unconditional mean of Z and the conditional mean E(Z|D = 1); S1 is a measure of the extent to which preferences Z of people found on D = 1 jobs differ from the average preferences among workers in the population. We can derive an expression for S1 as: S1 = Z¯ − E(Z|D = 1) 
 φ ∆w−Z = σ   σ ¯ Φ ∆wσ−Z (23) (24) Above, we defined rents as excess returns relative to what would be required to change an individual’s decision. The average rent accruing to individuals choosing D = 1 jobs is then: R1 = ∆w − E(Z|D = 1) 
 ∆w−Z φ = ∆w − Z¯ + σ   σ ¯   Φ ∆wσ−Z = ∆w − Z̄ + S1 (25) (26) (27) The average rent among D = 1 workers is thus increasing in ∆w and in the variance of preferences. Note that as the variance in preferences σ 2 goes to 0, R1 goes to zero: with no heterogeneity, there are no inframarginal workers, implying there are no rents. More generally, the average rent is a function of σ, and thus depends on the distribution of preferences. 2.3 Rosen (1974): Continuous case The binary choice model illustrates the key issues that arise in models of equalizing differentials. Rosen (1974) presents a more general version of the model incorporating a continuous measure D of disamenity - say, parts per million of particulates as a continuous measure of pollution. The main insight is to point out that the market-clearing equilibrium W (D) wage-amenities locus is determined by the distributions of workers’ tastes and producers’ costs. Rosen refers to the Θ curves as bid functions (a formulation that had been and continues to be widely used in urban economics), and to the Φ curves as offer functions. Rosen’s theoretical framework has been extremely influential, but the identification approach 8 proposed in his original paper was heavily criticized. Kahn and Lang (1988) and Ekeland, Heckman and Nesheim (2004) discuss some of the major issues that have been raised; Bartik (1987) and Epple (1987) provide additional in-depth discussions. A variety of other econometric issues also arise when estimating compensating differentials; we’ll discuss these more below. 3 Estimating compensating differentials There has been a tremendous amount of interest in empirically measuring compensating differentials for at least two reasons: first, as a test of this theoretical model; and second, because in many contexts estimates of compensating differentials are quite policy-relevant. 3.1 Version 1.0: Cross-section estimates of compensating differentials As best I know, Lucas (1977) was the first paper to use a data source that subsequently became central to this literature: the Dictionary of Occupational Titles. The introduction of this paper has a great opening anecdote from Adam Smith’s observation that public hangmen received higher wages for their “obnoxious task.” The work underlying this paper was part of his 1973 PhD dissertation at MIT, supervised by Frank Fisher. At the time Lucas wrote this paper, the DOT data described occupational attributes for over 14,000 jobs - whether the job requires working in high temperatures, dust or fumes, high noise levels, hazardous conditions, the level of strength required, the length of vocational preparation required, whether the job is repetitive, whether the job requires making judgements, etc. Lucas combined this DOT data with the Survey of Economic Opportunity (SEO) data on hourly earnings in order to estimate cross-sectional regressions of how wages varied with job characteristics. That is, his goal was to investigate how wages vary with indicators of the quality of working life by including these job characteristics in standard wage equations. 9 Reading the paper, you can get the sense that combining these two data sources wasn’t easy. The SEO data didn’t use the DOT classification system for occupations, but rather used the US 1960 five-digit Census classification scheme (with 294 categories). Lucas was able to use a cross-walk compiled by the US Department of Labor to match these two types of codes. This highlights a useful general theme: before embarking on a month-long (or many-month long!) effort to generate a cross-walk file, try to ascertain that no such cross-walk exists! 10 Rosen’s cross-sectional specification regresses log wages on a piecewise linear function for age, an indicator for union status, and a series of job characteristics. The results are presented in Table 1. Lucas notes that the union wage premium (which he and others had documented in earlier work) persists after conditioning on these job attributes. Jobs requiring significant vocational preparation (denoted SVP in the table) are associated with a wage differential on the order of 25 percent. Although some of the coefficients for other job characteristics are of the expected sign - for example, repetitive work appears to be compensated (“Repetitive”) others are not: jobs requiring physical strength are generally associated with lower (not higher) wages (“Nonsedentary”), which Lucas notes “suggests the omission of some skill associated with sedentary job holders.” Courtesy of Robert Lucas and the American Economic Association. Used with permission. 11 3.2 Version 2.0: Panel estimates of compensating differentials In 1980, Brown (1980) reviewed the cross-sectional evidence of the type presented by Lucas and concluded there was surprisingly limited support for the theory of equalizing wage differentials: “The overall pattern that emerges...is one of mixed results: some clear support for the theory but an uncomfortable number of exceptions. Among the studies that fail to find equalizing differences, the most common explanation is the omission of important worker abilities, biasing the coefficients of job characteristics.” (See Brown’s Table 1 for a summary of references and estimates.) The goal of the Brown paper was to move beyond these cross-sectional estimates to develop “a more appropriate test of the theory.” He notes that the focus of the Rosen model is on choices made by individuals with given personal characteristics (X) among jobs with different wages (w) and non-wage attributes (Z). In order to attract labor of a given quality, an employer offering jobs that are hazardous or otherwise undesirable must pay higher wages than employers offering jobs with more desired nonwage characteristics. Individuals face a set of jobs with differing combinations of w and Z and chooses among these jobs to maximize utility. Let larger values of Z - a vector of non-wage attributes - represent less desired jobs, where “less desired” reflects the preferences of the marginal individual. The empirical goal is to relate log wages to non-wage job characteristics, conditioning on X’s that isolate “similar workers” in terms of productivity, to test the theoretical prediction that δw δZj > 0 for all j. However, we will likely have some unobserved individual productivity attributes, such as (say) unmeasured ability Ai . The true equation would therefore be: ln(wi ) = α + βXi + γZi + δAi + ui (28) We expect, based on the Rosen framework, that worse job characteristics should be associated with higher wages, so that γ > 0. We can also assume by construction that higher ability Ai is associated with higher wages, so that δ > 0. If workers with higher ability use some of their higher earnings capacity to “purchase” better working conditions, then cov(Zi , Ai ) < 0. We would thus expect the estimate of γ to be biased downwards. As noted above, the cross-sectional empirical estimates suggest this direction of bias may be a problem, because many estimates of compensating differentials are “wrong-signed” (negative). Hwang et al. (1992) assess how much unobserved productivity biases estimates of compensating wage differentials, and conclude that these biases may be large. Brown (1980)’s insight was that some progress can be made on this issue to the extent that most of the omitted variables are fixed within individuals over time, in which case longitudinal data (an individual fixed effects model) can be used to control for unobserved individual differences in productivity. He uses the NLS Young Men’s survey, combined with the same DOT data used by Lucas, to estimate how earnings change when job characteristics change, conditional on individual-level fixed effects. Of course, some determinants of earnings capacity do change over time: age, work experience, job training, marital status, etc. The hope is that controlling for 12 individual fixed effects reduces the omitted variables bias problem even if it does not eliminate it. In terms of his regression specifications, in addition to year fixed effects (to capture e.g. the effect of price inflation on wages) Brown includes variables for human capital investment (company training, part-time school courses, and “other”), cumulative work experience, tenure with current employer, three measures of unionization, marital status, three geographic controls, a measure of involuntary job separation, and a measure of health. He first shows four specifications (Columns 1-4) that include these controls in addition to a ‘standard’ set of individual covariates: e.g. race, SES, schooling; he then shows these same specifications replacing individual covariates with individual fixed effects. Columns 1-4 differ based on the job characteristics they include. Table II presents Brown’s results, which still give many “wrong-signed” estimates relative to what is predicted by the theory. In particular, including individual fixed effects does not result in a marked improvement in the correspondence between these coefficients and a priori predictions. He concludes: “The hypothesis that the inconsistent support for the theory of equalizing differences that characterized previous studies was due to the omission of important dimensions of worker quality was not supported by the data.” © Oxford University Press. All rights reserved. 13 © Oxford University Press. All rights reserved. 14 Brown summarizes his results as: “The impacts of the intercepts on the coefficients of job characteristics vary considerably, and there is no marked improvement in the correspondence between these coefficients and a priori predictions.” In his conclusion, Brown discusses several potential explanations for his results: 1. Labor markets are not as competitive as the theory of equalizing differences assumes. Are the assumptions of perfect information profit- and utility-maximization that underlie Rosen’s model too strong? Brown argues that considerably weaker assumptions still imply such differences. 2. Marginal workers’ tastes may differ from those assumed in a priori sign of the coefficients. Brown argues this is plausible for the physical strength variable, but less so for other variables. That is, while some workers may wish to avoid physical labor, other workers may prefer physical labor to sitting at a desk all day, in which case jobs requiring physical strength may not be unpleasant for the marginal worker and no equalizing difference may be required. However, this story seems unlikely to hold for alternative factors like the risk of death on the job (!). 3. Job characteristics are not well-measured. This is likely a huge issue. Use of the DOT data introduces measurement error because workers are assigned the mean job characteristics for her occupation/industry, not the characteristics specific to her job. Some other studies use self-reported survey measures of job characteristics, but these measures are thought to be quite unreliable (see the discussion in Rosen (1986)). If this measurement error is accurately characterized as classical measurement error, this will bias γ towards zero. Note that this measurement error problem will likely be exacerbated in individual fixed effects models. 4. Omitted variables - both individual- and job-level - may be biasing the estimates. Brown highlights freedom to work overtime as one potentially important omitted job characteristic. In terms of individual-level characteristics, an individual fixed effects specification solves some problems but introduces others: why do some individuals change jobs but not others? Could job changes be driven by changes in unobserved individual-level productivity? Are job changes driven by poor “matches,” which could induce bias? 5. Testing the model on twenty-somethings is inappropriate. He limits the sample to older workers, and the estimates don’t change remarkably, so this doesn’t appear to be a major issue. The Murphy and Topel (1987) chapter on the syllabus applies an individual fixed effects approach to explore inter-industry wage differentials. Several big-picture issues come up in Version 1.0 and Version 2.0 of this empirical literature on equalizing wage differentials. First, unobserved factors are a huge issue. The Brown paper attempts to address unobserved individual-specific factor by controlling for individual factors 15 that don’t vary over time, but obviously this approach doesn’t address all omitted variables bias concerns and it introduces several new problems: measurement error may be exacerbated in firstdifferenced or fixed effect models (you’ll work through this on a problem set), and endogenous mobility may introduce problems. Second, measurement error in the job characteristics data is likely a major issue. Both of these concerns motivated researchers to search for alternative methodologies to estimating compensating wage differentials. 3.3 Version 3.0: Estimating compensating differentials using policy variation The cross-section and panel methods for estimating compensating differentials produced a variety of non-sensical and “wrong-signed” estimates on how workers value non-wage job characteristics. While we previously focused on risks such as risk of death on the job, a different type of non-wage job characteristic is a non-wage benefit such as health insurance. Do workers value health insurance one-for-one for wages? What about other types of benefits, such as workers’ compensation insurance? In a cross-section, these questions are hard to answer for the standard omitted variables reasons. To address those challenges, this literature moved towards examining mandated employer-provided benefits - such as for health insurance and workers’ compensation - as a source of quasi-experimental variation. 3.3.1 Employer-mandated benefits: Summers (1989) A famous American Economic Review Papers & Proceedings article by Larry Summers (Summers, 1989) laid out a framework clarifying how to think about the incidence and welfare costs of employer-mandated benefits. Importantly, Summers showed that mandated benefits may have different effects on equilibrium wages and employment than do taxes: the labor market equilibrium depends on how workers value the benefits. Let labor demand be Ld = fd (W + C) where W is the wage and C is the cost of the mandated health insurance. Let labor supply be Ls = fs (W + αC), where αC is the monetary value that employers place on health insurance. In equilibrium, Ls = Ld . Letting η d denote the labor demand elasticity and η s denote the labor supply elasticity, you can show by totally differentiating the equilibrium conditions that if α = 1 (that is, if employees value benefits as if they were wages) then wages fall by the full cost of the mandated benefit, while there is no change in employment. The problem set will have you work through a version of this model. In public finance, this is referred to as the tax-benefit linkage. Importantly, there could be no deadweight loss from this if workers fully value the benefit. This Summers framework spurred an empirical literature analyzing how wages and employment respond to changes in mandated employer-provided benefits. Gruber and Krueger (1991) analyze the impact of increases in mandated workers’ compensation insurance, finding evidence of substantial wage offsets; Fishback and Kantor (1995) analyze the initial passage of workers’ compensation laws, finding similar evidence. Gruber (1994) analyzed the impact of state and federal requirements mandating insurance coverage of maternity care, finding evidence of wage 16 offsets among workers who would have generated the increased insurance costs. Gruber (1997) analyzes the incidence of payroll tax changes, incorporating the Summers (1989) insight that payroll tax revenues are often used to finance programs that benefit only workers - such as Social Security or workers compensation - implying a tax/benefit linkage with payroll taxes. Gruber finds that reductions in payroll taxes appear to have been fully passed along to workers, with little effect on employment levels; although this is consistent with full employee valuation of benefits, Gruber is unable to test this possibility against competing explanations. 3.3.2 Workers’ compensation: Fishback and Kantor (1995) I want to discuss the Fishback and Kantor (1995) paper as an example, which is an excellent example of how economic history can again be a good laboratory for labor economics research questions. State-level workers’ compensation was introduced in the 1910s. Fishback and Kantor give a useful discussion of the political economy of these policies: social reformers widely touted the introduction of these programs as bonuses to workers, whereas the literature cited above suggests that the de facto redistribution of income caused by employer-mandated benefits may differ in important ways from that view. In order to investigate the extent to which these mandated benefits were passed along to workers in the form of lower wage rates, Fishback and Kantor collect pre- and post- wage data from three (relatively dangerous) industries: coal mining, lumber milling, and the unionized building trades. 17 Table 2. Table 2 of their paper shows the variation in expected workers’ compensation as a percentage of annual earnings by state, 1910-1923. The main thing to note is the staggered timing, although adoption was relatively rapid. Within a year, the level of the benefits also varied across states. Not the most easy-to-interpret way of illustrating this variation, but it can be hard to summarize this type of variation. © Oxford University Press. All rights reserved. 18 © Oxford University Press. All rights reserved. 19 For each of their three industries, they estimate regressions of the following form for occupation i in state j in year t: wijt = β0 + Dijt β1 + Bijt β2 + W Tijt β3 + Aijt β4 + Uijt β5 + oi + sj + yt + eijt (29) where wijt is the real hourly wage, Dijt are measures of product market fluctuations and worker productivity (average), Bijt are postaccident benefits, W Tijt are working time restrictions, Aijt is the accident rate, Uijt are measures of strikes and union strength, sj are state fixed effects, yt are year fixed effects, and oi are occupation fixed effects. If we were estimating this today, we would show specifications both with and without these controls, and show some timing/event studies around the time of the policy change (but those graphs were not the ‘norm’ at the time this paper was written). 20 Table 3. The estimates are shown in Table 3. Looking at a 0/1 variable for any workers’ compensation, we see wage declines that are statistically significant for coal mining and lumber mills. The measure of compensation generosity pains a similar picture; given that a coefficient of -1 implies that workers fully paid for increases in the expected benefits, Fishback and Kantor can’t reject full pass through for the non-unionized industries. These wage offsets are roughly similar in magnitude to Gruber and Krueger (1991)’s analysis of workers compensation in more recent data. Interestingly, wage offsets are much smaller for the unionized sector. © Oxford University Press. All rights reserved. 21 Source: Fishback and Kantor (1995). © Oxford University Press. All rights reserved. 3.4 Version 4.0: Estimating compensating differentials using job offers A variant of Version 2.0 - the individual fixed effects approach - which makes progress on holding unobserved individual productivity fixed is a method developed by Stern (2004). Stern’s insight was that in entry-level labor markets many individuals receive multiple job offers before accepting any specific offer. Each offer is composed of a wage and a series of job characteristics. By observing individuals’ accepted job relative to the wage-amenity characteristics of other jobs in their offer set, it is possible to estimate compensating differentials in a way that overcomes many of the problems inherent in the individual fixed effects approaches that rely on job changes. We will talk more about this approach in an upcoming lecture. 22 4 Application: Spatial equilibrium (Roback, 1982) As an application, we are going to focus on the Roback (1982) model (sometimes referred to as the Rosen-Roback model), which is motivated by wanting to understand the implicit prices of various urban attributes. The Roback model is arguably the model of urban economics, and one that has been very widely applied to a range of economic questions. 4.1 Motivating facts Roback motivates her paper by saying that a long history of (at the time) previous work had observed intercity and regional wage differences among workers with the same measured productivity (as proxied by education and experience). For example, Table 4 of Fuchs (1967) documents regional differences in wages across observably similar workers, with wages in the Southern US tending to be lower than in other regions.4 © Victor Fuchs. All rights reserved. You may look at this and think: why is it surprising that workers in different parts of the country with the “same” potential earnings (in a human capital sense) earn different wages? Surely the amenities available differ across areas, and it’s not like we expect earnings to equalize across areas, we expect utilities to equalize across areas. My sense from the Fuchs book and the Roback paper was that that was not what people were thinking at the time – they instead saw these regional wage differences as a “puzzle.” Similarly, Table 8 of Fuchs documents significant differences in wages across observably similar workers between cities of different sizes. 4 This text is available on the NBER website: http://www.nber.org/books/fuch67-1. 23 © Victor Fuchs. All rights reserved. My sense is that prior to Roback’s analysis, most investigations of these types of regional and city-size wage differentials focused solely on the consumer side of the market without taking into account the behavior of firms. In contrast, Roback’s contribution was to conceptualize the equilibrium decisions of both firms and workers. Her motivating example is the following: if workers require a compensating wage differential to live in a big, polluted, or otherwise unpleasant city, the firms in that city must have some productivity advantage that allows them to pay the higher wage. Her model applies Rosen’s key insight: the implicit prices of attributes represent both the marginal valuation to consumers and the marginal cost to firms; focusing only on the consumer side will provide an incomplete picture. The contribution of Roback’s model relative to the previous literature was two-fold. First, the idea that in a spatial equilibrium both land and labor markets must clear had not previously been well-understood. Second, the factors that influence the decomposition of the implicit prices of attributes into wage gradients vs. rent gradients had not been analyzed. Intuitively, in order for utility to be equalized across locations, either wages must be lower or costs of living must be higher in more amenable locations; but in the absence of a model, it isn’t clear whether utility equalization across locations should show up as wage differences, or as differences in site-specific costs of land or housing, or both. The Roback model is designed to guide our thinking on both of these issues by presenting a general equilibrium model that incorporates both mobile factors (labor) and site-specific factors (land). The key difference between Roback’s model and Rosen’s model is that in Roback’s model, both land and labor markets must clear. In Rosen’s model, if firms are homogenous and workers are homogenous then exactly one wage-amenity bundle is offered in equilibrium. However, in Roback’s spatial allocation problem people cannot all occupy the same space even if they have identical preferences; the scarcity of land gives rise to an additional constraint on the problem. Roback solves for equilibrium wages and rents. 24 4.2 Model Let s index a vector of amenities that differ across cities but are constant (and fixed) within cities. The index s can include a variety of amenities - climate, pollution, crime - and is indexed such that a higher level of s is preferred by workers. Each city has a wage rate w and a price of land r. Roback focuses on the case where workers are identical and firms are identical. Workers and firms are perfectly mobile across cities (no moving costs), but must live and work in the same city. Land is fixed, although Roback notes that this can be relaxed to an assumption that there is a rising supply price of land (for example, due to within-city transportation costs). In the worker’s decision problem, leisure is ignored and each person supplies a single unit of labor (independent of the wage rate). Conditional on any given location with a particular level of s, workers choose consumption over a composite commodity x (sold at price 1) and residential land lc (sold at price r) subject to the budget constraint: maxx,lc U (x, lc ; s) subject to w + I = x + lc r, where I represents non-labor income. In spatial equilibrium, we can re-write this as an indirect utility function V (w, r; s) = k for a constant k. That is, wages and rents must adjust to equalize utility in all locations - otherwise individuals would have an incentive to move. Because a higher level of s is preferred by workers, ∂V ∂s > 0. In the firm’s decision problem, a constant returns to scale production function is assumed: x = f (lp , N ; s) where lp is land used in production and N is the total number of workers in the city. Note that this assumes that production requires physical space; it might be interesting to look at the location decisions of firms that allow employees to work remotely (or which by construction involve off-site work). Firms minimize costs subject to the production function and sell x on the world market at a fixed price, normalized to 1. Unit costs at each location must equal the price: C(w, r; s) = 1 = p. The unit cost function is increasing in both factor prices. If the amenity is unproductive, then Cs > 0 (unproductive amenities raise the cost of production).5 Roback’s example of an unproductive amenity is “clean air,” which workers value but which makes firm production more costly; her example of a productive amenity is “lack of severe snow storms,” because blizzards are both costly to firms and unpleasant to consumers (definitely true based on my former life as a North Dakotan, and winter 2015 in Boston). The indirect utility function and unit cost condition generate equilibrium levels of wages w(s) and rents r(s) for a given level of k, where k is determined by aggregate labor demand and aggregate labor supply. Roback illustrates the equilibrium in Figure 1. 5 Note that this seems to be a typo in the original paper, where Roback says that if the amenity is unproductive then Cs < 0. 25 © The University of Chicago Press. All rights reserved. In Figure 1, s is an unproductive amenity - say, clean air: workers prefer clean air but clean air raises firms’ costs because they must invest in pollution abatement. Two sets of curves are drawn: • Firm cost curves. The downward-sloping lines are combinations of wages w and rents r that equalize unit costs at a given level of s. • Worker indirect utility curves. The upward-sloping lines are combinations of wages w and rents r that equalize indirect utility at a given level of s. The results underlying Figure 1 can be obtained by deriving expressions for dw ds and dr 6 ds . Totally differentiating the indirect utility function V (w, r; s) = k gives vw dw + vr dr + vs ds = 0. Totally differentiating the cost function C(w, r; s) = 1 = p gives cw dw + cr dr + cs ds = 0. We can combine these equations to derive the following expressions for dw ds dr ds = = dw ds and dr ds : cs vr − cr vs cr vw − cw vr cw vs − cs vw cr vw − cw vr (30) (31) The denominator is the same in both expressions: cr vw − cw vr . We know that vw > 0, cr > 0, cw > 0, and vr < 0; taken together, these imply cr vw − cw vr > 0. For dw ds , we can sign the numerator because we know that cs > 0, vr < 0, cr > 0, and vs > 0, implying that – wages decline as s increases. For dr ds , dw ds <0 the numerator can’t be unambiguously signed: all four of cw , vs , cs , and vw are > 0, so the overall sign of the numerator will depend on the relative magnitude of cw vs relative to cs vw . Hence, rents may either increase or decline as s increases. 6 Note that in this derivation, Roback ignores elasticities of substitution because she focuses on small changes. 26 Thus, in locations with higher levels of unproductive amenities, wages should be lower, but the change in rents is uncertain. Intuitively, because s is unproductive, firms prefer low s locations whereas workers prefer high s locations. Because high rents discourage both firms and workers from locating in an area, worker equilibrium requires high rents in high s areas to discourage immigration, while firm equilibrium requires low rents in high s areas to induce firm location. On the other hand, low wages discourage workers and attract firms. The factor prices equalize to strike a balance between the conflicting locational preferences of firms and workers. If s were productive instead of unproductive, rents would rise whereas the change in wages would be ambiguous. Big-picture, the key take-away is the the extent to which wages vs. rents will adjust depend on whether workers, firms, both, or neither value a particular amenity. 4.3 Data Roback analyzes the 1973 May CPS data for the 98 largest US cities, and limits the sample to men over the age of 18 who reported earnings. She uses a publication called FHA Homes as a source of residential site prices - specifically, average site prices per square foot. This data is only available for 83 of the 98 cities, and over-samples lower income households (because the data only includes FHA-eligible families). 27 4.4 Results Table 1. Regressions of log(earnings) on city characteristics. Note the vintage table with t-statistics instead of standard errors :) The paper has notes on the data sources for these variables in Appendix B, but includes essentially zero discussion of how the covariates were chosen. Some of these are fixed area characteristics (climate variables), whereas others could be analyzed in a panel setting. She does not give much help on distinguishing which should be considered productive versus unproductive amenities (or how we might test that). It seems like more on that type of discussion would make it easier to learn something concrete from this table. © The University of Chicago Press. All rights reserved. 28 Table 2. Asks what influence city attributes have on regional differences in earnings. Although there are strong regional differences in wages, once city characteristics are conditioned on these regional differences largely disappear (except for the WEST coefficient, which is still statistically significant and negative). © The University of Chicago Press. All rights reserved. 29 Table 3. Analogous to Table 1, except the left-hand-side variable is now log(price per square foot) instead of log(earnings). As in Table 1, she does not give much help on distinguishing which should be considered productive versus unproductive amenities (or how we might test that). © The University of Chicago Press. All rights reserved. 30 Table 4. Combines coefficients from Tables 1 and 3 as well as the budget share of land in order to compute the implicit price for each attribute in percentage terms. Negative numbers indicate “bads” and positive numbers indicate “goods.” Table reports marginal price of the amenity evaluated at average earnings. For example, the average person would be willing to pay $69.55 per year for an additional clear day. © The University of Chicago Press. All rights reserved. 4.5 Take-aways The Roback model is an incredibly important contribution in terms of providing a concrete framework for analyzing spatial equilibrium, and for thinking about whether a given amenity is likely to be capitalized into rents, wages, or both. Practically, this model has been applied to thinking about quality of life indices across cities, as well as to a variety of applications in urban and labor economics. This is a useful model to have in your toolkit. References Bartik, Timothy, “The estimation of demand parameters in hedonic price models,” Journal of Political Economy, 1987, 95 (1), 81–88. Brown, Charles, “Equalizing differences in the labor market,” Quarterly Journal of Economics, 1980, 94 (1), 113–134. Ekeland, Ivar, James Heckman, and Lars Nesheim, “Identification and estimation of hedonic models,” Journal of Political Economy, 2004, 112 (S1: Papers in Honor of Sherwin Rosen), S60–S109. Epple, Dennis, “Hedonic prices and implicit markets: Estimating demand and supply functions for differentiated products,” Journal of Political Economy, 1987, 95 (1), 59–80. 31 Fishback, Price and Shawn Kantor, “Did workers pay for the passage of workers’ compensation laws?,” Quarterly Journal of Economics, 1995, 110 (3), 713–742. Fuchs, Victor, “Differentials in hourly earnings by region and city size, 1959,” 1967. NBER occasional paper #101. Gruber, Jonathan, “The incidence of mandated maternity benefits,” American Economic Review, 1994, 84 (3), 622–641. , “The incidence of payroll taxation: Evidence from Chile,” Journal of Labor Economics, 1997, 15 (S3), S72–S101. and Alan Krueger, “The incidence of mandated employer-provided insurance: Lessons from workers’ compensation insurance,” in David Bradford, ed., Tax Policy and the Economy Volume 5, 1991, pp. 111–143. Hwang, Haeshin, W. Robert Reed, and Carlton Hubbard, “Compensating wage differentials and unobserved productivity,” Journal of Political Economy, 1992, 100 (4), 835–858. Kahn, Shulamit and Kevin Lang, “Efficient estimation of structural hedonic systems,” International Economic Review, 1988, 29 (1), 157–166. Lucas, Robert, “Hedonic wage equations and psychic wages in the returns to schooling,” American Economic Review, 1977, 67 (4), 549–558. Murphy, Kevin and Robert Topel, “Unemployment, risk, and earnings: Testing for equalization,” in Kevin Lang and Jonathan Leonard, eds., Unemployment and the Structure of Labor Markets, 1987, pp. 103–140. Roback, Jennifer, “Wages, rents, and the quality of life,” Journal of Political Economy, 1982, 90 (6), 1257–1278. Rosen, Sherwin, “Hedonic prices and implicit markets: Product differentiation in pure competition,” Journal of Political Economy, 1974, 82 (1), 34–55. , “The theory of equalizing differences,” in Orley Ashenfelter and Richard Layard, eds., Handbook of Labor Economics, Volume 1, 1986, pp. 641–692. Stern, Scott, “Do scientists pay to be scientists?,” Management Science, 2004, 50 (6), 835–853. Summers, Lawrence, “Some simple economics of mandated benefits,” American Economic Review Papers and Proceedings, 1989, 79 (2), 177–183. 32