I. Conceptual Framework

The concept of a “skilled worker” is rather nebulous since the term “skill level” is often used as a portmanteau to refer to various aspects of human capital. The literature on human capital theory makes an important analytical distinction between general and firm-specific (or industry-specific) human capital (see Becker (1962)).³ General human capital increases the productivity of a worker in any firm. Such capital is assumed not to depreciate when a worker switches from one firm (or industry) to another. Firm-specific capital is, by definition, not transferable across firms. There is an element of risk to investment in such capital since a separation of the worker from the firm leads to the loss (or significant depreciation) of such capital. The real resource costs of investment in specific capital include training costs and forgone output from time spent training rather than directly producing output.⁴ Workers may bear a portion of these costs by accepting a wage below their marginal product when they join the firm. This joint investment provides an incentive for both the firm and the worker to avoid a separation.

Specific human capital is key to many of the models of labor market contracting that have direct implications for the cyclical behavior of employment, hours, and wages for workers of different skill levels. These models may be classified into two broad categories. The implicit contract models of Azariadis (1975, 1976), Baily (1974), and Gordon (1974) imply the existence of wage-smoothing arrangements provided by firms for their workers. The implicit contract models of Hashimoto (1981) and Raisian (1983), on the other hand, imply the existence of labor contracts with more procyclical compensation for workers with higher skill levels. The main differences between these two sets of models arise from their assumptions regarding the rules that determine how the costs of investment in specific capital and the returns from it are shared by firms and workers.

The Azariadis-Baily-Gordon class of models postulates that risk-neutral firms may implicitly offer insurance to risk-averse workers by guaranteeing them relatively stable employment and wages when faced with uncertain demand for the firm’s product. In its basic form, this theory suggests that all workers receive some form of employment or wage insurance, reflected in the weak response of employment and wage measures to demand shocks or other real shocks.

By incorporating firm-specific human capital, this theory can be extended to the case of risk-averse firms and heterogeneous workers. Risk-averse firms would, in general, not be willing to take on the risks posed by fluctuations in demand or productivity. However, if firms bear the cost of investment in specific human capital, they might be reluctant to lose the specific capital embodied in their workers. When faced with adverse demand or productivity shocks that they perceive to be transitory, firms then have an incentive to retain workers with high specific skill levels. Thus, skilled workers may be offered contracts that, at business cycle frequencies, provide them with smoother wages and more stable employment patterns than unskilled workers.⁵ Unskilled workers would simply be paid their marginal product in every period, and their employment would depend solely on current-period demand conditions. Hence, their employment and wages would tend to be strongly procyclical.

The Hashimoto-Raisian class of models has markedly different implications. The key underlying notion in these models is that specific capital involves a joint investment by the firm and the worker. Further, the returns from this form of human capital are assumed to be shared by the two parties. This gives both parties, the worker and the firm, an incentive to avoid a separation that would cause specific capital to be lost. Thus, workers with more firm-specific human capital would face a trade-off between employment and wage variability. Since such workers typically have higher incomes than unskilled workers, they are likely to have better access to asset markets where they could insure against income fluctuations. Consequently, skilled workers would be willing to accept more procyclical variation in wages than unskilled workers, in return for a greater degree of employment stability.

It follows from the above argument that the larger their share in the returns from specific capital, the more employment stability the skilled workers would value (or the more variability in wages they would accept in return for not being laid off).⁶ However, it is likely that skilled workers face a greater degree of adjustment at the intensive margin (that is, weekly hours worked). In other words, firms may respond to downturns by laying off unskilled workers and reducing the hours worked by skilled workers. This implication follows from the assumption that specific human capital is unaffected by hours variation but depreciates if a firm and a worker are separated for one or more periods.

However, the Hashimoto-Raisian model could also be consistent with countercyclical hours variation for skilled workers. Since highly skilled workers in many industries typically earn salaries rather than hourly wages, working longer hours in a downturn may be a means of taking a temporary cut in hourly wages.

Thus, the Azariadis-Baily-Gordon and Hashimoto-Raisian models have similar predictions about employment variability: skilled workers should face less cyclical variation in employment. Neither set of models has a definitive implication regarding relative cyclical variability of hours for workers of different skill levels, although, in its simplest form, the Hashimoto-Raisian model implies that skilled workers face more procyclical variation at the intensive margin (hours worked).

The predictions regarding differences in wage variability across skilled and unskilled workers are diametrically opposed in these two sets of models. The Azariadis-Baily-Gordon class of models posits the existence of contracts that assure skilled workers of smoother wages and employment, thereby implying a countercyclical wage premium for skilled labor. This is consistent with the findings of Reder (1955, 1962), who uses aggregate data to show that skilled-unskilled wage differentials narrow in booms and widen in recessions. The Hashimoto-Raisian model, on the other hand, implies that workers with more specific capital face more variable wages. That is, skilled workers may have more stable employment but are subject to greater wage variability than unskilled workers. This implies procyclical wage differentials between skilled and unskilled workers. Raisian (1983) provides evidence in support of this view.

We attempt to provide an empirical resolution of this issue using a detailed set of micro survey data to correct for various potential sources of bias. In our empirical work, we use different proxies for skills in order to disentangle the various effects of human capital on wage and employment variability. The primary determinant of general human capital is education. The costs of such human capital investment are usually considered to be borne entirely by the worker, as are the returns accruing from it. It is also likely that measures of general human capital, such as education, are highly positively correlated with levels of specific human capital. For instance, workers with a college degree presumably have attributes that make them more efficient at acquiring specific capital. Thus, education levels are a good measure of general as well as specific human capital.

A more direct proxy for specific human capital is tenure on the current job (see Altonji and Shakotko (1987)). On-the-job training and learning-by-doing are likely to enhance skills that are of particular value to a specific firm or industry. Further, longer tenure arguably indicates the existence of specific capital that the firm and the worker are reluctant to lose.⁷ Another useful proxy for human capital is total labor market experience, although it is not obvious what aspect of human capital this best measures. We use all three variables as proxies for skills and estimate a series of models that independently analyze their effects on wage and employment variability.

II. Econometric Framework

The basic regression model is as follows:

The real hourly wage rate of individual i at time t is represented by W_it. X_it is a vector of observed individual-specific variables that affect this wage rate, with associated coefficient vector β. In our application, we use the aggregate unemployment rate in the economy, U_t, as an indicator of the cycle.⁸ α indicates the relation between the real wage and the business cycle. For instance, a negative estimate of α would imply that the average real wage declines when the aggregate unemployment rate rises (or that the average real wage is procyclical). μ_i stands for a vector of unobserved individual-specific characteristics that are fixed over time. The elements of μ_i may be correlated with X_it. The regression error €_it is assumed to be independently and identically distributed (i.i.d.).

We are interested in estimating the effects of observed measures of skills on the cyclicality of a worker’s real wage. This is accomplished by including an appropriate interaction term as follows:

The variable E_it is a measure of skill level (it should also be included in X_it). The coefficient γ on the interaction term U_t E_it captures differences in the cyclicality of wages for workers with different skill levels. A positive estimate of γ would indicate a countercyclical wage premium for skills—that is, the skill premium increases when the unemployment rate rises. Conversely, a negative γ would indicate a procyclical skill premium.

Estimating equation (2) by ordinary least squares (OLS), with μ_i + €_it being the composite error term, would yield biased estimates of β and γ unless the variables in μ_i were uncorrelated with the regressors. In general, this is not likely to be true. Workers with a high (unobserved) value of μ_i are high-ability workers. If high-ability workers were less likely to be laid off in a recession than were low-ability workers, the mean level of μ_i among employed workers would covary positively with the aggregate unemployment rate. The correlation between such unobserved individual fixed effects and the unemployment rate would induce an upward (countercyclical) bias in the estimated coefficient on the unemployment rate. Similarly, if an unobserved component of ability were positively correlated with, say, the observed level of education, the estimated coefficient on the education variable would be biased upward.

The interaction coefficient γ is subject to similar bias. For instance, if an increase in the aggregate unemployment rate caused the average unobserved ability of workers in lower skill categories (those with lower values of E_it) to rise relative to the average unobserved ability of workers in higher skill categories, γ would be biased downward. This procyclical bias in the estimated cyclical variation of the skill premium would spuriously indicate a narrowing (or understate the increase) of the skill premium in a recession.

To deal with such unobserved individual fixed effects, we employ a fixed effects estimator by using OLS to estimate the following transformed equation:

This transformation subtracts out individual means over time for each variable, causing the individual fixed effects to drop out. The error term €_it is i.i.d. and is uncorrelated with the regressors. Note that to implement the fixed effects model we need to leave out control variables that are constant over time or collinear with the time trend.

To estimate the cyclical behavior of wages and skill premia at the industry level, we could include interactions of U_t and U_t E_it with industry dummies. Specifically, the following counterpart to the OLS model in equation (2) may be employed:

I_ijt is a binary indicator variable that takes the value one if worker i locates in industryj at time t. Otherwise, I_ijt equals zero. The coefficients α and γ are now indexed by industry. For instance, γ_j is an estimate of the cyclical variation in the skill premium in industry j. With appropriate transformations of the variables as described in equation (3), a similar pooled regression could be used to estimate the fixed effects model at the industry level:

A potential problem with specification (5) is that it restricts individual fixed effects to be the same across all industries. This would bias the coefficients of industry-level estimates if there were industry-specific unobserved fixed effects that were correlated with any of the regressors.⁹ Further, both equations (4) and (5) restrict the coefficient vector β to be the same across industries. This implies the strong assumption that the returns to observed worker characteristics are the same in all industries.

Apart from unobserved individual fixed effects and industry-specific effects, there remains another potential source of bias. All of the above discussion assumes that the mean of €_it conditional on individual i being employed in period t is zero. But notice that wages are observed only for those individuals who are employed in a given period. If an unobserved component of ability that affected the wage rate for an individual was correlated with the unobserved component that affected that individual’s probability of employment, we would be faced with a typical selection bias problem.¹⁰ For instance, changes in U_t may cause workers with systematically high or low values of the time-varying unobserved productivity component (reflected in high or low values of €_it) to enter or leave employment in an industry. The effect of changes in average labor force quality resulting from the inflow or outflow of high or low productivity workers would then bias the unemployment rate coefficient. If, in addition, the magnitude of this effect differed by skill level, a fixed effects estimate of γ_j would be a biased estimate of the change in the mean offer-wage differential in industry j.

To eliminate the effect of cyclical changes in the composition of the work force induced by such systematic selection, we estimate the variability of mean industry offer wages for different classes of workers using a maximum-likelihood version of Heckman’s (1974) self-selection model. This model estimates a wage equation for each industry jointly with a probit choice equation that determines whether a worker locates in that industry. The model is written as follows:

Here $I_{ijt}^{*}$ is the latent index of a probit employment equation that determines whether worker i is employed in industry j at time t. Z_it is a vector of individual-specific regressors that affect the probability of employment in industry j at time t. The corresponding coefficient vector is denoted by θ_j Typically, Z_it contains elements that enter into X_it as well as some additional variables that may affect labor supply propensity but not worker productivity. Since our data set does not contain any variables that fall clearly into this category, we include the same set of controls in the wage and employment choice equations. Further, our results were not sensitive to the overidentifying restrictions of omitting variables from X_it. Individual fixed effects in the employment choice equation are represented by ψ_ij

We estimate binomial selection models separately for each industry. This allows fixed effects to vary across industries and, thus, obviates the potential bias from restricting the fixed effects for any given individual to be the same across all industries. Further, it allows the coefficient vector β to vary across industries.¹¹

The error terms €_ijt and ω_ijt are assumed to have a bivariate normal distribution with correlation ρ_jj and respective standard deviations σ_ij and 1. The latter variance is normalized to one for identification of the probit choice equation. The parameter ρ_j, estimated from the cross-equation correlation between wage equation residuals and employment equation residuals, is crucial for correcting the selection bias. The sign of this correlation coefficient is informative. A negative estimate of ρ_j for instance, would indicate that workers with a high transitory wage component are more likely to be laid off in a downturn. This would, if a selection correction is not employed, bias the estimated effect of the cycle on the real wage and the skill premium.¹²

The source of selection bias can now be demonstrated fairly easily. For instance, consider the coefficient α_j in the above wage equation. Under the distributional assumptions made above (and ignoring fixed effects for the moment), we have

where m_ijt = λ_ijt (λ_ijt + Z_it θ_j + U_t δ_j) and λ_ijt is the Mill’s ratio. It can be shown that m_ijt > 0. Hence, the estimate of α_j is biased downward if ρ_j and δ_j have the same sign and is biased upward if they have opposite signs. The other coefficients in the wage equation are also affected by selection bias.

The selection-corrected fixed effects estimator in this paper is implemented by full-information maximum likelihood.¹³

III. Data

The data set used in this paper is the National Longitudinal Survey of Young Men (NLS), comprising a nationally representative sample of 5,225 young males. The participants were between 14 and 24 years of age in 1966 and were interviewed in 12 of the 16 years from 1966 to 1981. Data were collected on their employment status, wage rates, and sociodemo-graphic characteristics. We screened the sample to include only those persons who, as of the interview date, were at least 21 years of age, had completed their schooling and military service, and had available data for all variables used in our analysis. The final sample contained 4,439 males and a total of 23,927 person-year observations. This included years when the person being interviewed was unemployed. The employment status dummy is nonzero in 21,203 of these person-year observations. Table A1 in the Appendix reports sample means for the individual specific variables used in the estimation. Workers are classified into 11 broadly defined industries on the basis of the three-digit census industrial classification (CIC) codes. The list of industries, their CIC codes, and the sample size for each industry are reported in the Appendix in Table A2.

The wage measure we use is the hourly straight-time earnings reported by workers for the survey week. This measure is deflated by the consumer price index (CPI) to provide a real wage measure normalized in terms of 1967 dollars. It is important to note that this is a point-in-time wage measure taken at the date of the interview. This obviates the recall bias that may contaminate annual measures that are obtained by dividing annual earnings by annual hours worked. The NLS does not include data on overtime earnings in all of the interview years. Hence, we restrict ourselves to a straight-time wage measure rather than attempt to impute overtime earnings for years in which they were not available. To adjust for nonwage compensation, such as variation in fringe benefits across industries, the hourly wage rate for each worker is multiplied by the ratio of total labor costs to wages in the corresponding industry. Data on total labor costs are obtained from the National Income and Product Accounts. The log of this adjusted real wage measure, denoted by W_CPI is used in all of our analysis. This variable has a sample mean of 1.065 in 1967 dollars.

The HOURS variable we use in this study is a measure of the weekly hours worked reported by the worker for the survey week. In four of the survey years, this variable was unavailable, and we use the “usual weekly hours worked” instead.¹⁴ The three variables used as proxies for human capital are DEGREE, EXPERIENCE, and TENURE. DEGREE is a dummy variable that takes the value one if the worker has a college degree and zero if he or she does not. EXPERIENCE is defined as the total number of years of labor market experience; it is calculated as the interview date minus the completion date of a worker’s schooling or military service, whichever is later. It is important to note that the EXPERIENCE variable is a measure of labor force participation rather than of actual work experience. TENURE is defined as the length of uninterrupted tenure (in years) on the current job.

We use the aggregate civilian unemployment rate in the economy as an indicator of the business cycle. The variable URATE is defined as the seasonally adjusted monthly unemployment rate for all civilian workers aged 16 years and older. We also experimented with other indicators of the business cycle, such as real GNP. Our main conclusions do not appear very sensitive to the particular choice of the business cycle indicator. We report below only the results using the aggregate unemployment rate.

IV. Empirical Results

This section presents empirical results on employment, hours, and wage variability across skill levels and concludes with an interpretation of the findings.

Employment Variability

Table 1 contains estimates from a set of linear employment probability models. The first two columns report results from regressions with the URATE * DEGREE interaction term.¹⁵ The second set of columns contains results from a similar regression using the URATE * EXPERIENCE term. In this table, the estimated coefficients on the interaction terms measure the differential impact of the cycle on employment probabilities for skilled and unskilled workers.¹⁶ In the top row, the significant positive coefficient on URATE * DEGREE for all workers (0.0161, with a standard error of 0.0035) implies that workers with higher levels of education have lower procyclical employment probabilities. In fact, the sum of the coefficients on URATE and URATE * DEGREE (-0.0164 + 0.0161) is close to zero indicating that workers with a college degree have essentially acyclical employment patterns. EXPERIENCE, on the other hand, does not seem to affect employment probabilities over the cycle, as the interaction term URATE * EXPERIENCE for all workers has a coefficient (-0.0005, s.e. 0.0003) that is not significantly different from zero.

Table 1.

Estimated Correlations of Business Cycle with Employment Probabilities

Notes: Standard errors are in parentheses. Double asterisks (^**) indicate significance at 5 percent level. A single asterisk (*) indicates the 10 percent level. Sample size = 23,927. Controls are a time trend; education; experience and its square; four dummies for types of college degrees; five dummies for fields of degree; an SMSA dummy: a south dummy; a race dummy; a marriage dummy; number of children; and interactions of experience with education, a college degree dummy, and a race dummy.

^aFinancial services, insurance, and real estate.

Table 1.

Estimated Correlations of Business Cycle with Employment Probabilities

		URATE *		URATE *
	URATE	DEGREE	URATE	EXPERIENCE
All workers	-0.0164*	0.0161**	-0.0099**	-0.0005
	(0.0020)	(0.0035)	(0.0028)	(0.0003)
Durable manufacturing	-0.0137**	0.0109**	-0.0170**	0.0008**
	(0.0025)	(0.0O44)	(0.0035)	-0.0004
Construction	-0.0029	0.0075**	-0.0003	-0.0001
	(0.0018)	(0.0032)	(0.0026)	(0.0003)
Transportation and utilities	-0.0015	0.0046	0.0001	-0.0001
	(0.0017)	(0.0030)	(0.0024)	(0.0003)
Wholesale trade	-0.0006	-0.0028	-0.0012	0.0000
	(0.0013)	(0.0033)	(0.0018)	(0.0002)
Retail trade	0.0022	0.0079**	0.0088**	-0.0007**
	(0.0019)	(0.0033)	(0.0027)	(0.0003)
F.I.R.E.a	0.0012	-0.0043**	0.0001	0.0000
	(0.0012)	(0.0020)	(0.0016)	(0.0002)
Services	0.0017	-0.0137**	-0.0026	0.0002
	(0.0020)	(0.0035)	(0.0028)	(0.0003)
Government	-0.0016	0.0046*	0.0029	-0.00O5**
	(0.0015)	(0.0026)	(0.0021)	(0.0002)
Agriculture	0.0002	0.0000	0.0005	-0.0001
	(0.0009)	(0.0016)	(0.0013)	(0.0001)
Mining	0.0001	0.0011	0.0007	-0.0001
	(0.0007)	(0.0013)	(0.0010)	(0.0001)

Notes: Standard errors are in parentheses. Double asterisks (^**) indicate significance at 5 percent level. A single asterisk (*) indicates the 10 percent level. Sample size = 23,927. Controls are a time trend; education; experience and its square; four dummies for types of college degrees; five dummies for fields of degree; an SMSA dummy: a south dummy; a race dummy; a marriage dummy; number of children; and interactions of experience with education, a college degree dummy, and a race dummy.

^aFinancial services, insurance, and real estate.

View Table

Table 1.

Estimated Correlations of Business Cycle with Employment Probabilities

		URATE *		URATE *
	URATE	DEGREE	URATE	EXPERIENCE
All workers	-0.0164*	0.0161**	-0.0099**	-0.0005
	(0.0020)	(0.0035)	(0.0028)	(0.0003)
Durable manufacturing	-0.0137**	0.0109**	-0.0170**	0.0008**
	(0.0025)	(0.0O44)	(0.0035)	-0.0004
Construction	-0.0029	0.0075**	-0.0003	-0.0001
	(0.0018)	(0.0032)	(0.0026)	(0.0003)
Transportation and utilities	-0.0015	0.0046	0.0001	-0.0001
	(0.0017)	(0.0030)	(0.0024)	(0.0003)
Wholesale trade	-0.0006	-0.0028	-0.0012	0.0000
	(0.0013)	(0.0033)	(0.0018)	(0.0002)
Retail trade	0.0022	0.0079**	0.0088**	-0.0007**
	(0.0019)	(0.0033)	(0.0027)	(0.0003)
F.I.R.E.a	0.0012	-0.0043**	0.0001	0.0000
	(0.0012)	(0.0020)	(0.0016)	(0.0002)
Services	0.0017	-0.0137**	-0.0026	0.0002
	(0.0020)	(0.0035)	(0.0028)	(0.0003)
Government	-0.0016	0.0046*	0.0029	-0.00O5**
	(0.0015)	(0.0026)	(0.0021)	(0.0002)
Agriculture	0.0002	0.0000	0.0005	-0.0001
	(0.0009)	(0.0016)	(0.0013)	(0.0001)
Mining	0.0001	0.0011	0.0007	-0.0001
	(0.0007)	(0.0013)	(0.0010)	(0.0001)

Notes: Standard errors are in parentheses. Double asterisks (^**) indicate significance at 5 percent level. A single asterisk (*) indicates the 10 percent level. Sample size = 23,927. Controls are a time trend; education; experience and its square; four dummies for types of college degrees; five dummies for fields of degree; an SMSA dummy: a south dummy; a race dummy; a marriage dummy; number of children; and interactions of experience with education, a college degree dummy, and a race dummy.

^aFinancial services, insurance, and real estate.

The remaining rows of the table break down the sample by industry. In our sample, durable manufacturing accounts for the bulk of the aggregate cyclical variation in employment. In this industry, workers with college degrees face little cyclical variation in employment probabilities. The employment probabilities of workers with a degree are procyclical in services and F.I.R.E. (finance, insurance, and real estate) and countercyclical in construction and retail trade. For workers without degrees, employment probabilities are highly procyclical in durable manufacturing and acyclical in all other industries.

Increased experience, by contrast, seems to lead to less procyclical employment in durable manufacturing. However, in retail trade and government, the URATE * EXPERIENCE coefficient is significantly negative, indicating that workers with more experience are more likely to be laid off in a recession. Experience does not have a significant effect in the remaining industries, where the weak employment effect may be due to the substitution of younger workers for older workers in a recession. Given the young age of our sample, this may offset the overall negative employment effect in those industries during a recession.

To sum up, Table 1 indicates that having a college degree significantly reduces the cyclical variation in a worker’s employment probability. On the other hand, the total labor force experience possessed by a worker does not have a similar effect on the cyclicality of his or her employment probability except in durable manufacturing, where increased experience significantly reduces procyclical variation in employment.

Variation in Weekly Hours

Table 2 provides estimates of the cyclical variability of weekly hours worked. The estimation framework is identical to that used for real wages, as discussed earlier. As in the case of wages, unobserved individual fixed effects could potentially bias measures of cyclical variation in weekly hours worked. For instance, it is likely that low-ability workers, on average, work fewer hours and are more likely to get laid off in a recession. This compositional change over the cycle might induce an upward bias in the URATE coefficient and indicate countercyclicality in weekly hours worked. However, OLS estimates for virtually all of the relevant coefficients were of the same sign and generally close in magnitude to the fixed effects (FE) estimates. Hence, we report only FE estimates in Table 2.

Table 2.

Estimated Correlations of Business Cycle with Weekly Hours Worked: Fixed Effects Estimates

(Dependent variable: Log weekly hours)

Notes: Standard errors are in parentheses. Double asterisks (^**) indicate significance at 5 percent level. A single asterisk (*) indicates the 10 percent level. Sample size = 21,004. For the tenure equations, sample size = 20,309 since the tenure variable is not available for all employed workers. The same set of controls is used as in Table 1, except that tenure is added as an additional control for the second set of equations.

^aFinancial services, insurance, and real estate.

Table 2.

Estimated Correlations of Business Cycle with Weekly Hours Worked: Fixed Effects Estimates

(Dependent variable: Log weekly hours)

		URATE *		URATE *		URATE *
	URATE	DEGREE	URATE	TENURE	URATE	EXPERIENCE
All workers	-0.0061**	0.0082**	-0.0049**	0.0002	0.0006	-0.0007**
	(0.0014)	(0.0022)	(0.0018)	(0.0003)	(0.0023)	(0.0003)
Durable manufacturing	-0.0077**	0.0131**	-0.0057*	0.0003	0.0020	-0.0009**
	(0.0023)	(0.0029)	(0.0029)	(0.0004)	(0.0036)	(0.0003)
Nondurable manufacturing	-0.0098*	0.0088**	-0.0058*	0.0000	-0.0035	-0.0007**
	(0.0029)	(0.0034)	(0.0036)	(0.0004)	(0.0042)	(0.0003)
Construction	-0.0035	0.0019	-0.0063*	0.0004	0.0003	-0.0006*
	(0.0032)	(0.0039)	(0.0036)	(0.0004)	(0.0047)	(0.0003)
Transportation and utilities	-0.0076**	0.0097**	-0.0083	0.0004	-0.0029	-0.0006*
	(0.0035)	(0.0039)	(0.0044)	(0.0004)	(0.0050)	(0.0003)
Wholesale trade	-0.0048	0.0050	-0.0O28	-0.0002	0.0027	0.0009**
	(0.0045)	(0.0035)	(0.0051)	(0.0004)	(0.0057)	(0.0004)
Retail trade	-0.0058*	0.0048	-0.0066*	0.0003	0.0008	-0.0008**
	(0.0030)	(0.0033)	(0.0035)	(0.0004)	(0.0043)	(0.0003)
F.I.R.E.a	-0.0032	0.0091**	-0.0020	0.0002	0.0083	-0.0011**
	(0.0056)	(0.0039)	(0.0059)	(0.0005)	(0.0071)	(0.0004)
Services	-0.0047	0.0082**	-0.0036	0.0004	-0.0039	-0.0002
	(0.0031)	(0.0027)	(0.0032)	(0.0004)	(0.0037)	(0.0003)
Government	-0.0037	0.0062*	0.0001	0.0000	0.0090*	-0.0011**
	(0.0044)	(0.0035)	(0.0052)	(0.0004)	(0.0055)	(0.0004)
Agriculture	-0.0127**	0.0082	-0.0146**	0.0005	0.0062	-0.0014**
	(0.0061)	(0.0084)	(0.0072)	(0.0006)	(0.0089)	(0.0004)
Mining	0.0059	0.0137*	-0.0031	0.0006	0.0225**	-0.0013**
	(0.0079)	(0.0083)	(0.0090)	(0.0006)	(0.0106)	(0.0005)

Notes: Standard errors are in parentheses. Double asterisks (^**) indicate significance at 5 percent level. A single asterisk (*) indicates the 10 percent level. Sample size = 21,004. For the tenure equations, sample size = 20,309 since the tenure variable is not available for all employed workers. The same set of controls is used as in Table 1, except that tenure is added as an additional control for the second set of equations.

^aFinancial services, insurance, and real estate.

View Table

Table 2.

Estimated Correlations of Business Cycle with Weekly Hours Worked: Fixed Effects Estimates

(Dependent variable: Log weekly hours)

		URATE *		URATE *		URATE *
	URATE	DEGREE	URATE	TENURE	URATE	EXPERIENCE
All workers	-0.0061**	0.0082**	-0.0049**	0.0002	0.0006	-0.0007**
	(0.0014)	(0.0022)	(0.0018)	(0.0003)	(0.0023)	(0.0003)
Durable manufacturing	-0.0077**	0.0131**	-0.0057*	0.0003	0.0020	-0.0009**
	(0.0023)	(0.0029)	(0.0029)	(0.0004)	(0.0036)	(0.0003)
Nondurable manufacturing	-0.0098*	0.0088**	-0.0058*	0.0000	-0.0035	-0.0007**
	(0.0029)	(0.0034)	(0.0036)	(0.0004)	(0.0042)	(0.0003)
Construction	-0.0035	0.0019	-0.0063*	0.0004	0.0003	-0.0006*
	(0.0032)	(0.0039)	(0.0036)	(0.0004)	(0.0047)	(0.0003)
Transportation and utilities	-0.0076**	0.0097**	-0.0083	0.0004	-0.0029	-0.0006*
	(0.0035)	(0.0039)	(0.0044)	(0.0004)	(0.0050)	(0.0003)
Wholesale trade	-0.0048	0.0050	-0.0O28	-0.0002	0.0027	0.0009**
	(0.0045)	(0.0035)	(0.0051)	(0.0004)	(0.0057)	(0.0004)
Retail trade	-0.0058*	0.0048	-0.0066*	0.0003	0.0008	-0.0008**
	(0.0030)	(0.0033)	(0.0035)	(0.0004)	(0.0043)	(0.0003)
F.I.R.E.a	-0.0032	0.0091**	-0.0020	0.0002	0.0083	-0.0011**
	(0.0056)	(0.0039)	(0.0059)	(0.0005)	(0.0071)	(0.0004)
Services	-0.0047	0.0082**	-0.0036	0.0004	-0.0039	-0.0002
	(0.0031)	(0.0027)	(0.0032)	(0.0004)	(0.0037)	(0.0003)
Government	-0.0037	0.0062*	0.0001	0.0000	0.0090*	-0.0011**
	(0.0044)	(0.0035)	(0.0052)	(0.0004)	(0.0055)	(0.0004)
Agriculture	-0.0127**	0.0082	-0.0146**	0.0005	0.0062	-0.0014**
	(0.0061)	(0.0084)	(0.0072)	(0.0006)	(0.0089)	(0.0004)
Mining	0.0059	0.0137*	-0.0031	0.0006	0.0225**	-0.0013**
	(0.0079)	(0.0083)	(0.0090)	(0.0006)	(0.0106)	(0.0005)

Notes: Standard errors are in parentheses. Double asterisks (^**) indicate significance at 5 percent level. A single asterisk (*) indicates the 10 percent level. Sample size = 21,004. For the tenure equations, sample size = 20,309 since the tenure variable is not available for all employed workers. The same set of controls is used as in Table 1, except that tenure is added as an additional control for the second set of equations.

^aFinancial services, insurance, and real estate.

The first two columns of Table 2 contain results from regressions with the URATE * DEGREE interaction term. In the top row, the coefficient on URATE for all workers is -0.0061 (s.e. 0.0014) and the coefficient on URATE * DEGREE is 0.0082 (s.e. 0.0022).¹⁷ Since the sum of these coefficients is near zero, weekly hours are essentially acyclical for workers with a college degree. On average, however, weekly hours worked are estimated to decline by 0.42 percent for every 1 percentage point increase in the unemployment rate.¹⁸ At the industry level, these results corroborate the aggregate finding that educated workers face little cyclical variation in weekly hours. In fact, weekly hours for educated workers appear to be weakly countercyclical in durable manufacturing, F.I.R.E., and services. On average, hours are procyclical in durable and nondurable manufacturing, transportation and utilities, retail trade, and agriculture.

The fact that, for workers with a degree, weekly hours are much less cyclical at the aggregate level, and are even countercyclical in industries such as durable manufacturing, is an important result. Since educated workers typically earn higher wages than other workers, our findings imply a countercyclical bias in those measures of hourly wages that simply divide the aggregate wage bill by aggregate hours in an industry.¹⁹ However, the bias induced by differentials in hours variability across skill levels turns out to be rather small, once differentials in employment variability are accounted for.²⁰ Turning to the second set of columns in Table 2, it is clear that tenure has virtually no effect on weekly hours either at the aggregate level or in any of the industries. In the third set of columns, the coefficient on URATE * EXPERIENCE is significantly negative for all workers and in virtually every industry. This indicates that older workers are more likely to work fewer hours in a recession.

While Table 2 shows that average weekly hours worked are clearly procyclical, the magnitude of the variation in weekly hours appears to be much less important than variation in employment in accounting for cyclical fluctuations in aggregate hours.²¹ The aggregate results from all three panels indicate that weekly hours decline by about 0.4 percent when the unemployment rate goes up by 1 percent. In other words, variation in average weekly hours accounts for about 30 percent of the total variation in aggregate hours. This echoes the findings of Hansen (1985) and others that a substantial fraction of the variation in aggregate hours in the postwar U.S. economy is explained by employment variation rather than variation in weekly hours.

However, in a few industries that exhibit weak employment responses to the cycle, it appears that procyclical hours variation may account for a relatively larger share of the variation in total labor input. Nondurable manufacturing, transportation and utilities, retail trade, and agriculture fall into this category.

Wage Variability

Table 3 presents results from a series of estimated wage equations that incorporate the URATE * DEGREE interaction term. The first two columns contain results from OLS regressions (specification (4)). The next two columns contain results from a fixed effects estimator (specification (5)). The last two columns report results from a selection-corrected fixed effects model (specification (6)).²²

Table 3.

Estimated Correlations of Business Cycle with Real Wages: Degree Interactions

(Dependent variable: Log real wage)

Note: Standard errors are in parentheses. Double asterisks (^**) indicate significance at 5 percent level. A single asterisk (*) indicates the 10 percent level. Sample size = 21,004. Controls are a time trend; education; experience and its square; four dummies for types of college degrees; five dummies for fields of degree; four dummies for occupation; an SMSA dummy; a south dummy; a race dummy; a marriage dummy; number of children; and interactions of experience with education, a college degree dummy, and a race dummy. Estimates for the selection models use the full sample of 23,927 person-year observations. The probit employment choice equation estimates from the selection models are not reported here.

a Financial services, insurance, and real estate.

Table 3.

Estimated Correlations of Business Cycle with Real Wages: Degree Interactions

(Dependent variable: Log real wage)

			Fixed effects		Selection-corrected
	OLS estimates		estimates		fixed effects
		URATE *		URATE *		URATE *
	URATE	DEGREE	URATE	DEGREE	URATE	DEGREE
All workers	0.0031	-0.0294**	-0.0062**	0.0013	-0.0057**	0.0012
	(0.0024)	(0.0041)	(0.0017)	(0.0026)	(0.0019)	(0.0013)
Durable manufacturing	0.0073**	-0.0348**	-0.0033	-0.0028	-0.0049*	-0.0059*
	(0.0036)	(0.0048)	(0.0026)	(0.0034)	(0.0030)	(0.0026)
Nondurable manufacturing	0.0082*	-0.0231**	-0.0044	-0.0004	-0.0029	-0.0170**
	(0.0046)	(0.0051)	(0.0034)	(0.0039)	(0.0033)	(0.0030)
Construction	-0.0118**	-0.0468**	-0.0112**	-0.0015	-0.0135**	0.0010
	(0.0049)	(0.0056)	(0.0037)	(0.0046)	(0.0040)	(0.0045)
Transportation and utilities	0.0222**	-0.0393**	0.0060	-0.0057	0.0037	0.0067
	(0.0055)	(0.0053)	(0.0041)	(0.0045)	(0.0043)	(0.0049)
Wholesale trade	-0.0O05	-0.0091	0.0007	0.0104**	-0.0020	0.0046
	(0.0007)	(0.0057)	(0.0052)	(0.0041)	(0.0051)	(0.0038)
Retail trade	-0.0O79 *	-0.0253**	-0.0065 *	0.0055	-0.0057	0.0136**
	(0.0047)	(0.0052)	(0.0035)	(0.0038)	(0.0036)	(0.0041)
F.I.RE.a	-0.0070	-0.0181**	-0.0087	0.0113**	-0.0111 *	-0.0044
	(0.0086)	(0.0057)	(0.0065)	(0.0046)	(0.0061)	(0.0043)
Services	-0.0039	-0.0245**	-0.0214**	0.0079**	-0.0161**	0.0152**
	(0.0049)	(0.0048)	(0.0036)	(0.0031)	(0.0044)	(0.0030)
Government	0.0176**	-0.0300**	-0.0004	0.0027	0.0028	0.0004
	(0.0067)	(0.0051)	(0.0051)	(0.0041)	(0.0049)	(0.0030)
Agriculture	0.0102	-0.0132	0.0117 *	-0.0019	0.0168 *	-0.0130
	(0.0092)	(0.0080)	(0.0071)	(0.0097)	(0.0101)	(0.0167)
Mining	-0.0061	-0.0387**	-0.0156	-0.0022	0.0021	-0.0673**
	(0.0118)	(0.0094)	(0.0092)	(0.0096)	(0.0100)	(0.0080)

Note: Standard errors are in parentheses. Double asterisks (^**) indicate significance at 5 percent level. A single asterisk (*) indicates the 10 percent level. Sample size = 21,004. Controls are a time trend; education; experience and its square; four dummies for types of college degrees; five dummies for fields of degree; four dummies for occupation; an SMSA dummy; a south dummy; a race dummy; a marriage dummy; number of children; and interactions of experience with education, a college degree dummy, and a race dummy. Estimates for the selection models use the full sample of 23,927 person-year observations. The probit employment choice equation estimates from the selection models are not reported here.

a Financial services, insurance, and real estate.

View Table

Table 3.

Estimated Correlations of Business Cycle with Real Wages: Degree Interactions

(Dependent variable: Log real wage)

			Fixed effects		Selection-corrected
	OLS estimates		estimates		fixed effects
		URATE *		URATE *		URATE *
	URATE	DEGREE	URATE	DEGREE	URATE	DEGREE
All workers	0.0031	-0.0294**	-0.0062**	0.0013	-0.0057**	0.0012
	(0.0024)	(0.0041)	(0.0017)	(0.0026)	(0.0019)	(0.0013)
Durable manufacturing	0.0073**	-0.0348**	-0.0033	-0.0028	-0.0049*	-0.0059*
	(0.0036)	(0.0048)	(0.0026)	(0.0034)	(0.0030)	(0.0026)
Nondurable manufacturing	0.0082*	-0.0231**	-0.0044	-0.0004	-0.0029	-0.0170**
	(0.0046)	(0.0051)	(0.0034)	(0.0039)	(0.0033)	(0.0030)
Construction	-0.0118**	-0.0468**	-0.0112**	-0.0015	-0.0135**	0.0010
	(0.0049)	(0.0056)	(0.0037)	(0.0046)	(0.0040)	(0.0045)
Transportation and utilities	0.0222**	-0.0393**	0.0060	-0.0057	0.0037	0.0067
	(0.0055)	(0.0053)	(0.0041)	(0.0045)	(0.0043)	(0.0049)
Wholesale trade	-0.0O05	-0.0091	0.0007	0.0104**	-0.0020	0.0046
	(0.0007)	(0.0057)	(0.0052)	(0.0041)	(0.0051)	(0.0038)
Retail trade	-0.0O79 *	-0.0253**	-0.0065 *	0.0055	-0.0057	0.0136**
	(0.0047)	(0.0052)	(0.0035)	(0.0038)	(0.0036)	(0.0041)
F.I.RE.a	-0.0070	-0.0181**	-0.0087	0.0113**	-0.0111 *	-0.0044
	(0.0086)	(0.0057)	(0.0065)	(0.0046)	(0.0061)	(0.0043)
Services	-0.0039	-0.0245**	-0.0214**	0.0079**	-0.0161**	0.0152**
	(0.0049)	(0.0048)	(0.0036)	(0.0031)	(0.0044)	(0.0030)
Government	0.0176**	-0.0300**	-0.0004	0.0027	0.0028	0.0004
	(0.0067)	(0.0051)	(0.0051)	(0.0041)	(0.0049)	(0.0030)
Agriculture	0.0102	-0.0132	0.0117 *	-0.0019	0.0168 *	-0.0130
	(0.0092)	(0.0080)	(0.0071)	(0.0097)	(0.0101)	(0.0167)
Mining	-0.0061	-0.0387**	-0.0156	-0.0022	0.0021	-0.0673**
	(0.0118)	(0.0094)	(0.0092)	(0.0096)	(0.0100)	(0.0080)

Note: Standard errors are in parentheses. Double asterisks (^**) indicate significance at 5 percent level. A single asterisk (*) indicates the 10 percent level. Sample size = 21,004. Controls are a time trend; education; experience and its square; four dummies for types of college degrees; five dummies for fields of degree; four dummies for occupation; an SMSA dummy; a south dummy; a race dummy; a marriage dummy; number of children; and interactions of experience with education, a college degree dummy, and a race dummy. Estimates for the selection models use the full sample of 23,927 person-year observations. The probit employment choice equation estimates from the selection models are not reported here.

a Financial services, insurance, and real estate.

In the first set of columns in Table 3, the estimated OLS coefficient on URATE * DEGREE for all workers is -0.0294 (s.e. 0.0041) indicating that, at the aggregate level, workers with a college degree face relatively more procyclical real wages.²³ A similar pattern holds at the industry level. The coefficients on the interaction terms URATE * DEGREE are significantly negative in a majority of the industries. Thus, the OLS estimates imply that, in a downturn, skilled workers find their wages falling relative to the wages of unskilled workers. Recall that Tables 1 and 2 showed that workers with college degrees are protected from cyclical variation in employment and hours. In conjunction, these results suggest that skilled workers accept steeper wage cuts in a downturn but have more stable employment and hours than unskilled workers. This accords with the predictions of the Hashimoto-Raisian model.

We turn next to the fixed effects estimates in Table 3. The change in the estimated coefficients relative to the OLS estimates is substantial, with some of the coefficients even reversing signs. Controlling for unobserved individual fixed effects changes the URATE coefficient for all workers from 0.0031 (s.e. 0.0024) to -0.0062 (s.e. 0.0017). This indicates that, among workers without a college degree, low-ability workers are more likely to be laid off in a downturn, thereby increasing labor force quality. This induces a positive (countercyclical) bias in the OLS URATE coefficient.

On the other hand, the coefficient on URATE * DEGREE changes from -0.0294 (s.e. 0.0041) to 0.0013 (s.e. 0.0026), indicating that the OLS estimate of the interaction coefficient is procyclically biased.²⁴ The FE estimate implies that the skilled wage falls at about the same percentage rate as the unskilled wage. In other words, after the fixed effects correction, the skilled-unskilled relative wage differential appears to be essentially acyclical.²⁵

At the industry level, the URATE * DEGREE coefficient is insignificant for most industries except wholesale trade, F.I.R.E., and services. In these three industries, the interaction coefficient becomes positive and significant, indicating a countercyclical skill differential (that is, the relative wage of skilled workers rises in a downturn). In most industries, however, the fixed effects estimates show that, after controlling for unobserved time-invariant measures of ability, the relative wages of skilled and unskilled workers remain essentially unchanged over the cycle.

The last set of columns in Table 3 contains selection-corrected fixed effects (SCFE) estimates. The estimated parameter ρ (not reported here) was small and insignificantly different from zero in the aggregate and for all industries. This indicates that the correlation between the transitory components of workers’ wages and their employment probabilities is small, once fixed effects are accounted for. It appears that most of the compositional changes over the cycle can be measured by the combination of observed characteristics of workers and unobserved individual fixed effects. Thus, the selection correction has little impact on the estimates. For all workers, the coefficient on URATE goes from -0.0062 (s.e. 0.0017) in the FE estimates to -0.0057 (s.e. 0.0019) in the SCFE estimates while the URATE * DEGREE coefficient remains insignificantly different from zero. This confirms the result that, at the aggregate level, the relative offer-wage differential between skilled and unskilled workers is essentially acyclical.²⁶

At the industry level, the coefficients for most industries change from the FE estimates. Since the estimated p is insignificant for all industries, this change is attributable to the potential bias in the FE estimates resulting from restricting the fixed effects and the returns to observed characteristics (that is, the coefficient vector β) to be the same across all industries.²⁷ The selection models are estimated separately for each industry, thereby controlling for general as well as industry-specific individual fixed effects. The industry estimates in the selection models also allow the returns to observed worker characteristics to vary across industries.²⁸

In durable and nondurable manufacturing and in mining, the coefficient on URATE * DEGREE turns significantly negative. This means that in a recession degreed workers in those industries face a larger decline in offer wages than nondegreed workers. The difference between the FE and SCFE results may arise because skilled workers who leave manufacturing and mining in a recession have a lower level of industry-specific individual fixed effects than those who remain (in other words, unobserved quality rises). Alternatively, this difference may arise because the returns to observable quality are higher than average in manufacturing and mining, and high-(observed) quality workers move into these industries during a recession. In any case, these three industries provide strong support for the Hashimoto-Raisian hypothesis that skilled workers take larger wage cuts in a downturn than unskilled workers. On the other hand, in retail trade and services, the interaction term is positive and significant. In those industries, workers with a college degree do better, in relative terms, when the aggregate unemployment rate in the economy goes up.

Next, we look at the effect of another human capital variable, TENURE. Table 4 contains OLS and fixed effects estimates of wage equations that include the URATE * TENURE interaction term.²⁹ The coefficient on URATE * TENURE for all workers is close to zero in both the OLS and FE estimates. This seems to indicate that specific human capital levels, as measured by tenure on the current job, have virtually no effect on relative wage variability.

Table 4.

Estimated Correlations of Business Cycle with Real Wages: Tenure Interactions

(Dependent variable: Log real wage)

Notes: Standard errors are in parentheses. Double asterisks (^**) indicate significance at 5 percent level. A single asterisk (^*) indicates the 10 percent level. Sample size = 20.309. Same controls as in Table 3 are used, except that tenure is added as an additional control variable.

^aFinancial services, insurance, and real estate.

Table 4.

Estimated Correlations of Business Cycle with Real Wages: Tenure Interactions

(Dependent variable: Log real wage)

	OLS estimates		Fixed effects
		URATE *		URATE *
	URATE	TENURE	URATE	TENURE
All workers	-0.0036	-0.0001	0.0052**	-0.0004
	(0.0027)	(0.0005)	(0.0021)	(0.0004)
Durable manufacturing	-0.0029	-0.0002	-0.0019	-0.0002
	(0.0041)	(0.0005)	(0.0043)	(0.0005)
Nondurable manufacturing	-0.0030	-0.0003	0.0032	-0.0003
	(0.0051)	(0.0005)	(0.0053)	(0.0006)
Construction	-0.0053	-0.0024**	-0.0042	-0.0025**
	(0.0050)	(0.0006)	(0.0053)	(0.0006)
Transportation and utilities	0.0071	-0.0003	0.0092	-0.0002
	(0.0059)	(0.0005)	(0.0062)	(0.0006)
Wholesale trade	0.0081	-0.0O02	0.0054	0.0003
	(0.0072)	(0.0006)	(0.0076)	(0.0006)
Retail trade	-0.0093**	0.0003	-0.0127**	0.0003
	(0.0050)	(0.0005)	(0.0052)	(0.0006)
F.I.R.E.a	-0.0159*	0.0008	-0.0171**	0.0008
	(0.0082)	(0.0007)	(0.0086)	(0.0007)
Services	-0.0103**	-0.0004	-0.0125**	-0.0005
	(0.0045)	(0.0005)	(0.0047)	(0.0005)
Government	-0.0039	0.0001	0.0011	-0.0002
	(0.0069)	(0.0006)	(0.0073)	(0.0006)
Agriculture	0.0367	-0.0040**	0.0355**	-0.0042**
	(0.0096)	(0.0007)	(0.0101)	(0.0008)
Mining	0.0025	-0.0020**	0.0007	-0.0021**
	(0.0121)	(0.0006)	(0.0127)	(0.0009)

Notes: Standard errors are in parentheses. Double asterisks (^**) indicate significance at 5 percent level. A single asterisk (^*) indicates the 10 percent level. Sample size = 20.309. Same controls as in Table 3 are used, except that tenure is added as an additional control variable.

^aFinancial services, insurance, and real estate.

View Table

Table 4.

Estimated Correlations of Business Cycle with Real Wages: Tenure Interactions

(Dependent variable: Log real wage)

	OLS estimates		Fixed effects
		URATE *		URATE *
	URATE	TENURE	URATE	TENURE
All workers	-0.0036	-0.0001	0.0052**	-0.0004
	(0.0027)	(0.0005)	(0.0021)	(0.0004)
Durable manufacturing	-0.0029	-0.0002	-0.0019	-0.0002
	(0.0041)	(0.0005)	(0.0043)	(0.0005)
Nondurable manufacturing	-0.0030	-0.0003	0.0032	-0.0003
	(0.0051)	(0.0005)	(0.0053)	(0.0006)
Construction	-0.0053	-0.0024**	-0.0042	-0.0025**
	(0.0050)	(0.0006)	(0.0053)	(0.0006)
Transportation and utilities	0.0071	-0.0003	0.0092	-0.0002
	(0.0059)	(0.0005)	(0.0062)	(0.0006)
Wholesale trade	0.0081	-0.0O02	0.0054	0.0003
	(0.0072)	(0.0006)	(0.0076)	(0.0006)
Retail trade	-0.0093**	0.0003	-0.0127**	0.0003
	(0.0050)	(0.0005)	(0.0052)	(0.0006)
F.I.R.E.a	-0.0159*	0.0008	-0.0171**	0.0008
	(0.0082)	(0.0007)	(0.0086)	(0.0007)
Services	-0.0103**	-0.0004	-0.0125**	-0.0005
	(0.0045)	(0.0005)	(0.0047)	(0.0005)
Government	-0.0039	0.0001	0.0011	-0.0002
	(0.0069)	(0.0006)	(0.0073)	(0.0006)
Agriculture	0.0367	-0.0040**	0.0355**	-0.0042**
	(0.0096)	(0.0007)	(0.0101)	(0.0008)
Mining	0.0025	-0.0020**	0.0007	-0.0021**
	(0.0121)	(0.0006)	(0.0127)	(0.0009)

Notes: Standard errors are in parentheses. Double asterisks (^**) indicate significance at 5 percent level. A single asterisk (^*) indicates the 10 percent level. Sample size = 20.309. Same controls as in Table 3 are used, except that tenure is added as an additional control variable.

^aFinancial services, insurance, and real estate.

At the industry level, TENURE has little effect on the cyclicality of wages in a majority of the industries. In construction, agriculture, and mining, workers with longer tenure have marginally more procyclical wages, since the URATE * TENURE coefficients are negative but not very large in magnitude in both the OLS and FE estimates. We conclude from Table 4 that workers with different amounts of tenure face essentially the same degree of cyclical variation in real wages. Interpreting specific capital levels as a measure of skill, this bolsters the conclusion drawn from the previous table that, at the aggregate level, the relative wage differential between skilled workers and unskilled workers is acyclical.

Finally, we examine the effect of total labor market experience on the cyclicality of wages. The first set of columns in Table 5 contains OLS estimates of the wage equation with the URATE * EXPERIENCE interaction term. For all workers, the coefficient on URATE * EXPERIENCE is −0.0010 (s.e. 0.0004), implying a tenth of a percent increase in the cyclical sensitivity of wages for every added year of experience.

Table 5.

Estimated Correlations of Business Cycle with Real Wages: Experience Interactions

(Dependent variable: Log real wage)

Notes: Standard errors are in parentheses. Double asterisks (^**) indicate significance at 5 percent level. A single asterisk (^*) indicates the 10 percent level. Sample size = 21,004. See note for Table 3 for list of controls.

^aFinancial services, insurance, and real estate.

Table 5.

Estimated Correlations of Business Cycle with Real Wages: Experience Interactions

(Dependent variable: Log real wage)

					Selection-corrected
	OLS estimates		Fixed effects estimates		fixed effects
		URATE *		URATE *		URATE *
	URATE/	EXPERIENCE	URATE	EXPERIENCE	URATE	EXPERIENCE
All workers	0.0033	-0.0010**	0.0121**	-0.0026**	0.0122**	-0.0026
	(0.0034)	(0.0004)	(0.0017)	(0.0003)	(0.0018)	(0.0002)
Durable manufacturing	0.0067	-0.0009**	0.0145**	-0.0036**	0.0076*	-0.0017**
	(0.0048)	(0.0004)	(0.0041)	(0.0003)	(0.0046)	(0.0004)
Nondurable manufacturing	0.0206**	-0.0017**	0.0194	-0.0030**	0.0056	-0.0015**
	(0.0058)	(0.0005)	(0.0049)	(0.0004)	(0.0045)	(0.0004)
Construction	-0.0174**	-0.0006	-0.0036	-0.0019**	-0.0006	0.0015**
	(0.0064)	(0.0005)	(0.0055)	(0.0004)	(0.0062)	(0.0006)
Transportation and utilities	0.0233**	-0.0012**	0.0195**	-0.0024**	0.0326**	-0.0039**
	(0.0068)	(0.0005)	(0.0058)	(0.0004)	(0.0061)	(0.0006)
Wholesale trade	0.0067	-0.0009*	0.0193**	-0.0024**	0.0056	0.0010
	(0.0084)	(0.0005)	(0.0066)	(0.0004)	(0.0069)	(0.0008)
Retail trade	-0.0136**	-0.0002	0.0119**	-0.0025**	-0.0268**	-0.0042**
	(0.0060)	(0,0004)	(0.0049)	(0.0004)	(0.0048)	(0.0005)
F.I.R.E.a	-0.0221**	0.0003	0.0064	-0.0021**	0.0183**	-0.0055**
	(0.0097)	(0.0006)	(0.0082)	(0.0005)	(0.0077)	(0.0010)
Services	-0.0044	-0.0015**	0.0060	-0.0032**	0.0187**	-0.0047**
	(0.0051)	(0.0004)	(0.0043)	(0.0003)	(0.0034)	(0.0005)
Government	0.0198**	-0.0016**	0.0159**	-0.0026**	0.0301**	-0.0047**
	(0.0075)	(0.0005)	(0.0064)	(0.0004)	(0.0056)	(0.0007)
Agriculture	0.0251**	-0.0015**	0.0307**	-0.0026**	0.0032	0.0013
	(0.0014)	(0.0006)	(0.0104)	(0.0005)	(0.0156)	(0.0014)
Mining	0.0074	-0.0019**	0.0111	-0.0031**	0.0284**	-0.0037**
	(0.0140)	(0.0007)	(0.0122)	(0.0006)	(0.0137)	(0.0012)

Notes: Standard errors are in parentheses. Double asterisks (^**) indicate significance at 5 percent level. A single asterisk (^*) indicates the 10 percent level. Sample size = 21,004. See note for Table 3 for list of controls.

^aFinancial services, insurance, and real estate.

View Table

Table 5.

Estimated Correlations of Business Cycle with Real Wages: Experience Interactions

(Dependent variable: Log real wage)

					Selection-corrected
	OLS estimates		Fixed effects estimates		fixed effects
		URATE *		URATE *		URATE *
	URATE/	EXPERIENCE	URATE	EXPERIENCE	URATE	EXPERIENCE
All workers	0.0033	-0.0010**	0.0121**	-0.0026**	0.0122**	-0.0026
	(0.0034)	(0.0004)	(0.0017)	(0.0003)	(0.0018)	(0.0002)
Durable manufacturing	0.0067	-0.0009**	0.0145**	-0.0036**	0.0076*	-0.0017**
	(0.0048)	(0.0004)	(0.0041)	(0.0003)	(0.0046)	(0.0004)
Nondurable manufacturing	0.0206**	-0.0017**	0.0194	-0.0030**	0.0056	-0.0015**
	(0.0058)	(0.0005)	(0.0049)	(0.0004)	(0.0045)	(0.0004)
Construction	-0.0174**	-0.0006	-0.0036	-0.0019**	-0.0006	0.0015**
	(0.0064)	(0.0005)	(0.0055)	(0.0004)	(0.0062)	(0.0006)
Transportation and utilities	0.0233**	-0.0012**	0.0195**	-0.0024**	0.0326**	-0.0039**
	(0.0068)	(0.0005)	(0.0058)	(0.0004)	(0.0061)	(0.0006)
Wholesale trade	0.0067	-0.0009*	0.0193**	-0.0024**	0.0056	0.0010
	(0.0084)	(0.0005)	(0.0066)	(0.0004)	(0.0069)	(0.0008)
Retail trade	-0.0136**	-0.0002	0.0119**	-0.0025**	-0.0268**	-0.0042**
	(0.0060)	(0,0004)	(0.0049)	(0.0004)	(0.0048)	(0.0005)
F.I.R.E.a	-0.0221**	0.0003	0.0064	-0.0021**	0.0183**	-0.0055**
	(0.0097)	(0.0006)	(0.0082)	(0.0005)	(0.0077)	(0.0010)
Services	-0.0044	-0.0015**	0.0060	-0.0032**	0.0187**	-0.0047**
	(0.0051)	(0.0004)	(0.0043)	(0.0003)	(0.0034)	(0.0005)
Government	0.0198**	-0.0016**	0.0159**	-0.0026**	0.0301**	-0.0047**
	(0.0075)	(0.0005)	(0.0064)	(0.0004)	(0.0056)	(0.0007)
Agriculture	0.0251**	-0.0015**	0.0307**	-0.0026**	0.0032	0.0013
	(0.0014)	(0.0006)	(0.0104)	(0.0005)	(0.0156)	(0.0014)
Mining	0.0074	-0.0019**	0.0111	-0.0031**	0.0284**	-0.0037**
	(0.0140)	(0.0007)	(0.0122)	(0.0006)	(0.0137)	(0.0012)

Notes: Standard errors are in parentheses. Double asterisks (^**) indicate significance at 5 percent level. A single asterisk (^*) indicates the 10 percent level. Sample size = 21,004. See note for Table 3 for list of controls.

^aFinancial services, insurance, and real estate.

A similar pattern holds at the industry level. The coefficients on the interaction term URATE ^* EXPERIENCE are significantly negative in a majority of the industries. In industries where the average real wage is highly procyclical, experience does not have much effect. This includes construction, retail trade, and F.I.R.E. On the other hand, the coefficient on URATE is significantly positive in nondurable manufacturing, transportation and utilities, government, and agriculture, indicating a countercyclical average real wage for new entrants in those industries. The URATE ^* EXPERIENCE coefficient is negative and significant in each of those four industries, implying that added experience reduces the countercyclicality of a worker’s real wage.³⁰

The fixed effects estimates in the second set of columns reinforce this picture. The coefficient on URATE becomes significantly positive in the aggregate and for most industries. The coefficient on URATE ^* EXPERIENCE for all workers goes from −0.0010 (s.e. 0.0004) in the OLS estimates to −0.0026 (s.e. 0.0003) in the fixed effects estimates. The URATE ^* EXPERIENCE coefficient is negative and significant in all industries. The estimated interaction coefficients at the industry level cluster around −0.0026. That is, the procyclicality of wages increases by about a quarter of a percent for every added year of labor market experience. Thus, both the OLS and FE estimates show that in a downturn workers with more labor market experience find their wages falling relative to the wages of workers with little or no experience.

The selection-corrected fixed effects estimates in the third set of columns show that the selection correction has virtually no effect on the results for all workers. The coefficients on URATE (0.0122, s.e. 0.0018) and URATE ^* EXPERIENCE (−0.0026, s.e. 0.0002) are almost identical to the FE estimates.³¹ This is as expected since the parameter ρ was estimated to be close to zero in the aggregate and for all industries. At the industry level, many of the point estimates change from the FE estimates, again indicating the bias in the industry-level FE estimates that results from restricting individual fixed effects and returns to observed ability to be the same across all industries.

Despite the change in the point estimates, most of the interaction coefficients at the industry level remain significantly negative. Thus, workers with higher levels of labor market experience face relatively greater procyclical variation in wages than recent entrants into the labor force. Since the EXPERIENCE variable is defined as current age minus age at entry into the labor force, it is possible that the above results are dominated by age effects rather than by any aspect of human capital.

An important point to note about the aggregate and industry-level estimates in Table 5 is that they clearly show that older workers (those with more labor market experience) have more procyclical wages. Since our sample contains only young men, this implies that our estimates of the procyclicality of wages may significantly understate the true degree of procyclical variation in average wages in the U.S. economy.³²

V. Summary and Conclusion

The results of this paper indicate that, at the aggregate level, offer-wage differentials between skilled and unskilled workers are essentially acyclical. Using education and tenure as proxies for skill levels, we find no evidence to support the view that relative wage differentials between skilled and unskilled workers are countercyclical, a view that has long been accepted as a stylized fact. Our OLS estimates suggest, in fact, that relative wage differentials may actually be procyclical: the difference between skilled and unskilled workers’ wages widens in booms and narrows in recessions.

However, the OLS estimates are subject to bias resulting from systematic compositional changes that are not captured by observed worker characteristics. To provide correct measures of the cyclically of wage differentials, we estimate a series of models that control for unobserved individual fixed effects as well as selectivity bias. The magnitude of selectivity bias turns out to be not very significant in our sample, but unobserved fixed effects are found to bias the OLS coefficients severely. Correcting for observed and unobserved worker heterogeneity, we conclude that, at the aggregate level, correctly measured relative offer-wage differentials do not have any consistent procyclical or countercyclical tendencies. However, after controlling for other characteristics, older workers do appear to have more procyclical wages than younger workers.

Variation in employment appears to be more important quantitatively than variation in weekly hours worked in explaining cyclical fluctuations in total labor input. On average, weekly hours worked are procyclical, with an average decline of about 0.4 percent for every percentage point increase in the aggregate unemployment rate. Variation in weekly hours per worker accounts for about 30 percent of the variation in total hours.

We find that workers with a college degree have acyclical employment probabilities and weekly hours, while uneducated workers have highly procyclical variation in employment and hours. These results imply that the average quality (in terms of skill level) of labor input per manhour rises in recessions and falls in booms. It follows that measures of the real wage that simply divide the total wage bill by total manhours are likely to be substantially countercyclically biased.

Our industry-level estimates reveal striking differences across industries in the cyclical variation of employment, hours, and wages. For example, in our sample, durable manufacturing displays the strongest procyclical employment variation, although workers with a college degree face much less relative variation in employment in that sector. In both durable and nondurable manufacturing, workers with a degree have weakly countercyclical hours, but face much more procyclical variation in wages than other workers. Thus, the results for manufacturing support the Hashimoto-Raisian hypothesis that skilled workers accept more wage variation in return for employment stability, thereby also implying pro-cyclical wage differentials between skilled and unskilled workers.

On the other hand, retail trade and services reveal a different pattern. In retail trade, college-educated workers have countercyclical wages while nondegreed workers have procyclical wages. In services, uneducated workers have strongly procyclical wages while workers with a degree have acyclical wages. Thus, for this measure of skill level, the relative skilled-unskilled wage differential is estimated to be countercyclical in these two industries.

In some industries such as nondurable manufacturing and retail trade, a large fraction of the variation in total hours worked appears to be accounted for by variation in average weekly hours rather than in the number of persons employed. In most other industries, as in the aggregate, employment variation seems more important than hours variation in accounting for cyclical fluctuations in total hours.

Our disaggregated results indicate that worker heterogeneity and sectoral differences interact in crucial ways over the business cycle. Analyzing this issue is beyond the scope of this paper, but our results suggest that both kinds of heterogeneity and the interactions between them are likely to play an important role in modeling and understanding business cycle fluctuations.