Differences in state unemployment rates: the role of labor and product market structural shifts.

I. Introduction

State unemployment rates diverged dramatically in the 1980s. Differing unemployment rates across states are an important public policy concern because of equity concerns and the pure human consequences of higher unemployment. Moreover, at the national level, a greater dispersion in state and regional unemployment rates can increase the natural rate of unemployment and shift the Phillips curve out because of an inefficient allocation of the labor force [3; 38].

One possible cause for the divergence in unemployment rates is national industrial restructuring. U.S. industrial restructuring was suggested by Lilien [26] as a cause for the rise in the national unemployment rate. Moreover, national restructuring can have differential impacts across states because states can have dramatically different mixes of industries from each other. One of the difficulties of national restructuring from a regional perspective is that it is difficult to avoid. On the other hand, if regional differences in unemployment are due to state idiosyncratic causes, state and local public policy can potentially play a more active role. For example, the economic problems that plagued the Northeast and California during the early 1990s have been typically attributed to national downsizing in the defense industries. Conversely, economic problems in these states may have been driven by factors that are located within the state, such as the collapse of their real estate markets.

Even beyond the issue of national restructuring, the 1980s represented a structural break in relative regional economic performances. Though Neumann and Topel [35] show that state unemployment rates can be very persistent, even over the course of decades, state unemployment rates for the mid-1980s are not strongly correlated with their respective unemployment rates from the 1970s or the 1990s. For example, based on data introduced later in the paper, the correlation of the 1977 and 1987 state unemployment rates is -0.055, while the correlation between 1977 and 1992 state unemployment rates is 0.74.(1) Thus, the high correlation of state unemployment rates between the 1970s and 1990s seems to represent a return to historical patterns observed by Neumann and Topel. Moreover, annual measures of the coefficient of variation of the 48 contiguous state unemployment rates rise sharply beginning in the early 1980s, peak in the late 1980s, and then fall back to the levels experienced during the 1970s. This also indicates that there was a structural break in the 1980s and that long-run patterns reemerged in the early 1990s.

This study advances previous state unemployment research by unifying several possible causes of unemployment into one empirical framework. First, by using shift-share analysis, we separate a state's employment growth rate into the portion due to national factors, such as industrial restructuring, and a separate portion due to state-specific idiosyncratic factors. This allows us to assess the relative roles of national restructuring and state-specific factors as causes for unemployment rates to vary cross-sectionally across states within a given time period and over time within a given state. Another closely associated issue examined is whether workers who lose their jobs to national restructuring have less options for finding work elsewhere than if they lose their jobs to state-specific causes. With the exceptions of Gallaway [12] and Treyz et al. [40], there has been little examination of both interindustry labor mobility and interregional mobility together, and none related to unemployment.

Second, we also separate earnings into a wage-mix component that arises from the industrial composition of the state and a wage-competitiveness component. This allows us to better separate wait unemployment effects from cost competitiveness effects. Third, while carefully controlling for employment growth and wage structure, we also utilize the within state variance of sectoral employment growth rates to account for matching difficulties for workers and firms that can occur when a large number of workers are forced to change sectors. This is closely associated with the analysis of Neumann and Topel [35] at the state level and Lilien [26] at the national level. Fourth, we consider data from 1972 to the early 1990s to fully evaluate the industrial restructuring of the 1980s.

Finally, to avoid spurious relationships between the above variables and state unemployment rates, and to identify other state-specific factors behind unemployment, we appraise a multitude of other possible causes for unemployment identified in previous studies including unionization, unemployment insurance, and demographic differences across states that can influence a given state's natural rate of unemployment [6; 9; 17; 21; 27; 28; 30; 32; 35; 37]. Thus, by simultaneously considering more factors than has typically been the case, the results are less subject to an omitted variable bias.(2)

In what follows, section II presents the underlying theoretic model and empirical methodology. Section III contains the empirical results, while specification issues are addressed in section IV. The final section presents some concluding thoughts.

II. Theoretical Considerations

The unemployment rate is a reduced form function of the factors that affect labor demand and labor supply. These factors can broadly be categorized as industry or product market variables (IND), non-demographic labor market variables (LABOR), demographic variables (DEMOG), and regional characteristics (REGION). Thus, unemployment for state i in time period t may be written as

[U.sub.it] = [Alpha] + [Beta]IND + [Gamma]LABOR + [Omega]DEMOG + [Psi]REGION + [[Sigma].sub.i] + [[Sigma].sub.t] + [e.sub.it], (1)

where [Alpha] is the constant term, [[Sigma].sub.i] and [[Sigma].sub.t] are state and time fixed effects and [e.sub.it] is an error term. [Beta], [Gamma], [Omega], and [Psi] represent coefficient vectors. However, the error term [e.sub.it] could exhibit first-order autocorrelation. This can be represented as

[e.sub.it] = [Rho][e.sub.it-1] + [v.sub.it], (2)

where [v.sub.it] is i.i.d. and [Rho] is the degree of first-order autocorrelation that is assumed constant across all states.

Variable definitions and sources are shown in Table I. Unemployment rates and the independent variables are measured for the 48 contiguous states from 1972-91. The independent variables and their hypothesized influence on unemployment are described below.

A primary factor in determining unemployment differences is employment growth. Indeed, Abraham [1] and Medoff [31] argue that increased regional dispersion in employment demand is one reason that the national natural rate of unemployment drifted higher in the 1970s. For instance, if a state has a mix of industries that are faring relatively better at the national level, then state employment should increase relatively. However, employment growth at the regional level may not reduce the unemployment rate. This can occur because in-migrants may absorb all the new jobs [7], leaving unemployment unaffected.(3)

Moreover, different sources of employment growth or decline may have different effects on unemployment. To investigate this, employment growth is broken into two parts using the shift-share method. Stevens and Moore [36] provide a thorough description of the shift-share method. First, state employment growth that would occur if its industries grew at their respective national rates is calculated (INDMIX).(4) Second, the remaining employment growth, or state idiosyncratic change (COMPETE) is calculated.(5) This component may reflect employment growth due to state competitiveness or state-specific restructuring. To allow for more complex dynamics, the lags of INDMIX and COMPETE are included (INDMIX_LAG, COMPETE_LAG).

INDMIX and COMPETE may have differing impacts on unemployment to the extent that worker mobility across industries differs from their mobility across regions. If workers have limited mobility across industries [26], the impact on unemployment will be greater if the loss of employment occurs because of national restructuring (i.e., the magnitude of the INDMIX coefficient will be larger). This follows because there is less incentive for workers to migrate out of the region if their industry is in national decline and this causes unemployment to rise. On the other hand, workers displaced by state-specific employment changes have a better chance of obtaining employment in their industry outside the region. Thus, there is greater incentive to migrate, leaving unemployment relatively unchanged.

Table I. Variable Definitions and Sources

VARIABLE          DEFINITION AND SOURCE

UNEMP. RATE       The civilian unemployment rate for each state.
                  Source: [46].

INDMIX            The state's employment growth rate if employment
                  in the state's two-digit industries were growing
                  at their respective average national industry
                  employment growth rates for the non-farm private
                  sector. Source: Bureau of Economic Analysis (BEA).

COMPETE           The difference between the state's actual non-farm
                  private sector employment growth rate and the
                  industrial mix employment growth rate (INDMIX).
                  (COMPETE = ACTUAL STATE GROWTH RATE - INDMIX)
                  Source: BEA.

INDVAR            The employment weighted variance of two-digit
                  non-farm private sector employment growth rates.
                  To avoid possible wild fluctuations due to very
                  small industries in small states, industries with
                  employment less than 500 were excluded from the
                  calculation. Source: BEA.

WAGEMIX           The real average annual earnings per worker
                  (1000s) in the state if each of the non-farm
                  private sector industries in the state paid their
                  respective national average industry earnings per
                  worker. The earnings were deflated by the U.S.
                  Consumer Price Index (1982-84 = 100). Source: BEA
                  and [45].

WAGE_COMP        The difference between real actual non-farm
                  private sector average annual earnings per worker
                  (1000s) in the state and the real-wage mix wage
                  (WAGEMIX).

HERFIN            The two-digit industry employment Herfindahl index
                  which is the sum over all non-farm private sector
                  industries of the square of employment share in
                  percent accounted for by that industry. Source:
                  BEA.

INDUSTRY SHARES   The shares of total civilian employment in the
                  state that are accounted for by agriculture
                  (FARM); manufacturing (MANU); mining (MINING);
                  construction (CONST); transportation and public
                  utilities (TRANS); finance, insurance, and real
                  estate (FIRE); and all government (GOVT). Source:
                  BEA.

MALE_LF%          The percent of total civilian employment in the
                  state that is accounted for by males. For some of
                  the smaller states, male employment was
                  unavailable for 1972-75. In these cases, the
                  earliest year where exact data was available was
                  interpolated with data from the 1970 census.
                  Source: BEA and [42; 46].

BLACK_POP%        The percent of the population in the state that is
                  black. Data was interpolated using 1970, 1980, and
                  1990 as endpoints. Source: [45].

MARRIED%          The percent of women over the age of 16 that are
                  married. Data was interpolated using 1970, 1980,
                  and 1990 as endpoints. Source: [42].

CHILD6_MAR        The percent of women over the age of 16 that are
                  married with children under the age of 7. Data was
                  interpolated using 1970, 1980, and 1990 as end
                  points. Source: [42].

CHILD6_NONMAR     The percent of women over the age of 16 that are
                  not married with children under the age of 7. Data
                  was interpolated using 1970, 1980, and 1990 as end
                  points. Source: [42].

AGE65%            The percent of the population that is 65 years old
                  or older. Source: [43].

AGE14%            The percent of the population that is 14 years old
                  or younger. Source: [43].

AGE15_19%         The percent of the population that is between the
                  ages of 15 and 19. Source: [43].

HS_GRAD           The percent of the population over the age of 24
                  that are high school graduates. Data was
                  interpolated using 1970, 1980, and 1990 as end
                  points. Source: [45].

COLL_GRAD         The percent of the population over the age of 24
                  that are college graduates. Data was interpolated
                  using 1970, 1980, and 1990 as end points. Source:
                  [45].

BLUE_COLL         The percent of civilian employment that are blue
                  collar employees. For 1972-75, because exact
                  figures were unavailable, the data was derived by
                  interpolating the 1976 figure with the 1970 Census
                  figure. Source: [42; 46].

UNION%            The percent of the civilian labor force that are
                  union members. Source: [16; 45].

UI_BEN            The real average weekly unemployment insurance
                  benefit for each state using the CPI as the
                  deflator. Source: [45; 47].

UI_COV            The percent of civilian employment that is covered
                  by unemployment insurance in each state. Source:
                  [45; 47].

POPDEN            Annual population density of the state. Source:
                  [45].

METRO%            Annual percent of the state's population that
                  resides in a metropolitan area. Source: [45].

The wage rate in a region may also affect its unemployment rate. For example, there may be a hedonic wage-unemployment tradeoff [14; 39] or there may be wait unemployment where workers queue for high-wage jobs [15; 37]. However, Blanchard and Katz [7] point out that increases in labor demand can simultaneously increase wage and reduce unemployment. Thus, the association between wages and the unemployment rate is ambiguous. It should be noted though; to the extent that employment growth can be identified as increased labor demand, inclusion of employment growth in the equation controls for labor demand's effect on wages and the unemployment rate.

Since the wage rate for a region is influenced by its mix of industries, the wage rate is separated into two parts. First, the wage rate that would result if each of the state's industries paid their respective national average wage rate is calculated (WAGEMIX).(6) WAGEMIX, for the most part, should capture wait unemployment influences. Next, the state competitiveness wage effect (WAGE_ COMP) is calculated as the difference between the state's actual earnings and WAGEMIX. WAGE_COMP, for the most part, reflects cost effects on labor demand. Finally, the lags of WAGEMIX and WAGE_ COMP are included to capture more complex dynamics (WAGEMIX_LAG, WAGE_COMP_LAG).

Not only does national and state-specific restructuring affect relative state employment growth rates, it influences the dispersion of industry growth rates within each state. Even when employment is not declining, matching difficulties for employers and employees can arise when there is high dispersion in industry growth rates. In turn, the large number of employees that must change sectors increases frictional unemployment [13; 17; 35].(7) In contrast to previous studies, which based their dispersion measures on one-digit categories, a two-digit employment weighted variance of employment growth across the state's industries (INDVAR) is calculated. Two-digit categories are used because one-digit categories can mask significant restructuring within each one-digit category.(8)

Similarly, the diversity of employment opportunities in a state may affect the unemployment rate [27; 35]. The more diverse an economy, the more readily employment reductions in any given sector can be absorbed into other sectors. Diversity of employment is measured by a two-digit industry employment Herfindahl index (HERFIN).

Employment concentrations in particular sectors may have an additional influence on the unemployment rate. Cyclically sensitive industries such as manufacturing can have greater variations in employment [29] or employment mobility may be limited between certain sectors. Consistent with previous studies [6; 27; 32; 37], this is controlled for by including employment shares by sector: agriculture (FARM); manufacturing (MANU); mining (MINING); construction (CONST); transportation and public utilities (TRANS); finance, insurance, and real estate (FIRE); and government (GOVT); where trade and services are omitted to avoid perfect collinearity.

The percent of the labor force that is unionized (UNION) may also affect the unemployment rate by increasing wages and wait unemployment [6; 37]. Offsetting these impacts is the possibility that union workers are more productive [11] making the net effect ambiguous [13]. Similarly, the availability and generosity of unemployment insurance may be positively related to the unemployment rate because it reduces the worker's cost of unemployment or increases their reservation wage.(9) The availability and generosity of state unemployment insurance programs are measured by the real average weekly unemployment benefit (UI_BEN) and the percent of the labor force covered by unemployment insurance (UI_COV).

There are several demographic variables included that influence labor demand and labor supply. First, we include the percent of the labor-force that is male (MALE_LF%), which is expected to be negatively related to the unemployment rate because males tend to have a greater attachment to the labor-force.(10) The percent of the population that is black (BLACK_POP%) is included to capture potential effects caused by discrimination or by differences in tastes and preferences for market work. The percent of females above the age of 16 that are married (MARRIED%) is included to capture the possibility that married women withdraw from the labor-force. Similarly, women with young children may be more likely to withdraw from the labor-force. This is captured by the percent of women over 16 years old that are married and have children under 7 years old or who are not married and have children under the age of 7 (CHILD6_MAR, CHILD6_NONMAR). Percent of the population that is over 64 years old, under 15 years old, and between 15 and 19 years old are also included in the specification (AGE65%, AGE14%, AGE15_19%) to control for the age structure and quality of the labor-force.

Three variables are included to control for the human capital of the labor-force. The percent of the population above the age of 24 who are high school graduates (HS_GRAD) and who are college graduates (COLL_GRAD) are included. In particular, the share of the labor-force that are college graduates should be inversely related to the unemployment rate through its positive influence on labor demand. Also, college graduates are geographically more mobile. The percent of the labor-force that are blue collar workers (BLUE_COLL) is also included as a measure of human capital.

To control for urbanization, population density (POPDEN) and the percent of the state's population that lives in a metropolitan area (METRO%) are included. For instance, amenities can differ between urban and rural areas where local amenities and the unemployment rate are positively related [28]. Alternatively, employer and employee matching problems may be mitigated in urban areas because of their greater diversity of employment opportunities.

Dummies for the Northeast, Midwest, and West (the South is the omitted category) are also included to account for differences in regional labor markets or in total regional amenities which influence migration patterns (i.e., people tradeoff amenities with a greater likelihood of employment). In fact, industry mix and labor markets can differ so much across states that Murphy and Hofler [34] suggest that it is improper to pool all of the states into one regression specification. Fortunately, our specification controls for these and a multitude of other factors making it unlikely that unaccounted for differences across states will bias our results. Annual fixed effects are included to measure national cyclical effects on state unemployment rates. Thus, the resulting coefficients measure the impact on state unemployment rates net of aggregate cyclical conditions.

III. Empirical Results

Column (1) of Table II presents the means and standard deviations for the variables. Column (2) presents the ordinary least square (OLS) estimates for state unemployment rates. The OLS results assume that there are no state fixed effects and no autocorrelation ([[Sigma].sub.i] = 0, [Rho] = 0). The OLS results capture both cross-sectional and time-series differences in the independent variables. However, the first-order autocorrelation of the OLS residuals was 0.67 suggesting the possibility of unmeasured state fixed effects. Column (3) adds state dummies to the specification ([[Sigma].sub.i] [not equal to] 0, [Rho] = 0).(11) Nonetheless, the first-order autocorrelation of residuals in this specification [TABULAR DATA FOR TABLE II OMITTED] still equalled 0.52 suggesting that even after including state dummies, significant autocorrelation persists. Thus, column (4) allows for state-specific fixed effects and corrects for first-order autocorrelation by dropping the first observation for each state during generalized differencing ([[Sigma].sub.i] [not equal to] 0, [Rho] [not equal to] 0). The panel results in columns (3) and (4), however, only account for within state variations in the independent variables. Yet, for some independent variables, almost all of the variation of the variables are cross-sectional suggesting that only accounting for within state variation may be overly restrictive.

The national restructuring and state idiosyncratic employment coefficients (INDMIX, COMPETE) are robust across all three specifications.(12) Using column (4), a one-year 1% (0.01) increase in a state's industrial mix employment growth rate implies an approximately 0.44 lower unemployment rate that year followed by an unemployment rate that remains 0.08% lower the next year. A one-year 1% increase in idiosyncratic state-specific employment results in an approximately 0.15% lower unemployment rate that year and about a 0.10% lower unemployment rate the following year. Both results are consistent with Murphy's [33] finding that product market variations between states are a significant determinant of differences in state unemployment rates.

These results support our hypothesis that if a worker's industry declines in one state, and this decline is national, there is less incentive for the worker to migrate.(13) However, if the decline in the worker's industry is specific to that state, then the worker is more likely to migrate to states where the industry is faring well, which mitigates the increase in the original state's unemployment rate. Together, these points imply that limited interindustry mobility of the labor-force plays a role in affecting the unemployment rate. In addition, these results show why specific localities hit by layoffs induced by national restructuring seem to undergo such distress (e.g., defense industry communities of the 1990s).

The coefficient magnitudes indicate that a 1% current period change in employment due to industrial mix has a larger influence than a 1% change in employment due to competitiveness. Nonetheless, it would be incorrect to assume that national restructuring employment changes were more important, in general, than idiosyncratic state employment changes. In particular, variations across states or movements over time in industrial mix employment are relatively small. For example, the average annual 1983-91 cross-sectional standard deviation in the industry mix employment growth rate was only 0.004 (0.4%) compared to 0.021 (2.1%) for competitiveness. Thus, the difference in unemployment rates between two states, one with an industrial mix employment growth rate that has been two standard deviations above the mean for two years and another state with an industrial mix employment growth rate that is two standard deviations below the [TABULAR DATA FOR TABLE III OMITTED] mean is only 0.84%. The analogous difference for competitiveness is 2.01%. Consequently, state idiosyncratic differences in employment growth were more than twice as important in explaining annual cross-sectional differences in state unemployment rates.

A similar analysis can be conducted for time-series within state unemployment rate variations. The average 1983-91 annual time-series within state standard deviation in industry mix employment growth rate was only 0.003 (0.3%) compared to 0.018 (1.8%) for competitiveness. As before, two different economic scenarios will be used to illustrate the relative differences in unemployment rate behavior. Scenario one is where the state has an industrial mix employment growth rate that has been two standard deviations above the mean for two years and scenario two is where the state has an industrial mix employment growth rate that is two standard deviations below the mean. The difference in unemployment rates between these two scenarios is only 0.66%. The difference for analogous scenarios for competitiveness employment growth rate is 1.75%. In this case, state idiosyncratic differences in employment growth are almost three times more important in explaining time-series variations of a given state's unemployment rate.

Overall, variations in national industrial restructuring are either too small or move too slowly to significantly explain why unemployment rates vary in a fixed time period across all the states or why a given state's unemployment rate varies relative to the national average over time. In addition, the simple correlation between COMPETE and INDMIX is 0.0, which supports the notion of an independent regional component. One implication of these results is that to understand the performance of the U.S. economy, regional dimensions should not be ignored.

For added perspective, Table III decomposes average regional time-series industrial mix and competitiveness variations in the unemployment rate for nine regions relative to the U.S. average for the 1983 to 1986 and the 1988 to 1991 time periods. These two periods have the advantage of considering unemployment over two different phases of the business cycle. The U.S. unemployment rate fell from 9.6% to 7.0% between 1983 and 1986 for a total decrease of 2.6%. Similarly, it rose from 5.5% in 1988 to 6.7% in 1991 for a total increase of 1.2%. For example, column (3) of Table III shows that New England's unemployment rate fell 0.42% more than the U.S. average between 1983 and 1986 (i.e., it fell by 3.02%) while column (6) shows that it increased 3.28% more than the U.S. unemployment rate between 1988 and 1991 (i.e., it rose by 4.48%).

One finding that stands out in Table III is that changes in national industrial structure played very little role in explaining why regional unemployment rates changed relative to the national average. In fact, the West North Central region during 1988-91 represented the only case where national restructuring influenced regional unemployment rates more than 0.2% (i.e., -0.22%).

Competitiveness employment effects played a larger role during the downturn in 1988-91 than during the expansion of 1983-86. Also, regions that had a positive or negative competitiveness component during 1983-86 had the opposite effect during 1988-91 in five cases. In the South Atlantic region, the contribution of the competitiveness term to the unemployment rate increased by over 0.3%. Thus, taken together, this suggests a regional cyclical component. Moreover, for individual states, the swing in competitiveness components can be rather large. For instance, in the case of the energy dependent states Louisiana, Colorado, and Texas (not shown), the relative swings in their competitiveness effects between the two periods worked to reduce their relative unemployment rates 1.93%, 2.18%, and 1.83%, respectively (after accounting for national industry mix effects such as oil production). The overall cycle appears to be that a state with a greater (less) than average competitiveness employment growth rate undergoes a pattern of rising (falling) input prices - e.g., wages, real estate - which plants the seeds for the subsequent downturn (upturn) in the next period.

To further investigate the cyclical nature of competitiveness, we regressed the current period COMPETE on nine of its lag terms and on a set of state dummies for the 1971-91 period. We then dropped the longest lag term if we found that it had a statistically insignificant coefficient. The fourth lag was the first case where the longest lag turned out to be significant:

COMPETE = 0.74 COMPETE L1 - 0.08 COMPETE L2

(22.63) (1.71)

-0.03 COMPETE L3 - 0.11COMPETEL4 + STATE, (0.57) (2.95)

[R.sup.2] = 0.55, F = 1.02,

where in parentheses are the t-statistics, STATE is a vector of state dummies, and the F-statistic is for the joint null hypothesis that the state dummies are all equal to zero. The results show a clear cyclical pattern where the effect of an exogenous, one-time, one-percent increase in current period competitiveness employment is almost completely eliminated by the fourth year and turns negative by the fifth year. This is consistent with our cyclical pattern shown in Table III. The F-statistic indicates that the null hypothesis for the state dummies cannot be rejected at any conventional level of significance, which suggests that after controlling for its lags, competitiveness employment changes do not vary across the states. Thus, it may be difficult for state and local policies to indirectly affect the unemployment rate by influencing state-idiosyncratic employment growth.

Table III also presents interesting evidence regarding the defense build-down in New England from 1988-91 (where the same pattern follows for California). An unfavorable industry mix raised New England's unemployment rate by only 0.03% more than the U.S. average, but competitiveness or state-specific restructuring raised the average unemployment rate by 1.02%. Despite the concentration of defense industries in these areas, overall national restructuring was not particularly severe in New England. Although changes in defense spending may have signaled the unsustainable nature of their economic boom, other state-specific factors played a much larger role during the actual downturn (at least initially). These state-specific factors can include changes in business expectations, differences in credit conditions, real estate markets, and different economic conditions in the state's prevalent foreign markets.

The relative wage structure of a state's industry mix (WAGEMIX) is negatively related to the unemployment rate in all three specifications, but it is insignificant at the 5% level. The lagged wage structure (WAGEMIX_LAG) is also insignificant when state fixed effects are controlled for. Thus, either the prevalence of high-wage jobs in a state is not important or within state variation of this variable is not great enough for the panel estimates to adequately capture the effect of WAGEMIX with any degree of precision. Nonetheless, Blanchard and Katz [7] also found that wage levels only had a small influence on employment.

The deviation of a state's wages after controlling for its mix of industries (WAGE_COMP) is insignificant in column (4), but its lag (WAGE_COMP_LAG) is significant in all three specifications. Furthermore, joint F-tests in all three specifications reject the null hypothesis that both coefficients are equal to zero at the 0.001% level of significance (F_WAGE_COMP at the bottom of Table II). The positive response of the unemployment rate to the lag of the state-specific wage effect suggests that higher than average earnings begin to depress employment in the second year by making businesses in the state uncompetitive nationally, which is consistent with the regional cyclical pattern found above.

The sectoral employment shares suggest that states with high shares of employment in farming and construction have lower unemployment rates. States with higher shares of employment in government have higher unemployment rates. Nevertheless, there were no statistically significant differences in the effects of relatively high-wage manufacturing and relatively low-wage service producing sectors, i.e., transportation, FIRE, trade, and services.

The variance of within state employment growth rates (INDVAR) is insignificant in all three specifications. This suggests that, whether caused by national or state-specific restructuring, within state employment mismatch difficulties induced by industries growing at differential rates are not a significant cause of greater unemployment rates. This result is consistent with Holzer's [17] results and Abraham's [1] results for smaller metropolitan area labor markets.(14) Similarly, the coefficient for the concentration of employment in particular sectors (HERFIN) varies widely between columns (2)-(4), which suggests that diversity of employment opportunity plays an ambiguous role.

The percent of the labor-force that is male (MALE_LF%) is negatively related to the unemployment rate while there is weak evidence that the share of women with young children (CHILD6_MAR, CHILD6_NONMAR) is also negatively related to unemployment. The percent of the population that is black and married (BLACK_POP%, MARRIED%) appears to have no strong association with state unemployment rates. With the exception of the share of senior citizens in the population, there appears to be no strong relationship between the age structure of the population and the state's unemployment rate (AGE65%, AGE14%, AGE15_19%).

There is strong evidence that a greater share of the population that has a college degree (COLL_GRAD%) is negatively related to the unemployment rate. Using column (4)'s coefficient, a one standard deviation increase in the percent of the population that has a college degree reduces the unemployment rate by about 2.0%. This relatively large effect is likely caused by greater education or human capital increasing both labor demand and the worker's attachment to the labor market. In addition, college graduates are also more likely to migrate in response to changing unemployment rates and labor market conditions. However, the results for the share of the population with a high school degree (HS_GRAD) and for the share of the labor-force who are blue collar workers (BLUE_COLL) were somewhat mixed.

The unionization coefficient (UNION%) is negative but insignificant in the panel specifications. Thus, unionization, per se, may have very little impact on unemployment rates, which is consistent with the findings of cross-sectional studies of industrial nation unemployment rates [5]. The results weakly suggest that unemployment insurance generosity (UI_BEN) and the share of employment covered by unemployment insurance (UI_COV) are positively associated with unemployment rates, but the coefficients are typically insignificant.(15) However, there may not be enough within state variation in our measures to determine their precise effects.

The relationship between population density and the unemployment rate is erratic across specifications. However, in all three cases, the percent of the population that lives in a metropolitan area is negatively related to the unemployment rate.

IV. Other Specification Issues

One empirical concern is the possibility that the error term [v.sub.it] from equation (2) is spatially correlated (e.g., a shock in Indiana can spillover to Illinois). Following Kelejian and Robinson [22; 23; 24], we assume the spatial correlation is of the error component specification:

v = M[Theta] + [Epsilon], (3)

where M is a 960 x 960 matrix of weights that correspond to the neighboring states' average value(16) [Theta] is a 960 x 1 vector, and [Theta] and [Epsilon] are i.i.d with mean and variance-covariance matrix of [Mathematical Expression Omitted] and [Mathematical Expression Omitted]. The model shown in (3) assumes that two different disturbances are formed in each region during period t, where [Theta] does spillover to neighboring states while [Epsilon] is state-specific and does not spillover to neighboring states.

Kelejian and Robinson [23; 24] suggest the following method to test for the possibility of spatial correlation. Run an OLS regression of the square of the estimated value of [v.sub.it] on a constant term and [d.sub.it], where [d.sub.it] is the diagonal element of M x M[prime] corresponding to year t for state i. The null hypothesis is that there is no spatial correlation. Kelejian and Robinson suggest that the null hypothesis should be accepted if the t-statistic on [d.sub.it] is smaller than 1.645. For the autocorrelation corrected specification in column (4), the t-statistic on [d.sub.it] was -0.22 suggesting little spatial correlation of the error terms. The results for the specifications in columns (2) and (3) were also very insignificant.

Another empirical concern for all three specifications is potential endogeneity of some of the variables. Specifically, greater unemployment rates could depress current employment growth and potentially reduce wages. This should not be a problem for the lagged employment terms (INDMIX_LAG, COMPETE_LAG) and lagged wage terms (WAGEMIX_LAG, WAGE_COMP_LAG). However, endogeneity may arise for the current period wage and employment variables. By construction, the industrial mix variable is determined by national forces, not by state unemployment rates. In fact, Bartik [4] and Blanchard and Katz [7] argued that the industrial mix variable is exogenous to local labor market conditions and used the industrial mix variable as an instrument for total employment growth. Similarly, WAGEMIX should be influenced by average earnings determined in the national labor market, not by an individual state's unemployment rate. However, the state-specific competitiveness employment effect (COMPETE) and the state-specific wage effect (WAGE_COMP) could be influenced by the state unemployment rate.

To test for the possibility of endogeneity of the COMPETE and WAGE_COMP coefficients, a Hausman test was conducted.(17) The Hausman test statistic for these two variables is shown at the bottom of Table II (HAUS.-COMPETE, HAUS.-WAGE COMP). The null hypothesis is that potential endogeneity of the variable is not biasing the coefficients. For COMPETE, the Hausman test was clearly insignificant in all three specifications suggesting that potential endogeneity of COMPETE is not biasing the coefficients. For WAGE_COMP, the Hausman test was only significant in column (3)'s results suggesting that potential endogeneity of WAGE_COMP was not too serious.(18)

A final empirical concern is that some of the variables are highly correlated (e.g., industry mix and its lag). One implication is that the coefficients could be imprecisely measured and sensitive to the choice of variables. To examine this issue, the autocorrelation-corrected specification was reestimated after omitting variables that were not statistically significant at the 10% level in a two-tail test (with the exception of the lag of industry mix). These results are shown in column (5) of Table III. The coefficients in column (5) are essentially the same as those in column (4) and the statistical significance of the individual coefficients improved in every case. This suggests that our findings are robust to possible multicollinearity concerns and insensitive to variable selection.

V. Discussion and Concluding Thoughts

Employment shifts in the 1980s were significant for explaining state unemployment rate differences. A decomposition of state employment growth, however, revealed that most of the changes in employment were state-idiosyncratic, not the result of national restructuring. Moreover, the state-idiosyncratic employment growth was primarily cyclical, not persistent in the long run. Thus, in spite of the limited industry mobility implied by the large effects on unemployment of a unit change in industry mix, national restructuring played little role in unemployment differences in the 1980s. This result is consistent with the conclusion by Abraham and Katz [2] that national restructuring is not markedly influencing the natural rate of unemployment. On the other hand, since a state's industry mix and state-idiosyncratic growth components were not correlated, our results do not support national aggregate demand (through cyclical industry mix effects) as being responsible for regional unemployment differences.

The effect of other factors for unemployment differences were also examined. A state's wage structure had a limited effect on unemployment. This is good news for policymakers because they can concentrate on creating and attracting high-wage industries without dramatically changing their unemployment rate. Consistent with cross-national studies [5], union densities played little role in determining cross-state unemployment rate differences. Similarly, there was not strong evidence that the availability and generosity of state unemployment insurance increased unemployment rates. Nevertheless, the relationship between the unemployment insurance system and unemployment is complex suggesting that more research is still needed [5]. However, states with a higher proportion of college graduates had lower unemployment rates. Thus, states may either directly increase their number of college graduates or promote industry that largely employs college graduates to reduce their unemployment rate.

In conclusion, the results suggest that more research should be undertaken regarding the factors that determine state-idiosyncratic growth. Much prior research has focussed on persistent long-run factors such as the level of state and local taxation, but little has been undertaken regarding cyclical factors. Such factors could include region-specific product cycles and differing vintages of regional capital stocks. Alternatively, by changing business expectations, changes in industry mix employment can indirectly influence regional business cycle turning points. An understanding of the nature and timing of the cycles may help states moderate or shorten the downturns, or at the very least better mitigate their impacts.

1. Blanchard and Katz [7, 12] also report similar trends in unemployment rate correlations.

2. Other approaches in exploring regional unemployment rates include examining the time series characteristics of regional unemployment rates [4; 7], state-specific Phillips curves [8; 19; 20; 38], and the influence of the variance-covariance structure of state-specific sectoral employment shocks on the state's unemployment rate [13; 35].

3. In the case of Georgia, a nationally prominent economic forecaster predicted that a large in-migration would cause the unemployment rate to rise from 5.8 to 6.2 percent in 1994 even while employment was forecasted to grow by 4.2 percent (Atlanta Constitution, May 25, 1994, p. E1).

4. Formally, INDMIX for state i in year t equals:

[INDMIX.sub.it] = ([summation of][g.sub.UStk] * [E.sub.it-1k])/[E.sub.it-1],

where [g.sub.Stk] is the national growth rate in industry k, [E.sub.it-1k] is state i's employment in industry k in year t - 1, and [E.sub.it-1] is state i's total non-farm private sector employment in year t - 1. The summation is over all non-farm private sector industries.

5. Formally, COMPETE for state i in year t equals

[COMPETE.sub.it] = ([E.sub.it]/[E.sub.it-1] - 1) - INDMIX,

where the notation is defined in the previous footnote.

6. [WAGEMIX.sub.it] for state i in year t equals

[WAGEMIX.sub.it] = [summation of][S.sub.itk][W.sub.UStk],

where [S.itk] equals industry k's share of state i's employment, [W.sub.UStk] equals national average annual earnings in industry k, and the summation is over all non-farm private sector industries.

7. Lilien [26] argues that the dispersion of employment growth rates across industries increases unemployment because of greater matching costs, search and hiring costs, and employer adjustment costs; and it was a cause for the rise in the natural unemployment rate. However, Abraham and Katz [2] show that dispersion in industry growth rates may have ultimately been caused by cyclical considerations, though Hosios [18] presents a model with implications inconsistent with Abraham and Katz.

8. The lag of INDVAR was also experimented with in some preliminary specifications, but it was never statistically significant and the other results were virtually identical including the results for INDVAR.

9. Levine [25] discusses the total effect of unemployment compensation on workers who are covered by unemployment insurance and by those who are uncovered.

10. Ehrenberg and Smith [10] summarize the demographic characteristics related to worker turnover and unemployment rates.

11. The F-statistic for the joint null hypothesis that the state dummies equal zero was 19.55.

12. Several alternative specifications that consisted of adding and dropping variables were conducted to examine the robustness of the results and to analyze the possibility of multicollinearity. The INDMIX and COMPETE coefficients were robust to these changes. For example, one alternative specification considered the possibility that exogenous changes in farm employment or defense spending had a significantly larger effect on unemployment than the average industrial mix effect. In this case, effects that we are attributing to state-specific competitiveness factors could really have been caused by exogenous shocks to the defense or agriculture sectors. We examined this possibility by including the annual percent change in employment due to the farm sector and the annual change in the percentage of gross state product that is accounted for by defense contracts as additional variables in the autocorrelation corrected panel specification. Yet, the coefficients for these two variables were either the wrong sign (i.e., negative) or insignificant and the competitiveness and the industry mix coefficients and t-statistics were virtually identical. Another possibility is that the sectoral share variables (e.g., FARM, MANU) are masking national restructuring employment effects. Fortunately, the industry mix coefficients were almost identical and their t-statistics were larger when the sectoral share variables were omitted.

13. The null hypotheses that the current period coefficients for industrial mix and competitiveness were equal were rejected at or below the 2% level in all three cases (F_INDMIX = COMP). However, the null hypotheses that the lagged period industrial mix and competitiveness coefficients were equal could not be rejected at reasonable levels of significance.

14. Like our results, Holzer found that the within labor-market variance in sales growth across industries played a much smaller role than between labor market variations in employment growth in determining local labor market unemployment rates.

15. The percent of welfare expenditures as a share of personal income and the percent of taxes as a share of personal income were also added as additional variables to test the sensitivity of the results. However, both coefficients were unexpectedly negative or insignificant and the other results remained virtually identical.

16. The matrix M is ordered by state and by time period (i.e., 48 states for 20 time periods). The weights were based on the average of 1972 and 1991 non-farm employment in state i's neighboring states. Non-farm employment was chosen because it appears to be a reasonable measure of the degree of spillover that a state could have on its neighbors' labor markets. See Kelejian and Robinson [22; 23; 24] for further details of the weighting matrix. Suppose that state i is bordered by states 1, 2, and 3. The weights in M for state i are [m.sub.ij] = [e.sub.j]/([e.sub.1] + [e.sub.2] + [e.sub.3]), where j = 1, 2, 3, [m.sub.ij] is state j's weight for state i for every time period, and [e.sub.j] is the average of state j's 1972 and 1991 non-farm employment.

17. The Hausman test creates predicted values for the possible endogenous variables from an instrumental variable regression. These predicted values are then included as additional variables in the regression specification where the t-statistics from these predicted variables are the test statistics. The exogenous instruments include the remaining independent variables from the specification and the following variables: annual international migrants as a share of the state's population; the percent of personal income that is accounted for by state and local taxes; regional employment growth in the state's region (i.e., Northeast, Midwest, West, South) net of the state's employment growth; the lag of regional employment growth in the state's region; the state's relative fuel costs; the state's relative housing costs; the annual change in the share of employment accounted for by farm employment; the annual change in military expenditures as a share of Gross State Product; the annual change in Federal civilian expenditures as a share of Gross State Product; and separate one-digit employment shares for durable, non-durable, and traded goods. The tax variable is from Government Finances [44] and the other variables are discussed in Treyz, Rickman, and Shao [41].

18. We also estimated an autocorrelation-corrected two-stage least squares model treating COMPETE and WAGE_COMP as endogenous to examine the sensitivity of our results. Although there were modest differences in the industrial mix and competitiveness coefficients, our policy conclusions remained unchanged. For example, the difference in unemployment between two states, one with an industrial mix employment growth rate that has been two standard deviations above the mean for two years and another state with it two standard deviations below the mean is 0.86%. The comparable difference for competitiveness is 1.62%. In both cases, they are about the same as before. Both industry mix coefficient t-statistics were well above 2.0. The competitiveness coefficients were imprecisely estimated where a joint F-test regarding the significance of the two coefficients could not be rejected until the 0.115% level. Regardless, the 2SLS estimates are inefficient and the Hausman test indicated that they should not be used in this case.

References

1. Abraham, Katharine G., "Help Wanted Advertising, Job Vacancies, and Unemployment." Brookings Papers on Economic Activity, 1, 1987, 207-48.

2. ----- and Lawrence F. Katz, "Cyclical Unemployment: Sectoral Shifts of Aggregate Disturbances." Journal of Political Economy, June 1986, 507-22.

3. Archibald, G. C., "Wage-Price Dynamics, Inflation, Unemployment." American Economic Review, March 1969, 124-34.

4. Bartik, Timothy J. Who Benefits From State and Local Economic Development Policies? Kalamazoo, Michigan: Upjohn Institute, 1991.

5. Bean, Charles R., "European Unemployment: A Survey." Journal of Economic Literature, June 1994, 573-619.

6. Blackley, Paul, "The Measurement and Determination of State Equilibrium Unemployment Rates." Southern Economic Journal, October 1989, 440-56.

7. Blanchard, Olivier J., and Lawrence F. Katz, "Regional Evolutions." Brookings Papers on Economic Activity, 1, 1992, 1-75.

8. Brechling, F. P. R. "Wage Inflation and the Structure of Regional Unemployment," in Inflation and the Labor Market, edited by D. Laidler and D. Purdy. Manchester, U.K.: Manchester University Press, 1973, pp. 197-226.

9. Campbell, Mike, "A Spatial and Topological Disaggregation of Unemployment as a Guide to Regional Policy." Regional Studies, 1975, 157-68.

10. Ehrenberg, Ronald G. and Robert S. Smith. Modern Labor Economics. 5th ed. New York: Harper Collins, 1994.

11. Freeman, Richard, and James Medoff. What Do Unions Do? New York: Basic Books, 1984.

12. Gallaway, Lowell E. "Geographic Labor Mobility in the United States 1957 to 1960." Research Report #28. Social Security Administration, Office of Research and Statistics, 1969.

13. Gruben, William C. and Keith R. Phillips, "Unionization and Unemployment Rates: A Reexamination of Olson's Labor Cartelization Hypothesis." The Review of Regional Studies, Winter 1990, 1-10.

14. Hall, Robert, "Turnover in the Labor Force." Brookings Papers on Economic Activity, 3, 1972, 709-64.

15. Harris, John R. and Michael P. Todaro, "Migration, Unemployment and Development: A Two Sector Analysis." American Economic Review, March 1970, 126-42.

16. Hirsch, Barry T. and David A. Macpherson, "Union Membership and Coverage Files from the Current Population Surveys: Note." Industrial and Labor Relations Review, April 1993, 574-78.

17. Holzer, Harry J., "Employment, Unemployment, and Demand Shifts in Local Labor Markets." Review of Economics and Statistics, February 1991, 25-32.

18. Hosios, Arthur J., "Unemployment Vacancies with Sectoral Shifts." American Economic Review, March 1994, 124-44.

19. Hyclak, Thomas and Geraint Johnes, "Real Wage Rigidity in Regional Labor Markets in the U.K., the U.S., and West Germany." Journal of Regional Science, August 1989, 423-32.

20. -----. Wage Flexibility and Unemployment Dynamics in Regional Labor Markets. Kalamazoo, Michigan: Upjohn Institute, 1992.

21. Hyclak, Thomas, and Gerald Lynch, "An Empirical Analysis of State Unemployment Rates in the 1970s." Journal of Regional Science, August 1980, 377-86.

22. Kelejian, Harry H. and Dennis P. Robinson, "Spatial autocorrelation: A New Computationally Simple Test With an Application to Per Capita County Police Expenditures." Regional Science and Urban Economics, September 1992, 317-31.

23. -----, "A Suggested Method of Estimation for Spatial Interdependent Models with Autocorrelated Errors, and an Application to a County Expenditure Model." Papers in Regional Science, July 1993, 297-312.

24. -----. "Infrastructure Productivity: A Razor's Edge." Mimeo, University of Maryland, February 1994.

25. Levine, Phillip B., "Spillover Effects Between the Insured and Uninsured Unemployed." Industrial and Labor Relations Review, October 1993, 73-86.

26. Lilien, David, "Sectoral Shifts and Cyclical Unemployment." Journal of Political Economy, August 1982, 777-93.

27. Malizia, Emil E. and Shanzi Ke, "The Influence of Economic Diversity on Unemployment and Stability." Journal of Regional Science, May 1993, 221-35.

28. Marston, Stephen T., "Two Views of the Geographic Distribution of Unemployment." Quarterly Journal of Economics, February 1985, 57-79.

29. McGee, Robert T., "State Unemployment Rates: What Explains the Differences?" Federal Reserve Bank of New York Quarterly Review, Spring 1985, 28-35.

30. McHugh, Richard and Richard Widdows, "The Age of Capital and State Unemployment Rates." Journal of Regional Science, February 1984, 85-92.

31. Medoff, James L., "U.S. Labor Markets: Imbalance, Wage Growth, and Productivity in the 1970s. "Brookings Papers on Economic Activity, 1, 1983, 87-120.

32. Moore, Thomas S. and Aaron Laramore, "Industrial Change and Urban Joblessness: An Assessment of the Mismatch Hypothesis." Urban Affairs Quarterly, June 1990, 640-58.

33. Murphy, Kevin J., "Geographic Differences in U.S. Unemployment Rates: Variance Decomposition Approach." Economic Inquiry, January 1985, 135-58.

34. ----- and Richard A. Hofler, "Determinants of Geographic Unemployment Rates: A Selectivity Pooled-Simultaneous Model." Review of Economics and Statistics, May 1984, 216-23.

35. Neumann, George, and Robert H. Topel, "Employment Risk, Diversification, and Unemployment." Quarterly Journal of Economics, November 1991, 1341-65.

36. Stevens, Benjamin H., and Craig Moore, "A Critical Review of the Literature on Shift-Share as a Forecasting Technique." Journal of Regional Science, 1980, 4-37.

37. Summers, Lawrence H., "Why is the Unemployment Rate so very High Near Full Employment?" Brookings Papers on Economic Activity, 2, 1986, 339-83.

38. Tiller, Richard B. and Robert W. Bednarzik, "The Behavior of Regional Unemployment Rates Over Time: Effects on Dispersion and National Unemployment." Journal of Regional Science, August 1983, 479-99.

39. Topel, Robert H. "Local Labor Markets." Journal of Political Economy, June 1986, S111-43.

40. Treyz, George I., Dan S. Rickman, Gary L. Hunt, and Michael J. Greenwood, "The Dynamics of U.S. Internal Migration." Review of Economics and Statistics, May 1993, 209-14.

41. Treyz, George I., Dan S. Rickman, and Gang Shao, "The REMI Economic-Demographic Forecasting and Simulation Model." International Regional Science Review, #3, 1992, 221-53.

42. U.S. Bureau of the Census. Census of Population, Decennial. Washington, DC: U.S. Government Printing Office (1970, 1980, 1990).

43. -----. Current Population Report. Washington, DC: U.S. Government Printing Office, Various Years.

44. U.S. Department of Commerce. Government Finances. Washington, DC: U.S. Government Printing Office, Various Years.

45. -----. Statistical Abstract of the United States. Washington, DC: Government Printing Office, Various Years.

46. U.S. Department of Labor. Geographic Profiles of Employment and Unemployment. Washington, DC: Government Printing Office, Various Years.

47. U.S. Employment and Training Administration. Unemployment Insurance Data Summary. Washington, DC: Government Printing Office, Various Years.