New evidence on workers' willingness to pay for job attributes.

Reed, W. Robert

I. Introduction

This paper reports the results of an unorthodox approach to measuring workers' willingness to pay for job attributes. For a wide variety of reasons, researchers have long been interested in attaching monetary units to the values that workers associate with the nonwage portions of their work. With some exceptions, these studies have generally come to the conclusion that nonwage job attributes are of relatively minor importance to workers.(1)

The bulk of this work has relied on the theory of compensating wage differentials. The attraction of this theory is that it holds the promise of inferring workers' willingness to pay by observing wage differences across jobs having different attributes.(2) Unfortunately, empirical application of this theory has been plagued by a number of anomalies. Surveying the literature over a decade ago, Brown wrote "The overall pattern that emerges . . . is one of mixed results: some clear support for the theory but an uncomfortable number of exceptions [3, 118]." Dickens [6] echoes this evaluation in a more recent survey.(3)

There are no lack of explanations for this. Hypotheses range from unobserved heterogeneity [10; 16], omitted variables [23], measurement error [7], and misspecification [17]. Several recent studies have shown just how severe these problems can be in real labor market data. Hwang, Reed, and Hubbard [16] decompose the bias associated with unobserved worker productivity into three components. Using reasonable estimates for the sizes of these components, they show that estimates of workers' willingness to pay for job attributes based on compensating wage studies are likely to be downwardly biased by 50 percent or more, even resulting in wrong-signed estimates. In another paper, Hwang, Mortenson, and Reed [17] establish that unobserved firm productivity also generates a bias in the downward direction when estimates of workers' willingness to pay are based upon compensating wage models. Gronberg and Reed [11] demonstrate that this latter source of bias can also be quantitatively large in actual labor market data. These studies raise the possibility that workers attach greater value to nonwage job attributes than is measured by the compensating wages approach.

The shortcomings of the conventional methodology underscore the importance of alternative approaches to valuing workers' willingness to pay for job attributes. Accordingly, this paper turns to a series of questions in the National Longitudinal Survey, Youth Cohort (NLSY) in which workers were asked to report the minimum wage they would have to receive to accept various types of work. Like others, we have our reservations about data based on what workers say, as opposed to the job choices workers make. Nevertheless, there are mitigating factors which recommend this particular data set.

First, the NLSY surveyed all 12686 individuals about their willingness to accept a variety of jobs. This is a very large sample. It is rare for willingness to pay studies to have more than one or two thousand observations, most have only several hundred. Second, a number of other researchers have used these self-reported values from the NLSY and found them useful. These include Borus [2] and Holzer [15].(4) And third, each worker in the NLSY survey supplied reservation wage data for a number of different jobs. Thus, one can look at differences in reservations wages for different kinds of work for the same worker. This eliminates the unobserved heterogeneity problem which is so problematic in conventional compensating wage studies. In summary, given the unsettled state of the literature on workers' willingness to pay for job attributes, we believe that the data here provide a useful alternative approach to this subject.

Two questions form the core of our analysis. First, how large are the monetary premia/penalties that workers attach to differences in alternative kinds of work? And second, how much taste heterogeneity exists across workers in their job evaluations? We proceed as follows. Section II uses the theory of job search to derive the relationship between reservation wages and the monetary value of nonwage job attributes. Section III presents our methodology for using the categorical data of the NLSY to make inferences about continuous reservation wage data. Section IV discusses the data. Section V investigates the reliability of the self-reported reservation wages. Section VI discusses our empirical findings and section VII concludes.

II. Theory

Khandker [18] develops a model of job search where jobs differ in their nonwage characteristics. Suppose a worker follows an optimal sequential search strategy with the purpose of maximizing her expected lifetime utility. Suppose further that there are J different types of jobs, where each type is defined by its nonwage attributes. Let the instantaneous utility of the worker be given by

u = {b, if searching {[w.sub.j] + [v.sub.j], if working in a Type j job; j = 1, 2, . . ., J;

where b is the value of leisure net of search cost, [w.sub.j] is the wage paid in a Type j job, and [v.sub.j] is a parameter that imputes a dollar value to the utility associated with the nonwage attributes in a Type j job. If one defines [Rho] as the interest rate, [V.sup.u] as the (discounted present) value of unemployment, and [r.sub.j] as the jth reservation wage, then Khandker [18] shows that the equilibrium reservation wage for a Type j job is given by

[Mathematical Expression Omitted];

where [([Rho][V.sup.u]).sup.*] is uniquely determined by the parameters of the J wage distributions, the vector of separation rates for the J job types, the arrival rate of job offers, the probabilities of the different kinds of jobs, the interest rate ([Rho]), the value of leisure net of search costs (b), and the vector of nonwage parameters (the [v.sub.j]'s). It follows that

[Mathematical Expression Omitted].

The importance of equation (3) is that it states that the (negative of the) difference in reservation wages between jobs j and k provides a monetary value of the difference in utilities that the worker receives from the nonwage components of work at those jobs. For example, if a worker revealed that her reservation wage for working at a hamburger place was $3.85/hour, while her reservation wage for washing dishes was $4.45/hour, we would know that she viewed working at a hamburger place 60 cents/hour more attractive than washing dishes. The remainder of this paper is concerned with estimating differences in reservation wages across job categories for the workers in this data set.

III. Predicting Job- and Worker-Specific Reservation Wages

The reservation wage data used in this study derives from a series of questions contained in the 1979 NLSY survey.(5) Each individual in that survey was asked whether they would accept a given type of work if it offered $2.50 an hour in pay. They responded either positively or negatively. If negatively, they were asked if they would accept it at $3.50 an hour. If they once again said no, they were asked one last time about accepting the job at $5.00 an hour. Thus, for any given job, each respondent in the NLSY identified their reservation wage as belonging to one of four categories: (1) less than $2.50/hour, (2) between $2.50 and $3.50/hour, (3) between $3.50 and $5.00/hour, and (4) greater than $5.00/hour.

This procedure was used to acquire reservation wage information for six different jobs: (i) "working at a check-out counter in a supermarket" (SUPRMRKT), (ii) "working away from home in a national forest or park" (PARKS), (iii) "working at a hamburger place" (BURGERS), (iv) "washing dishes" (DISHES), (v) "working as a cleaning person" (CLEANING), and (vi) "working cleaning up neighborhoods" (NEIGH).(6) A desirable feature of the data set is that all six jobs are narrowly defined and would be reasonably well-known to the NLSY respondents.

Given that the reservation wage data is categorical in nature, we are left with the task of using this information to estimate continuous differences in reservation wages across job types as in equation (3). Accordingly, we estimate categorical log wage equations in order to obtain predicted values of the reservation wage for each individual on each job. Let the distribution of reservation wages for individual i at job j be given by

log [r.sub.ij] = [X[prime].sub.i][[Beta].sub.j] + [[Sigma].sub.j][[Epsilon].sub.ij], (4)

where [X.sub.i] represents a vector of personal and labor market characteristics; [[Beta].sub.j] is the associated vector of coefficients; and [[Epsilon].sub.ij] is an error term having the logistic distribution with positive scale parameter [[Sigma].sub.j]. The observed categorical data take on the values:

[C.sub.ij] = {1, if log [r.sub.ij] [less than] log (2.50) {2, if log (2.50) [less than] log [r.sub.ij] [less than] log (3.50) {3, if log (3.50) [less than] log [r.sub.ij] [less than] log (5.00) {4, if log (5.00) [less than] log [r.sub.ij].

Maximum likelihood estimation of the parameters [[Beta].sub.j] and [[Sigma].sub.j] based on data ([C.sub.ij], [X.sub.i]) for a single job is straightforward.(7) However, if [[Epsilon].sub.ij] and [[Epsilon].sub.ik] are not independent for j [not equal to] k, estimating six independent, categorical log wage equations ignores potentially useful information. For example, if an individual with an above average reservation wage on job k (given [X.sub.i]) is also likely to have an above average reservation wage on job j, knowledge of [C.sub.ik] can improve our prediction of [r.sub.ij]. Efficient estimation of [[Beta].sub.j] and [[Sigma].sub.j] would require specifying the joint distribution of the variables [C.sub.i1], [C.sub.i2], . . ., [C.sub.i6]. Estimation involving anything of higher dimension than a bivariate distribution is computationally difficult and likely to be plagued by numerical problems. As a result, we take a simpler approach: we include dummy variables indicating the reservation wage category for the five other job types in each categorical log wage equation. Thus our full model is:

log [r.sub.ij] = [X[prime].sub.i][[Beta].sub.j] + [[Delta][prime].sub.ij][[Gamma].sub.j] + [[Sigma].sub.j][[Epsilon].sub.ij] (5)

where [[Delta][prime].sub.ij] = ([[Delta].sub.i12], [[Delta].sub.i13], [[Delta].sub.i14], . . ., [[Delta].sub.i62], [[Delta].sub.i63], [[Delta].sub.i64]) - (0, . . ., 0, [[Delta].sub.ij2], [[Delta].sub.ij3], [[Delta].sub.ij4], 0, . . ., 0); and [[Delta].sub.ijk] = 1 if [C.sub.ij] = k, and 0 otherwise, k = 1, 2, 3, 4. While not equivalent to joint estimation, this has the advantage of using all the available information.(6)

Once we obtain estimates of [[Beta].sub.j], [[Gamma].sub.j], and [[Sigma].sub.j], we use them to predict reservation wages for each individual on each job. Let [Mathematical Expression Omitted] represent our initial predicted (log) reservation wage,

[Mathematical Expression Omitted].

Note that [Mathematical Expression Omitted] is not the best prediction of the ith individual's (log) reservation wage for the jth job. This is because the actual value of a given reservation wage can (and frequently does) fall in a category other than the predicted one. For example, [Mathematical Expression Omitted] might fall in the interval [log(2.50), log(3.50)] - category 2 - even though the individual's responses indicate that her reservation wage actually lies in the interval [log(3.50), log(5.00)] - category 3. We want to incorporate this latter information in our final prediction.

We do this by predicting individuals' reservation wages as follows:

[Mathematical Expression Omitted],

where f([center dot]) and F([center dot]) denote the p.d.f. and c.d.f. of the logistic distribution. This gives the expected value of the dollar reservation wage ([X[prime].sub.i][[Beta].sub.j] + [[Delta][prime].sub.ij][[Gamma].sub.j] + [[Sigma].sub.j][[Epsilon].sub.ij]), given the reservation wage category originally reported by the respondent and our estimates of [[Beta].sub.j], [[Gamma].sub.j], and [[Sigma].sub.j]. [Mathematical Expression Omitted] and [Mathematical Expression Omitted] give the lower and upper bounds on values of [Epsilon] which are consistent with a reservation wage in the reported category. For example, if the individual reported a reservation wage in category 3, [Mathematical Expression Omitted] and [Mathematical Expression Omitted]. For category 1, L equals -[infinity]; and for category 4, U equals [infinity].

The preceding estimation procedure guarantees that every predicted reservation wage lies within the original reservation wage category reported by the respondent. If an individual originally reported a reservation wage between $2.50 and $3.50/hour, then the predicted reservation wage will also lie between these two values. The only difference is that we have used additional information - namly, the categorical reservation wage data reported by this respondent for other jobs, and the reservation wage data reported by other (observationally similar) respondents - to obtain a continuously-valued, reservation wage prediction.

IV. Data

The 1979 NLSY consists of 12686 female and male respondents between the ages of 14 and 22. The survey oversamples civilian Hispanic, black, and economically disadvantaged non-Hispanic, non-black youth and those in military service. From this overall cohort we obtained a sample composed of 7839 individuals (henceforth referred to as the "general" sample). The large drop in sample size was primarily due to missing data on labor force participation, which had the effect of eliminating all 14 and 15 year-olds.(9) In addition, there were a large number of missing value codes in the local labor market characteristics data.(10) We present results both for this general sample, and a subsample of 1483 white, non-Hispanic males who worked at least 50 percent of the weeks between the 1979 and 1980 survey dates. The reason for looking at this subsample is to examine whether a relatively homogeneous group of workers has more homogeneous valuations of jobs.

Table I presents summary statistics for these two samples. The characteristics reported in the first two parts of this table are used as predictors of reservation wages within the reported intervals. They are also of interest in giving a feel for the composition of the two samples. The data used are the usual demographic and human capital variables (age, schooling, labor force experience, marital status and sex, race, and family background); attitudinal variables (willingness to accept welfare, scores on "Rotter" and knowledge of work tests, and educational goals); school enrollment and labor force status variables at the time of the survey; hourly wage data; and characteristics of the respondents' local labor market (unemployment rates, size of city, poverty rates, per capital income, and regional identification).

Compared to the general sample, members of the white males sample are slightly older and have slightly higher educational attainment. They have greater employment experience and appear to be a little more knowledgeable about the labor market. Their parents have greater educational attainment and their families show lower incidences of welfare recipiency. A greater percentage of them were employed at the time of the survey and they earned higher wages.

Figure 1 presents histograms representing the categorical reservation wage data as reported by the respondents in both the "general" and "white males" samples. As one can see, there is a considerable amount of dispersion in the original data. Further, substantial differences are evident in the shapes of the respective distributions. The job types (SUPRMRKT, PARKS, etc.) are roughly ranked in order of desirability, with those having the lowest reservation wages ranked first. Finally, one can clearly see a difference in the distributions of reservation wages between the general and white males samples. White males generally report higher reservation wages for all job types other than PARKS.

The last section in Table I reports the mean values of the predicted reservation wages for each of the six types of jobs. Predicted reservation wages are calculated according to equation (6). Once again, the jobs are ranked in terms of overall desirability, as measured by the mean reservation wage in the general sample. For the general sample, the mean ranges from a low of $3.75/hour for the supermarket job (SUPRMRKT) to a high of $5.15/hour for cleaning up neighborhoods (NEIGH). For the white males sample, the mean ranges from $3.82/hour (PARKS) to $5.55/hour (CLEANING). In general, the predicted reservation wages evidence the same patterns across job types and samples seen in the original reservation wage data.

V. Are the Self-Reported Reservation Wages Reliable?

It is a straightforward exercise to derive the following results from the equilibrium reservation wage equation in (2),

[Mathematical Expression Omitted],

where [[Mu].sub.j] is the mean of the distribution of market wages for job Type j, b is the value of leisure net of search costs, and [w.sub.k] is the wage an employed worker receives at her current job. Equation (7) predicts that workers with more attractive employment opportunities - either due to greater human capital possessed by the worker or a more favorable labor market environment - should [TABULAR DATA FOR TABLE I OMITTED] have higher reservation wages. Further, less (more) attractive nonworking activities should give rise to lower (higher) reservation wages. Finally, employed workers with higher wages will have higher reservation wages than those employed at lower wages. In this section we test whether workers' self-reported reservation wages behave in accordance with these predictions.

Table II reports the results of estimating equation (4) for each of the six jobs.(11) This discussion focuses on those variables that have a significant coefficient of the same sign in at least 5 of the 6 estimated reservation wage equations. These variables are PCAPINC, AGE, GRADE, HSGRAD, PCTWRK78, ROTTER, MARRMEN, NOWELF, and LOGWAGES. (We note that the coefficients can generally be interpreted as percentage changes in the respective log reservation wage associated with a unit increase in the independent variable.)

Consider first the labor market variable PCAPINC. PCAPINC is a measure of per capita income in the respondent's county of residence. Theory predicts that workers will have higher reservation wages in labor markets characterized by a higher mean distribution of wages. To the extent that increases in the variable PCAPINC identify high wage markets, we expect its coefficient to be positive in the reservation wage equations. Indeed, it is positive and significant in all 6 equations.

Age (AGE), education (GRADE and HSGRAD), and previous work experience (PCTWRK78) represent traditional human capital variables. Theory predicts that workers with greater human capital will have higher reservation wages. This is because the opportunity costs of accepting a job is greater for these workers. Consistent with this theory, each of these 4 variables has the predicted positive coefficient in at least five of the six reservation wage equations.

Similarly, the variables ROTTER and MARRMEN are also likely to be capturing differences in the wage distributions available to workers. The variable ROTTER is a psychological instrument used to measure the respondent's belief that life's economic outcomes lie within the respondent's control. Respondents affirming these beliefs are likely to behave in ways that are rewarded in the labor market. The variable MARRMEN is a dummy variable that compares reservation wages for married males to single males (the omitted category). Being married is commonly identified as proxying for unobserved labor market productivity for men [14; 19]. For both of these variables, increases are associated with higher reported reservation wages. This is consistent with the prediction of theory.

NOWELF is a dummy variable indicating the respondent's unwillingness to accept welfare. In the context of (7), this is capturing the respondent's evaluation of the utility associated with the state of not working. Theory predicts that workers with less attractive nonwork activities will have lower reservation wages. Consistent with this theory, the variable NOWELF has a negative coefficient and is significant in 5 of the 6 regression equations.

Finally, LOGWAGES measures the wages of those respondents who are employed in the sample. Theory predicts that workers with higher wages at their current jobs will require higher wages at alternative jobs in order to induce them to take those jobs. As predicted, the coefficient for LOGWAGES is positive and significant in the equations for all 6 job types.

The only other variables that are significant in at least 5 of the 6 equations are the dummy variables for married and single women (MARRWMN and SINGWMN - where the omitted category [TABULAR DATA FOR TABLE II OMITTED] is single men) and for blacks (BLACK).(12) Unlike the previous variables, these do not consistently take the same sign in each of the equations. Accordingly, these variables may be picking up different preferences within each group towards the alternative jobs. For example, the fact that single women report a higher reservation wage for moving away from home and working in a national park or forest (PARKS) may simply reflect the fact that this type of work is less appealing to single women than single men.

In summary, we find that the self-reported reservation wages in the NLSY behave remarkably consistent with the predictions of theory.(13) This of course does not prove that the self-reported reservation wage data is accurate. It does suggest, however, that these data contain useful information about workers' reservation wages. The remainder of this study is concerned with exploiting this information to learn about workers' valuations of job differences.

VI. Empirical Results

Characterizing Differences in Reservation Wages

We wish to characterize both (i) the size of the monetary premia/penalties that workers attach to differences in alternative kinds of work, and (ii) the extent to which taste heterogeneity exists across workers in their evaluations of job differences. The subsequent empirical work uses a variety of measures to do this. We choose to focus on the median as opposed to the mean as our measure of central tendency since this is likely to be less sensitive to distributional assumptions about the error term in equation (5). For the same reason, we report interquartile ranges rather than variances to characterize the dispersion of the reservation wage distribution.

Let [r.sub.ij] be the predicted reservation wage obtained from equation (6), noting that [r.sub.ij] represents a dollar value estimate of the reservation wage for a given individual and job type. Let the difference between the ith worker's reservation wages for the jth and kth job types be given by [d.sub.ijk] where

[d.sub.ijk] = -([r.sub.ij] - [r.sub.ik]).

Given 6 job types, there will be 15 distinct pairings (jk pairs). Note that [d.sub.ijk] is constructed to be positive when the jth job is viewed as more attractive than the kth job. For an individual, this measure provides a dollar valuation of the worker's willingness to pay (forego wages) to have job j rather than job k.

Let [Mathematical Expression Omitted] represent the median value of the distribution of [d.sub.ijk]. A positive value for [Mathematical Expression Omitted] suggests that workers generally view job j as more desirable than job k. Note that [Mathematical Expression Omitted] could take a value of 0 because all workers were indifferent between the two jobs. But it also could equal zero if half of the workers strongly preferred j to k, and the other half just as strongly preferred k to j. To address this ambiguity, we also report the median of the distribution of absolute values of [d.sub.ijk], [[absolute value of [d.sub.jk]].sup.med]. In words, this is the median of the distribution of monetary values that workers would be willing to pay to have their more preferred job of the two jobs in the (j, k) job pair.

We use one other related measure in our study: the difference in monetary values that workers ascribe to their highest and lowest ranked job types. Let [Mathematical Expression Omitted] and [Mathematical Expression Omitted] be defined by

[Mathematical Expression Omitted].

[Mathematical Expression Omitted].

Then the value of the difference between [R.sup.max] and [R.sup.min] for the ith respondent, call it [D.sub.i] provides a measure of the monetary value that the respondent attaches to the differences in nonwage attributes between her most and least preferred job types. We present median values of this measure, [D.sup.med], along with alternative measures which are less sensitive to any over-estimates of [R.sup.max].(14)

Workers' Monetary Valuations of Differences in Jobs

Table III contains the main empirical results of this study. Columns (2) and (5) report the median value of estimated differences in reservation wages across the respective pairs of jobs, [Mathematical Expression Omitted]. It is intended to reflect the following thought experiment. Suppose each individual in our samples was placed in the same job j. Suppose next that they were asked how much they would be willing to pay to move to job k. Some would have to be compensated to induce them to move to k. Their corresponding [d.sub.ijk] value would be negative. Others would prefer to move to k. Their corresponding [d.sub.ijk] value would be positive. [Mathematical Expression Omitted] is the median value of this distribution of [d.sub.ijk] values. Note that the [Mathematical Expression Omitted] values in Table III are arranged so that the median worker would prefer the first job to the second job.(15)

Columns (3) and (6) report the estimated values of [[absolute value of [d.sub.jk]].sup.med] in descending order for each possible job pair and each sample. It is intended to reflect a somewhat different thought experiment. Suppose each individual in our samples were placed at her least preferred choice of a pair of jobs. [absolute value of [d.sub.ijk]] measures how much she would be willing to pay (forego in wages) in order to work at the more preferred job instead. [[absolute value of [d.sub.jk]].sup.med] is the median value of the distribution of [absolute value of [d.sub.ijk]] values for the respective sample of respondents.

The first thing we note about the [[absolute value of [d.sub.jk]].sup.med] values in Table III is that they are generally much larger than the corresponding [Mathematical Expression Omitted] values. For example, the median difference in the reservation wages for PARKS and DISHES jobs is approximately 27 cents in the general sample ($0.266). In contrast, we estimate that the median individual would be willing to pay about $1.42 to move from her less to her more preferred job of the pair. If each of the sample respondents preferred PARKS to DISHES jobs, the [Mathematical Expression Omitted] and [[absolute value of [d.sub.jk]].sup.med] values would be the same. The substantial differences [TABULAR DATA FOR TABLE III OMITTED] between these two measures indicate that a significant degree of heterogeneity in tastes exists within each of the two samples. This subject will be discussed in greater detail below.

The second thing we note about the estimates of [[absolute value of [d.sub.jk]].sup.med] is that they are large. For the general sample, they range from a high of approximately $1.42/hour (for PARKS-DISHES) to a low of $0.43/hour (for DISHES-CLEANING). It should be emphasized that these estimates are median differences, suggesting that half of the respondents valued the differences in these job pairs in excess of these amounts.(16)

It is helpful to place these dollar values in the context of the labor market of 1979, the year of the survey. In 1979, the minimum wage was $2.90/hour. Further, if we look at employed white females in the NLSY, the average hourly wage at the time of the survey was around $3.40. The corresponding figure for white males was about $4.40. The general sample in this study contains not just a mixture of females and males, but blacks and whites, employed and not employed workers. Accordingly, we suppose that $4.00 an hour would represent a generous estimate for the average market wage available to the workers in this sample. By that calculation, these differences amount to a substantial portion of workers' average wage opportunities. The differences for 2 of the job pairs are estimated as being valued by half of the respondents at a third or more of their average going wage (PARKS-DISHES and PARKS-CLEANING). 10 of the 15 differences between job pairs are estimated to be valued by half of the respondents at 25 percent or more of the prevailing working wage. By any standard, these are substantial differences.

Turning now to the sample of white males, we see that the estimates of [[absolute value of [d.sub.jk]].sup.med] range from $1.72/hour (PARKS-BurGERS) to $0.53/hour (BURGERS-DISHES). They are generally smaller than their counterparts for the general sample: white males have smaller [[absolute value of [d.sub.jk]].sup.med] values for 11 of the 15 job pairs. Thus, these results suggest that the individuals in the white males sample generally value the nonpecuniary differences between job types less than those in the general sample.

In fact, this can be tested directly. As discussed above, the reason for creating the subsample of white males was to determine whether a relatively homogeneous group of workers - as defined by their observed socioeconomic characteristics - had more homogeneous valuations of jobs. We find this to be the case. In each of the 6 logit regressions, the scale parameter ([[Sigma].sub.j]) was estimated to be smaller for the white males sample than for the general sample. The differences in estimated values were all significant at the 5 percent level (two-tailed test).

Even for this group, however, the size of the estimates are substantial. For the sake of comparison, let us assume that the average market wage for the white males sample at the time of the survey was $4.50 an hour, again erring on the large side given what we know about workers' actual wages. Column (6) reports that 5 of the 15 median differences between job pairs are measured to be 25 percent or more of this average market wage.

Perhaps the most striking result concerns the median value that the workers in the samples attached to the nonwage differences between their most and least preferred job types, [D.sup.med]. This is reported in the last row of Table III. For the individuals in the general sample, [D.sup.med] is estimated to be approximately $3.69/hour. For the individuals in the white males sample, the corresponding value is $2.89/hour.

The large size of these values made us question whether this result might not be the result of our specification assumptions. However, consider the following. Over a quarter of the sample report reservation wages in both the lowest and the highest reservation wage categories. Clearly, for these workers, the difference in reservation wages between their least and most preferred job is greater than $2.50/hour.

An alternative way to minimize the effect of prediction errors is to truncate the reservation wage distribution at what seems like reasonable bounds. Accordingly, we imposed upper and lower bounds on the predicted reservation wage distribution of $5.50 and $2.00 respectively. Using this approach, the new [D.sup.med] values for the general and white males samples are $2.09 and $1.94, respectively. While these are smaller than the [D.sup.med] values calculated from the nontruncated distributions, they are still substantial monetary values. We shall have more to say below about the interpretation of these large estimates.

The Extent of Taste Heterogeneities

The preceding analysis was concerned with the central tendency of the distribution of values workers attached to differences across jobs. This section focuses on the dispersion of that distribution.

One of the distinctive aspects of this investigation is that it allows one to identify differences in tastes across individuals. While the reservation wage levels for any given job reflect both human capital and taste factors, theory states that the differences in reservation wages across jobs for any given worker reflect only tastes. The NLSY data provide a unique opportunity to investigate the importance of taste differences across workers.

The question we want to address is how quantitatively important are differences in tastes across workers? Define [d.sub.jk,25] and [d.sub.jk,75] as the 25th and 75th percentile values of the distribution of differences in workers' reservation values between jobs j and k (i.e., the distribution of the [d.sub.ijk]'s). In the following analysis we look at the difference between [d.sub.jk,25] and [d.sub.jk,75] for each of the 15 job pairs.

Figure 2 represents the distributions of the estimated [d.sub.ijk]'s for each of the 15 possible job pairs. In the interests of brevity, only the results for the general sample are represented. The graphs are presented in descending order of the size of the differences between the 75th and 25th percentile values of the empirical distributions. These differences are reported inside each of the respective graphs ("[d.sub.75] - [d.sub.25]"). As is easily seen from the graphs, workers' job evaluations differ dramatically. For example, for the pair of job types PARKS and DISHES, the 25th and 75th percentile values differ by $3.57. This means that the two workers represented by these values differed in their evaluations of the relative attractiveness of these two jobs by almost $3.60/hour. Looking down the rest of the figure, we see that the dollar differences between the 25th and 75th percentile values of the empirical distributions gradually diminish to a minimum of $0.89 for CLEANING and DISHES jobs. Given the types of jobs included in this analysis, we do not find it unreasonable that the respondents would be in greatest agreement concerning the differences between these two types of jobs.

VII. Conclusion

This study finds evidence that nonwage differences between jobs are very important to workers. Perhaps the most striking finding in this regard is the median value that workers in the samples attached to the differences in nonwage attributes between their most and least preferred job types. Using conservative calculations, we estimate this value to be approximately $2.09/hour for individuals in the general sample, and approximately $1.94/hour for the individuals in the white males sample. Two questions raised by these findings are (i) What are these nonwage differences between jobs that workers are valuing so highly? and (ii) Are workers' willingness to pay values really as large as the self-reported reservation wage data indicate?

With regard to the first question, it is possible that the differences in reservation wages are picking up differences in other, pecuniary dimensions of the job. Some jobs might offer superior benefits (e.g., pensions, health insurance, etc.), promotion opportunities and faster wage growth than other jobs. While this might be true in general, we think it is unlikely given the set of jobs studied here. "Washing dishes," "working in a check-out counter in a supermarket," "working at a hamburger place," etc., are not the kinds of jobs one associates with handsome pecuniary fringe benefits. This leads us to the conclusion that the nonwage differences that the workers are valuing reflects their evaluations of the nonpecuniary dimensions of the respective jobs. That is, working as a check-out clerk and washing dishes are very different kinds of activities. These results suggest that workers have strong preferences about the kinds of work they do.

The second question touches on how much weight one should attach to the large monetary values reported here. The fact that the underlying data are self-reported certainly suggests caution.(17) Nevertheless, we know of no theoretical reason why self-reported data should necessarily lead to large willingness to pay values. Nor has this been the general experience of previous studies [8; 9]. In contrast, the compensating wages approach is known to lead to downwardly biased willingness to pay estimates. It is also known that this bias may be quite severe [16; 17]. The important thing to note is that there is a lot of room between the large values of workers' willingness to pay for job attributes estimated here, and the more conventional findings of small and insignificant values estimated by the compensating wages approach. Accordingly, we interpret these results as evidence that workers' valuations of nonpecuniary dimensions of work are substantially larger than previous research has indicated.

Finally, the existence of substantial taste heterogeneities between workers underscores the importance of the labor market problem of matching workers and jobs. In particular, it suggests that the sorting of workers across job types and occupations will have important welfare implications even if the marginal worker views alternative job prospects as close substitutes. Phrased differently, it is evidence that occupational rents are an important factor in labor market outcomes. This has obvious implications for length of job search, losses associated with job displacement, and the inter-occupational mobility of labor.

[TABULAR DATA FOR APPENDIX A OMITTED]

[TABULAR DATA FOR APPENDIX B OMITTED]

Steve McClaskie of the Center for Human Resource Research and Jeannie Bennetti of the U.S. Census Bureau both provided data assistance and we are appreciative of their help. We wish to thank James Brown, Frank Stafford, and session participants at the 1995 American Economic Association meetings for their insightful comments. We are also indebted to an anonymous referee whose comments resulted in a number of substantial improvements.

1. Recent work on this subject include, Butler and Worrall [4], Low and McPheters [22], Topel [29], Viscusi and Moore [30], Hamermesh and Wolfe [12], Hersch and Viscusi [13], and Kostiuk [20].

2. Valuable summaries of the compensating wages literature are contained in Linnerooth [21], Smith [28], Brown [3], Rosen [27] and Dickens [6].

3. In a particularly pessimistic review of the contribution of compensating wage theory to valuing life, Dickens writes, "Although intuitively appealing, the use of hedonic wage estimates to determine people's willingness to pay to avoid the risk of fatal hazards is fraught with problems. . . . [This] paper concludes that it is unlikely that economics has much to contribute to the public policy debate over the value of a life [6]." His criticisms of the theory of compensating wages to estimate workers' willingness to pay for job safety apply with equal force to other job attributes.

4. Dunn [8; 9] and Cavalluzzo [5] have also used self-reported values from survey data to attach monetary values to job attributes.

5. The actual question was, "If . . . you were offered a full-time job at (hypothetical hourly wage rate) do you think you would accept it if it were (a given type of work)?" If the respondent was enrolled in school at the time of the survey, she was asked if she would be willing to accept the job if it were offered the following summer. If not enrolled, she was asked if she would accept the job immediately.

6. A seventh job type was "working in a factory." Comments on an earlier version of this manuscript suggested that this job type be dropped because of ambiguities in the content of this job.

7. The log likelihood contribution for individual i is:

ln [L.sub.i] = [summation of] [[Delta].sub.ijk] where k = 1 to 4 ln[[F(([u.sub.k] - [X[prime].sub.i][[Beta].sub.j])/[[Sigma].sub.j] - F(([l.sub.k] - [X[prime].sub.i][[Beta].sub.j])/[[Sigma].sub.j])]

where [u.sub.k] = log(2.50) for k = 1, log(3.50) for k = 2, . . ., and [infinity] for k = 4; [l.sub.k] = -[infinity] for k = 1, log(2.50) for k = 2, etc.; [[Delta].sub.ijk] = 1 if person i indicates her reservation wage falls in interval k on job j, and 0 otherwise; and F(x) = [[1 + exp(-x)].sup.-1].

8. An alternative single equation approach might be to estimate [[Epsilon].sub.ij] from a first round of estimation using equation (4), and then to include the estimated [[Epsilon].sub.ij] from each of the other jobs as explanatory variables in the second round. We experimented with this technique. With [X.sub.i] and the predicted [[Epsilon].sub.ij] as explanatory variables, we obtained lower values of the log likelihood than with the model characterized by equation (5). With the [[Delta].sub.ij] included as well, we would sometimes reject the null hypothesis that the coefficients on the [[Epsilon].sub.ij] were all equal to zero using a likelihood ratio test. However, this augmented model gave predicted values very similar to those obtained from equation (5).

9. Those respondents younger than 16 were not surveyed about their labor force participation.

10. We deleted observations on the basis of missing values if there weren't very many associated with a particular variable, or in cases where the reservation wage answers were missing. For those variables characterized by a significant number of missing values, we retained the respective observations by including dummy variables to indicate instances of missing values.

11. Note that this equation omits the vector of reservation wage categories reported by the worker for the other jobs (the [[Delta].sub.ij]'s). This is done for the purposes of this section only. The inclusion of the [[Delta].sub.ij] complicates the interpretation of the coefficients of the included variables. This makes it difficult to determine whether the self-reported reservation wage data behave in accordance with the predictions in equation (7). We emphasize that the subsequent empirical work uses predicted reservation wages which do employ these variables. These latter equations are reported in the Appendix.

12. We note that our finding that blacks generally have lower reservation wages than white confirms the same finding reported by Borus [2].

13. Of the other included variables only the previous experience variables, OCCEXPER and INDEXPER, require some explanation. Workers with bad previous experiences in a given job type are likely to have little tenure in that job and higher reservation wages. We thus would expect these variables to have negative coefficients. The empirical results are consistent with this hypothesis. (We thank an anonymous referee for suggesting the inclusion of these variables.) The other significant coefficients in Table II are - in almost every case - consistent with the predictions of theory.

14. Conversion of predicted log wages to dollar wages mitigates prediction errors in category 1 (the lowest reservation wage category) because moderate differences in log wages result in small differences in dollars. In contrast, the opposite is true for reservation wages in category 4 (the highest reservation wage category) where conversion to dollars exacerbates positive prediction errors.

15. For example, the reservation wage for SUPRMRKT is estimated to be (slightly) lower than the estimated reservation wage for PARKS type jobs for the median worker in the general sample, [Mathematical Expression Omitted]. However, the median worker in the white males sample has a higher predicted reservation wage for SUPRMRKT type jobs, [Mathematical Expression Omitted]. As a result, Table III switches the order of the job pair for the white males sample indicating that the median worker in this sample views PARKS type work more attractive than SUPRMRKT type work: [Mathematical Expression Omitted].

16. Not surprisingly, the workers in the sample seem to value the differences between PARKS type jobs and everything else as greater than almost all other job pairs. This is even more evident in the white males sample, Recall that this job explicitly involved "working away from home." This locational move thus had to be factored in along with any other differences in job tasks associated with working in "national parks and forests."

17. It should be noted that these findings are also consistent with the industrial psychology literature on this subject [1; 25; 26]. That research reports that job components are highly correlated with job satisfaction.

References

1. Barnowe, Thad J., Thomas W. Mangione, and Robert P. Quinn. "The Relative Importance of Job Facets as Indicated by an Empirically Derived Model of Job Satisfaction," in The 1969-70 Survey of Working Conditions: Chronicles of an Unfinished Enterprise, edited by Thad J. Barnowe, Thomas W. Mangione, and Robert P. Quinn. Ann Arbor: Institute for Social Research, 1974.

2. Borus, Michael, "Willingness of Youth to Work." Journal of Human Resources, Fall 1982, 581-93.

3. Brown, Charles, "Equalizing Differences in the Labor Market." Quarterly Journal of Economics, February 1980, 113-34.

4. Butler, Richard J. and John D. Worrall, "Workers' Compensation: Benefit and Injury Claims Rates in the Seventies." Review of Economics and Statistics, November 1983, 580-89.

5. Cavalluzzo, Linda C., "Nonpecuniary Rewards in the Workplace: Demand Estimates Using Quasi-Market Data." Review of Economics and Statistics, November 1991, 508-12.

6. Dickens, William T., "Assuming the Can Opener: Hedonic Wage Estimates and the Value of Life." Journal of Forensic Economics, Fall 1990, 51-60.

7. Duncan, Greg J. and Bertil Holmlund, "Was Adam Smith Right After All? Another Test of the Theory of Compensating Wage Differentials." Journal of Labor Economics, October 1983, 366-79.

8. Dunn, L. F., "Quantifying Nonpecuniary Returns." Journal of Human Resources, Summer 1977, 347-59.

9. -----. "Nonpecuniary Job Preferences and Welfare Losses among Migrant Agricultural Workers." American Journal of Agricultural Economics, May 1985, 257-65.

10. Garen, John, "Compensating Wage Differentials and the Endogeneity of Job Riskiness." Review of Economics and Statistics, February 1988, 9-16.

11. Gronberg, Timothy J. and W. Robert Reed, "Estimating Workers' Marginal Willingness to Pay for Job Attributes Using Duration Data." Journal of Human Resources, Summer 1994, 911-31.

12. Hamermesh, Daniel S. and John R. Wolfe, "Compensating Wage Differentials and the Duration of Wage Loss." Journal of Labor Economics, January 1990, S175-97.

13. Hersch, Joni and W. Kip Viscusi, "Cigarette Smoking, Seatbelt Use, and Difference in Wage-Risk Tradeoffs." Journal of Human Resources, Spring 1990, 202-27.

14. Hill, Martha S., "The Wage Effects of Marital Status and Children." Journal of Human Resources, Fall 1979, 579-94.

15. Holzer, Harry J. "Black Youth Nonemployment: Duration and Job Search," in The Black Youth Unemployment Crisis, edited by Richard Freeman and Harry Holzer. Chicago: University of Chicago Press, 1986.

16. Hwang, Hae-shin, W. Robert Reed, and Carlton Hubbard, "Compensating Wage Differentials and Unobserved Productivity." Journal of Political Economy, August 1992, 835-58.

17. -----, Dale T. Mortensen, and W. Robert Reed. "Hedonic Wages and Labor Market Search." Mimeo, University of Oklahoma, 1995.

18. Khandker, Rezaul, "Offer Heterogeneity in a Two State Model of Sequential Search." Review of Economics and Statistics, May 1988, 259-65.

19. Korenman, Sanders and David Neumark, "Does Marriage Really Make Men More Productive?" Journal of Human Resources, Spring 1991, 282-307.

20. Kostiuk, Peter F., "Compensating Differentials for Shift Work." Journal of Political Economy, October 1990, 1054-75.

21. Linnerooth, Joanne, "The Value of Human Life: A Review of the Models." Economic Inquiry, January 1979, 52-74.

22. Low, Stuart and Lee R. McPheters, "Wage Differentials and Risk of Death: An Empirical Analysis." Economic Inquiry, April 1983, 271-80.

23. Lucas, Robert E. B., "Hedonic Wage Equations and Psychic Wages in the Returns to Schooling." American Economic Review, September 1977, 549-58.

24. Maddala, G. S. Limited-Dependent and Qualitative Variables in Econometrics. Cambridge: Cambridge University Press, 1983.

25. Mobley, W. H.; R. W. Griffeth, H. H. Hand, and B. M. Meglino, "Review and Conceptual Analysis of the Employee Turnover Process." Psychological Bulletin, May 1979, 493-522.

26. O'Reilly, Charles A., III and David F. Caldwell, "Job Choice: The Impact of Intrinsic and Extrinsic Factors on Subsequent Satisfaction and Commitment." Journal of Applied Psychology, October 1980, 559-65.

27. Rosen, Sherwin. "The Theory of Equalizing Differences." In Handbook of Labor Economics, Volume 1, edited by Orley Ashenfelter and Richard Layard. Amsterdam: North-Holland, 1986.

28. Smith, Robert S., "Compensating Wage Differentials and Public Policy: A Review." Industrial and Labor Relations Review, April 1979, 339-52.

29. Topel, Robert H., "Equilibrium Earnings, Turnover, and Unemployment: New Evidence." Journal of Labor Economics, October 1984, 500-22.

30. Viscusi, W. Kip and Michael J. Moore, "Workers' Compensation: Wage Effects, Benefit Inadequacies, and the Value of Health Losses." Review of Economics and Statistics, May 1987, 249-61.