Legal minimum wages and employment duration.

Sicilian, Paul

1. Introduction

The effect of a legal minimum wage on employee turnover is an important consideration when evaluating the effectiveness of a minimum wage policy. If an increase in the minimum wage results in greater turnover, then firms' hiring costs are increased, and the direct benefit to workers who are paid the minimum wage may be lessened. A reduction in turnover, on the other hand, would mean more stability for employers and workers, though access to minimum wage jobs may be reduced for job seekers. Despite the recent proliferation of studies on the effects of legal minimum wages, their impact on turnover has received relatively little attention from economists. This may be due to the fact that much of the literature focuses on the "teen labor market" where there is more churning than in the "adult labor market" (Gardecki and Neumark 1998). The fact remains, however, that a significant proportion of workers hired at the minimum wage are adults, and the effects of the minimum wage on the turnover hazard are understudied. (1)

Standard economic theory suggests that a price floor such as a minimum wage results in fewer transactions and in rents for those suppliers who manage to sell. In the context of the labor market, this means workers in minimum wage jobs may be expected to receive rents and, therefore, may be less likely to quit. Another possible consequence of a minimum wage arises from imperfect information in the labor market combined with heterogeneity among workers and jobs. Prior to the start of an employment relationship, the worker does not know all the relevant aspects of the job and the employer does not know all the important characteristics of the worker. As a result, in the absence of a mandated minimum wage, job applicants may use wage offers to infer information about the nonpecuniary aspects of jobs, thereby facilitating the matching process. A binding minimum wage compresses the wage distribution so that a single wage is associated with a larger array of jobs. (2) This means that the wage offer becomes a poor indicator of a job's nonwage characteristics, which leads to a higher proportion of poor matches and, consequently, more turnover.

This paper uses data on young men and women in the 1988-1994 rounds of the National Longitudinal Survey of Youth (NLSY) to estimate the relationship between the minimum wage and the turnover hazard. The extent to which the minimum wage binds varies across states both because of differences in state laws and because of differences in wage levels across states. We exploit this variation to estimate turnover hazards as a function of the ratio of the nominal minimum wage to the state median wage. This variable measures the extent to which the minimum wage binds.

We find some evidence that in states where the minimum wage is low relative to the state's median wage, men hired at the minimum wage have lower employee-initiated separation hazards than other low-wage workers. As the minimum wage rises relative to the prevailing median wage--that is, as the minimum wage becomes more binding--men hired at the minimum wage are increasingly likely to leave their jobs at any given duration. When the minimum wage is high relative to the median wage we find that men hired at the minimum wage have significantly higher employee-initiated separation hazards than other low-wage workers. We interpret these findings as evidence that while rents may accrue to minimum wage workers, the job matching process is undermined when the minimum wage binds. Our results apply to the men in our sample, but not to the women. Indeed, for women we find no connection between employment duration and the starting wage rate, though we do uncover some evidence that for women in low-wage jobs, employment duration is more elastic with respect to the presence of family-friendly fringe benefits than it is for men.

The paper proceeds as follows. The next section explores the possible reasons minimum wages would affect turnover and discusses the existing empirical and theoretical research on this topic. Section 3 presents the theoretical framework we use to guide our estimation procedures and our interpretation of the empirical results. Section 4 describes our data and presents some descriptive evidence on the relationship between turnover and minimum wages. We present our econometric model and discuss our estimates in section 5. The conclusion discusses the implications of our findings.

2. Research on Minimum Wages and Turnover

The standard model of competitive labor markets predicts an increase in employer-initiated terminations when the minimum wage increases and employers move up along their labor demand curves. Economists have long suspected, however, that the relationship between turnover and a legal minimum wage is more complex. Mixon (1978), for example, argues that minimum wages reduce employee-initiated quits for two reasons. First, the net benefits of turnover are reduced because of the reduction in the difference between a worker's current wage and the expected wage on a new job. Second, since firms are likely to be more selective when hiring into minimum wage jobs, the costs to the worker of searching for a new job are expected to be higher. Using quarterly industry data from 1959-1975 he finds evidence that increases in the ratio of the minimum wage to the industry average wage reduces the monthly industry quit rate.

The standard model has, of course, been extended to recognize that compensation includes both monetary and nonmonetary benefits. Wessels (1980a, b) suggests that employers will respond to a binding minimum wage by reducing nonwage benefits so as to offset the mandated wage increase. If workers are not aware that the "full wage"--wages plus benefits--at other firms has also fallen, the minimum wage may prompt workers to change jobs, with the result that quit hazards are elevated, at least in the short-run, until information improves. He also argues that the impact on quit rates will be larger for workers earning relatively high wages prior to a change in the minimum wage, since it is more likely that these workers' nonwage benefits can be more fully adjusted in response to the minimum wage. Wessels, using industry-level quit rates from 1966, finds evidence supporting this story. Sicilian and Grossberg (1993), using data from the Employment Opportunities Pilot Project (EOPP), also find evidence consistent with this information distortion argument, as they show that workers hired at the minimum wage are significantly more likely to leave their jobs voluntarily than other low-wage workers. Also, several authors argue on theoretical grounds that employers are likely to reduce expenditures on job training when the minimum wage is binding (e.g., Leighton and Mincer 1981; Hashimoto 1982). If this is true, then minimum wage legislation can indirectly increase turnover since on-the-job training reduces turnover hazards (Mincer 1988; Parent 1999; Grossberg 2002). (3)

On the other hand, employers may be either unable or unwilling to reduce nonpecuniary job benefits so as to fully offset wage increases resulting from minimum wage legislation. In this case, minimum wage jobs offer rents, reducing voluntary turnover. Using data from the EOPP, Holzer, Katz, and Krueger (1991) find significantly longer queues of job applicants in minimum wage jobs than in other low-wage jobs. They take this as evidence of the existence of rents in minimum wage jobs. They do not, however, look at turnover. Similarly, Katz and Krueger (1992) find that a majority of fast-food restaurants in Texas did not report attempts to offset the 1990 minimum wage increase.

3. Rents and Job Matching on Minimum Wage Jobs

To date, no consensus has emerged on the relationship between minimum wages and turnover. There are theoretical arguments and empirical evidence supporting both the proposition that minimum wages reduce turnover and the proposition that minimum wages increase turnover. In this section we attempt to reconcile these opposing claims and to motivate our own empirical analysis.

Our point of departure is De Fraja (1999), who uses a model with asymmetric information to show, inter alia, that rents can result from the imposition of a minimum wage. In De Fraja's model each job is characterized by its wage (w) and by the amount of effort required on the job or, more generally, by the nonwage attributes of the job (e). Workers have heterogeneous preferences regarding effort exerted on the job, but employers cannot observe the workers' preferences. As a result, firms offer the same wage-effort package to all workers. One consequence is that, among those hired at the minimum wage, workers with relatively low disutility of effort receive rents. Thus, when the minimum wage is binding, turnover will be reduced due to the creation of rents for some workers.

De Fraja assumes that information regarding both wage and nonwage benefits is known by prospective employees. Typical search models of turnover, on the other hand, assume that while the wage on a job is readily observable, the job's nonwage attributes are not (e.g., Johnson 1978; Holmlund and Lang 1985). In these models, the labor market is characterized by the joint distribution (w, e) of existing job offers, though a worker must take a job to learn the value of e on any particular job.

We modify this approach by assuming that while the specific value of e associated with a particular job is not known prior to taking the job, workers know the joint distribution of w and e. As a result, knowledge of the wage offer for a particular job provides information about nonpecuniary characteristics, via knowledge of the conditional distribution (e|w). (4) If this conjecture is correct, and if minimum wage laws distort information by altering the conditional distribution (e|w), the result may be an increased propensity for bad matches to occur in minimum wage jobs and an increase in turnover.

To get a clearer fix on these issues, we make a more formal statement of our approach. Following De Fraja (1999), we start with the assumption that the jth job is characterized by a wage-work effort (or a wage-working conditions) pair ([w.sub.j], [e.sub.j]), and that the ith worker's utility depends positively on wages and negatively on work effort. The ith worker's utility from the jth job would then be given by

[U.sub.i] = [U.sub.i]([w.sub.j], [[gamma].sub.i][e.sub.j]),

where [[gamma].sub.i] indicates the ith worker's disutility of effort. Because firms are unable to observe [[gamma].sub.i], they offer the same wage-work effort pair ([w.sub.j], [e.sub.j]) to all workers in similar jobs. (5) As a consequence, workers learn that knowledge of [w.sub.j] reveals to them information about [e.sub.j], so they self-select based on their attitudes toward effort. Those with relatively low disutility of effort (low values of [gamma]) are more willing to accept minimum wage jobs, since they will be more likely to earn rents in those jobs. As a result, we expect the existence of rents to result in relatively low separation hazards on minimum wage jobs. Moreover, to the extent that these rents exist, we would expect them to be greatest in relatively low-wage states, where minimum wage laws have their most pronounced effect on starting wage rates.

It is important to note, though, that this conclusion requires an environment in which the conditional distribution ([e.sub.j]|[w.sub.j]) is defined clearly, so that workers are well informed about working conditions, or the amount of effort required in jobs offering the minimum wage. This turns out to be a problematic condition in low-wage states, however, since workers in these states are actually less likely to be well informed about the relationship between wages and effort because this is where wage compression is likely to be greatest. Recent studies have found, for example, that the minimum wage compresses the wage distribution and that the compression is greater in relatively low-wage states. This causes "bunching" at the minimum wage (Card and Krueger 1995; Lee 1999). This wage compression means that wages in many jobs are swept up to the minimum wage, thus altering the joint distribution of [w.sub.j] and [e.sub.j]. The result is that many different [e.sub.j]s are now associated with the same [w.sub.j] = [w.sub.min], so that knowledge of [w.sub.j] is less likely to convey information to workers about the expected value of [e.sub.j]. As a consequence, there are likely to be more mismatches between firms and workers, since workers now learn disappointing information regarding [e.sub.j] once they are on the job. This is likely to raise the separation hazard for workers hired at the minimum wage.

Thus, the minimum wage has potentially offsetting effects on turnover in low-wage states. If the creation of rents is more important we should see lower turnover on minimum wage jobs, but if the information distortion effect dominates, minimum wage workers in low-wage states will exhibit greater turnover propensities.

The net effect of minimum wage laws in high-wage states is clearer a priori than in low-wage states. Since the minimum wage is less likely to be binding in these states, we would expect minimum wage jobs to offer smaller rents. The existence of smaller rents suggests that the degree to which minimum wage jobs suppress turnover in high-wage states will be smaller than in low-wage states, although there may still be an effect. Moreover, since minimum wages are less likely to compress wage distributions in high-wage states, workers should be better informed about the relationship between wages and working conditions. Thus, even though minimum wage laws may create smaller rents in high-wage states, we expect workers in these states to be better able to sort themselves. The result is that the quality of job matches on minimum wage jobs should be better than in low-wage states, and turnover in minimum wage jobs should be reduced.

We have identified two possible effects of minimum wage laws on turnover: the "rents" effect suggests that turnover will be reduced, while the "information distortion/job matching" story suggests turnover will increase. These stories are not necessarily mutually exclusive. If both effects exist, minimum wage laws will affect turnover in potentially offsetting ways. The empirical issues, then, are to determine whether these effects can be observed at all and, if so, to determine the circumstances under which each effect dominates.

4. Data and Descriptive Evidence

In the analysis that follows we use event history data from the NLSY to examine the employment patterns of workers in minimum wage jobs versus workers in other jobs. The NLSY is a well-known source of data for studies on individual labor market behavior. The survey began in 1979 when 12,686 individuals between 14 and 21 years old were interviewed. Through attrition and elimination of some subsamples (6) the sample size had dropped to 10,465 by 1988 and to 8891 individuals by 1994.

The NLSY contains information on up to five jobs for each respondent in each interview year. The most recent or current job at each interview is referred to as the Current Population Survey (CPS) job, since many questions about that job are modeled on the CPS questionnaire. Starting in 1986, the individuals in the sample are asked about each job's starting wage as well as the current wage. The sample used for this paper employs data drawn from seven waves of the NLSY (1988-1994) and includes respondents who started new CPS jobs anytime after the 1987 interview. (7) We track respondents from the start of the first CPS job after the 1987 interview until the job ends or until it is censored. Interview years containing missing values are eliminated, although respondents are retained as long as we are able to observe the date on which a job starts, the starting wage, and when the job ends or is censored. Respondents eliminated from the sample include those in the military, those in agriculture, those who are self-employed, those who are enrolled in school any time during the job spell, those who report health problems affecting their ability to work any time during the job spell, and those for whom the observed job is a temporary job. The final sample consists of 1273 men, 566 of whom have complete job spells, and 1280 women, 708 of whom have complete job spells.

It is important for our purposes to differentiate the behavior of workers hired at the minimum wage from that of workers hired at other relatively low wages. Hence, we distinguish throughout our analysis between workers hired at wages below the prevailing legal minimum wage, those hired at the legal minimum wage, those hired at wages no more than 10% above the legal minimum, and those hired at wage rates more than 10% above the minimum wage. (8) About 4.4% of the men in our sample and about 10.2% of the women have starting wages less than or equal to the minimum wage, with about 2.3% of men and about 6.5% of women starting at the minimum wage. Another 2.6% of men and 8.8% of women have starting wages no more than 10% above the prevailing minimum wage. These percentages concur with Bureau of Labor and Statistics research for 1988 that estimated that 2.2% of workers aged 25-34 years earned the minimum wage and 1.7% earned less than the minimum wage (Haugen and Mellor 1990).

We supplement the NLSY data with wage data from the Outgoing Rotation Groups (ORGs) of the 1988-1994 Current Population Surveys. The ORG data are used to generate information on the wage distributions of each state, so that a state-by-state measure of the importance of the minimum wage can be created. (9) The measure that we use is the ratio of the state's effective minimum wage to its median wage. (l0) The highest value of the minimum wage-median wage ratio is 0.5667 ($4.25/$7.50) for both Arkansas and Mississippi in 1992. The lowest value is 0.3182 ($3.35/$10.53) for New Jersey in 1989. As these values suggest, the ratio tends to be highest--so that the minimum wage is most likely to be a binding constraint on behavior--in relatively low-wage states, and takes on it lowest values in relatively high-wage states.

Starting Wages and Job Duration

To get a sense of the behavior of minimum wage workers versus other workers, we begin our analysis by exploring the data using Kaplan-Meier survivor functions. Note that we investigate turnover for men and women separately. While a pooled sample would allow us to cluster minimum wage observations, there is substantial evidence that turnover behavior differs across gender. Research by Lynch (1991), Light and Ureta (1992), and Grossberg (2000), for example, finds that turnover for men and women is determined by different sets of variables, so that turnover hazards should be estimated separately. Our estimates are consistent with that conclusion.

Since we expect the effects of legal minimum wages to differ in high-wage versus low-wage states, we present survivor functions separately for respondents whose minimum wage-median wage ratios are in the top quartile of all such ratios, so that the minimum wage is high relative to the prevailing local wage distribution, and for those whose minimum wage-median wage ratios are in the bottom quartile, so that the minimum wage is relatively low. One problem with disaggregating the data in this manner is the small number of low-wage workers, especially in the bottom quartile, where the minimum wage represents a very low rate of pay. (11) Nevertheless, a pattern starts to emerge, at least for men. In high-wage states, as shown in Figure 1, men in minimum wage jobs appear to have longer employment durations than those in other low-wage jobs. In low-wage states, on the other hand, as shown in Figure 2, men in minimum wage jobs have relatively short durations even when compared to men in other low-wage jobs. For women, as shown in Figures 3 and 4, no such pattern emerges. Job durations for women in minimum wage jobs are similar to those for women in other low-wage jobs, regardless of how the minimum wage compares to the prevailing local wage distribution.

[FIGURES 1-3 OMITTED]

5. Duration Analysis

In this section of the paper we use multivariate duration analysis to explore the behavior of workers hired at the minimum wage. This approach allows us to more effectively isolate the effects of the minimum wage on employment duration. Our discussion in section 3 of the possible effects of minimum wages on turnover suggests that minimum wages are most likely to affect employee-initiated turnover, either by creating rents or by distorting information. Fortunately, the NLSY contains information concerning the reasons that respondents leave their jobs, and this allows us to focus our empirical analysis on job spells terminated by the employee. Accordingly, we estimate duration models using a competing risks framework for employee- and employer-initiated turnover.

Our models include controls for the starting wage, human capital, job and firm characteristics, personal and family characteristics, demographic and labor market characteristics, occupation, and industry. (12) As in the preceding descriptive analyses, we use dummy variables to distinguish between workers who have starting wages below the prevailing minimum wage, equal to the minimum wage, above the minimum wage by no more than 10%, and those with starting wages more than 10% above the minimum wage (the omitted category). To control for the extent to which the minimum wage binds, we interact the minimum wage-median wage ratio with the starting wage dummy variables. This allows us to measure how the relationship between starting at the minimum wage versus starting at other wage rates varies as the importance of the minimum wage itself varies.

The human capital variables include the logged sum of all hours of formal training reported in the 1988 interview or later and occurring prior to the start of the current job, as well as measures of formal training funded by the employer, formal training not funded by the employer, actual labor market experience, and education. To account as much as possible for variations in job characteristics, a dummy variable equal to one if the employer makes health insurance available is included, along with a dummy variable equal to one if the respondent reports being covered by a union contract, a dummy variable equal to one if the respondent works part time (fewer than 35 hours per week), and the log of the number of employees in the firm. The difference between the log of the current hourly real wage and the log of the starting hourly real wage is also included and is meant to represent the difference between the respondent's current earnings and the value of the next best job offer. (13)

Regressions also control for the respondent's percentile score on the Armed Forces Qualification Test, the number of children in the household, and the log of annual real family income net of the respondent's annual earnings. In an effort to control for "movers" versus "stayers," the number of jobs ever held divided by the number of years of labor market experience is included. South and urban dummy variables are included along with the local unemployment rate. Finally, both the calendar year and the state of residence are included as a time-varying covariates.

The basic analytical tool we use is the semiparametric Cox proportional hazard model. The hazard function for the ith respondent at time t, [h.sub.i](t), can be written as

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII],

where [[lambda].sub.0](t) is the baseline hazard, [beta] is a parameter vector, and [X.sub.it], is a vector of covariates, some of which may vary with time. The principal advantage of the Cox proportional hazard model is that [[lambda].sub.0](t) is not estimated and is therefore not constrained to any particular functional form. As a result, estimates of [beta] are consistent in the absence of unobserved individual-specific heterogeneity. If unobserved individual-specific heterogeneity is present, the Cox model yields estimates of [beta] that are biased toward zero but that have approximately correct standard errors (Ridder and Verbakel 1983; Gail, Wieand, and Piantadosi 1984, as reported in Lancaster 1990, p. 304). Thus, the principal effect of unobserved individual-specific heterogeneity in the Cox model is to reduce the likelihood of rejecting null hypotheses concerning [beta].

We allow for the possibility that the risks of employee- and employer-initiated terminations are interdependent by estimating marginal distributions in the manner suggested by Wei, Lin, and Weissfeld (1989). This method, which allows estimation of failure-specific partial likelihoods, is attractive because it accounts for potential interdependence of the competing risks without imposing a specific functional form. We focus on estimated competing risk models for employee-initiated turnover in the remainder of the paper. Estimated competing risk models for employer-initiated turnover are contained in an Appendix available from the authors upon request.

Table 1 reports competing risk estimates for employee-initiated terminations for men and women. To facilitate the exposition we report only the effects of the starting wage variables. For men, we find that the conditional probability of ending a job spell at any point in time is affected significantly by the interaction between the minimum wage-median wage ratio and whether the starting wage in a job was equal to the minimum wage. Specifically, we find that the likelihood a man in a minimum wage job will terminate a job spell increases as the minimum wage-median wage ratio increases. To understand the implications of this, we evaluate the net effects of each of the starting wage variables at the 90th percentile, median, and 10th percentile of the minimum wage-median wage ratio.

When the ratio is set to its value at the 90th percentile, 0.509--Table 1B--we find that the conditional weekly quit hazard of men hired at the minimum wage is about 51 times that of high-wage workers. (14) More importantly, however, we find that their job spells have a significantly higher probability of termination than do those of other low-wage workers. We do not find a specific minimum wage effect when we evaluate the regression with the minimum wage-median wage ratio set to its median of 0.4308 (Table 1C). (15) At this level men in both minimum wage jobs and in jobs starting below the minimum wage have greater conditional quit hazards than men in high-wage jobs, though these effects are not significantly different from one another. Finally, when the minimum wage-median wage ratio is set to its value at the 10th percentile, 0.3691 (Table 1D), men in minimum wage jobs are estimated to have lower conditional quit hazards than men in high-wage jobs, although the difference is not statistically significant. However, men in minimum wage jobs do have (marginally) significantly lower conditional quit hazards than men in other low-wage jobs.

Taken together, the estimates in Table 1B, C, and D are consistent with the hypothesis that while legal minimum wages may create rents for some workers, they also disrupt the job matching process when the price floor is high relative to the local wage distribution. These estimates suggest that in low-wage states the net effect of a minimum wage is consistent with the information distortion hypothesis. In high-wage states, the point estimates suggest that voluntary turnover is lower for men in minimum wage jobs than for men in other low-wage jobs. This is consistent with the hypothesis that rents in minimum wage jobs are most likely to occur in states where the effective minimum wage is low relative to the prevailing wage distribution.

For women, no such story emerges. Indeed, for women we find no evidence of a relationship between the starting wage rate and employment duration. As discussed above, it is common to find gender differences in the causes and consequences of turnover. While we can speculate on the cause of the difference we observe, we are unable to state its source with certainty. There are several possible explanations.

Loprest (1992) investigates gender differences in job mobility and wage growth using early waves of the NLSY. She finds that men and women experience similar wage growth when not changing jobs and that they demonstrate similar rates of job mobility. Men, however, experience much greater wage growth when changing jobs than do women. Loprest hypothesizes that gender differences in job preferences may explain the difference in the wage effects of job mobility. Specifically, she argues that, "women may on average switch to jobs that have more flexible schedules or lower hours than men because of different household responsibilities" (Loprest 1992, p. 528). In support of this hypothesis, she shows that the women in her sample more often switch from full-time to part-time work and that this switch results in lower wages for both genders. Keith and McWilliams (1997), using a longer panel from the NLSY, find mixed evidence. They find no gender differences in the wage effects of aggregated turnover. When they distinguish between employer- and employee-initiated turnover, they find that both men and women experience wage gains from quitting, but that men's wage growth exceeds women's wage growth by 35% (8.5% vs. 6.3%). When the data are further disaggregated by reason for quitting, they find that the return to family-related quits is much lower than for non-family-related quits and that women are much more likely than men to quit for family-related reasons. Similarly, using data from a single firm, Sicherman (1996) finds that women are more apt to quit for nonmarket reasons (family illness, change of residence, etc.) than are men. This is true for our data as well: among those who leave jobs voluntarily, the men in our sample report doing so for family-related reasons about 2.2% of the time, whereas the women leave for family reasons 33.8% of the time. Hence, it may be that the rent-forming and/or information-distorting effects of the minimum wage for the women in our sample are overwhelmed by these other concerns. (16)

We are unable to observe personal attributes such as labor market attachment or job attributes such as the location of a job, although the NLSY does contain data on the availability of some family-friendly fringe benefits. (17) Table 2 shows how median employment durations vary with the starting wage and with the presence of these fringe benefits for men and women. The data show that the presence of these amenities is associated with longer employment durations for women, particularly women in low-wage jobs, although the relationship for men is much more tenuous. In general, these data appear to indicate that, on average, employment duration is relatively elastic for women in low-wage jobs with respect to the presence family-friendly benefits. (18)

Perhaps, as Royalty (1996) suggests, investments in job-specific human capital for women are impeded by their higher turnover propensities, and this tends to undermine their prospects for achieving lasting job matches, particularly in low-wage jobs. Finally, the possibility that discrimination may bar women from certain higher paying occupations and jobs would clearly be expected to affect turnover decisions and may cause observable differences in turnover behavior.

6. Conclusion

This paper provides evidence on the impact of legal minimum wages on employment duration. The standard price floor approach to this topic emphasizes the rents created by the minimum wage and implies minimum wages will reduce turnover. A second approach concentrates on the information loss that results from the compression of the wage distribution as the minimum wage increases and as employers attempt to offset the minimum wage by altering nonpecuniary aspects of the jobs.

Our analysis recognizes that these two lines of reasoning are not mutually exclusive, and we argue that the net impact depends on the value of the minimum wage in relation to local market wages. For men we find that where the minimum wage is low relative to the state median wage, employee-initiated turnover hazards are reduced. This finding is consistent with the argument that minimum wages create rents in the low-wage labor market. We also find clear evidence that when the minimum wage is high relative to the state median wage, minimum wages increase employee-initiated turnover. We suggest that this occurs because when the minimum wage binds, it distorts information in the labor market, thereby decreasing the likelihood of achieving good job matches. For women, we find no evidence of a connection between the starting wage and turnover behavior. While we discuss possible explanations for the gender differences in the impact of the minimum wage, future research is required to provide a more definitive explanation.

The implications of this research regarding the welfare effects of minimum wage laws are mixed. On one hand, where minimum wages are relatively low, holding aside employment effects of minimum wage laws, our findings suggest that minimum wage legislation results in more stable employment relationships. While this may limit access to minimum wage jobs for some applicants, it is also likely to reduce hiring costs for firms. On the other hand, our findings also suggest that large increases in the minimum wage--or increases when the wage floor is already high relative to local wages--may make the job matching process more difficult, thus increasing turnover in low-wage labor markets. These conclusions are tempered, however, by the lack of evidence that minimum wage laws significantly impact women's turnover behavior, since women make up a significant proportion of minimum wage earners.

Table 1. Duration Model Estimates for
Employee-Initiated Separations (a)

 Men Women

A. Estimated Coefficients
 (asymptotic t statistics
 in parentheses) (b)

Started at the minimum wage -14.3666 0.2992
 (-3.29) (0.23)
Started at the minimum wage x 35.9728 -1.4632
 (Minimum wage/state median wage) (3.55) (-0.50)
Started below the minimum wage 9.6854 3.3469
 (2.63) (0.79)
Started below the minimum wage x -20.8465 -7.2001
 (Minimum wage/state median wage) (-2.52) (-0.74)
Started no more than 10% above the minimum wage 4.2497 1.1608
 (1.96) (0.73)
Started no more than 10% above the minimum wage x -9.7356 -3.1884
 (Minimum wage/state median wage) (-2.06) (-0.84)
(Minimum wage/state median wage) 0.8007 0.0621
 (0.46) (1.38)

B. When (minimum wage/state median
 wage) is at the value of its
 90th percentile, 0.509

 Net effect of "started at the minimum wage" 3.9445 -0.4456
 Net effect of "started below the minimum wage" -0.9245 -0.3179
 Net effect of "started no more than 10% above
 the minimum wage" -0.7067 -0.4621
p values for hypothesis tests
 [H.sub.0]: Started at the minimum wage = 0 0.00001 0.1063
 [H.sub.0]: started below the minimum wage = 0 0.1559 0.6193
 [H.sub.0]: started no more than 10% above the
 minimum wage = 0 0.1196 0.2274
 [H.sub.0]: started at the minimum wage =
 started below the minimum wage <0.00001 0.8742
 [H.sub.0]: started at the minimum wage =
 Started no more than 10% above the
 minimum wage <0.00001 0.9660

C. When (minimum wage/state median wage)
 is at its median sample value of 0.4308

 Net effect of "started at the minimum wage" 1.1315 -0.3312
 Net effect of "started below the minimum wage" 0.7047 0.2451
 Net effect of "started no more than 10% above
 the minimum wage" 0.0548 -0.2128
p values for hypothesis tests
 [H.sub.0]: started at the minimum wage = 0 0.0029 0.0777
 [H.sub.0]: started below the minimum wage = 0 0.0510 0.3248
 [H.sub.0]: started no more than 10% above the
 minimum wage = 0 0.8837 0.1799
 [H.sub.0]: started at the minimum wage =
 started below the minimum wage 0.4306 0.0241
 [H.sub.0]: started at the minimum wage =
 Started no more than 10%n above the
 minimum wage 0.0266 0.5621

D. When (minimum wage/state median
 wage) is at the value of its
 10th percentile, 0.3691

 Net effect of "started at the minimum wage" -1.0881 -0.2409
 Net effect of "started below the minimum wage" 1.9910 0.6894
 Net effect of "started no more than 10% above
 the minimum wage" 0.6556 -0.0160

p values for hypothesis tests
 [H.sub.0]: started at the minimum wage = 0 0.1288 0.3860
 [H.sub.0]: started below the minimum wage = 0 0.0045 0.2925
 [H.sub.0]: started no more than 10% above
 the minimum wage = 0 0.2168 0.9451
 [H.sub.0]: started at the minimum wage =
 Started below the minimum wage 0.0021 0.1529
 [H.sub.0]: started at the minimum wage =
 started no more than 10% above the
 minimum wage 0.0594 0.4964

(a) The estimated equations also include controls for training
prior to the start of the current job as well as on the current
job, previous experience, education, availability of medical
insurance, union membership, employer size, wage growth on the
current job, whether the job is part time, race, Armed Forces
Qualifications Test (AFQT) score, number of children, marital
status, number of jobs per year of experience, net family
income, residence in the South or in an urban area, local
unemployment rate. calendar year, state of residence, industry,
and occupation.

(b) t statistics are calculated using robust standard errors.

Table 2. Fringe Benefits and Median Employment
Durations for Men and Women

 Median Employment Duration
 When the Starting Wage Is

 Below
 Minimum Wage Minimum Wage
Variable (sample size) (sample size)

Women Flexible work 36 49
 hours (13) (31)
 Inflexible work 16.5 24
 hours (24) (19)
Men Flexible work 14 17
 hours (9) (8)
 Inflexible work 39 17.5
 hours (13) (12)
Women Unpaid parental 33.5 61
 leave available (18) (36)
 No unpaid parental 20 28
 leave available (28) (44)
Men Unpaid parental 15.5 42
 leave available (6) (9)
 No unpaid parental 26.5 29
 leave available (18) (18)
Women Medical insurance 33.5 61
 available (10) (34)
 No medical insurance 23.5 33
 available (38) (49)
Men Medical insurance 14.5 48.5
 available (14) (10)
 No medical insurance 39.5 18
 available (12) (20)

 Median Employment Duration
 When the Starting Wage Is

 [less than
 or equal to]
 10% above >10% above
 Minimum Wage Minimum Wage
Variable (sample size) (sample size)

Women Flexible work 38 65.5
 hours (45) (386)
 Inflexible work 30 52
 hours (37) (321)
Men Flexible work 21 64.5
 hours (6) (386)
 Inflexible work 19 64
 hours (13) (432)
Women Unpaid parental 61 77
 leave available (56) (675)
 No unpaid parental 25 31
 leave available (49) (275)
Men Unpaid parental 30 80
 leave available (9) (471)
 No unpaid parental 28 54
 leave available (21) (543)
Women Medical insurance 46 73
 available (51) (723)
 No medical insurance 30 38
 available (61) (313)
Men Medical insurance 20.5 78
 available (18) (874)
 No medical insurance 28 43
 available (15) (309)

The authors are grateful to Patrick Nolen for his able research assistance and to Robert Margo and an anonymous referee, whose comments helped to improve this paper.

(1) The Bureau of Labor and Statistics (BLS) estimates that 60.3% of minimum wage workers were 20 years old or older in 1988, while 38.2% were 25 or older (Haugen and Mellor 1990, table 1).

(2) For evidence on the effects of minimum wage laws on the wage distribution see Card and Krueger (1995) and Lee (1999).

(3) Whether minimum wages do, in fact, reduce on-the-job training is an unresolved question. Early research by Hashimoto (1982) and Leighton and Mincer (1981) suggested a negative impact. More recently, Neumark and Wascher (1998) also find a negative impact. However, Acemoglu and Pischke (1999), Grossberg and Sicilian (1999), and Fairris and Pedace (2004) find little or no effect.

(4) Bac (2002) shows that in the absence of information asymmetries (and with a high enough ratio of "good jobs" to "bad workers"), wage offers act as efficient signals to applicants and as screens for employers, with the result that labor market equilibria are characterized by efficient matching.

(5) Of course, there are also legal constraints, like the Equal Pay Act, that prevent employers from setting differential pay for substantially similar work.

(6) The military subsample was reduced from its original 1280 to fewer than 200 respondents as of 1985, while about 1000 low-income whites were eliminated as of 1991.

(7) One of the advantages of the NLSY is its collection of detailed information on formal training within jobs. This is useful because, as mentioned above, one of the unresolved issues pertaining to the minimum wage is whether those hired at the minimum wage receive less training than other low-wage workers and to what extent any reduction in training affects employment duration. While the NLSY survey has contained questions regarding formal training since its inception, the data collected in the survey improved dramatically in 1988. Prior to 1988 information was collected only on training spells lasting at least 4 weeks. With the 1988 survey, the NLSY began collecting information on up to four training spells per year, regardless of the length of those training spells.

(8) As alternatives to the "no more than 10% above the minimum wage" dummy variable we also experimented with "no more than 25 cents above the minimum wage" and "no more than 50 cents above the minimum wage." Our results were unaffected by these changes.

(9) We describe the method for calculating hourly wages from the ORG data in an Appendix available from the authors upon request.

(10) We experimented with other reference points from the state wage distributions--for example we used the minimum wage as a percentage of the wage representing the 20th percentile in the state, but our findings were unaffected. The median wage is computed for each calendar year for each state, using workers aged 18-64. To control for regional cyclical fluctuations we include controls for the calendar year of each observation and for state of residence. Following Acemoglu and Pischke (1999), we also computed estimates using the median nominal wage for the entire 1988-1994 period for each state. Our results were unaffected by this alternative specification.

(11) This is a common problem in the literature on minimum wages using the NLSY (e.g.. Acemoglu and Pischke 1999).

(12) Tables containing descriptive statistics for all variables used in the analysis that follows are available from the authors.

(13) We recognize the potential endogeneity of this variable. All of the models presented below were also estimated with the wage growth variable omitted. This change affects none of our conclusions.

(14) To calculate the relative hazard, take the exponential of the point estimate: exp 3.9445 = 51.65.

(15) States with minimum wage-median wage ratios between 0.425 and 0.435 include California in 1989; Colorado in 1993 and 1994; Connecticut in 1991; Georgia in 1988; Michigan in 1993 and 1994; Minnesota in 1992 and 1993; Missouri in 1988 and 1989: Ohio in 1993; Oregon in 1990, 1991, 1992 and 1993; Texas in 1988; Virginia in 1992; and Wisconsin in 1990.

(16) In regressions not reported here, we estimate competing risk models for women with "employee-initiated terminations" separated into two categories: "employee-initiated terminations for family reasons" and "employee-initiated terminations for other reasons." We find no relationship between starting wages and the quit hazard in either case.

(17) These include whether the job offers flexible work hours, whether unpaid parental leave is available, and whether medical insurance is available. In addition to these fringes, the NLSY contains data on the availability of child care, dental insurance, on-the-job training or education, retirement benefits, and profit sharing. We do not include child care in our analysis because only one woman with a sub-minimum wage and one woman at the minimum wage report an employer that made child care available. Six women with wages between the minimum wage and 10% above the minimum wage report having child care available.

(18) In regressions not reported here, we estimate duration models for employee-initiated turnover for women, with dummy variables indicating the presence of specific fringe benefits included, and interacted with a low-wage dummy variable. We find that child care availability significantly reduces the probability of turnover for low-wage workers. Also. the availability of parental leave significantly reduces the conditional quit hazard at all wage levels.

References

Acemoglu, Daron, and Jorn-Steffen Pischke. 1999. Minimum wages and on-the-job training. NBER Working Paper No. W7184. Bac, Mehmet. 2002. On the informational content of wage offers. International Economic Review 43:173-92.

Card, David, and Alan B. Krueger. 1995. Myth and measurement: The new economics of the minimum wage. Princeton, NJ: Princeton University Press.

De Fraja, Gianni. 1999. Minimum wage legislation, productivity and employment. Economica 66:473-88.

Fairris, David, and Roberto Pedace. 2004. The impact of minimum wages on job training: An empirical exploration with establishment data. Southern Economic Journal. 70:566-83.

Gail, M. H., S. Wieand, and S. Piantadosi. 1984. Biased estimates of treatment effect in randomized experiments with nonlinear regressions and omitted covariates. Biometrika 71:431-44.

Gardecki, Rosella, and David Neumark. 1998. Order from chaos? The effects of early labor market experiences on adult labor market outcomes. Industrial & Labor Relations Review 51:299-322.

Grossberg, Adam J. 2000. The effect of formal training on employment duration. Industrial Relations: A Journal of Economy and Society; 39:578-99.

Grossberg, Adam J., and Paul Sicilian. 1999. Minimum wages, on-the-job training, and wage growth. Southern Economic Journal 65:539-56.

Hashimoto, Masanori. 1982. Minimum wage effects on training on the job. American Economic Review 72:1070-87.

Haugen, Steven E., and Earl F. Mellor. 1990. Estimating the number of minimum wage workers. Monthly Labor Review 113:70-4.

Holmlund, Bertil, and Harald Lang. 1985. Quit behavior under imperfect information: Searching, moving, learning. Economic Inquiry 23:383-93.

Holzer, Harry J., Lawrence F. Katz, and Alan B. Krueger. 1991. Job queues and wages. Quarterly Journal of Economics 106:739-68.

Johnson, William R. 1978. A theory of job shopping. Quarterly Journal of Economics 92:261-77.

Katz, Lawrence F., and Alan B. Krueger. 1992. The effect of the minimum wage on the fast-food industry. Industrial and Labor Relations Review 46:6-21.

Keith, Kristen, and Abagail McWilliams. 1997. Job mobility and gender-based wage growth differentials. Economic Inquiry 35:320-33.

Lancaster, Tony. 1990. The econometric analysis of transition data. Cambridge, UK: Cambridge University Press.

Lee, David S. 1999. Wage inequality in the United States during the 1980s: Rising dispersion or falling minimum wage? Quarterly Journal of Economics 114:977-1023.

Leighton, Linda, and Jacob Mincer. 1981. The effects of minimum wages on human capital formation. In The economics of legal minimum wages, edited by Simon Rottenberg. Washington, DC: American Enterprise Institute, pp. 155-73.

Light, Audrey, and Manuelita Ureta. 1992. Panel estimates of male and female job turnover behavior: Can female nonquitters be identified? Journal of Labor Economics 10:156-81.

Loprest, Pamela J. 1992. Gender differences in wage growth and job mobility. American Economic Review 82:526-32.

Lynch, Lisa M. 1991. The role of off-the-job vs. on-the-job training for the mobility of women workers. American Economic Review 81:151-6.

Mincer, Jacob. 1988. Job training, wage growth, and labor turnover. NBER Working Paper No. 2690.

Mixon, J. Wilson, Jr. 1978. The minimum wage and voluntary labor mobility. Industrial and Labor Relations Review 32: 67-73.

Neumark, David, and William Wascher. 1998. Minimum wages and training revisited. NBER Working Paper No. 6651.

Parent, Daniel. 1999. Wages and mobility: The impact of employer-provided training. Journal of Labor Economics 17: 298-317.

Ridder, Geert, and Wim Verbakel. 1983. On the estimation of the proportional hazard model in the presence of heterogeneity. University of Amsterdam, Faculty of Actuarial Science and Econometrics, A&E Report 22/83.

Royalty, Anne Beeson. 1996. The effects of job turnover on the training of men and women. Industrial and Labor Relations Review 49:506-21.

Sicherman, Nachum. 1996. Gender differences in departures from a large firm. Industrial and Labor Relations Review 49:484-505.

Sicilian, Paul, and Adam J. Grossberg. 1993. Do legal minimum wages create rents? A re-examination of the evidence. Southern Economic Journal 60:201-9.

Wei, L. J., D. Y. Lin, and L. Weissfeld. 1989. Regression analysis of multivariate Incomplete failure time data by modeling marginal distributions. Journal of the American Statistical Association 84:1065-73.

Wessels, Walter J. 1980a. Minimum wages, fringe benefits, and working conditions. Washington, DC: American Enterprise Institute.

Wessels, Walter J. 1980b. The effect of minimum wages in the presence of fringe benefits: An expanded model. Economic Inquiry 18:293-313.

Received April 2002; accepted April 2003.

Adam J. Grossberg * and Paul Sicilian ([dagger])

* Department of Economics, Trinity College, 300 Summit Street, Hartford, CT 06106 USA; E-mail adam.grossberg@ trincoll.edu.

([dagger]) Department of Economics, 480C DeVos Center, Grand Valley State University, 401 W. Fulton Street, Grand Rapids. MI 49504 USA; E-mail siciliap@gvsu.edu; corresponding author.