I.  Worker Transitions over the Business Cycle

I follow a well-known methodology using matched files from the Current Population Survey to construct a measure of the flow of workers between employment (E), unemployment (U), and not in the labor force (N); see, for example, Shimer (2012). In each month, this gives me the share of workers in one state transitioning to another state; so, for example, ${\lambda }_t^{\mathit{EU}}$ is the share of employed workers in month $t-1$ who are unemployed in month t. I let Λ denote the resulting 3 × 3 transition matrix:

I do two exercises with this transition matrix. First, if the economy were in steady state, with constant flows from month to month, ${\mathrm{\Lambda }}_t=\mathrm{\Lambda }$, the share of workers in each state would converge to a constant, given by the eigenvector associated with the matrix Λ’s unit eigenvalue. I compare that implied steady state unemployment rate $u_t^{*}$ with the actual unemployment rate u_t in each month. Figure 1 shows that the two are very similar, with $u_t^{*}$ slightly larger than u_t during the two recession periods and slightly smaller during the subsequent expansions. The gaps reflect the fact that unemployment rises (falls) when the flows push it toward a higher (lower) rate. But more importantly for my purposes here, the two lines are similar because the underlying transition rates change relatively slowly over time.

Second, I construct counterfactual unemployment rates to examine the effect of each of the six transition rates in turn. More precisely, I allow just one of the six transition rates in equation (1) to vary over time and hold the other five fixed at their average value. I then look at the eigenvector associated with the unit eigenvalue of the resulting matrix to construct a counterfactual unemployment rate. The black lines in figure 2 show that the bulk of the increase in unemployment during and after the 2008–9 recession is accounted for by (the decline in) the unemployment-to-employment transition probability ${\lambda }_t^{\mathit{UE}}$, which I refer to also as the job finding rate. Interestingly, however, an early increase in the employment-to-unemployment transition probability ${\lambda }_t^{\mathit{EU}}$, which I call here the job loss rate, accounts for additional upward pressure on the unemployment rate, pushing it up to 6.3% by 2009Q1, before a slow and steady decline in job loss pulled it back down to 4.4% 10 years later. Changes in the other flows, that is, movements in and out of the labor force, had less of an impact on the steady state unemployment rate.

The story of the pandemic recession in 2020 is dramatically different but equally informative. In April 2020, the job loss rate, ${\lambda }_t^{\mathit{EU}}$, spiked, explaining virtually all of the increase in unemployment. It declined almost as rapidly, so by June 2021, elevated job loss contributed less to the unemployment rate than did the low job finding rate. Still, as I write this comment, it appears that the economy has settled into a fairly normal (and comparatively slow) recovery, with unemployment hovering near 6%.

I will argue that the difference between the pandemic recession and “normal” recessions is that during the pandemic, most workers understood that they could return to their old job, or at least their old occupation, after the worst of the pandemic had passed. In contrast, other recessions are associated with more long-term displacement as workers leave their occupation and gradually retrain for other jobs. This slow process leads to a depressed job finding rate for years after the initial shock. I conjecture that a similar phenomenon will likely arise for those workers lingering in unemployment as the United States emerges from this most recent recession.

II.  Two Types of Unemployment

I now describe a mechanical model where workers can be either employed or unemployed and can have either high or low human capital. Idiosyncratic shocks move workers between the two employment statuses and between the two human capital statuses, with a positive correlation between job loss and human capital loss, as in the Ljungqvist and Sargent (1998) model of turbulence. I think of human capital as being something occupation-specific, so a high human capital worker is someone who has found an occupation that they excel at, whereas a low human capital worker is searching for such an occupation (Kambourov and Manovskii 2009; Alvarez and Shimer 2011).

I assume that unemployed workers with high (low) human capital find a job at rate f_h (f_l), whereas employed workers with high (low) human capital lose their jobs at rate s_h (s_l), all constant and exogenous. Finally, employed workers with low human capital acquire human capital at rate γ, and a fraction γ of job losses for high human capital workers are associated with a loss of human capital, the Ljungqvist and Sargent (1998) turbulence shock.

Let u_h(t) and u_l(t) denote the fraction of workers who are unemployed at time t with high and low human capital, respectively. Let e_h(t) and e_l(t) denote the corresponding employment rates. The worker flows in the previous paragraph imply these evolve according to

I want to think quantitatively about what kinds of dynamics this model can generate, but to be clear, the numbers I use here are only suggestive. I stress that there are two critical quantitative assumptions: human capital status is much more persistent than employment status, and high human capital workers typically have shorter unemployment durations and longer employment durations than low human capital workers. I think of a time period as a month. I assume that 10% of low human capital workers switch employment status each month, $s_l=f_l=0.1$, whereas only 2% of employed high human capital workers lose their job, $s_h=0.02$, and 60% of unemployed high human capital workers find a job, $f_h=0.6$. These transition rates ensure that workers with high human capital have much lower unemployment duration, much higher employment duration, and consequently much lower unemployment rates. Finally, I also assume $\gamma =0.02$, so occupation-specific human capital is highly persistent.

In steady state, 95.9% of workers have high human capital, and the remaining 4.1% of workers have low human capital. The unemployment rate is 3.2% for the high human capital workers and 54.5% for the low human capital workers, so despite the paucity of low human capital workers, they account for almost half of the 5.3% unemployment rate. The job finding rate for the typical unemployed worker is 38.8% per month, whereas the job loss rate is 2.2% per month, both reflecting the mixture of human capital in the two pools.

Now suppose the model economy is initially in steady state but is hit by a one-time shock that adversely affects workers’ human capital and throws some of them into unemployment. This might be, for example, a sharp and permanent decline in the industry demanding the worker’s skills. More precisely, I assume that 10% of high human capital workers, employed and unemployed, lose their human capital. In addition, half the employed workers who lose their human capital also immediately lose their job. This combination of shocks immediately nearly doubles the unemployment rate to 10.0% and, more importantly, more than triples the share of workers with low human capital to 13.7%.

More interesting are the subsequent dynamics, which largely reflect the slow recovery of the stock of human capital. On impact, the job finding rate falls by 50 log points to 23.6%, whereas the job loss rate increases less dramatically, by 18 log points to 2.6%. Each then recovers monotonically but very slowly. As a result, we get a prolonged period of high unemployment with a very low job finding rate and a slightly elevated job loss rate. Figure 3 depicts the resulting counterfactual unemployment rates, showing persistence not unlike the response of the United States labor market to the financial crisis depicted in figure 2.

Figure 4 looks at the same dynamics from a different perspective. The dotted line shows the share of the labor force that is unemployed immediately after the shock at time 0 and also unemployed at time t (possibly with one or more intervening employment spells). The black line shows the share that is employed immediately after the shock but unemployed at t. Notably, the two lines cross after 6 months, so workers who lost their job at the time of the shock subsequently account for less than half of all unemployment.

If I handed data generated from this model to Hall and Kudlyak, they would note the high unemployment rate at t of workers who remained employed through the initial shock, together with the low job finding rate for the typical unemployed worker. Putting those two facts together, I believe they would reach the same conclusion as in Subsection IV.C of their paper, “The excess job loss accounts for the magnitude of the initial increase in unemployment, but not its persistence. The persistence is too large to be explained as reflecting only the personal experiences of the extra job losers dating from the spike.” Their syllogism would then lead them to conclude that unemployment is contagious.

But of course there is no contagion in this simple mechanical model. The persistent dynamics reflect the poor labor market experience of the 10% of workers who lost their human capital due to the initial shock, only some of whom immediately experienced unemployment. Unemployment persistence is a manifestation of the compositional shift in the skill distribution caused by the initial adverse shock.

III.  Relationship to Other Labor Market Facts

My simple model with two types of unemployment builds on a long tradition in macroeconomics that emphasizes that sectoral shifts are concentrated in recessions (Lilien 1982). One can think of two versions of such a model. In the first, a new sector of the economy opens up, pulling workers out of old sectors and causing a brief recession due to frictional unemployment. This version of the sectoral shifts model is empirically implausible because we do not see evidence of that growth sector during any post-World War II recession, at least in the United States (Abraham and Katz 1986). In the second version of the model, an adverse shock, say a financial crisis, leads to a recession. The pace of reallocation then accelerates during the recession, possibly due to the low opportunity cost of time (Caballero and Hammour 1994). This is the model of sectoral shifts that I have in mind.

I will offer three pieces of evidence supporting this view of recessions. First, Jaimovich and Siu (2020) show that since 1990, employment in routine tasks has looked like a step function, flat during expansions and dropping precipitously during recessions. If a worker employed in a routine task wants to find a new stable job, they typically have to retrain and look for work in a nonroutine (manual or cognitive) occupation. In a similar vein, key manufacturing industries such as steel and automobile fabrication declined sharply during the deep recessions in 1974 and 1982 with little recovery during the subsequent expansion. The road back to stable employment was arduous for those workers.

Second, Davis and von Wachter (2011) look at the subsequent earnings of men who lose a job where they had 3 or more years of tenure. There are two important conclusions in that paper. First, on average these men lose 1.4 years of discounted predisplacement earnings during the 20 years after the initial job loss. As Davis and von Wachter (2011) emphasize, this is far more than can be explained by a Diamond-Mortensen-Pissarides model where unemployed workers suffer no loss in human capital and quickly recover their employment status. It is consistent with the model I have described here, where job loss is correlated with human capital loss, because human capital and wages are positively correlated. The second conclusion is that losing a job when the unemployment rate is above 8% is twice as bad as losing a job at an average point in the business cycle. It leads to 2.8 years of lost predisplacement earnings. This is consistent with the idea that these turbulence shocks, linking job loss to skill loss, are concentrated during periods with elevated unemployment. Most job loss during expansions is not associated with human capital loss, whereas the reverse is true during recessions.

The third piece of evidence is the “dog that didn’t bark,” the 2020 pandemic recession. Why did unemployment recover so quickly from its peak? This recession featured a temporary stop in activity, not a period of cleansing and rapid adjustment. Many restaurant workers were laid off during the peak of the pandemic, but no one expected the restaurant industry to disappear. Indeed, many of the workers went on temporary layoff and have now returned to the same or similar establishments. Without the loss in human capital, rapid recovery was possible and has occurred. Still, even in response to this unusual shock, some jobs may be lost forever. For example, the pandemic likely accelerated a preexisting decline in brick-and-mortar retail, forcing those workers to look for new stable employment, an arduous process. My expectation is that it will take years for the employment rate to return to its prepandemic level, as in past recoveries.

I want to close by recognizing that the model I have described here probably does not account for all fluctuations in unemployment. For example, it would have a hard time explaining why the number of unemployed new entrants to the labor force gradually rises during recessions and the early stages of recoveries (Hall and Kudlyak’s fig. 15). A more satisfactory model would probably have a persistent underlying shock. It would probably also have uncertainty about which growing sectors of the economy will absorb the workers leaving the declining sectors, which may propagate the shock and delay the recovery. And it may have a somewhat inelastic supply of job vacancies, which would slow expansions (even if it would not explain the observed decline in vacancies during a recession). But a proper accounting will give a large role to the importance of human capital losses during downturns and may not cede any role to contagious or infectious unemployment.

Whether unemployment is contagious matters for economic policy. According to the theories mentioned by Hall and Kudlyak, high unemployment for some workers causes high unemployment for other workers, an externality. This may create a role for programs that moderate the cyclicality of unemployment, for example, by subsidizing work sharing during recessions. In contrast, if efficient sectoral reallocation is concentrated during downturns due to the low opportunity cost of time, work-sharing programs would hinder that necessary reallocation and hence create long-run productive inefficiencies. Hall and Kudlyak have identified an interesting possibility. Future research, with better data and more serious models, will help us understand whether the possibility is correct.