Jobless Recovery and Structural Change: A VAR Approach

Abstract

This study uses a vector autoregression approach to examine the link between jobless recoveries and the fast employment expansion in finance, health and education (FHE) sectors. Both reduced-form estimates and impulse responses indicate a negative effect of the expansion on aggregate employment. While the expansion Granger causes aggregate employment fluctuations, up to 40 per cent of the error variance of those fluctuations can be explained by innovations in the expansion. Moreover, movements in aggregate employment are reduced by 25 per cent when the expansion is accounted for. Therefore, the fast expansion of the FHE sectors is shown to have notably contributed to the onset of jobless recoveries.

JEL Classification: E24, E32, C32

Keywords

Jobless recovery structural change sectoral shift VAR

Introduction

The jobless recovery phenomenon refers to the slow recovery of employment despite a quick rebound of total output after a recession. The USA has experienced three jobless recoveries since 1990. Some studies relate the occurrence to labour market structural changes (Garin, Pries, & Sims, 2013; Groshen & Potter, 2003). However, those studies scant the application of rigorous econometric methods. Also, the existing literature in general has mixed results on accounting the contribution of structural change to jobless recoveries. Therefore, this article adopts an econometric approach to examine, and thus to offer some conclusive evidence on the link between the structural change and jobless recovery.

We document three structural changes in the post-1990 US labour market. First, finance, health and education (FHE) sectors have surpassed manufacturing to have the largest employment share in the private sector. Second, while the business cycle dynamics of FHE employment has stayed relatively stable, that of manufacturing employment reveals the need for laid-off manufacturing workers to find jobs outside their sector after 1990. Last, the disproportionate concentration of college-educated workers in the FHE sectors has widened the education requirement gap for workers across sectors.

Since these changes coincide with the emergence of jobless recoveries, we hypothesize that the fast expansion of the FHE sectors is one of the key triggering factors. In particular, an individual laid off in another sector, especially a declining sector, during a post-1990 recession might lack the qualification for a job in the FHE sectors where new openings are most abundant. This could result in a long period of unemployment from which a jobless recovery ensues.

We then examine our hypothesis using reduced-form estimates as well as recursive vector autoregressions (VARs). The reduced-form estimates suggest that the expansion of the FHE sectors has no significant effect on aggregate employment before 1990, but a statistically significant and negative effect after 1990. The Granger causality test supports this finding. The impulse responses for the recursive VAR show that an unexpected positive shock to FHE employment growth only has a prolonged negative effect on aggregate employment growth after 1990. This shock accounts for up to 40 per cent of the forecast error variance of the aggregate employment growth after 1990, but only 23 per cent before 1990. Compared to the pre-1990 era, including the expansion of the FHE sectors reduces the post-1990 aggregate employment fluctuation by 25 per cent. Thus, our VAR results substantiate the link between jobless recoveries and the expansion of the FHE sectors, indicating a notable contribution of structural change to jobless recoveries.

The remainder of the article is organized as follows. The next section relates this study to the existing literature. The third section describes our data. The fourth section documents recent jobless recoveries and then relates them to the structural change of the labour market. The penultimate section discusses the methodology and the results. The last section concludes the article.

Related Literature

A number of studies have tried to explain jobless recoveries via structural changes. Among them, Groshen and Potter (2003) and Garin et al. (2013) are arguably the most prominent. Unfortunately, neither of the studies has undertaken a rigorous econometric approach.

Groshen and Potter (2003) reported an absence of employment recovery after the 2001 recession. Their structural change is defined as the permanent job relocation from one industry to another. Using payroll data from 70 industries, the study compares each industry’s job growth rate during the recession with that during the recovery phase. The comparison reveals that job gains after 1990 tend not to occur in the same industries where job losses initially take place. The study infers that this structural change leads to the 2001 jobless recovery.¹

Garin et al. (2013) showed a much delayed recovery in unemployment and total hours worked after the past three recessions. They conjecture that the slow recovery is caused by a reallocation shock that raises the productivity of one sector relative to another, but leaves the aggregate productivity intact. A two-island model is developed in which a reallocation shock during a recession motivates workers in the relatively less productive island to move to the relatively more productive island. Before those workers join production on the other island, they have to experience a period of unemployment, thereby generating a jobless recovery.

Aside from the above studies, findings on the contribution of structural change to jobless recoveries are inconclusive. For example, Sahin, Song, Topa and Violante (2014) created a mismatch index using the difference between an actual unemployment rate and a counterfactual unemployment rate indicated by a theoretical social planner’s problem. The index suggests that mismatch across two-digit industries explains up to 23 per cent of the rise in unemployment during the Great Recession. DeNicco and Laincz (2014) showed that unemployment rates had a structural break in the fourth quarter of 1984. Including the changing industry composition of the labour force amplifies this structural break by more than 13 per cent. However, Aaronson et al. (2004) constructed a measure of sectoral reallocation by decomposing the growth rate of industry employment share into three components (i.e., long-term trend, cyclical pattern and idiosyncratic movement). Their measure suggests a much lower reallocation level for both the 1990 and the 2001 recessions, relative to the previous recessions.

Against this background, the present study adopts an econometric approach to investigate the link between structural changes and jobless recoveries. It also tries to quantify the contribution of structural change to the jobless recovery, thereby offering a resolution to the ongoing debate.

Although this article considers structural change as a promising theory, it does not claim that it is ‘the’ cause of jobless recoveries, especially given the existence of many alternative explanations. For example, Koenders and Rogerson (2005) qualitatively generated a jobless recovery using a model in which organizations wait for the recessions to eliminate the excess labour they have hoarded during their long expansions.² Bachmann (2012) attributed jobless recoveries to the trade-off between the intensive (i.e., hours worked) and extensive margins (i.e., number of workers) of firms. We do not compare the structural change explanation with these. Instead, we try to offer some new empirical evidence in favour of the structural change theory.

Data Description

Our analysis uses the following macroeconomic data series. First, we approximate total output using quarterly real gross domestic product (GDP) in chained 2009 dollars from the Bureau of Economic Analysis. Second, we approximate aggregate employment using the quarterly averages of monthly total private employment from the Current Employment Statistics (CES) survey published by the Bureau of Labor Statistics (BLS). Last, we obtain each sectoral employment by taking quarterly averages of monthly employment in the corresponding sector from the CES survey. All data series are seasonally adjusted and span from the first quarter of 1948 (1948Q1) to the fourth quarter of 2015 (2015Q4).

The next section also uses the annual person-level data from the 1968–2014 Integrated Public Use Microdata Series–Current Population Survey (IPUMS-CPS) (King et al., 2014). In 1968, the IPUMS-CPS started to report the industry in which the respondent worked in the previous year. We adopt the BLS classification and categorize each respondent’s industry into one of the following sectors: mining, construction, manufacturing, transportation and utilities, wholesale trade, retail trade, financial activities, health and education, and other services. Following Autor, Katz and Kearney (2008), we restrict the sample to full-time employees who aged from 18 to 64. We define a person who has completed no more than 12 years of schooling as non-college educated (i.e., high school dropouts and high school graduates) and at least 13 years of schooling as college educated (i.e., some college and college plus). All reported statistics are weighted.

Jobless Recovery and Structural Change

Jobless Recovery

In the literature, a jobless recovery refers to the divergent recovery paths of total output and aggregate employment after a recession (Aaronson et al., 2004; Bachmann, 2012; Groshen & Potter, 2003). It is a time where the aggregate employment continues to plunge despite a simultaneous improvement in the total output after a recession trough date decided by the National Bureau of Economic Research (NBER). We assume this definition and show that jobless recoveries are a new phenomenon in the post-1990 US economy.

Table 1 reports the number of quarters in which a positive GDP growth rate is accompanied by a negative employment growth rate after each trough date. We observe at most one such quarter following a pre-1990 recession, but at least three after a post-1990 recession. Table 1 also reports the number of quarters for total private employment to return to its end-of-recession level after GDP has already done so. It takes the employment at most two quarters to return to its trough level before 1990, but at least six quarters after 1990. Hence, there is strong evidence indicating the onset of jobless recoveries in the post-1990 era.

Table 1.

Total Private Employment Recovery Timeline

NBER Recession	Number of Quarters
NBER Recession	+ GDP Growth and – Employment Growth	Employment Back to Trough Level after GDP Recovery
1948Q4–1949Q4	0	0
1953Q2–1954Q2	1	1
1957Q3–1958Q2	0	0
1960Q2–1961Q1	0	0
1969Q4–1970Q4	0	0
1973Q3–1975Q1	1	2
1980Q1–1980Q3	0	0
1981Q3–1982Q4	0	0
1990Q3–1991Q1	4	6
2001Q1–2001Q4	3	9
2007Q4–2009Q2	3	6

Sources: Authors’ own.

Structural Change

This subsection documents three US labour market changes that coincide with the start of jobless recoveries. We then form a hypothesis that links these changes to the jobless recovery phenomenon.

Figure 1 shows various sectoral shares of total private employment over the past six decades. Among all the sectors, manufacturing and FHE sectors have undergone the most dramatic change. While the former sees its share falling from 37 per cent in 1948 to 10 per cent in 2015, the share of the latter has increased from 10 per cent to 25 per cent. Since 1990, the FHE sectors have not only experienced the highest growth but overtaken manufacturing sectors to become the biggest employer in the private sector.

Figure 1.

Sources: Authors’ own.

Although 1990 can be considered as a landmark, the decline of manufacturing sector and the rise of FHE sectors are a gradual process that started much earlier. Therefore, we decided to further examine the business cycle dynamics of manufacturing and FHE employment for the past six recessions.³ In particular, both employment series are logged, HP-filtered and normalized to zero at each trough date. Figure 2 plots the data from one year before the peak date to one year after the trough date, and the grey bar represents a recession.

Since recessions usually hit the goods sectors harder than the service sectors, FHE employment is much less sensitive to business cycle fluctuations than manufacturing employment. The FHE employment also typically fails to return to its trough level within a year after a recession ends.⁴ However, its pre-and post-1990 business cycle behaviours do not differ drastically. If anything, the FHE employment seems to have become more acyclical after 1990; it actually rose during the 2001 recession.

On the contrary, the business cycle dynamics of manufacturing employment has clearly changed after 1990, despite the fact that the sector started shrinking much earlier. The manufacturing employment rises above the trough level within one year following a recession before 1990, but fails to recover within the same timeline after 1990.⁵ Manufacturing sector seems to have difficulty gaining back the jobs that it has lost during a recession after 1990. As a result, more unemployed manufacturing workers have to find jobs outside the sector. The coincidental occurrence of jobless recoveries suggests that other sectors, especially the expanding FHE sectors, were not absorbing those unemployed workers quickly. Thus, we next propose a potential reason for the slow absorption that also applies to sectors beyond manufacturing.

Figure 2.

Sources: Authors’ own.

Figure 3.

Source: Authors’ own.

Figure 3 shows the last change in the labour market using the 1968–2014 IPUMS-CPS data. The previous section details the sample selection criteria, the classification of industry and the definition of college education. For each year, we first obtain the percentage of college-educated workers in every sector. Then the annual standard deviation of those percentages is computed.

The cross-sector standard deviation, fluctuating around 12.5 per cent, is quite stable prior to 1990. This indicates a relatively constant difference in the education requirement for workers across sectors. The series has a clear upward trend after 1990, especially up to 2008. This indicates a widening gap in the education requirement for workers across sectors. It is evident that certain sectors over those years have become increasingly harder for unskilled workers to enter. Specifically, college workers consist on average 75 per cent of the FHE sectoral workforce during the past decade. Meanwhile, barely half of the workforce in the other sectors on average attain the same education. Thus, it is more likely for a typical worker laid off in another sector to lack the qualification for a job in the FHE sectors after 1990.

To sum up, FHE sectors have replaced manufacturing sectors to become the biggest private employer since 1990. A closer look at the business cycle dynamics of manufacturing employment reveals the need for laid-off manufacturing workers to find jobs outside their own sector after 1990. However, the onset of jobless recoveries suggests that those workers are having a hard time relocating. The fact that college workers have become disproportionately concentrated in the expanding FHE sectors could indicate a rising barrier to entry. Taking the three changes together, we conjecture that if an individual loses her job in one sector, especially a declining sector, and lacks the qualification for a job in the FHE sectors where new openings are most abundant, structural unemployment will result and linger, causing a jobless recovery.

Before we formally test our conjecture, the following fact merits some discussion. As shown in Figure 2, while the adsorption capacity of the FHE sectors has always been slow, job losses in manufacturing sectors have become more permanent after 1990. Therefore, all the facts indicate that the post-1990 slow job trend is characteristically different from the pre-1990 trend. More specifically, in addition to the fast expansion of the FHE sectors, the anaemic absorption of these sectors compounded with the failure of manufacturing to rebound can only aggravate a jobless recovery. But why is manufacturing employment unable to recover? And how does this inability to recover contribute to jobless recoveries? We offer two potential explanations, namely job polarization and population aging.

Job polarization refers to the disappearance of middle-skill or routine jobs—jobs that can be replaced by computers or be relocated overseas (i.e., assemblers, machine operators, bookkeepers).⁶ According to Autor, Katz and Kearney (2006), job polarization started to emerge in the USA in the late 1980s. Jaimovich and Siu (2014) found that job polarization is mostly concentrated in recessions when the opportunity cost of firm restructuring is the lowest. Given the nature of manufacturing jobs, it is reasonable to expect a bigger negative effect of job polarization on manufacturing employment. Thus, the recovery of the US labour market becomes increasingly dependent on the absorption capacity of the FHE sectors.

Americans are also getting older, albeit more slowly, than most other advanced economies (Goldin, 2016). In general, older people have lower job turnover rates and are more resistant to a career change. As a result, it becomes more difficult to get an aging workforce back to employment after a recession, especially if the greater number of them also require a change of fields (i.e., from the declining manufacturing to the fast growing FHE sectors).

Methodology and Results

Motivated by the analysis in the previous section, we are interested in three variables. The first variable is total output denoted as GDP. The second variable is aggregate employment denoted as EMP. The third variable is the FHE sectoral share of aggregate employment denoted as FHE. The source and the retrieval of the data are described in the third section.

The main specification of our VAR is as follows.

Y_{t} = c + A_{1} Y_{t - 1} + ... + A_{p} Y_{t - p} + u_{t},

(1)

where Y _t is a vector of three variables (i.e., ΔlnGDP_t, ΔFHE_t and ΔlnEMP_t), c is a vector of constants, u _t is a vector of error terms and p denotes the number of lags. Specifically, ΔlnGDP_t is the log difference between GDP at quarters t and t – 1. ΔFHE_t is the difference between FHE at quarters t and t – 1. ΔlnEMP_t is the log difference between EMP at quarters t and t – 1. Our focus on jobless recovery makes it more appropriate to use differences than levels.

We split our data into two time periods, 1948Q1–1989Q4 and 1990Q1–2015Q4. Figures 4 and 5 plot the time series of ΔlnGDP, ΔFHE and ΔlnEMP for the two time periods, respectively. None of the six series exhibits any growth trend over time, and all their means are statistically significantly different from zero. Therefore, we apply the augmented Dickey–Fuller test with a constant and four lags to these series, and they all reject unit root at the 5 per cent level. We conclude that VAR analysis can be performed on these six series without further modification.

Figure 4.

Sources: Authors’ own.

Figure 5.

Sources: Authors’ own.

Table 2.

VAR Lag Order Selection Criteria

Lag	1948Q1–1989Q4			1990Q1–2015Q4
Lag	AIC	SIC	HQ	AIC	SIC	HQ
0	–29.863	–29.805	–29.839	–31.325	–31.249	–31.295
1	–31.033	–30.803*	–30.940	–33.950*	–33.645*	–33.826*
2	–31.143	–30.741	–30.980*	–33.871	–33.337	–33.655
3	–31.178	30.604	–30.945	–33.938	–33.175	–33.629
4	–31.184*	–30.437	–30.880	–33.945	–32.953	–33.543
5	–31.123	–30.204	–30.750	–33.907	–32.687	–33.413
6	–31.090	–29.999	–30.647	–33.793	–32.343	–33.205

Sources: Authors’ own.

Notes: AIC: Akaike information criterion; SIC: Schwarz information criterion; HQ: Hannan–Quinn information criterion; * indicates lag order selected by the criterion.

Table 2 reports the lag order selection criteria of AIC, SIC and HQ for up to six lags. For the first time period, the results are inconsistent, so we also run the Wald test which indicates an optimal lag length of four, in line with the AIC selection. Thus, considering the quarterly frequency of the data, we decide to include four lags in our VAR. For the second time period, we include one lag as indicated unanimously by all three criteria.⁷ For both time periods, we fail to reject no serial correlation in residuals at the selected lag length, and the VAR system is stable with all the eigenvalues smaller than one.

Reduced-form Estimates

Given the limitations of reduced-form VAR, we are mainly interested in the signs and significance levels of the coefficients of lagged ΔFHEs in the regression where ΔlnEMP_t is the dependent variable. Table 3 reports the VAR estimates for the pre-1990 period. None of those coefficients is statistically significant and only that of the fourth lag is negative. Thus, the FHE employment expansion at best has a weak positive effect on the aggregate employment growth before 1990. The Granger causality test fails to indicate that ΔFHE helps predict ΔlnEMP (the p value is 0.120).

Table 3.

VAR Estimates, 1948Q1–1989Q4

	ΔlnGDP	ΔFHE	ΔlnEMP
ΔlnGDP (–1)	–0.039 (0.115)	–0.011 (0.017)	0.071 (0.063)
ΔlnGDP (–2)	0.230** (0.117)	–0.046*** (0.017)	0.174*** (0.064)
ΔlnGDP (–3)	0.082 (0.117)	–0.011 (0.017)	0.048 (0.064)
ΔlnGDP (–4)	0.076 (0.106)	–0.022 (0.016)	0.068 (0.058)
ΔFHE (–1)	0.424 (1.045)	0.223 (0.153)	0.882 (0.569)
ΔFHE (–2)	–0.638 (1.162)	–0.227 (0.170)	0.219 (0.633)
ΔFHE (–3)	1.438 (1.176)	0.044 (0.172)	0.778 (0.641)
ΔFHE (–4)	–1.643 (1.074)	0.350** (0.157)	–0.640 (0.585)
ΔlnEMP (–1)	0.893*** (0.288)	–0.111*** (0.042)	1.000*** (0.157)
ΔlnEMP (–2)	–1.054*** (0.334)	0.077 (0.049)	–0.544*** (0.182)
ΔlnEMP (–3)	0.412 (0.342)	–0.028 (0.050)	0.394** (0.186)
ΔlnEMP (–4)	–0.659** (0.277)	0.145*** (0.041)	–0.419*** (0.151)
Constant	0.004** (0.002)	0.001** (0.000)	–0.001 (0.001)
N	163	163	163
RMSE	0.0041	0.0006	0.0022
R²	0.3087	0.5512	0.6541
χ²	72.783***	200.154***	308.283***

Sources: Authors’ own.

Notes: AIC = –31.21; SIC = –30.47; HQ = –30.91. * p < 0.1; ** p < 0.05; *** p < 0.01; SE in parentheses.

Table 4 reports the VAR estimates for the post-1990 period. In contrast to the pre-1990 estimates, the coefficient of ΔFHE_t-₁ is statistically significant and negative. The FHE employment expansion in the previous quarter slows down the current aggregate employment growth after 1990. The Granger causality test also suggests that ΔFHE helps predict ΔlnEMP (the p value is 0.078). Hence, the reduced-form estimates seem to support our hypothesis.

Table 4.

VAR Estimates, 1990Q1–2015Q4

	ΔlnGDP	ΔFHE	ΔlnEMP
ΔlnGDP (–1)	0.256** (0.108)	–0.099*** (0.023)	0.196*** (0.041)
ΔFHE (–1)	0.224 (0.528)	0.858*** (0.111)	–0.354* (0.201)
ΔlnEMP (–1)	0.394 (0.266)	0.061 (0.056)	0.620*** (0.101)
Constant	0.001** (0.001)	0.000** (0.000)	0.000 (0.000)
N	104	104	104
RMSE	0.0024	0.0005	0.0009
R²	0.2144	0.7930	0.8511
χ²	28.387***	398.349***	594.374***

Sources: Authors’ own.

Notes: AIC = –33.95; SIC = –33.64; HQ = –33.83. * p < 0.1; ** p < 0.05; *** p < 0.01; SE in parentheses.

Impulse Responses

We also examine our hypothesis by looking at impulse responses. It is widely accepted that employment as an economic indicator lags behind GDP. Our conjecture also identifies the expansion of the FHE sectors as an important trigger of the jobless recovery. Thus, we use Cholesky decomposition and order the three variables as follows: ΔlnGDP, ΔFHE and ΔlnEMP. Other set-ups are kept the same as the reduced form.

Figures 6 and 7 plot the pre-1990 and the post-1990 impulse responses, respectively. The first row shows the effect of an unexpected one-percentage-point drop in GDP growth rate on the three variables. The second row shows the effect of an unexpected one-percentage-point growth in FHE employment share. The third row shows the effect of an unexpected one-percentage-point increase in aggregate employment growth rate. Dashed lines indicate the 95 per cent confidence interval for each impulse response. We are mainly interested in the effects shown in the first two rows. Table 5 reports the pre-and post-1990 variance decompositions. We are mainly interested in the decomposition of ΔlnEMP shown in Panel C.

Figure 6.

Sources: Authors’ own.

Figure 7.

Sources: Authors’ own.

A comparison of the first row between Figures 6 and 7 confirms the jobless recovery phenomenon. In Figure 6, the first row shows a quick and concurrent recovery of GDP and aggregate employment growth rates after a negative shock to the GDP growth rate. In Figure 7, the GDP growth rate is back to the steady state after about 15 quarters when the employment growth rate is still below its steady state.

The above observation has a caveat: the persistence of the GDP shock has changed. It takes 5 quarters before 1990 (but 15 quarters after 1990) for the GDP growth rate to return to the steady state in response to a shock to itself. Thus, jobless recoveries could be caused by the changing nature of aggregate shocks instead of (or in addition to) the FHE employment expansion. Based on the variance decomposition, the amount of variance of aggregate employment growth explained by the GDP shock almost halves on impact after 1990, but is four-percentage-point higher compared to the pre-1990 era in the later quarters. Thus, to tease out the effect of the aggregate shock, one need to compare impulse responses of a dynamic structural model featuring a GDP shock of the same persistence, but different FHE employment shares. Unfortunately, it is outside the empirical scope of this article. Also, as mentioned earlier, the present study concedes the possibility that jobless recoveries are driven by multiple factors rather than a single factor. Our goal is to assess whether the expansion of the FHE sectors makes a significant contribution.

Table 5.

Variance Decomposition

(a) Variance Decomposition of ΔlnGDP
Horizon	Variance Decomposition as in Percentage Points
	1948Q1–1989Q4			1990Q1–2015Q4
	ΔlnGDP	ΔFHE	ΔlnEMP	ΔlnGDP	ΔFHE	ΔlnEMP
1	100 (0.0)	0 (0.0)	0 (0.0)	100 (0.0)	0 (0.0)	0 (0.0)
5	87 (4.3)	6 (3.5)	7 (3.2)	97 (2.3)	1 (2.1)	2 (2.3)
10	84 (4.8)	6 (3.4)	10 (4.0)	96 (3.4)	2 (3.2)	2 (2.7)
15	84 (4.9)	6 (3.4)	10 (4.0)	96 (3.8)	2 (3.6)	2 (2.7)

(b) Variance Decomposition of ΔFHE
Horizon	Variance Decomposition as in Percentage Points
	1948Q1–1989Q4			1990Q1–2015Q4
	ΔlnGDP	ΔFHE	ΔlnEMP	ΔlnGDP	ΔFHE	ΔlnEMP
1	43 (5.8)	57 (5.8)	0 (0.0)	17 (6.7)	83 (6.7)	0 (0.0)
5	53 (6.8)	41 (6.1)	6 (3.3)	47 (10.7)	52 (10.6)	1 (0.8)
10	51 (6.7)	36 (5.9)	13 (5.2)	50 (11.8)	50 (11.5)	0 (0.8)
15	51 (6.8)	36 (6.0)	13 (5.2)	50 (12.1)	50 (11.8)	0 (0.7)

(c) Variance Decomposition of ΔlnEMP
Horizon	Variance Decomposition as in Percentage Points
	1948Q1–1989Q4			1990Q1–2015Q4
	ΔlnGDP	ΔFHE	ΔlnEMP	ΔlnGDP	ΔFHE	ΔlnEMP
1	52 (5.4)	23 (4.2)	25 (3.4)	29 (7.5)	40 (6.8)	31 (5.0)
5	59 (7.1)	16 (4.3)	25 (6.0)	61 (9.0)	28 (8.7)	11 (4.5)
10	59 (7.2)	16 (4.9)	25 (6.2)	63 (9.6)	28 (10.4)	9 (4.7)
15	59 (7.2)	16 (4.9)	25 (6.2)	63 (9.8)	29 (10.8)	8 (4.7)

Sources: Authors’ own.

Note: SE in parentheses.

A comparison of the second row between Figures 6 and 7 seems to support our hypothesis. In Figure 7, a positive shock to the growth of FHE employment share has a prolonged negative effect on the aggregate employment growth. In Figure 6, this effect is short-lived and quickly reversed. Thus, a faster FHE employment expansion impedes the aggregate employment recovery after 1990.

This observation has the same caveat: the more persistent FHE shock could have triggered jobless recoveries. However, we want to emphasize the following three findings in favour of our hypothesis. First, the reduced-form estimates show that the coefficient of lagged ΔFHE is only statistically significant and negative for the post-1990 era. Second, the Granger causality test indicates that past values of ΔFHE only aid in the prediction of current ΔlnEMP after 1990. Last, variance decomposition suggests that shocks to ΔFHE explain 16–23 per cent of the variance of aggregate employment growth rates before 1990, but 28–40 per cent after 1990. Unfortunately, one might argue that the improvement in the variance decomposition is not surprising as it roughly corresponds to the increase in the FHE employment share during the time period. As shown in Figure 1, the share rose from 10 per cent in 1948 to 25 per cent in 2015. A similar argument can also be made for the first two findings.

As a result, we estimate the following VAR to gain additional insights on the issue.

Y_{t} = c + A_{1} Y_{t - 1} + ... + A_{p} Y_{t - p} + B X_{t} + u_{t}

(2)

where X _t is a vector of three exogenous dummies indicating the first quarter of 1990, and all else are the same as in Equation (1). Hence, we use the data of the entire time span instead of breaking them into two periods. We compare the result of this VAR with that of the same VAR, but excluding ΔFHE_t from Y _t. Based on the unit root test and lag order selection procedures described earlier, all data series are stationary, and an optimal lag length of four is selected for both VAR with and without ΔFHE.

Tables 6 and 7 report the reduced-form estimates of the two VARs, respectively. We focus on the coefficient of the date dummy in the regression where ΔlnEMP is the dependent variable. First, the coefficient is negative in both VAR analyses, indicating that changes in aggregate employment have become smaller after 1990. Second, with the inclusion of ΔFHE, the coefficient not only becomes statistically significant but decreases by 0.0001 (or about 25 per cent). In other words, without the change in FHE employment share, the change in aggregate employment would have been bigger relative to the pre-1990 era. Thus, it supports our hypothesis that the expansion of the FHE sectors contributes to the slow labour market recovery after 1990.

Table 6.

VAR Estimates Without ΔFHE, 1948Q1–2015Q4

	ΔlnGDP	ΔlnEMP
90Q1 (date dummy)	–0.0011** (0.0005)	–0.0004 (0.0002)
ΔlnGDP (–1)	0.0001 (0.0847)	0.0526 (0.0455)
ΔlnGDP (–2)	0.1907** (0.0847)	0.1045** (0.0455)
ΔlnGDP (–3)	0.0363 (0.0837)	–0.0146 (0.0450)
ΔlnGDP (–4)	0.1011 (0.0764)	0.0223 (0.0410)
ΔlnEMP (–1)	0.7910*** (0.1542)	0.9072*** (0.0828)
ΔlnEMP (–2)	–0.7784*** (0.1836)	–0.4666*** (0.0986)
ΔlnEMP (–3)	0.0657 (0.1870)	0.2472** (0.1004)
ΔlnEMP (–4)	–0.2578** (0.1311)	–0.2127*** (0.0704)
Constant	0.0031*** (0.0005)	–0.0006** (0.0003)
N	267	267
RMSE	0.0036	0.0019
R²	0.2842	0.6680
χ²	105.991***	537.140***

Sources: Authors’ own.

Notes: AIC = –18.71; SIC = –18.44; HQ = –18.60. * p < 0.1; ** p < 0.05; *** p < 0.01; SE in parentheses.

Table 7.

VAR Estimates with ΔFHE, 1948Q1–2015Q4

	ΔlnGDP	ΔFHE	ΔlnEMP
90Q1 (date dummy)	–0.0012*** (0.0005)	0.0001 (0.0001)	–0.0005* (0.0003)
ΔlnGDP (–1)	0.0054 (0.0852)	–0.0113 (0.0144)	0.0510 (0.0457)
ΔlnGDP (–2)	0.1821** (0.0846)	–0.0331** (0.0143)	0.1022** (0.0454)
ΔlnGDP (–3)	0.0365 (0.0837)	0.0090 (0.0142)	–0.0167 (0.0449)
ΔlnGDP (–4)	0.0924 (0.0765)	–0.0104 (0.0130)	0.0205 (0.0410)
ΔFHE (–1)	0.0193 (0.6043)	0.7776*** (0.1024)	–0.1461 (0.3241)
ΔFHE (–2)	–0.6179 (0.7293)	–0.2302* (0.1236)	0.0791 (0.3911)
ΔFHE (–3)	0.5298 (0.7271)	0.3259*** (0.1232)	–0.1503 (0.3899)
ΔFHE (–4)	–0.4382 (0.5959)	–0.0478 (0.1010)	–0.1848 (0.3196)
ΔlnEMP (–1)	0.7806*** (0.2096)	–0.0173 (0.0355)	0.8667*** (0.1124)
ΔlnEMP (–2)	–0.9130*** (0.2567)	0.0471 (0.0435)	–0.4393*** (0.1376)
ΔlnEMP (–3)	0.1987 (0.2588)	0.0239 (0.0439)	0.2067 (0.1388)
ΔlnEMP (–4)	–0.3654* (0.1939)	0.0440 (0.0329)	–0.2554** (0.1040)
Constant	0.0037*** (0.0008)	0.0001 (0.0001)	0.0010** (0.0004)
N	267	267	267
RMSE	0.0036	0.0006	0.0019
R²	0.2890	0.6225	0.6710
χ²	108.502***	440.205***	544.440***

Sources: Authors’ own.

Notes: AIC = –31.64; SIC = –31.08; HQ = –31.41. * p < 0.1; ** p < 0.05; *** p < 0.01; SE in parentheses.

The Great Moderation

To test the robustness of our results, we also split the data based on the Great Moderation, namely 1948Q1–1984Q4 and 1985Q1–2015Q4. We repeat all of our VAR analyses using the same set of procedures, and almost all of our previous findings hold except the following. When we run the VAR with the date dummy—indicating the first quarter of 1985—the coefficient of the dummy is negative, but statistically insignificant regardless of the inclusion of ΔFHE. Also, including ΔFHE only lowers the coefficient by 0.00002 (or about 9 per cent).

The Spread of Jobless Recovery

The fast expansion of the FHE sectors is not unique to the US economy. Almost all advanced economies today are service oriented, and the FHE sectoral workforce constitutes a considerable, if not the biggest, proportion of total employment. Therefore, do we see jobless recoveries in other advanced economies, say the European Union (EU)? Our answer is affirmative based on the evidence we can gather.⁸ In 2013, the European Commission cautioned that the recovery in economic activities in the EU from the last global recession would result in more gradual, not rapid, job creation (Hewitt, 2013). Confirmed by the 2014 report from the International Labour Organization (ILO), the recovery in the EU has remained in economic activities, not in jobs. These warning signs of jobless recovery in the EU are not surprising given the fact that job polarization and population aging are also pervasive in the post-1990 Europe (Goos, Manning, & Salomons, 2014; International Labour Organization [ILO], 2014).

However, the US economy still differs from that of the EU in many aspects. These differences might help explain why jobless recoveries are much more pronounced in the USA. For instance, the USA granted permanent normal trade relations to China in 2001, eliminating potential tariff increases on Chinese imports. Pierce and Schott (2012) found that this policy change causes a sharp decline in the US manufacturing employment, and this trend is absent in the EU where no such policy change was implemented. Also, it is well known that the EU has greater employment protection laws and a stronger union presence than the USA, both of which tend to dampen job losses during recessions.

Conclusion

This article investigates the link between jobless recoveries and the fast FHE employment expansion. We use micro-level data to show that jobs in the FHE sectors have a higher education requirement, potentially creating a barrier to entry for unemployed workers in other sectors. As a result, a long period of structural unemployment could occur and lead to a jobless recovery. Subsequently, we examine this proposition using both reduced-form and recursive VARs.

Despite a few caveats, we consider our empirical findings mostly in favour of the proposition, thereby asserting the notable contribution of structural change to the onset of jobless recoveries. These findings can be summarized as follows.

Based on the reduced-form estimates, the FHE employment expansion only has a statistically significant and negative effect on aggregate employment after 1990. The Granger causality test suggests that the former only helps predict the latter after 1990.

Based on the impulse responses, positive shocks to FHE employment growth only have a prolonged negative effect on aggregate employment growth after 1990. These shocks explain 28–40 per cent of the error variance of the aggregate employment growth.

Including the change in FHE employment share reduces the fluctuation in aggregate employment growth rate by 25 per cent after 1990 relative to the pre-1990 era.

Footnotes

Acknowledgements

This work is supported by the summer research grant from the College of Business and Economics, California State University, Los Angeles.

Notes

References

Aaronson

, Rissman

, & Sullivan

(2004). Can sectoral reallocation explain the jobless recovery? Federal Reserve Bank of Chicago Economic Perspectives, 28(2), 36–49.

Autor

, Katz

, & Kearney

(2006). The polarization of the US labor market. American Economic Review, 96(2), 189–194.

Autor

, Katz

, & Kearney

(2008). Trends in U.S. wage inequality: Revising the revisionists. Review of Economics and Statistics, 90(2), 300–323.

Bachmann

(2012). Understanding the jobless recoveries after 1991 and 2001. Retrieved 9 February 2017, from http://www3.nd.edu/~rbachman/bachmann_jlr_2012.pdf

Berger

(2012). Countercyclical restructuring and jobless recoveries. Retrieved 9 February 2017, from http://economics.mit.edu/files/7498

DeNicco

, & Laincz

(2014). Jobless recovery: A time series look at the United States. Retrieved 9 February 2017, from http://www.pages.drexel.edu/~jpd48/Jobless%20Recovery.pdf

Garin

, Pries

, & Sims

(2013). Reallocation and the changing nature of economic fluctuations. Retrieved 9 February 2017, from http://www3.nd.edu/~esims1/gps_june27_2013.pdf

Goldin

(2016). How Japan and the US can reduce the stress of aging? (NBER Working Paper No. 22445). Retrieved 5 August 2016, from http://www.nber.org/papers/w22445

Goos

, Manning

, & Salomons

(2014). Explaining job polarization: Routine-biased technological change and offshoring. American Economic Review, 104(8), 2509–2526.

10.

Groshen

, & Potter

(2003). Has structural change contributed to a jobless recovery? Federal Reserve Bank of New York Current Issues in Economics and Finance, 9(8), 1–7.

11.

Hewitt

(2013). Europe: A jobless recovery. BBC News. Retrieved 9 February 2017, from http://www.bbc.com/news/world-europe-24821783

12.

International Labour Organization (ILO). (2014). Global employment trends 2014: Risk of a jobless recovery? Retrieved 5 August 2016, from http://www.ilo.org/wcmsp5/groups/public/–dgreports/–dcomm/–publ/documents/publication/wcms_233953.pdf

13.

Jaimovich

, & Siu

H.E.

(2014). The trend is the cycle: Job polarization and jobless recoveries. NBER Working Paper No. 18334. Retrieved 5 August 2016, from http://www.nber.org/papers/w18334

14.

King

, Ruggles

, Alexander

, Flood

, Genadek

, Schroeder

, . (2014). Integrated public use microdata series: Current population survey (Machine-readable database). Minneapolis, MN: University of Minnesota.

15.

Koenders

, & Rogerson

(2005). Organizational dynamics over the business cycle: A view on jobless recoveries. Federal Reserve Bank of St. Louis Review, 87(4), 555–579.

16.

Pierce

, & Schott

(2012). The surprisingly swift decline of US manufacturing employment American Economic Review, 106(7), 1632–1662.

17.

Sahin

, Song

, Topa

, & Violante

(2014). Mismatch unemployment. American Economic Review, 104(11), 3529–3564.