Understanding declining fluidity in the U.S. labor market.

Molloy, Raven ; Smith, Christopher L. ; Trezzi, Riccardo 等

ABSTRACT In this paper, we first document a clear, downward trend in labor market fluidity that is common across a variety of measures of worker and job turnover. This trend began in the early 1980s, if not somewhat earlier. Next, we present evidence for a variety of hypotheses that might explain this downward trend, which is only partly related to population demographics and is not due to the secular shift in industrial composition. Moreover, this decline in labor market fluidity seems unlikely to have been caused by an improvement in worker-firm matching or by mounting regulatory strictness in the labor or housing markets. Plausible avenues for further exploration include changes in the worker-firm relationship, particularly with regard to compensation adjustment; changes in firm characteristics, such as firm size and age; and a decline in social trust, which may have increased the cost of job searches or made both parties in the hiring process more risk averse.

*********

There is mounting evidence that the U.S. labor market has experienced marked declines in fluidity along a variety of dimensions. Examples include the rate of job-to-job transitions (Bjelland and others 2011; Molloy, Smith, and Wozniak 2014), the formation of new firms (Davis and Haltiwanger 2014), hires and separations (Hyatt and Spletzer 2013), and geographic movement across labor markets (Kaplan and SchulhoferWohl 2012; Molloy, Smith, and Wozniak 2014). This emerging consensus centering on a general set of concurrent trends raises obvious questions about whether these trends are related and what is causing them. Moreover, these trends could have substantial implications for the performance of the aggregate economy. On one hand, the declines in labor market fluidity could signal a rise in the costs of making labor market transitions, which are likely to have negative effects on aggregate productivity and economic performance. On the other hand, lower labor market fluidity could be a sign that there is less need to make such transitions, in which case the implication for aggregate economic performance may well be positive. The goals of this paper are to determine whether the trends in various measures of labor market fluidity are related, to establish when this fluidity began to decline, and to make progress in understanding the likely causes of this decline.

Because we seek to determine the date of the beginning of the decline in fluidity, we need to examine patterns of labor market fluidity that extend for a lengthy period of time, during the past several decades at a minimum. Consequently, our analysis focuses on data series that are available since at least the early 1980s. We combine information on labor market flows as measured from the perspective of workers (transitions into and out of employment and job-to-job transitions), flows as measured from the perspective of firms (job creation and job destruction), and flows as measured using interstate migration, a transition that is frequently associated with a job change or a change in labor force participation. Bringing together evidence from a variety of sources and methods of measurement is helpful because it reduces the influence of factors that might be idiosyncratic to a particular measure of fluidity and also smooths measurement error that might be specific to a particular data source.

A related body of literature on business dynamism documents that the formation of new firms, or start-ups, has declined for several decades, and that the reallocation of jobs across firms and establishments has declined during a similar period (Davis and others 2007; Davis, Faberman, and Haltiwanger 2012; Davis and Haltiwanger 2014; Karahan, Pugsley, and Sahin 2015). Our research complements this literature by exploring declines in transitions from the worker perspective, focusing on workers' labor market transitions such as changes in employers and changes in labor force participation. By examining how the propensity of workers to alter their employment situations has changed during the last several decades, we can gain new insights into the decline in business dynamism. Although it is possible that the decline in firm-side dynamism might explain the general decline in transitions among U.S. workers, our worker-side focus allows us to consider a number of other explanations as well. (1)

We begin with an analysis of trends in aggregate data. Using time series techniques, we isolate the low-frequency movements in each of eight labor market flows. Most of these trends present a clear, downward trajectory. Moreover, these downward trends appear to be related; using principal component analysis, we identify a single component that explains a large portion of the variation of the low-frequency movements of these series and puts a positive weight on all of them. The downward trajectory seems to have begun at least in the 1980s, and possibly earlier. Thus, the downward trend in fluidity predates the early 1990s, which highlights an important limitation of analyses of the fluidity trends that are based on data sources only available starting in the 1990s. Using our unified measure, labor market fluidity decreased by 10 to 15 percent during the period that we are studying. However, this single measure smooths changes across several separate measures of fluidity, which individually decline by as much as 25 percent. A decline of this magnitude implies a marked change in the labor market, and suggests that the effects of the trend in fluidity could be substantial.

Having established that the decline in labor market fluidity is sizable and appears to be a phenomenon that has been ongoing for three to four decades, we next turn to the question of why. This analysis is composed of three main sections. First, we examine the role of population demographics to see if changes in the distribution of worker characteristics can explain the declines or if the downward trends are concentrated among certain types of workers. Previous research has shown that shifts in the age distribution of the population, as well as other characteristics of workers--including health care-related job lock among those covered by employer-provided insurance--do not explain a substantial portion of the decline in various measures of fluidity (Kaplan and Schulhofer-Wohl 2012; Molloy, Smith, and Wozniak 2014). Other research has ruled out a compositional role for some firm characteristics (Decker and others 2014a; Hyatt and Spletzer 2013). Similarly, we find that changes in the distribution of age, sex, and marital status explain no more than half the trends in labor market flows as measured from the worker perspective. Trends in transitions into and out of employment appear to mirror trends in labor force participation. Thus, demographic groups with a secular increase in labor force participation (such as prime-age women) have experienced larger declines in transitions out of employment and increases in transitions into employment, whereas the reverse has been true for demographic groups with a secular decrease in labor force participation. Meanwhile, trends in job-to-job flows and interstate migration are similar for most demographic groups. Putting this all together, although demographics go some way toward explaining some labor market flows, they do not seem to account for the bulk of the decline in transitions that is common across all measures. Therefore, explanations for the general downward trend should apply to a wide range of workers.

Our second method of narrowing down explanations is to examine state-level trends in labor market fluidity. Local labor markets vary along many dimensions, so it seems natural to expect whatever is causing the aggregate decline in fluidity to have a larger influence in some locations relative to others. Following a strategy similar to the one we used for the aggregate data, we create a measure of general decline in labor market fluidity for each state that is based on both worker and job reallocation. Although labor market fluidity has decreased in all states, it has fallen much more in some states than others. There is a clear geographic pattern, in that fluidity has declined more in the Mountain and Pacific census divisions than in other locations. Surprisingly, this geographic pattern persists even after we control for a wide array of state characteristics, indicating that it is not related to the standard observable attributes of the population or to the industrial composition of firms. It is not obvious to us what might be driving this result, and we think it is worth exploring in future research.

Another outcome of the state-level analysis is that states with a larger share of workers in administrative support occupations and machine operators in the late 1970s experienced smaller subsequent declines in labor market fluidity. Workers in these occupations were particularly hard-hit by the secular decline in demand for workers who perform routine-intensive tasks, so their labor market transition rates may have been boosted as they left old jobs and searched for new ones. Thus, these results suggest that the decrease in fluidity would have been larger absent the secular decline in demand for middle-skilled workers. It therefore seems unlikely that the secular change in demand for skills and the accompanying widening of wage inequality could have caused the decline in labor market fluidity.

Finally, we directly assess a variety of specific theories for the decline in labor market fluidity by assembling evidence from existing research as well as new analysis. As we consider these hypotheses, we find it helpful to divide them into two general categories: those that have benign implications for general economic activity, and those with less benign implications. The benign explanations imply a reduced need for reallocation, such as reasons for improved worker-firm matches. The less benign explanations generally involve an increase in some cost that has caused labor market transitions to become more difficult.

Regarding the benign explanations, one hypothesis is that the match quality between workers and firms has improved. This trend would likely result in either larger returns to staying in the firm, or higher wages in the initial match. Using three cohorts from the National Longitudinal Surveys (NLS), we show that after controlling for returns to industry and occupation tenure, returns to employer tenure are small and have not changed noticeably from the late 1960s to the late 2000s. We also examine long-run trends in starting wages in the NLS and the Panel Study of Income Dynamics (PSID), and we find no evidence of a secular increase in match quality as reflected in higher initial wages. Consequently, it seems unlikely that the decline in labor market fluidity can be explained by better matching between workers and firms. A related hypothesis that could explain less labor market fluidity is that workers and firms have been investing more in job-specific training, since this type of investment is associated with reduced separations from employers (Cairo and Cajner 2014). Research on the long-run trends in this type of firm-specific training has had mixed results, and more studies on this topic would be helpful. Finally, a decrease in worker turnover might result from a greater ability of compensation to adjust to changes in the productivity of the worker-firm match. Again, evidence supporting this theory is rather mixed, but further investigation, particularly using matched employer-employee data, seems worthwhile.

Turning to the less benign explanations, we consider a number of factors that may have caused changes in the labor market to become more costly: a general decrease in the liquidity of the labor market resulting from a reduction of young workers; a decrease in job searches or willingness to take new jobs arising from decreases in social capital; and increased regulations in the housing or labor markets that inhibit labor market transitions. We find little support for any of these hypotheses, with the exception of the social capital channel, where we find weakly suggestive evidence of a role for declining trust. In particular, states with larger declines in the fraction of people who think that strangers are trustworthy have also experienced larger declines in labor market fluidity. This correlation is provocative, and more research is needed to explore the mechanism.

In the final portion of our analysis, we discuss the potential implications of the decline in labor market transitions. With fewer workers making these transitions, we might expect firms and workers to renegotiate wages less frequently. We find that in the 1980s and 1990s, wages were most strongly correlated with the best labor market conditions since the worker-employer relationship began, suggesting that wages were renegotiated when outside labor market conditions improved. In the 2000s, wages have become more closely tied to conditions in the worker's first year of employment. Thus, workers appear to be renegotiating wages less frequently.

I. Time Series Analysis

The goals of this section are (i) to identify long-run trends in various measures of labor market fluidity; (ii) to determine whether these trends are related; and finally (iii) to determine when declines in fluidity began. To do this, we identify eight aggregate time series on flows in the labor market and use time series techniques to estimate low-frequency trends in each of these series. We then assess the comovement of these low-frequency trends and discuss what these trends suggest about the magnitude and timing of declining fluidity.

Labor market flows can be measured from the perspective of workers making a transition or from the perspective of firms changing the number or composition of their employees. For example, the new employees at a firm must consist of workers who were formerly unemployed (coded UE), out of the labor force (NE), or working for another firm (JtJ). Similarly, workers flow out of a firm by transitioning to unemployment (EU), leaving the labor force (EN), or leaving to work for a different firm (JtJ). These worker flows are sometimes grouped into "hires" and "separations," defined as follows:

(1) Hires = NE + UE + JtJ,

(2) Separations = EU + EN + JtJ.

These transitions are gross flows, in that someone moving directly from one firm to another will be counted both as a separation (from the old firm) and a hire (to the new firm). Meanwhile, job flows (from firms' perspectives) are usually measured as a net flow. Specifically, job creation is usually defined as the net new jobs in new firms and expanding firms, whereas job destruction is usually defined as the net job loss from contracting firms and firms that have shut down. Notably, the sum of aggregate job creation and job destruction is much lower than the sum of aggregate hires and separations because many transitions do not necessarily lead to a change in the number of filled jobs (Davis and Haltiwanger 2014; Hyatt and Spletzer 2013; Fallick and Fleischman 2004).

In our analysis, we simultaneously consider flows as measured from both the worker and firm perspectives because both sets of variables are measured with error and are subject to idiosyncratic influences that are unrelated to the secular decline in fluidity. By combining them, we think we are more likely to identify a common component that accurately reflects general changes in labor market fluidity.

We start our analysis by considering EU, UE, NE, and EN, because these four flows are available at a quarterly frequency over a span of more than 40 years. Relative to annual data, the quarterly frequency makes it easier to isolate business cycle fluctuations from those located at lower-than-business-cycle frequencies. The long time series is essential for determining when the low-frequency movements began to turn down. Following the analysis of these four quarterly series, we extend the analysis to include job-to-job flows in order to complete the picture of reallocation from the worker's perspective. Doing so requires switching to an annual frequency and considering a shorter time period. Finally, we add in three additional annual series: job creation (JC), job destruction (JD), and interstate migration (IM). Although IM does not measure labor market flows directly, more than half of all interstate migrants report having moved for a reason related to the labor market. (2) Also, because we measure IM using a separate data source from the other worker flows, we think that including this measure helps to mitigate concerns that the measured declines in fluidity are due to mismeasurement in a particular data source.

In our time series analysis, we adopt a two-step procedure. First, we estimate the (smooth) low-frequency movement of the series, using a biweight filter. (3) As James Stock and Mark Watson (2012) point out, the local means estimated using the biweight kernel are approximately the same as those computed as the average of the series over a centered moving window, except that the biweight filter means are less noisy because they avoid the sharp cutoff of a moving window. Endpoints are handled by truncating the kernel and renormalizing the truncated weights to sum to 1. (4) The resulting low-frequency trends capture the long-run fluctuations of the series. In the second step, we run a principal component analysis (PCA) on the estimated low-frequency series. PCA is a statistical method that uses orthogonal transformation to convert a set of possibly correlated variables (in our case, time series) into a set of linearly uncorrelated variables called principal components. The idea is to identify one or more components that explain the largest possible portion of the variance of the underlying series. If a single component is associated with an eigenvalue greater than 1 and explains a large fraction of the underlying variance, this component can be interpreted as a common factor driving variation in all series. In our case, we interpret the first principal component as a measure of the long-run decline in labor market fluidity. With an estimate of the long-run trend in labor market fluidity in hand, we can then assess the magnitude of this decline and when it began.

An alternative, more formal approach would require testing each series for a unit root and, conditional on finding that the series are nonstationary, testing for a "common trend" among them (that is, cointegration). However, each series that we examine consists of rates, and because they are naturally bounded between 0 and 1, they are stationary by definition. Moreover, owing to the small number of observations--especially when considering annual series--unit root and cointegration tests would have very low power. A second alternative could be to test for a common cyclical component, assuming that each series contains two cycles--one at a business cycle frequency, and the other a lower frequency. However, once again the small number of observations and the use of annual series prevents us from taking this approach. Consequently, we prefer to first isolate the trend in each series and then use PCA to consider how they are related. We use PCA for several reasons. First, PCA provides a statistically based way to combine worker flows and job flows into a single measure of fluidity. Because worker flows are gross flows but job flows are net flows, adding or averaging these flows is not appropriate. Second, PCA gives equal weight to each series. Another way to combine the worker flows would be to add up the number of individuals making each transition and divide by total employment--a measure called "worker reallocation" by Steven Davis and John Haltiwanger (1999). This method, although appropriate for quantifying aggregate reallocation patterns, heavily weights NE and EN transitions because these flows are about twice larger in magnitude than EU and UE flows. Because our purpose is to search for a trend that is common across all types of transitions, we prefer a method that does not weight some flows more than others a priori.

I.A. The Data Series

The four quarterly series reflecting transitions into and out of employment (EU, UE, EN, and NE) are derived from the Current Population Survey (CPS), and are available from 1967:Q2 to 2015:Q3, for a total of 194 observations. (5) All flows are expressed as a share of persons in the initial labor market state (for example, EN shows the number of transitions from employment to not in the labor force as a share of employment). (6)

Regarding the annual series, we calculate aggregate job-to-job transitions from micro data for the Current Population Survey's Annual Social and Economic Supplement (CPS-ASEC), as provided by the Unicon Research Corporation. (7) Specifically, we calculate these transitions as the fraction of employed workers who report having had more than one employer in the previous year (respondents are explicitly instructed not to count multiple jobs held at the same time). We use data from 1975 to 2012 as provided by the Unicon Research Corporation, and extend through 2014 using data from the Integrated Public Use Microdata Series (IPUMS) (Ruggles and others 2015), for a total of 40 annual observations. Although this measure is a count of job transitions within a year, and therefore is not conceptually identical to more common measures of month-to-month transitions, it is highly correlated with measures created by matching CPS cross sections across months (Fallick and Fleischman 2004), which can be calculated from 1994 onward, as well as with job-to-job flows, as measured in the Quarterly Workforce Indicators published by the Census Bureau, which are available from 2000 onward. (8) Interstate migration from 1975 to 2010 is from the Internal Revenue Service's (IRS) migration data. Because the methodology for measuring migration changed in 2011, we extend the IRS data post-2010 with growth rates of migration rates from the American Community Survey (ACS). (9) Finally, the job creation and job destruction data are from the Census Bureau's Business Dynamics Statistics, recorded from 1977 to 2013, for a total of 37 observations.

I.B. Results

Figure 1 shows the four quarterly series (EU, UE, EN, and NE) and the extracted low-frequency components, and figure 2 shows the series recorded at annual frequency (JtJ, IM, JD, and JC) and the extracted trends. (10) All four annual series show clear evidence of downward trends during the sample period. UE also declines for most of its (somewhat longer) sample period. EU increases from the mid-1960s to mid-1980s, but then falls for much of the remaining period. EN falls from the mid-1960s to the late 1990s, and then flattens out. Finally, of all these measures, NE shows the least evidence of a downward trend--although, as described in the next section, this is because declines for prime-age men and younger persons are offset by a rise in NE for prime-age females, consistent with trends in labor force participation for these groups. Broadly, the evidence emerging from figures 1 and 2 suggests a long-run decline in fluidity, with all trends at the end of the sample being below or well below their levels in 1975.

Because the series in figures 1 and 2 have different scales, it is hard to compare the magnitude of the declines. In table 1, therefore, for the low-frequency component of each series we report the sample mean, the sample standard deviation, the minimum and the year in which it occurred, and the maximum and the year in which it occurred. In each case, the minimum is located at the very end of the series, whereas the maximum is at the beginning of the sample period (with the exception of NE, for which it is in the middle). On average, the size of the decline in fluidity measures is substantial. If we compare the deviation between maximum and minimum, the drop amounts to almost a fourth of the initial level for JtJ and to about 20 percent for IM, EN, and JC, whereas it is smaller for EU and NE. Also, these long-run fluctuations seem to be highly correlated. In table 2, we report the pairwise correlation coefficients among the eight estimated trends. Although these correlations are computed from a relatively small number of observations (37 annual data points), the evidence emerging from table 2 suggests a high degree of comovement across the low-frequency components of these labor market fluidity measures, with the exception of NE, which appears less correlated with the other trends.

[FIGURE 1 OMITTED]

[FIGURE 2 OMITTED]

Next, we formalize these correlations using PCA. Table 3 reports results from three distinct PCAs: one using the trends based on the four quarterly flows, one using the trends in the five annual measures of worker flows, and one using the trends in all eight annual series. In all cases, the first principal component explains the majority of the variance of the underlying series: 52 percent in the first case (when considering only EU, UE, EN, and NE), 77 percent in the second case, and 86 percent of the total variance when running the PCA including all series. For this reason, we focus on the first component as our common component of interest. As the right-hand side of table 3 shows, the PCAs put a positive weight on nearly every variable in the analysis (again, NE is the exception), indicating that the common component identifies a factor that is positively correlated with seven of the eight flows. Finally, we plot the first components of the three PCAs in figure 3. These three measures convey the same message: There is a clear, downward, long-run trend that is common across virtually all measures of labor market fluidity. Because the components generated by PCA are normalized to have a mean equal to 0 and a variance equal to 1, the result does not give much insight into the magnitude of the decline in fluidity.

Taking a simple average of the eight individual long-run trends suggests that general labor market fluidity decreased by 10 to 15 percent during the period that we examine. (11)

Finally, we address the timing of the decline in labor market fluidity. Although various idiosyncratic factors may have caused each series to have peaks in different time periods, the PCAs shown in figure 3 suggest that the decline in labor market fluidity began at least in the early 1980s. Notably, these declines appear to be fairly constant for most of the period we are considering. To a large extent, we obtain this result because we choose trends that filter out all but the very low-frequency movements in each series. Filters that allow for higher-frequency movements, such as using a 60-quarter window for the quarterly series, are more difficult to interpret because they tend to be correlated with the severe business cycles of the early 1980s and the late 2000s. Thus, it is difficult to distinguish possible changes in the long-run trend from the fact that there were two severe business cycles toward the beginning and end of the period. Consequently, we focus on the lowest-possible frequency movements in labor market fluidity, smoothing through possible inflection points in the data. (12)

[FIGURE 3 OMITTED]

All this evidence is robust to a large set of robustness checks (shown in appendix A). As mentioned above, methods that include higher-frequency movements lead to qualitatively similar results, but they are more difficult to interpret because they pick up the severe recessions of the early 1980s and late 2000s. We also tried a wider window, and we obtained almost identical results compared with the baseline. (13) Second, results are similar when we use two alternative filters to estimate the low frequency movements: (i) the low-pass version of the band pass filter developed by Lawrence Christiano and Terry Fitzgerald (2003), and (ii) the low-frequency cosine projection method suggested by Ulrich Muller and Watson (2015). (14) As a final check, we also reversed the order of our two-step procedure, running the PCA first on the raw series and then estimating the low-frequency trend of the first principal component. In this case, we obtained similar results to the baseline--although, without first smoothing out the cyclical fluctuations in these series, it is more difficult to identify a component that has a positive weight on all measures of fluidity.

Overall, these robustness checks largely confirm the baseline evidence of a long-run decline in fluidity that tends to be positively related to virtually all separate transition measures.

II. Worker Demographics and the Decline in Mobility

Declines in labor market fluidity during the past three to four decades coincide with other demographic and economic trends that seem, on their face, like logical explanations for a substantial portion of the secular decline in fluidity. Examples include the aging of the population and rising female labor force participation. Previous research has found that these demographic shifts account for only a little of the secular decline in some measures of fluidity. For example, Greg Kaplan and Sam Schulhofer-Wohl (2012) show that changes in the age distribution, changes in the types of occupations and industries, rising income inequality, and increased numbers of dual-earning households only explain a small amount of the decline in cross-state migration. In Molloy, Smith, and Wozniak (2014), we show that, in addition to being unable to explain much of the decline in interstate migration, shifts in these and other demographic factors (for example, education and geography) cannot explain much of decline in employment transitions across firms, occupations, or industries.

Similarly, Henry Hyatt and James Spletzer (2013) show that changes in a variety of worker characteristics (for example, age, gender, and education), and firm characteristics (for example, size and age) explain only a small fraction of the changes in worker flows, including job-to-job transitions, hiring rates, and separation rates. Regarding flows measured from the firm's perspective, Ryan Decker and others (2014a) find that the shift in the age distribution of firms (toward older firms) can account for no more than a third of the decline in job creation and destruction since the late 1980s.

In this section, we revisit these questions, using the labor market transitions that make up our fluidity measure. We assess how much of the aggregate change in fluidity can be explained by changes in the distribution of demographic characteristics (for example, aging of the population), and we identify important differences in fluidity across demographic groups. For ease of exposition, we examine job finding rates (UE and NE transitions as a share of nonemployment), job separation rates (EU and EN as a share of employment), and job-to-job transitions (as a share of employment), although we also note where findings are different for the separate UE, NE, EU, and EN flows.

Demographic shifts should affect movements in labor market fluidity to the extent that average levels of fluidity vary across demographic groups. Indeed, figure 4 reveals a number of important demographic differences in job finding and separation rates, as well as demographic differences in job-to-job transitions and interstate migration (as measured in the CPS-ASEC). Job separation rates tend to be higher for younger workers (ages 16-24). Job finding rates and job-to-job transitions are higher than average for younger workers and lower than average for workers nearer retirement age (ages 55 and older). Migration rates are also higher for younger persons, and lower for older persons. Because age appears to be an important determinant of the level of many of these measures, the gradual aging of the labor force offers one potential, cohesive explanation for the decline in these measures of fluidity.

II.A. The Role of Changes in Demographics

To assess the contribution of shifts in the distribution of characteristics to the aggregate movement in these measures of fluidity, we estimate the following regression, which follows the approach of Robert Moffitt (2012):

(3) [y.sub.ikt] = [[beta].sub.0] + [X.sub.ikt][[beta].sub.k] + [[THETA].sub.t] + [[epsilon].sub.ikt].

Here, k is an age-sex-education-marital status category. For ease of computation, we collapse the data to k-level cells by age, sex, four education groups, and marital status, by year. (15) Included covariates are, depending on the specification, a full set of age dummies, sex dummies, education group dummies (no high school degree, high school degree, some college but less than a 4-year degree, 4-year degree or more), and marital status dummies (ever married or not). When we only include year fixed effects in the regression, the fixed effects estimate the average fluidity in each year. When other demographic controls are included, the year fixed effects represent the annual average fluidity in each year after controlling for the included demographic controls. We then normalize the year fixed effects to 0 in the start of the sample (1976 for labor market flows, and 1981 for interstate migration).

Figure 5 plots these fixed effects for job finding and job separation rates, job-to-job transitions, and interstate migration. The solid line plots year fixed effects without controlling for any demographics; the trends in these series correspond to the aggregate trends shown in figures 1 and 2. The dashed line plots the year fixed effects after controlling for age, sex, and marital status. The dotted line shows the year fixed effects after also controlling for education, another characteristic of the workforce that has displayed a considerable secular change during the past four decades.

For all series, the first set of demographic controls explains at most half the decline in all measures of fluidity during this period. When we include education, we can overexplain the decline in job separations. Mechanically, persons without a high school degree tend to have high job separation rates, and persons with a 4-year degree or more have low job separation rates; these differences are so big that the secular rise in educational attainment would be expected to reduce job separation rates by even more than actually occurred, all else being equal. Meanwhile, shifts in the distribution of education do not do much to help explain any of the movements in job finding rates, job-to-job transitions, and migration. (16)

These figures may be somewhat misleading in that the dashed and dotted lines (fixed effects with demographic or education controls) are notably above the solid lines (fixed effects with no controls) at the end of the sample. However, in most cases the divergence between these lines occurs during the late 1980s or early 1990s, and the subsequent trends in the lines are similar. Appendix table A.1 shows the percent of the change in each fluidity measure (since the early 1980s or early 1990s) that is explained by the age-gender distribution as estimated by this regression-based approach. (This is calculated as the change in the dashed lines as a share of the solid lines.) The lower panel of table A.1 shows that for all measures but EN and EU, little of the change in fluidity since the early 1990s can be explained by changes in the age-gender distribution.

[FIGURE 4 OMITTED]

[FIGURE 5 OMITTED]

We also estimate the contribution of demographics with a traditional "shift-share" analysis; that is, we hold demographic shares fixed for a particular period and allow fluidity measures within demographic cells to evolve as they did (Moffitt 2012; Aaronson and others 2014). Appendix table A.1 shows results for changes since the early 1980s and early 1990s. For the most part, these findings are broadly consistent with our regression-based approach; changes in the age-sex distribution can explain no more than half the decline in UE, NE, JtJ, or IM, but can explain a good bit more of the decline in EN and EU.

II.B. Changes in Fluidity for Particular Demographic Groups

As shown above, demographic shifts appear to explain some, but not all, of the secular decline in fluidity. To understand what may be responsible for the remainder, it is useful to consider differences in fluidity trends across demographic groups.

Returning to figure 4, declines in the job finding rates have been much steeper for prime-age males and younger persons; job finding rates for prime-age females rose for much of this sample period. By contrast, separation rates fell more for older workers and prime-age females. The demographic differences in these trends mirror well-documented differences in labor force participation (Aaronson and others 2014; Council of Economic Advisers 2014). Female labor force participation rates rose steadily through the late 1990s, reflecting changes in social and workplace norms and increased job opportunities for women. This pattern likely contributed to the secular increase in job finding rates and secular decline in job separation rates for prime-age women. Meanwhile, participation rates for prime-age men have been in a prolonged decline, likely in part due to technology- and globalization-driven shifts in labor demand away from male-dominated occupations and industries. This decline in participation is consistent with the secular decline in job finding rates and rise in job separation rates for this group. Participation rates for older persons have also risen as retirement ages have increased, and consequently their job separation rate has fallen. Finally, the decline in the job finding rate of younger persons is consistent with the decline in participation rates for this group, likely reflecting increased rates of college enrollment. (17)

Trends in job-to-job transitions and interstate migration are much more similar across groups than job finding and separation rates (figure 4). In particular, job-to-job transitions have declined for all groups except those ages 55 and older, and interstate migration rates have declined for all groups.

To summarize, there are important differences in labor market fluidity across groups of workers according to sex and age, which in many cases reflect secular trends in labor force attachment during this period. Shifts in the composition of the population toward groups that tend to make labor market transitions less often explain some, but no more than half, of the aggregate trends in fluidity. Thus, although demographic shifts clearly matter, there remains considerable room for other explanations.

III. State-Level Differences in Labor Market Fluidity

As is true for any large country, the U.S. labor market is a collection of smaller, local labor markets that differ along many dimensions. Geographic movement of workers and firms helps to integrate these markets, although this integration is far from perfect because long-distance migration is costly. The decline in the national average of labor market fluidity that we have documented thus far must therefore also occur at the local level, but perhaps to varying degrees across states. (18) In this section, we analyze variation in trends in labor market fluidity across states in hopes of shedding light on the factors that are behind the decline in the national average. State-level trends in fluidity are likely similar to those at finer levels, making them a useful approximation of subnational labor markets (Molloy, Smith, and Wozniak 2011). Moreover, data concerns are also relevant: Publicly available data sets with annual data on labor market transitions are too small to be able to identify geographic areas any smaller than states. For this analysis, we create state-level measures of fluidity using the same eight measures that we used in the aggregate analysis. Most measures are from the CPS: flows from employment to unemployment (EU), flows from employment to not in the labor force (EN), flows from unemployment to employment (UE), flows from not in the labor force to employment (NE), job-to-job transition rates (JtJ), and interstate migration rates (IM). Data on job creation (JC) and job destruction (JD) are from the Business Dynamics Statistics produced by the Census Bureau. Due to the availability of the migration and job creation and destruction variables, the eight measures combined are available from 1980 to 2013. (19) We focus on annual rather than quarterly data because we are concerned that many states are too small to reliably estimate labor market flows at a higher frequency.

Our state-level analysis follows the same two-step procedure we use with the aggregate data. We start by estimating a state-level trend for each measure of labor market fluidity. With only 34 annual observations for each state, it is not possible to employ our time series techniques to estimate these trends. Instead, for each state we estimate a linear time trend from an ordinary least squares regression that includes a trend and the state unemployment rate (contemporaneous and one-period lag). The coefficient on the linear trend reflects the average decline in each measure by state, after (roughly) accounting for the business cycle.

One concern with this method is that the linear trend assumes that declines in fluidity have been constant over time. To assess this assumption, we interact the linear trend with an indicator for the second half of the sample (post-1996). Because we estimate separate regressions for each state and each measure of labor market fluidity, this exercise yields 408 estimates (51 states times 8 measures) of trend breaks. In only about one-quarter of cases does the estimated trend change by more than 20 percent from the first half to the second half of the sample, with the difference being statistically significant at the (5) percent level or less. (20) Consequently, although there are clearly cases where the trend has not been constant during this 34-year sample period, we conclude that characterizing the general pattern with a linear trend is a reasonable approximation.

In the second step, we combine the data for all states and use PCA to identify the first principal component among all the eight measures of fluidity. Because we include the trends for all states in a single PCA, this method uses variation across states to determine the common patterns among the eight variables. The first principal component explains 70 percent of the variation among these eight trends and has a positive factor loading on each one. Thus, we use this component as the average decline in labor market fluidity by state.

[FIGURE 6 OMITTED]

Figure 6 reports the trend in labor market fluidity for each state. (21) The first point to take away from the figure is that the estimated trend is negative in all states. Nevertheless, there is a substantial amount of variation in declines in fluidity across states. Declines are relatively mild in a number of eastern states like North Carolina and Connecticut, averaging about 0.5 percent of the initial level of fluidity per year, whereas they average more than 1.5 percent per year in western states like New Mexico and Montana. (22)

The geographic pattern of declines in labor market fluidity is intriguing because states in the West differ from states in the East along a number of demographic and economic dimensions. With only 51 states and state-level characteristics that are highly correlated with one another, it is extremely difficult to tease out which state-level characteristics are robustly correlated with the decline in labor market fluidity. Nevertheless, as an attempt to examine this question, we estimate a series of regressions with the decline in labor market fluidity as the dependent variable and various sets of state-level characteristics as independent variables. (23) We consider the following sets of characteristics: population age, educational attainment, marital status, homeownership, industry, occupation, union membership, and class of worker (private, self-employed, or government). (24) For each set of variables, we consider correlations with the average level from 1977 to 1979, as well as the trend from 1980 to 2013. These trends are estimated using the same methodology as was used for estimating trends in labor market fluidity. The dependent and independent variables are all scaled to have a mean of 0 and a standard deviation of 1, so that the magnitudes of the coefficients can be interpreted in terms of standard deviations. For most sets of characteristics, one or two appear to be at least moderately correlated with the trend in labor market fluidity. One notable exception is homeownership--neither the initial level nor trend is correlated with the trend in fluidity--suggesting that changes in the cost of homeownership or the rise in homeownership are unlikely to explain the decline in fluidity. The remaining correlations are fairly difficult to distill into any clear explanations for the decline in fluidity, so we combine all the variables that appeared to be meaningful into a single regression and drop variables that do not maintain a significant coefficient with a magnitude of at least 0.1 (that is, a change of 1 standard deviation in the variable is associated with a change of at least 0.1 standard deviation in the trend in fluidity). Table 4 reports the results of this exercise. The coefficients for the full sets of variables are reported in appendix table A.6.

Four interesting patterns emerge. First, declines in fluidity are smaller in states with larger initial shares of workers in administrative support and operator or fabricator occupations. (25) This correlation is likely related to the secular decline in demand for middle-skilled workers, which was particularly prevalent for workers in these categories (Autor 2011). (26) The displacement of middle-skilled workers may generally contribute to additional churn in the labor market as these workers leave their old jobs and search for new jobs. It is worth noting that when we consider industry alone, we also find that declines in fluidity were smaller in states with a higher initial manufacturing share, a sector where the change in demand for skill was more pronounced. (27) However, this result disappears once we control for the occupation shares. In general, these relationships suggest that the displacement of middle-skilled workers has partly offset the general decline in fluidity in states with concentrated employment in routine-intensive jobs. However, they do not explain why the general decline occurred. In fact, they suggest that the decline in fluidity would have been more severe in the absence of the change in the demand for certain types of skill.

A second interesting correlation is that declines in labor market fluidity are marginally smaller in states with a larger decline in union membership, which is consistent with the notion that the decline of unions has reduced the frictions associated with hiring and firing workers. A third point to draw from table 4 is that accounting for these state characteristics reduces the coefficients on the census division indicators somewhat, but differences in the Middle Atlantic, Mountain, and Pacific divisions are still material. These regional differences also persist after controlling for state population growth from 1960 to 2010, or the change in population growth from 1960-70 to 2000-10. (28) Thus, the geographic patterns are not largely attributable to the wide array of observable state characteristics that we are considering here.

Finally, neither the levels nor trends of the states' distributions of age or education are related to the subsequent decline in labor market fluidity. Although we do find a positive correlation between the trend in the population ages 35 to 44 and the trend in fluidity, this result is entirely driven by the fact that declines in labor market fluidity were largest in Alaska, and this state also experienced the largest decrease in the population share for this age group. Thus, although these demographic factors were important in explaining the downward trends in some individual measures of fluidity, the result does not hold when combining all measures together. This difference makes sense because these demographics had opposite effects on different flows; for example, the rise in the labor force participation of older workers reduces job separation rates but contributes positively to job finding rates. Combining these flows and focusing on the first principal component across all measures of fluidity reduces the roles of such demographic effects.

IV. Why Is Fluidity Declining? Benign and Less Benign Explanations

In this section, we consider two classes of explanations for the decline in labor market fluidity: some that are not likely to imply adverse consequences for workers or economic activity, which we call "benign"; and some that are more likely to imply adverse consequences, which we call "less benign." By bringing together results from the literature and performing additional analyses, we assess several explanations in both categories.

IV.A. Benign Explanations

Reduced transitions may reflect improvements in the worker-firm relationship, and thus less need for workers to change jobs. A major source of such improvement may be better matching between workers and firms. As matching improves, it becomes less likely that another job exists where a worker would be more productive, and thus transitions in the labor market decline. A related but separate cause of improvements in the worker-firm relationship could be if firms are investing more in their workers through increased training, thereby strengthening workers' ties to their firms. Finally, compensation may have become more responsive to changes in productivity, reducing the need for the worker-firm match to dissolve in order for compensation to adjust. Although a full welfare analysis is outside the scope of this paper, all three of these explanations seem likely to be benign, if not beneficial, for the overall functioning of the economy.

EVIDENCE ON IMPROVED MATCHING If matching has improved and wages reflect match quality, then a worker's wage will be more closely aligned with his or her best possible match quality over the course of his or her career. We cannot directly test for changes in match quality because it is unobservable. However, if wages proxy for realized match quality, then trends in wages can provide some insight into the plausibility of the improved matching hypothesis.

To fix ideas, define match quality, [delta], in the following way. Let [[THETA].sub.F] be the set of all firms and [[THETA].sub.W] be the set of all workers. For simplicity of notation, assume that firms have only one worker. M([[THETA].sub.W], [[THETA].sub.W]) is a one-to-one allocation of workers to firms.

A role for match quality implies the existence of an allocation [M.sup.*], such that there is no Pareto-improving switch of workers across firms that would raise or hold constant match quality for all firms.

Specifically, under [M.sup.*], there is no change in worker-firm matches k, j, such that

(4) [[delta].sup.fj.sub.wk] > [[delta].sup.fj.sub.wk] and [[delta].sup.fk.sub.wk] [less than or equal to] [[delta].sup.fk.sub.wk],

where [[delta].sup.fj.sub.wk] is the match quality generated when worker k matches with firm j, and so on for the other terms. We define improved match quality to mean that the labor market has moved closer to [M.sup.*]. More firms employ their [M.sup.*] worker, and more workers are employed at their [M.sup.*] firm.

To consider the impact of improved matching on wages, assume that wages, [omega], equal match quality plus a base equal to the average level of human capital among workers, [bar.h], which we assume can be deployed for the same return in any firm:

(5) [[omega].sub.w] = [bar.h] + ([[delta].sup.f.sub.w]|M)

When matching falls short of [M.sup.*], swapping workers can result in Pareto improvements in match quality that raise wages for some without lowering wages for others. Improvements in matching should therefore result in higher average wages (all else being constant), given that these matches are more frequently made. To put it another way, under a better-matched allocation, more workers are employed by firms at which, if they were to change employment, the match quality for themselves, their replacement, or both would be lower.

The incidence of these higher wages over the course of a worker's career depends on when match quality is revealed in the worker-firm relationship. If match quality is revealed before starting employment, and if match quality is rising, then we should observe starting wages rising over time. Hyatt and Spletzer (2016) find no evidence that starting wages increased during the 1996-2014 period in the CPS or during the 1998-2008 period in the Longitudinal Employer-Household Dynamics.

To investigate further, we look at initial wages at a job in three cohorts from the National Longitudinal Surveys that span the years from the late 1960s to the present. Specifically, these data come from the young men's cohort of the National Longitudinal Survey of Older and Young Men (NLSM), the National Longitudinal Survey of Youth 1979 (NLSY79), and the National Longitudinal Survey of Youth 1997 (NLSY97). We focus on results for men because the labor force participation of women changed markedly during these decades, so the types of women starting jobs in recent years are likely quite different from the types of women starting jobs in the 1960s and 1970s. (29) Because respondents in the latest waves of the NLSY97 are still young, we restrict each sample to respondents ages 22 to 33 to maintain comparability across the samples.

To calculate starting wages, we regress the real wage of young male workers who have less than one year of tenure at their current employer on indicators for age, race, education, and year. The regression is estimated separately for three cohorts during the periods 1966-81, 1979-94, and 2002-13. The constant of this regression reveals the average starting real wage in each period. Table 5 shows that in these cohorts, starting wages rose somewhat from the first period to the second, but then decreased in the third period. If average starting wages constructed in this way reflect average match quality, this pattern suggests that matching was better in the 1980s and early 1990s than in more recent years, a result that is inconsistent with a rise in match quality. (30) The same pattern holds within broad skill groups, suggesting no trend improvement in initial match quality even among more educated workers, who faced rising demand for their skills during this period. We find similar results in the PSID, for which we can look at older workers as well as younger workers, but only from 1976 onward (see appendix figure A.5). One concern with interpreting the trends in starting wages as evidence of an initial match quality is that workers bargaining power may have declined over time, putting downward pressure on wages even as match quality improves. Indeed, labor's share of income has fallen substantially since the early 1980s. Nevertheless, given that aggregate productivity has been rising over time, changes in bargaining power would need to have been substantial to entirely offset these gains.

If match quality is only revealed after a worker has been with a firm for some amount of time, the quality of retained matches should rise across cohorts, even though the quality of new matches would not improve. As long as wages reflect match quality, returns to tenure with an employer should rise across cohorts of workers. We also test this hypothesis empirically using our panel of young workers from the three cohorts of NLS respondents. To examine changes in returns to employer tenure across cohorts, we estimate the following regression equation:

(6) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].

The dependent variable is log real hourly wages for respondent i at the main job in survey year t. The hourly wage is the "hourly rate of pay" variable constructed for each reported job by NLS administrators. [X.sub.it] is a set of additional background controls that includes age, age squared, and educational attainment dummies (high school dropout, high school graduate, 1-3 years of college, 4 or more years of college). [[theta].sub.t] is a set of survey-year dummies. [[THETA].sub.i] is a set of person fixed effects, which are included to mitigate the concern that higher-quality workers may stay longer with an employer, biasing up the estimated return to tenure. It is worth emphasizing that the regression includes controls for occupation and industry tenure, so the estimated return to employer tenure does not include returns to more general human capital that workers can take with them when they change employers.

We estimate this equation separately for each of our NLS cohorts. The results are reported in table 6, which shows estimates of the returns to a third year of tenure with an employer. We focus on the third year of experience because average tenure in each sample is about two to three years. Returns to employer tenure are economically small and insignificant for all three cohorts, providing no evidence for rising returns to employer tenure over time. (31)

CHANCES IN EMPLOYER-PROVIDED TRAINING Isabel Cairo (2013) and Cairo and Tomaz Cajner (2014) develop models in which job-related training reduces the propensity of workers to separate from employers. Thus, an increase in training requirements for broad groups of workers could contribute to the secular decline in labor market fluidity. Based on a model simulation, Cairo (2013) concludes that rising training requirements can account for about one-third of the decline in job creation and destruction from 1993 to 2011. A related hypothesis is that skills have begun to decay at a faster rate, as might be the case with firm- or job-specific skills. Shigeru Fujita (2015) proposes a model in which there is a secular increase in the risk of experience depreciation during an unemployment spell for all workers in an economy. Workers therefore become increasingly reluctant to separate from their firms and risk the loss of skills that would result from a failed transition to a new job. He argues that such a model can reconcile declining labor market turnover with stagnant wages and rising public anxiety about job security.

A challenge for the view that the decline in fluidity can be explained by an increase in job-related training is that evidence for a sustained increase in such training is limited. Cairo (2013) finds that the share of workers employed in occupations that require relatively more specific training, as classified by the Dictionary of Occupational Titles, increased from 1970 to 2005. Moreover, an update to the Dictionary's classification system reveals an increase in training requirements within occupations from 1977 to 1991 (Cairo 2013). Conversely, several studies that use direct evidence on the provision of training by employers find no evidence of a sustained upward trend. In fact, it appears to have declined after peaking in the mid-1990s, a period when policymakers were calling for increased training, and firms were providing it (Lerman, McKeman, and Riegg 2004). This evidence is reviewed by Peter Cappelli (2015) and the White House (2015). Cairo and Cajner (2014) also find that the incidence of formal on-the-job training was fairly flat, on net, from 1987 to 2007. Finally, it seems likely to us that a greater incidence of on-the-job training should result in greater returns to employer-specific tenure, for which we found no evidence in our analysis above.

ENHANCED COMPENSATION FLEXIBILITY A third benign possibility that we consider is that reduced transitions reflect fewer frictions from wage or compensation adjustment. If firms are able to adjust compensation to reflect changes in productivity, this could reduce the need for layoffs as well as decrease voluntary separations, whereby workers leave a job in search of a larger wage adjustment. (32)

There is some evidence for this kind of enhanced compensation flexibility in the literature. Peter Gottschalk and Moffitt (1994. 2002) demonstrate that the transitory component of men's earnings rose between the late 1970s and the late 1980s. They argue that enhanced compensation flexibility--perhaps arising from changes in worker protections or regulation, or from a more competitive product market--may have led to more rapid shifts in wages. Diego Comin, Erica Groshen, and Bess Rabin (2009) use Compustat data to test for an increasing correlation between firm-level volatility in total sales and firm average compensation over time. They find that firms with higher sales volatility also exhibit higher average wage volatility, and that this relationship became much stronger after 1980. They argue that this change over time reflects an enhanced pass-through of productivity fluctuations to worker wages. They further show that the compensation--sales volatility relationship is strongest in large firms. With a secular shift of employment toward larger firms, as documented by Davis and others (2007), one would expect the average pass-through of productivity to wage volatility to have increased.

However, other evidence raises questions about the potential for an increase in compensation flexibility to explain declining fluidity. First, there is little evidence of a sustained rise in the variance of transitory earnings. More recent analyses show a large increase in about the late 1970s and early 1980s, followed by a long, stable period and possibly even by a reversal, before it rose again into the 2000s and particularly in the Great Recession (Gottschalk and Moffitt 2009; Shin and Solon 2011; Koo 2016). One recent view holds that the increase in earnings volatility among men is related to severe recessions (Koo 2016), and is driven in large part by spells of unemployment (Ziliak, Hardy, and Bollinger 2011; Koo 2016). Additionally, any increase in earnings volatility is confined to male workers, given that earnings volatility has trended down for women since 1970 (Dynan, Elmendorf, and Sichel 2012).

IV.B. Less Benign Explanations

Another class of explanations associates declines in fluidity with an increase in some cost of making an employment transition. In general, we think that rising costs are unlikely to be benign in their overall impact on the economy. Not only does a cost require resources to surmount, but a rise in transition costs and the consequent reduction in reallocation will result in a less optimal allocation of resources.

THE ROLE OF AN AGING WORKFORCE REVISITED We begin by exploring the possibility that an aging workforce has led to fewer transitions in the labor market. Above, we showed that changes in the age composition of the population can explain a portion of the declines in some labor market flows--particularly those related to labor force participation--but that the age distribution alone could account for less than half of the general decline in fluidity. However, simple decompositions might not yield the entire effect of the age distribution, because if an aging workforce has broader general equilibrium effects on fluidity in the labor market, then aging could cause declines in fluidity even for older workers. For example, Robert Shimer (2001) develops a model in which a larger fraction of young workers generates more churning in the labor market, and older workers benefit from this churning as well. Similarly, Fatih Karahan and Serena Rhee (2014) develop a model in which an increase in the fraction of workers with higher moving costs (that is, older workers) causes firms to hire more local workers, reducing the migration rates of all types.

To evaluate the likelihood of such general equilibrium effects, we look to see whether states with a larger decline in the fraction of young people have also experienced a larger decline in the labor market fluidity of older workers. Although we found little evidence of this correlation in section III after controlling for other state characteristics, that analysis did not directly address the correlation of the youth share with the fluidity of older workers. Consequently, we slightly alter the method described in section III to measure declines in fluidity for older workers. We calculate state-level fluidity measures only for 35- to 64-year-olds, excluding job creation and destruction, given that those two variables are not available by age of worker. Next, we calculate the trends in these six measures using the same regression method described above, and then we combine the six trends using PCA. The results of the PCA are similar, in that all the variables have a positive loading and the first principal component explains a large fraction of the variation in the data.

Figure 7 graphs the estimated declines in youth share against the estimated decline in labor market fluidity among older workers. The correlation is very weak. And the correlations are similarly weak when we control for the state characteristics that were found to matter in section III, as well as when we examine each measure of labor market fluidity separately. (33) This evidence casts doubt on the idea that the decline in the population of young workers has had a general equilibrium effect on the labor market transition rate of older workers.

[FIGURE 7 OMITTED]

DECLINING social CAPITAL Social institutions, and social capital in particular, are positively related to economic performance (Knack and Keefer 1997). (34) Recent research argues that this relationship is causal, with greater aggregate social capital leading to improved long-run growth at the country level (Algan and Cahuc 2010). It is also possible that social capital is important for job and worker searches, as there is evidence that jobs are often found through personal networks (Bayer, Ross, and Topa 2008; Hellerstein, Mclnerney, and Neumark 2011; Hellerstein, Kutzbach, and Neumark 2014). Two major social capital measures for the United States, both taken from the General Social Survey (GSS), have been declining for the last several decades (Glaeser, Laibson, and Sacerdote 2002). Declines in social capital--particularly the extent and strength of social networks--may raise the cost of job searches by forcing workers to rely on more formal channels with less detailed information on the types of jobs available and the associated firm environments. In addition, reduced social capital may increase the cost of new hires because managers have less information about potential workers.

[FIGURE 8 OMITTED]

We use restricted-use GSS data with state identifiers to test for a relationship between social capital and fluidity in our state-level framework. The GSS has been widely used to measure social capital in the United States. Of several such measures that can be constructed, the indicator variable for agreement with the statement "Most people can be trusted" is available during the longest period, for almost all the years from 1972 to 2014. (35) As shown in figure 8, the fraction of respondents who agree that most people can be trusted has declined markedly during the past three decades.

[FIGURE 9 OMITTED]

According to Edward Glaeser, David Laibson, and Bruce Sacerdote (2002), this is a useful measure of aggregate social capital, for it indicates whether a community has a large share of members who are likely to behave in a trusting manner in their transactions. A second common measure of social capital from the GSS is the number of different types of membership organizations to which a respondent belongs. We focus on the trust measure because the memberships variable is not reported after 2004 and was not asked about consistently in the years before that, but the results given in figure 8 are broadly similar for the two measures.

Figure 9 shows the relationship between a state's trend in fluidity and its trend in social capital as measured by the trust share shown in figure 8. (36)

Due to gaps in state coverage in the GSS from year to year, we can only reliably estimate trends for 41 states. Nevertheless, the figure shows a roughly positive relationship between changes in a state's social capital and its change in fluidity. A regression using these 41 points shows that this relationship is not statistically significant and is small in magnitude. In particular, a more negative trend in trust of 1 standard deviation is associated with a larger decline in labor market fluidity of only 0.06 standard deviation. However, this correlation more than doubles, to 0.15, when two outliers where trust increased substantially are excluded. It is also worth noting that some of the states with the largest declines in trust were in the Western census region--the part of the country where declines in fluidity have been unusually large. The positive correlation between the trend in trust and the trend in fluidity is robust to controlling for the state characteristics that were found to matter in table 4. (37) Thus, this evidence is weakly suggestive that institutional changes, particularly a decline in social trust, may accompany the decline in fluidity. It is impossible to tell whether this reflects the role of a third factor vis-a-vis both trust and fluidity, or whether it reflects a causal relationship, but this question deserves further consideration by researchers.

REGULATION OF LAND USE AND BUSINESS PRACTICES A third candidate explanation we consider is regulations on businesses and land use. Specifically, we examine whether the regulation of land use, which restricts housing supply, or regulatory practices that affect the costs of hiring or firing workers are associated with declining fluidity. Restrictive land use regulations may be preventing the geographic reallocation of workers, and thus reducing labor market fluidity more generally (Ganong and Shoag 2015). Although this hypothesis may seem unlikely given that labor market fluidity has also declined substantially for transitions that do not require a change in residential location (Molloy, Smith, and Wozniak 2014), it is possible that it could be relevant if geographic reallocation is important for overall economic growth, as argued by Chang-Tai Hsieh and Enrico Moretti (2015).

Figure 10 displays the correlation of state-level declines in labor market fluidity with the average degree of regulation, as measured by the Wharton Residential Land Use Regulatory Index (Gyourko, Saiz, and Summers 2008), which is based on a survey conducted in the early 2000s. The figure shows no support for the hypothesis that declines in labor market fluidity are more concentrated in states with tighter land use regulation.

[FIGURE 10 OMITTED]

Regarding regulations that might affect labor market transitions more directly, Nathan Goldschlag and Alexander Tabarrok (2015) show that job creation and job destruction are not, in fact, lower in industries with a higher degree of federal regulation--including, but not limited to, labor regulations--in a panel of industries from 1999 to 2011. Moreover, federal regulation has been rising faster for manufacturing than for other broad industry categories since 1975, whereas fluidity has been declining by less in this sector (Decker and others 2014b). Meanwhile, in Molloy, Smith, and Wozniak (2015), we find no evidence that occupational licensing requirements, which have become considerably more common since the 1950s, have contributed to the secular decline in geographic or labor market transitions. Finally, we consider the role of the formalization of hiring practices using data on membership from the Society for Human Resource Management (SHRM), the major professional organization for human resource workers in the United States. Although the fraction of the labor force who are SHRM members has risen substantially since 1950 (see figure 11), those states that had larger increases in SHRM membership after 1998 (the earliest available year for state-level data) did not experience larger declines in labor market fluidity (figure 12). In sum, it seems unlikely that changes in regulatory practices that affect the housing and labor markets have been the primary driver of the secular decline in labor market fluidity.

[FIGURE 11 OMITTED]