"They're closing down the textile mill across the railroad tracks. Foreman says these jobs are going boys and they ain't coming back ..." Bruce Springsteen, My Hometown

I. INTRODUCTION

Each of the last three U.S. recessions has been followed by a period of declining employment during the early stages of recovery. While policymakers and the popular press have focused significant attention on jobless recoveries, economists have struggled to explain their occurrence. Within the academic literature competing explanations for jobless recoveries have emerged. On the one hand there are explanations that emphasize a less dynamic economy with slower growth, reduced labor market fluidity, a decline in startup activity, and even economic stagnation; while other studies emphasize the importance of dynamic structural change including offshoring and technological advances that replace middle skill (routine) labor. Determining the extent to which these various explanations contribute to jobless recoveries has been hampered by the small sample size of national business cycles. To overcome this obstacle our study utilizes U.S. state-level data to assess the degree that reduced economic dynamism and/or important structural changes in the economy contribute to the recent phenomenon of jobless recoveries.

Inspired by the finding of Mendez, Reber, and Schwartz (2016) that jobless recoveries are not a national phenomenon, we exploit variation in the propensity and duration of jobless recoveries across U.S. states and over time. The combination of cross-sectional and time series evidence allows us to evaluate the explanatory power of a number of competing hypotheses, and we find empirical evidence to support both the stagnation and structural change theories of jobless recoveries. On the stagnation side, we find that the rate of new business formation is a significant predictor of jobless recoveries. Even after controlling for the national downward trend in startup activity, we find that states with a lower proportion of startup firms are significantly more likely to experience a jobless recovery.

We also find evidence that links dynamic structural change to the jobless recovery phenomenon. Contrary to the stagnation theories, we find that states with higher trend productivity growth are more likely to experience jobless recoveries. Further, our most robust result suggests a link between jobless recoveries and labor market polarization (as defined by Autor and Dorn 2013). More precisely, we find that states experiencing a sharper trend decline in routine employment share are more likely to experience a jobless recovery. Our results therefore suggest that structural change in labor demand is causing certain types of jobs to be disproportionately slashed during recessions, and to the extent that these jobs are being automated, firms have less need to rehire during the expansion thus yielding a jobless recovery.

The link between labor market polarization and jobless recoveries is consistent with the national (United States) results of Jaimovich and Siu (2012) and the importance of new business formation supports the hypothesis of Pugsley and Sahin (2014), but our state-level data set allows for important new insights. For example, we find that jobless recoveries spread from a regional to a national phenomenon in a way that mirrors the geographic spread of labor market polarization and declining rates of business formation throughout the country. Each of these factors contributes to the likelihood and length of a jobless recovery and in combination have an economically significant, disruptive impact on cyclical recoveries in employment. Given that routine-replacing technological change is likely to continue to play a significant role in the future, one implication of our results would be to consider policies that reduce impediments to new business formation in the hopes that aggressive hiring by startups could reduce the likelihood and length of future jobless recoveries.

II. PROPOSED EXPLANATIONS FOR JOBLESS RECOVERIES

As an organizational framework we sort the proposed explanations for jobless recoveries into two broad categories: (1) those associated with reduced economic dynamism (stagnation) and (2) those associated with dynamic structural or technological change. (1)

First, we should acknowledge that, within the family of explanations involving a less dynamic economy, there is the school of thought that suggests jobless recoveries are actually a misnomer and would be better described as weak or slow recoveries (Gali, Smets, and Wouters 2012). Proponents of the "slow recoveries" view argue that the dominant characteristic of recent recoveries has been slow output growth and this is the root cause of the weak employment recoveries. One could even link the slow recovery view with concerns about secular stagnation (Summers 2014).

Most studies, however, point to a change in cyclical employment dynamics and seek to explore potential explanations. A prominent example of the potential role of labor market frictions is traditional labor hoarding theory. (2) During a recession firms may hoard labor if they wish to avoid the costs of firing workers during a recession and then rehire/retrain workers during the subsequent recovery. If firms retain excess labor during a recession, this can postpone the need for hiring during a subsequent recovery thereby generating a jobless recovery. Somewhat related to the concept of labor hoarding is the reduction in labor market fluidity documented by Davis and Haltiwanger (2014). Although their study is primarily focused on longer-run trends in labor force participation, the reduction in job creation and job destruction rates reported by Davis and Haltiwanger could also be linked to the phenomenon of jobless recoveries. A less dynamic labor market could be reflected in reduced separations during recessions (consistent with labor hoarding) and a slower pace of hiring during recoveries.

In a related literature, a number of studies have demonstrated the importance of young firms for employment dynamics and document an alarming decline in the rate of new business formation. Haltiwanger, Jarmin, and Miranda (2013) establish that young firms exhibit higher rates of job creation and destruction. Pugsley and Sahin (2014) demonstrate that the trend decline in firm entry has reduced job creation during recessions and recoveries thereby contributing to the observation of jobless recoveries.

In a separate branch of the literature a number of studies have suggested that jobless recoveries are not a symptom of reduced dynamism but rather a side effect of ongoing structural change. Models proposed by Berger (2015), Bachmann (2012), Koenders and Rogerson (2005), and Foster, Grim, and Haltiwanger (2015) all propose that firms have a tendency to grow fat during long expansions and then go through a period of intense restructuring during recessions. If this restructuring period is productivity enhancing it could delay the need for rehiring and therefore contribute to jobless recoveries.

Other models emphasize reallocation across sectors rather than restructuring within firms or industries. Garin, Pries, and Sims (2013) emphasize an increase in the importance of reallocative shocks (relative to aggregate shocks) and Groshen and Potter (2003) find evidence to suggest structural change and reallocation of workers across industries contributes to the phenomenon of jobless recoveries. One potential source for this structural change and sectoral reallocation is the job polarization trend documented by Autor and Dorn (2013). If technological change is causing jobs in the middle of the skill distribution to be disproportionately slashed during recessions and replaced by automation as the economy rebounds, firms have less need to rehire during the ensuing expansion, thus yielding a jobless recovery. Jaimovich and Siu (2012) and Gaggl and Kaufmann (2014) each find evidence that jobless recoveries can be traced to the polarization of U.S. labor markets, but Graetz and Michaels (2016) fail to find evidence of this phenomenon outside the United States.

III. DATA

A. Constructing State Business Cycle Data

One obstacle to analyzing the determinants of jobless recoveries is the small sample of recessions available at the national level. To overcome this limitation we exploit variation in business cycles among U.S. states, as documented by Owyang, Piger, and Wall (2003) and Mendez, Reber, and Schwartz (2016). While pursuing a state-level approach has the potential to help distinguish between competing theories of jobless recoveries, it comes with its own difficulties. First, the Bureau of Economic Analysis (BEA) only maintains an extended time series for state gross domestic product (GDP) at an annual frequency, which is inappropriate for business cycle analysis. (3) Thus we need an interpolation strategy, to create a quarterly series.

To generate an estimate of state-level quarterly GDP we first obtain data on personal income for each state from the BEA. This series is published at a quarterly frequency and is highly correlated with GDP. We deflate personal income using the GDP deflator and seasonally adjust using the X12 Census Bureau Method. Next, we assume that the proportion of GDP that occurs during each quarter follows a similar quarterly pattern to real personal income. It is well known, however, that strictly following this procedure can lead to abrupt changes in the interpolated series at the start of a new year. To ensure smooth transitions from year to year, we use the Denton (1971) methodology, which minimizes the transition from 1 year to the next, subject to the quarterly interpolated GDP series summing to the actual annual series reported by the BEA.

The next obstacle to using state-level data is the lack of a National Bureau of Economic Research (NBER) equivalent source for dating state business cycles, since state cycles may not conform to the national economy's peaks and troughs. We adopt the following procedure for determining the phases of state business cycles (4):

* Recessions are defined as all periods that include two or more consecutive quarters of negative GDP growth. These periods extend to earlier and later quarters, until two consecutive quarters of positive GDP growth is reached.

* Expansions are non-recessionary periods that are at least eight quarters long.

* Jobless recoveries begin when two quarters of negative job growth immediately follow a recession. Jobless recoveries end with a quarter of positive job growth or a recession. (5)

* Job growth recoveries begin immediately following a recession if either of the first two quarters of the expansion have positive job growth. Job growth recoveries end when the economy enters a recession.

* Short recoveries are non-recessionary periods that are shorter than eight quarters. These short expansions represent the intermediate periods between "double-dip" recessions and it is debatable whether they represent recoveries at all. We therefore exclude short recoveries from our primary analysis, although inclusion of these periods does not materially impact our results.

To implement our definitions of jobless and job growth recoveries we obtain state non-farm employment data from the Bureau of Labor Statistics. (6)

Applying these definitions yields 165 state-level recessions, of which slightly more than a third (64) end in jobless recoveries with an average length of 3.3 quarters (see Table A1 for a complete list of the state-level recession-recovery sample). Figure 1 illustrates the importance of using state-level definitions of the business cycle by documenting significant variation across states. The figure presents the percentage of states during each quarter that are in a recession or jobless recovery along with national NBER recessions shaded in gray. Our sample period includes five NBER recessions starting in 1980. (7) The 1982 and 2008 recessions were nationally synchronized with 80%-90% of states in a recession at the same time. By contrast, there was never a majority of states in a recession during the 1990-1991 and 2001 national recessions. Also noteworthy is the fact that during almost every quarter of national NBER expansions there are some states that are in a recession; and some NBER expansions include quarters where 40%-50% of states are in recession. To further quantify the relationship between state and national business cycles, we also examine the correlation between state and national GDP growth. The average correlation between state and national output growth is surprisingly low at 0.40, and ranges from 0.01 and 0.64. Almost a quarter of states have a correlation below 0.30.

It is also apparent from Figure 1 that state-level jobless recoveries are primarily clustered around national business cycle troughs and although they have become more prominent during recent recoveries they do not appear to be a broadly national phenomenon. For example, the recovery from the 1990 to 1991 recession is generally thought of as the first national jobless recovery (and would qualify under our definition based on three quarters of negative job growth), but the data in Figure 1 demonstrate that less than 20% of states experienced jobless recoveries at this time. It appears that during 1991 a larger number of states had recessions that extended beyond the official NBER trough. Some of the national jobless recoveries may therefore be combinations of states truly in jobless recoveries along with other states in the midst of an extended recession (more on the relationship between state and national jobless recoveries in Section V). It is abundantly clear that state business cycles are quite distinct from NBER cycles, and the differences between the phases of state business cycles provides variation that can be exploited to identify the determinants of jobless recoveries.

B. Explanatory Variables

In the empirical work that follows (Section IV) we use probit regressions to analyze the factors associated with an increased probability of a state experiencing a jobless recovery and ordered probits to analyze determinants of the length of a jobless recovery. In each case the sample includes a maximum of 165 state-level recoveries as detailed in Table A1. In order to test the various explanations of jobless recoveries put forward in the academic literature we gather a variety of empirical proxies at the state level. Table 1 organizes the proposed explanations into the broad categories of stagnation (or reduced dynamism) and structural change, as discussed in Section II, and lists our empirical proxies. Table 2 provides descriptive statistics for our key variables. Each variable is measured in the first quarter of a recovery with the exception of state and national GDP where we include both the first and second quarter of a recovery. The table provides statistics for all recoveries, as well as for jobless and job growth recoveries separately.

We use four measures to test the theories that emphasize economic stagnation as a source of jobless recoveries. First, to explore the theory that jobless recoveries are simply a manifestation of weak output growth, we include state GDP growth during the first two quarters of the recovery. Not surprisingly, in Table 2 we observe that state output growth is higher during job growth recoveries relative to jobless recoveries. To control for the potential impact of national conditions on state-level recoveries we also include national real GDP growth during the first two quarters of each state recovery, and find that aggregate conditions are also stronger during job growth recoveries.

Next, to determine whether the decline in labor market fluidity, as documented by Davis and Haltiwanger (2014), has led to jobless recoveries, we obtain the annual job reallocation rate by state from the Census Bureau's Business Dynamic Statistics database. The job reallocation rate is the sum of the job creation rate (the ratio of job creation to total employment) and the job destruction rate (the ratio of job destruction to total employment). Given that the measure is only available on an annual basis we develop a quarterly series, by assuming annual figures correspond to the first quarter of each year, and interpolate values linearly for quarters two, three, and four. To focus on longer term trends in job reallocation we follow Davis and Haltiwanger (2014) and construct a 12 quarter moving average. In our sample the average rate of job reallocation is about 30% and is lower for states experiencing jobless recoveries.

Pugsley and Sahin (2014) emphasize the decline in new business formation as a source of weak employment growth, so our final measure of stagnation is the proportion of firms that are start-ups which is also available from the Census Bureau's Business Dynamic Statistics Database. Since these data are also only available on an annual basis, we again utilize the same linear interpolation strategy to create a quarterly series. In our sample there has been a significant downward trend in the national startup rate but there is also significant heterogeneity among states in the level of startup activity and the extent and timing of its decline. For our empirical analysis we separately include the national and state-level startup rate in order to evaluate the extent to which variation in new business formation across states can help explain jobless recoveries. In our sample we observe a 1.8 percentage point higher startup rate in states experiencing job growth recoveries, which provides some initial evidence that fewer startups may be related to jobless recoveries.

Turning to theories emphasizing dynamic structural change, we would like to capture the extent of technological change, but this is a challenging task--especially at the state level. As a proxy, we estimate state-level productivity by dividing our interpolated GDP series by state employment. We are interested in structural, long-run movements in the technological environment of states, so we use the Hodrick Prescott (HP) filtered trend of productivity growth and create an annual growth rate by summing the current and last three quarters of the resulting HP filtered series. (8) We find that productivity growth is on average higher for states experiencing a jobless recovery which matches international evidence from Burger and Schwartz (2015) and provides some preliminary support for the structural change branch of the literature. (9)

A number of the explanations involving structural change suggest firms grow fat during long expansions and then go through a period of intense restructuring during recessions. To test these hypotheses we simply measure the length, in number of quarters, of the state expansion that precedes each state recession. Prior expansions average almost 7 years leading up to jobless recoveries and just three and a half years for job growth recoveries, suggesting there may be some merit to these theories.

Perhaps the two most prominent examples of structural change in the U.S. economy over the past few decades have been globalization/offshoring and routine-replacing technological change leading to a polarized labor market. Measures of vulnerability to offshoring have been developed by Firpo, Fortin, and Lemeiux (2011), who start by taking the average of the Face-to-Face Contact and OnSite Job variables from the U.S. Department of Labor's Occupational Information Network. These variables measure the degree an occupation requires one to work in close contact with others and be in a particular location. The sign of the mean of these variables is reversed, such that an increase in the index corresponds to greater offshorability, and the index is normalized to have a mean of zero.

To construct a measure of the potential of a state's workforce to be offshored we weight the offshorability index for each occupation by the proportion of a state's employment in each of the Census Bureau's Major Occupation categories, which we obtain from the Center of Economic Policy Research's (CEPR) outgoing rotation files for the Current Population Survey (CPS). (10) To construct quarterly measures we utilize all monthly files within the quarter. Similar to our measure of productivity growth, we HP filter the offshorability series to obtain the structural, or trend component, and then to eliminate any remaining noise or seasonality in the series related to small CPS samples we use a four quarter average. In our sample, offshorability is significantly higher for states experiencing a jobless versus job growth recovery.

Finally, we turn to the phenomenon of job polarization. Autor and Dorn (2013) document a striking decline in middle-skill jobs with growth at the high and low end of the skill spectrum. The leading explanation for polarization is automation of routine tasks. Jaimovich and Siu (2012) find evidence from national U.S. data that declines in routine employment are concentrated in recessions and contribute to jobless recoveries. Similar to our measure of offshorability we construct a measure of routine employment share using the CPS's outgoing rotation files. As in Jaimovich and Siu (2012) we use the following occupations for routine employment: (1) Production and Craft, (2) Transportation, Construction, Mechanic, Mining, and Farm, and (3) Machine Operators and Assemblers. For each state we define the routine employment share by summing the proportion of employment in the CPS across these three categories. To develop a proxy for structural change, as opposed to more transitory movements, we first HP filter the series, and then utilize the year over year change in the proportion of employment engaged in routine tasks in our analysis.

States which are experiencing a long-run downward trend in routine employment share are interpreted as experiencing polarization of the labor market. The routine employment share has been trending down nationally for several decades, but there is significant state by state variation in the pace and timing of polarization. In Figure 2 we see that Pennsylvania begins our sample period with a high share of routine employment but has seen it fall continuously for over 30 years. By contrast Texas experienced a notable rebound in routine employment during the 1990s and California began the period with far less routine employment than the other two states, and subsequently experienced a less dramatic decline during the sample period. The summary statistics in Table 2 reveal that on average routine employment share is declining more severely in states experiencing jobless recoveries, suggesting a potentially important role for routine-replacing technological change.

IV. EMPIRICAL RESULTS

Our empirical analysis is focused on evaluating the factors associated with an increased likelihood that a state will experience a jobless recovery. We seek to distinguish between the candidate causes of jobless recoveries using the empirical proxies listed in Table 1 and documented in Section III.

We conduct probit analysis of the form:

Pr([Y.sub.i] = 1|[X.sub.i]) = [PHI]([X.sub.i][beta])

where [Y.sub.i] is a dummy variable indicating whether the state recovery period, i, immediately following a state recession is jobless (first two quarters following a recession have non-positive employment growth) and [X.sub.i] is a vector of our independent variables measured contemporaneously with the first quarter of the recovery or in the case of state and national GDP the first two quarters of a recovery. (11)

Although our Sample includes variation across states and time, it is important to note that it is not a traditional panel because the units of analysis are state-level economic recoveries (as defined above). Including state-level fixed effects is possible but would significantly reduce the sample due to a handful of states having just one recovery and other states lacking time variation in the state-level experience with jobless recoveries (some states had no jobless recoveries and for others every recovery was jobless). We therefore choose not to include fixed effects in our estimated probit models.

Results from the probit analysis of the probability of a jobless recovery are presented in Table 3. To ease interpretation, we present the marginal effects for each variable at the means of our data set. Throughout our analysis standard errors are clustered at the state level and we exclude short expansions (less than eight quarters) as previously defined.

In column (1) of Table 3 we find support for a number of the proposed explanations for jobless recoveries. First, the negative and highly statistically significant coefficient on state GDP growth during the second quarter of a recovery indicates that "weak recoveries" at the state level are in fact more likely to be jobless recoveries. (12) The marginal effects reported in Table 3 indicate that state GDP growth that is 1 percentage point above average would decrease the probability of a jobless recovery by 7 percentage points. National conditions also matter as we estimate a highly significant negative coefficient on national GDP growth during the first quarter of a state-level recovery.

Additional evidence consistent with the reduced dynamism branch of the literature is provided by the negative and statistically significant coefficient on the state-level startup rate. Even after controlling for the national downward trend in new business formation we find that a state with a startup rate 1 percentage point lower than the national average increases its likelihood of a jobless recovery by approximately 14 percentage points.

To evaluate theories emphasizing dynamic structural change, we include the trend growth rate in labor productivity and find that states experiencing productivity growth 1 percentage point greater than the sample average are 22 percentage points more likely to experience a jobless recovery. This positive link between higher productivity growth and jobless recoveries runs counter to the stagnation theories and provides evidence in favor of structural change. Finally, we fail to find evidence that jobless recoveries are more likely following long expansions, which runs counter to theories of firm-level restructuring during recessions. In sum, the evidence from column (1) of Table 3 is mixed with some evidence supporting links from economic stagnation to jobless recoveries, but other results consistent with theories of dynamic structural change as a driver of jobless recoveries.

In column (2) of Table 3 we take the analysis a step further and include more specific proxies for the leading candidates of structural change. To evaluate the extent to which labor market polarization might be responsible for jobless recoveries we include the annual trend growth rate in routine employment share. As a proxy for the impact of globalization we include an index which measures the extent to which jobs in a particular state are offshorable.

The results in column (2) are similar to column (1) but we also estimate a negative and highly significant coefficient on trend growth in routine employment share which suggests routine-replacing technological change has contributed to the jobless recovery phenomenon. To quantify the economic significance of our results we use the estimates reported in column (2) of Table 3 to calculate the probability of a jobless recovery occurring at the mean for all explanatory variables, and if the growth rate in routine employment share were to fall by one standard deviation. Our calculations indicate that such a decline in routine employment growth would increase the probability of a jobless recovery from 27.5% to 44.4%. Similarly we find that a state with a startup rate one standard deviation below the sample average would see its probability of a jobless recovery increase to 55%. Our results clearly indicate that routine-replacing technological change and diminished startup activity have exerted an economically significant impact on the likelihood of jobless recoveries.

While our measures of productivity growth and polarization are each calculated as the annual growth rate of an HP filtered series with the intent of capturing long-run trends, one might still be concerned that the proxies are influenced by contemporaneous business cycle effects. As a robustness check we therefore include 1 year lags of the productivity and polarization variables for the specification reported in column (3). The results are quite similar suggesting that the link between productivity and jobless recoveries is not simply a mechanical relationship, but rather reflects the fact that higher long-run growth rates of productivity are linked to jobless recoveries. We also find that the impact of trend growth in routine employment share on the likelihood of a jobless recovery continues to be highly significant in a lagged specification.

To ensure our results are not driven by the limited sample sizes in the CPS for smaller states, in column (4) of Table 3 we estimate the same specification, but drop the five smallest states based on the number of CPS respondents, and we find very similar results. (13) As a further robustness check, in column (5) we exclude the top five energy-producing states from the analysis which, due to their sensitivity to world energy markets, may behave differently than other states. Again we find evidence that routine-replacing technological change and a decline in startup activity are robust predictors of jobless recoveries.

In Table 4 we present results from an ordered probit where the dependent variable is the number of quarters with negative job growth following each recession (which ranges from 0 to 10 quarters). (14) The results are remarkably similar to those from the binary probit reported in Table 3. We continue to find robust support for explanatory variables related to polarization and startups. In addition, the ordered probit reveals some support for global factors as we find states with more offshorable employment have significantly longer jobless recoveries. Productivity growth is also a significant factor in determining the length of a jobless recovery, but at a somewhat reduced significance level in comparison to the binary probit results in Table 3.

Figure 3 uses the ordered probit results from column (2) of Table 4 to demonstrate the impact of declines in routine employment share and startup activity, the two most robust and highly significant results, on the probability of various lengths of jobless recoveries. In Figure 3 the thick solid line presents the actual proportion of recoveries that were jobless for X quarters, while the dashed lines display the impact of a one standard deviation increase in either the growth rate of routine employment share or the startup rate on the probability of jobless recoveries of various lengths. Figure 3 reveals that a moderation in labor market polarization or more robust startup activity would significantly increase the probability of job growth recoveries (0 or 1 jobless quarters), and almost eliminate the possibility of a state experiencing a jobless recovery lasting four or more quarters. We conclude that routine-replacing technological change and the reduction in new business formation are important determinants of both the likelihood and length of jobless recoveries.

V. EXPLAINING NATIONAL JOBLESS RECOVERIES

Our probit and ordered probit analysis reveals two factors that appear to be the primary drivers of state-level jobless recoveries: the decline in new firm creation and routine-replacing technological change. In addition to being our most robust and highly significant findings, these factors also appear to be the most promising explanations of jobless recoveries in the existing literature. As a result, in this section we focus exclusively on how these phenomena may help explain national jobless recoveries.

Figure 4 compares national recoveries following the recessions of the early 1990s, 2000s, and the Great Recession of 2008-2009. The first column of Figure 4 classifies each state based on its business cycle phase two quarters after the end of the last three national recessions, the middle column displays the trend in the share of routine employment, and the last column indicates the share of new firms in each state (the same proxies we utilize in our regressions). Darker shades indicate sharper declines in the proportion of routine employment or lower startup rates, respectively.

Although the recovery in 1991 is generally thought of as the first national jobless recovery, the first column of Figure 4 reveals that jobless recoveries were concentrated in just a handful of eastern states. The perception of a national jobless recovery appears driven in part by California and the Northeast remaining in recession after the official NBER trough. To further elaborate on this point Figure 5 decomposes national job losses into employment declines coming from states that were in recession and those that were in jobless recoveries. Job losses coming from states in jobless recoveries are further decomposed into those coming from states with higher and lower than sample median growth in routine employment share and higher and lower than the mean startup rate. The figure shows that during quarter 3 of 1991 an overwhelming majority of job losses came from states that were still in recession. Those job losses were more than offset by the creation of 189,000 jobs in states that were experiencing job growth recoveries. During this same period, labor market polarization had yet to become a national phenomenon as routine employment share was growing or declining only moderately in much of the country (middle column of Figure 4). Stronger evidence of routine-replacing technological change was evident primarily in the Northeast and did appear to contribute to the handful of states experiencing jobless recoveries. The decomposition in Figure 5 indicates that employment declined by 49,500 jobs in the jobless recoveries of states with sharper drops in routine employment and lower than median startup activity.

The observation of a national jobless recovery following the 2001 recession appears driven primarily by the Midwest where six states were experiencing jobless recoveries and two others were still in recession. It is also apparent from Figure 4 that by 2001 declines in routine employment share had spread from the Northeast to the Midwest (and beyond). The decomposition in Figure 5 indicates that job losses associated with jobless recoveries were concentrated in states with sharper declines in routine employment share.

Finally, the jobless recovery that occurred in the aftermath of the Great Recession of 2008-2009 was truly a national phenomenon. In addition, the right two panels of Figure 4 make it clear that sharp declines in routine employment share and low startup rates had also become widespread. The decomposition in Figure 5 indicates that during quarter 4 of 2009 job losses were concentrated in states experiencing jobless recoveries. In fact all 238,000 of the job losses associated with state-level jobless recoveries occurred in states experiencing sharp declines in routine employment and lower than average startup rates.

Our investigation of the link between state and national jobless recoveries yields a number of insights. First, we find that the perception of a jobless recovery in the early 1990s was driven primarily by a number of highly populous states remaining in recessions after the NBER trough. Second, we find that sharp declines in routine employment share started in the Northeast and then spread to the Midwest and beyond. As the polarization phenomenon spread, so too did the observation of state-level jobless recoveries. Finally, we find that during the two most recent national jobless recoveries job losses were concentrated in states with sharper declines in routine employment share and lower startup rates.

VI. CONCLUDING THOUGHTS

We find empirical evidence to support both the stagnation and structural change theories of jobless recoveries. On the stagnation side, we find evidence that the rate of new business formation is a significant predictor of jobless recoveries. Even after controlling for the national downward trend in startup activity, we find that states with a lower startup share are significantly more likely to experience a jobless recovery. However, we also find evidence that links dynamic structural change to the jobless recovery phenomenon. Contrary to the stagnation theories, we find that states with higher trend productivity growth are more likely to experience jobless recoveries. Drilling deeper we find evidence linking routine-replacing technological change to the phenomenon of jobless recoveries: States experiencing a sharper trend decline in routine employment share are more likely to experience a jobless recovery. Our results linking labor market polarization to jobless recoveries are consistent with the national (United States) results of Jaimovich and Siu (2012), but our state-level data set allows for important new insights by leveraging the heterogeneity of state experiences with jobless recoveries and polarization (see Figures 1 and 4).

Our investigation of state-level conditions also reveals significant heterogeneity in the national experience with jobless recoveries over the past three business cycles. At the national level, the recovery from the 1990 to 1991 recession began with three quarters of negative job growth, which first inspired the term "jobless recovery." Our disaggregated analysis reveals that the perception of a national jobless recovery at this time was driven primarily by California and several populous northeastern states remaining in recession after the NBER trough. State-level jobless recoveries were more prevalent following the 2001 recession and became truly widespread following the Great Recession of 2008-2009. We also find that the spread of jobless recoveries from a regional to a national phenomenon mirrors the geographic spread of labor market polarization and declining rates of business formation throughout the country.

Our analysis of the variety of state-level experiences with jobless recoveries suggests that the onset of routine-replacing technological change, along with a decline in new business formation, have both generated a significant impact on labor markets during the early stages of cyclical recoveries. Each of these factors contributes to the likelihood and length of a jobless recovery, and in combination have an economically significant, disruptive impact on cyclical recoveries in employment. Given that routine-replacing technological change is forecast to continue to play a significant role in the future, one implication of our results would be to consider policies that reduce impediments to new business formation in the hopes that aggressive hiring by startups could reduce the likelihood and length of future jobless recoveries.

APPENDIX

TABLE A1
Jobless and Job Growth Recoveries by State

                                Jobless     Job Growth
State             Recoveries   Recoveries   Recoveries

Alaska                3            0            3
Alabama               3            0            3
Arkansas              3            2            1
Arizona               3            0            3
California            3            2            1
Colorado              2            2            0
Connecticut           2            1            1
Delaware              5            2            3
Florida               1            0            1
Georgia               2            1            1
Hawaii                2            1            1
Iowa                  4            2            2
Idaho                 4            1            3
Illinois              4            2            2
Indiana               4            2            2
Kansas                5            1            4
Kentucky              4            2            2
Louisiana             3            1            2
Massachusetts         4            3            1
Maryland              3            3            0
Maine                 2            1            1
Michigan              5            1            4
Minnesota             4            2            2
Missouri              5            0            5
Mississippi           5            2            3
Montana               4            1            3
North Carolina        3            1            2
North Dakota          1            0            1
Nebraska              5            1            4
New Hampshire         2            1            1
New Jersey            2            1            1
New Mexico            5            1            4
Nevada                1            0            1
New York              3            2            1
Ohio                  4            2            2
Oklahoma              2            1            1
Oregon                4            2            2
Pennsylvania          4            3            1
Rhode Island          3            1            2
South Carolina        3            2            1
South Dakota          6            1            5
Tennessee             4            3            1
Texas                 4            1            3
Utah                  3            1            2
Virginia              1            1            0
Vermont               3            0            3
Washington            5            2            3
Wisconsin             4            2            2
West Virginia         3            0            3
Wyoming               1            0            1
Total                165           64          101

ABBREVIATIONS

BEA: Bureau of Economic Analysis

CEPR: Center of Economic Policy Research

CPS: Current Population Survey

GDP: Gross Domestic Product

HP: Hodrick Prescott

NBER: National Bureau of Economic Research

doi: 10.1111/ecin.12535

Online Early publication December 4, 2017

REFERENCES

Autor, D., and D. Dorn. "The Growth of Low-Skill Service Jobs and the Polarization of the US Labor Market." American Economic Review, 103(5), 2013, 1553-97.

Bachmann, R. "Understanding the Jobless Recoveries after 1991 and 2001." Working Paper, 2012.

Berger, D. "Contercyclical Restructuring and Jobless Recoveries." Working Paper. 2015.

Biddle. J. "The Cyclical Behavior of Labor Productivity and the Emergency of the Labor Hoarding Concept." Journal of Economic Perspectives, 28(2), 2014, 197-212.

Bry, G., and C. Boschan. "Programmed Selection of Cyclical Turning Points," in Cyclical Analysis of Time Series: Selected Procedures and Computer Programs, edited by G. Bry and C. Boschan. Cambridge, MA: NBER, 1971, 7-63.

Burger, J. D., and J. S. Schwartz. "Productive Recessions and Jobless Recoveries." Contemporary Economic Policy, 33(4), 2015,636-48.

Davis, S. J., and J. Haltiwanger. "Labor Market Fluidity and Economic Performance." NBER Working Paper No. 20479, 2014.

Denton, F. T. "Adjustment of Monthly or Quarterly Series to Annual Totals: An Approach Based on Quadratic Minimization." Journal of the American Statistical Association, 66(333), 1971,99-102.

Firpo, S., N. M. Fortin, and T. Lemeiux. "Occupational Tasks and Changes in the Wage Structure." University of British Columbia Working Paper, 2011.

Foster, L., C. Grim, and J. Haltiwanger. "Reallocation in the Great Recession: Cleansing or Not?" Journal of Labor Economics, 34(S1), 2015, S293-331.

Gaggl, P., and S. Kaufmann. "The Cyclical Component of Labor Market Polarization and Jobless Recoveries in the US." Study Center Gerzensee Working Paper No. 14.03, 2014.

Gali, J., F. Smets, and R. Wouters. "Slow Recoveries: A Structural Interpretation." Journal of Money, Credit and Banking, 44, 2012,9-30.

Garin, J., M. Pries, and E. Sims. "Reallocation and the Changing Nature of Economic Fluctuations." University of Notre Dame Working Paper, 2013.

Graetz, G., and G. Michaels. "Is Modern Technology Responsible for Jobless Recoveries?" Working Paper presented at the AEA 2017 Annual Meetings, 2016.

Groshen, E. L., and S. Potter. "Has Structural Change Contributed to a Jobless Recovery?" Current Issues in Economics and Finance, 9(8), 2003, I -7.

Haltiwanger, J., R. Jarmin, and J. Miranda. "Who Creates Jobs? Small versus Large versus Young." Review of Economics and Statistics, 95(2), 2013, 347-61.

Harding, D., and A. Pagan. "Dissecting the Cycle: A Methodological Investigation." Journal of Monetary Economics, 49(2), 2002, 365-81.

Jaimovich, N., and H. E. Siu. "Job Polarization and Jobless Recoveries." NBER Working Paper No. 18334, 2012

Koenders, K., and R. Rogerson. "Organizational Dynamics over the Business Cycle: A View on Jobless Recoveries." Federal Reserve Bank of St. Louis Review, 87(4), 2005, 555-79.

Mendez, F., J. Reber, and J. Schwartz. "A New Approach to the Study of Jobless Recoveries." Southern Economic Journal, 83(2), 2016, 573-89.

Okun, A."Potential GNP: Its Measure and Significance." Cowles Foundation Paper 190 (Reprinted from the 1962 Proceedings of the Business and Economics Statistics Section of the American Statistical Association), 1963.

Owyang, M., J. Piger, and H. Wall. "Business Cycle Phases in U.S. States." Review of Economics and Statistics, 87(4), 2003,604-16.

Pugsley, B., and A. Sahin. "Grown-Up Business Cycles." Federal Reserve Bank of New York Staff Report No. 707, 2014.

Solow, R. "Draft of Presidential Address on the Short Run Relation of Employment and Output." Robert Solow papers, David M. Rubenstein Rare Book & Manuscript Library, Duke University. Box 70, Folder 1964, 1964.

Summers, L. "US Economic Prospects: Secular Stagnation, Hysteresis, and the Zero Lower Bound." Business Economics, 49(2), 2014, 65-73.

(1.) Although we readily admit that not every proposed theory in the literature falls neatly into one of these two broad categories. Furthermore, it may be that structural changes in the economy are causing the reduced dynamism observed in the U.S. economy. This link, however, has yet to be established in the literature and we leave establishing such causality to future research.

(2.) Biddle (2014) traces the labor hoarding concept back to Okun (1963) and the "labor hoarding" label to Solow (1964).

(3.) State GDP data at a quarterly frequency have recently become available from the BEA, but the series only goes back to 2005.

(4.) As a robustness check we also utilize the Bry and Boschan (1971) business cycle dating algorithm, which has been extended to quarterly data by Harding and Pagan (2002). Our results are qualitatively similar when using this alternative method for dating recessions.

(5.) In an earlier version of the paper we obtain similar results using a definition of jobless recoveries based on a single quarter of negative job growth, but upon further reflection we decided that a more restrictive definition requiring multiple quarters of negative job growth better matches the popular notion of jobless recoveries. In fact, at the national level the jobless recoveries of 1991, 2001, and 2009 each began with three or more quarters of negative job growth. We also consider alternatives to negative employment growth including employment growth lower than population growth and less than trend employment growth. These alternative definitions yield many more jobless recoveries and may also be influenced by demographic factors, consequently, we only report results defining jobless recoveries using negative employment growth.

(6.) Since the Bureau of Labor Statistics seasonally adjusts this series on a monthly frequency, we obtain quarterly non-seasonally adjusted data for this series and then perform a seasonal adjustment using the X12 Census Bureau program.

(7.) We are only able to construct time consistent routine employment and offshoring variables back to 1980. As a result we explore a sample from 1980 to 2012.

(8.) Throughout our work we use the standard smoothing parameter of 1,600 when HP filtering data.

(9.) Within a particular quarter one would expect higher productivity growth during a jobless recovery, essentially by definition. But we find the long-run trend in productivity growth, which does not have a mechanical relationship with job growth, is higher for states experiencing a jobless recovery.

(10.) Proportions in each occupation are calculated using CPS weights.

(11.) As a robustness check, in some specifications we lag the explanatory variables.

(12.) The significance of second quarter state GDP growth is logical given our requirement of two quarters of nonpositive job growth to qualify as a jobless recovery.

(13.) As additional robustness checks we estimated one probit where observations were weighted by the square root of the number of CPS respondents and another where each state received equal weight regardless of the number of recessions they experience and in each case we found qualitatively similar results.

(14.) Note that one quarter of jobless growth would not count as a jobless recovery based on our definition, but for the ordered probit analysis we are analyzing the number of jobless quarters in a continuous fashion. Approximately 95% of the state-level recoveries have between zero and four jobless quarters.

JOHN D. BURGER and JEREMY S. SCHWARTZ*

* The authors would like to thank Zhen Cui, James DeNicco, John Duca, Amy Guisinger, Fabio Mendez, Michael Owyang, Irina Panovska, and two anonymous referees for helpful comments. Schwartz acknowledges a summer research grant from the Sellinger School of Business.

Burger: Professor and Research Impact Fellow, Sellinger School of Business, Loyola University Maryland, Baltimore, MD 21210, Phone (410) 617-5831, Fax (410) 6172118, E-mail [email protected]

Schwartz: Associate Professor, Sellinger School of Business, Loyola University Maryland, Baltimore, MD 21210, Phone (410) 617-2919, Fax (410) 617-2118, E-mail [email protected]

Caption: FIGURE 1 Percentage of States in Recession or Jobless Recovery

Caption: FIGURE 2 Routine Employment Share for Select States

Caption: FIGURE 3 Probability of Jobless Recoveries by Length in Quarters

Caption: FIGURE 4 Jobless Recoveries, Change in Routine Employment and Firm Startup Rates by State

Caption: FIGURE 5 Decomposition of National Job Losses during Last Three Recoveries

TABLE 1
Proposed Explanations for Jobless
Recoveries and Their Empirical Proxies

             Proposed Explanation       Empirical Proxy

Stagnation   1. Weak recoveries         State GDP growth

             2. Labor market            Job reallocation rate
                frictions, adjustment
                costs (hoarding)

             3. Reduced labor           Ratio of startups
                fluidity                to total firms

             4. Decline in
                startups

Dynamic      1. Firm-level              Length of previous
structural      restructuring           expansion, trend
                                        productivity growth

change       2. Labor market            Trend growth in routine
                polarization            employment share

             3. Globalization/          Index of employment
                offshoring              vulnerability to
                                        offshoring

TABLE 2
Summary Statistics

                                           All Recoveries

                                              (N = 155)

                                            Mean      SD

Length of jobless recovery
Stagnation variables
  State GDP growth Q1 ([double dagger])    0.031    0.026
  State GDP growth Q2                      0.035    0.030
  National GDP growth Q1                   0.020    0.028
  National GDP growth Q2                   0.030    0.024
  Job reallocation rate                    0.290    0.035
  State startup firm share                 0.096    0.019
  National startup firm share              0.105    0.015
Structural change variables
  Growth in routine employment             -0.004   0.003
  Offshorability                           0.038    0.063
  Productivity growth                      0.012    0.010
  Length of prior expansion                19.667   20.564

                                              Jobless
                                             Recoveries

                                              (N = 64)

                                            Mean      SD

Length of jobless recovery                 3.297    1.411
Stagnation variables
  State GDP growth Q1 ([double dagger])    0.028    0.024
  State GDP growth Q2                      0.027    0.020
  National GDP growth Q1                   0.005    0.023
  National GDP growth Q2                   0.023    0.022
  Job reallocation rate                    0.276    0.027
  State startup firm share                 0.085    0.016
  National startup firm share              0.094    0.015
Structural change variables
  Growth in routine employment             -0.005   0.002
  Offshorability                           0.024    0.057
  Productivity growth                      0.016    0.008
  Length of prior expansion                26.375   25.537

                                            Job Growth
                                            Recoveries

                                             (N = 101)

                                            Mean      SD

Length of jobless recovery
Stagnation variables
  State GDP growth Q1 ([double dagger])    0.032    0.027
  State GDP growth Q2                      0.040    0.033
  National GDP growth Q1                   0.029    0.026
  National GDP growth Q2                   0.034    0.025
  Job reallocation rate                    0.298    0.038
  State startup firm share                 0.103    0.017
  National startup firm share              0.111    0.012
Structural change variables
  Growth in routine employment             -0.004   0.003
  Offshorability                           0.046    0.065
  Productivity growth                      0.009    0.010
  Length of prior expansion                15.416   15.350

Note: ([double dagger]) Means for jobless and job growth
recoveries subsamples are not significantly different at
the 10% level. For all remaining variables we reject
equality of the means at the 5% level.

TABLE 3 Jobless Recovery Probit--Marginal Effects

                               Full         Full         Lagged
                            Sample (1)   Sample (2)   Variables (3)

Length of prior expansion     0.001        0.001          0.000
                             (0.002)      (0.003)        (0.003)
Trend productivity growth   0.224 ***    0.198 ***
                             (0.066)      (0.073)
State GDP growth (Q1)         -0.007       0.002          0.004
                             (0.018)      (0.018)        (0.020)
State GDP growth (Q2)       -0.071***    -0.075 ***    -0.076 ***
                             (0.023)      (0.021)        (0.021)
National GDP growth (Q1)    -0.084 ***   -0.083 ***    -0.089 ***
                             (0.020)      (0.019)        (0.020)
National GDP growth (Q2)      0.021        0.021          0.013
                             (0.020)      (0.020)        (0.020)
State startup rate          -0.137 ***   -0.130 ***    -0.143 ***
                             (0.048)      (0.045)        (0.052)
National startup rate        -0.085 *    -0.121 **      -0.121 **
                             (0.048)      (0.052)        (0.053)
Job reallocation rate        0.037 *       0.026          0.026
                             (0.019)      (0.019)        (0.021)
Offshorability                            0.015 *       0.018 **
                                          (0.009)        (0.009)
Trend routine share                      -0.539 ***
                                          (0.207)
Lagged routine share                                   -0.659 ***
                                                         (0.184)
Lagged productivity                                     0.212 ***
                                                         (0.055)
Number of observations         167          165            163

                            Drop Small   Drop Energy
                            States (4)   States (5)

Length of prior expansion     -0.000        0.000
                             (0.003)       (0.003)
Trend productivity growth

State GDP growth (Q1)         -0.001        0.005
                             (0.020)       (0.024)
State GDP growth (Q2)       -0.094 ***   -0.077 ***
                             (0.027)       (0.022)
National GDP growth (Q1)    -0.107 ***   -0.084 ***
                             (0.022)       (0.020)
National GDP growth (Q2)      0.007         0.013
                             (0.023)       (0.022)
State startup rate          -0.142 **    -0.155 ***
                             (0.064)       (0.052)
National startup rate       -0.165 ***    -0.102 *
                             (0.056)       (0.056)
Job reallocation rate         0.028         0.031
                             (0.029)       (0.022)
Offshorability                0.022         0.014
                             (0.014)       (0.010)
Trend routine share

Lagged routine share        -0.579 ***   -0.658 ***
                             (0.188)       (0.188)
Lagged productivity         0.285 ***     0.212 ***
                             (0.049)       (0.056)
Number of observations         144           149

Notes: The dependent variable for each probit regression
is an indicator variable for jobless recoveries. Constants
are included but not reported. For Length of prior expansions
marginal effects are presented as the effect of the one quarter
change in the length of the prior expansion on the probability
of a jobless recovery at the mean of the all variables.
The remaining variables are presented as the marginal effect
of a percentage point (0.01) change on the probability of a
jobless recovery at the mean of all variables. Standard errors
are clustered at the state level and reported in parentheses.
Column (3) drops observations from the five smallest states
based on number of CPS respondents and column (4) drops the
five highest energy-producing states based upon 2014 data
from the U.S. Energy Information Agency.

***, **, and * denote statistical significance
at the 1%, 5%, and 10% levels, respectively.

TABLE 4
Jobless Recovery Length Ordered Probit

                                Full           Full
                             Sample (1)     Sample (2)

Length of prior expansion      0.006 *       0.007 *
                               (0.003)       (0.004)
Trend productivity growth    39.222 ***      24.885 *
                              (12.706)       (14.551)
State GDP growth (Q1)          -4.778         -3.934
                               (3.909)       (3.987)
State GDP growth (Q2)         -6.720 **     -7.353 **
                               (3.158)       (2.865)
National GDP growth (Q1)     -24.679 ***   -23.515 ***
                               (5.518)       (4.959)
National GDP growth (Q2)        0.115         -1.339
                               (4.174)       (4.195)
State startup rate           -35.035 ***   -32.866 ***
                               (9.779)       (10.178)
National startup rate          -10.649      -19.487 **
                               (7.913)       (8.609)
Job reallocation rate        12.627 ***      9.520 **
                               (4.678)       (4.648)
Offshorability                              4.446 ***
                                             (1.664)
Trend routine share                        -121.889 ***
                                             (41.079)
Lagged routine share

Lagged productivity

Number of observations           167           165

                                Lagged        Drop Small
                             Variables (3)    States (4)

Length of prior expansion       0.007 *         0.006
                                (0.004)        (0.004)
Trend productivity growth

State GDP growth (Q1)           -3.106          -2.959
                                (3.907)        (3.338)
State GDP growth (Q2)          -7.576 **     -10.088 ***
                                (3.005)        (3.601)
National GDP growth (Q1)      -23.015 ***    -24.867 ***
                                (5.031)        (6.058)
National GDP growth (Q2)        -2.599          -5.048
                                (4.607)        (4.842)
State startup rate            -36.092 ***    -43.840 ***
                               (10.839)        (13.193)
National startup rate         -21.018 **      -20.038 **
                                (8.787)        (9.818)
Job reallocation rate          10.753 **      16.009 **
                                (5.089)        (6.276)
Offshorability                 4.628 ***       3.114 *
                                (1.662)        (1.821)
Trend routine share

Lagged routine share         -126.354 ***    -115.196 ***
                               (35.471)        (37.763)
Lagged productivity            23.211 *       34.033 **
                               (13.720)        (14.378)
Number of observations            163            144

                             Drop Energy
                              States (5)

Length of prior expansion       0.005
                               (0.004)
Trend productivity growth

State GDP growth (Q1)           -2.461
                               (4.439)
State GDP growth (Q2)         -8.346 ***
                               (3.114)
National GDP growth (Q1)     -22.454 ***
                               (4.856)
National GDP growth (Q2)        -2.912
                               (4.574)
State startup rate           -36.219 ***
                               (11.338)
National startup rate         -18.546 **
                               (9.426)
Job reallocation rate         11.822 **
                               (5.238)
Offshorability                 4.240 **
                               (1.832)
Trend routine share

Lagged routine share         -130.266 ***
                               (35.618)
Lagged productivity            22.516 *
                               (13.562)
Number of observations           149

Notes: The dependent variable for each ordered probit
regression is jobless recovery length, which is a categorical
variable measuring the number of jobless quarters following each
state-level recession. In our sample, jobless recovery length
ranges from 0 (in the case of a job growth recovery) to a maximum
of 10 quarters. Constants are included but not reported. Standard
errors are clustered at the state level and reported in
parentheses. Column (4) drops observations from the live smallest
states based on number of CPS respondents and column (5) drops the
five highest energy-producing states based upon 2014 data from the
U.S. Energy Information Agency.

***, **, and * denote statistical significance at
the 1%, 5%, and 10% levels respectively.