The “Great Resignation” in perspective

A little over a year after the COVID-19 pandemic began, economists and other observers took note of a rising job quit rate, as measured by the U.S. Bureau of Labor Statistics (BLS) Job Openings and Labor Turnover Survey (JOLTS) program. JOLTS recorded a seasonally adjusted quit rate of 2.4 percent in the second month of the program’s existence (January 2001), and although this level was matched at other times, it was not surpassed until March 2021, when the quit rate reached 2.5 percent. This new record was quickly eclipsed in April 2021, when the quit rate stood at 2.8 percent; the current record is 3.0 percent, first reached in November 2021 and matched in December 2021. The rise in the quit rate has been called the “Great Resignation,” with many articles in the popular press speculating about why individuals have become more willing to leave their current employers.^[2] The fact that the labor force participation rate remains below its prepandemic high suggests that some of those who quit their jobs found new jobs and others exited the labor force.

In this article, I provide some additional perspectives on trends in quit rates. First, I address the variation in quit rates that might occur from month to month because of statistical chance, given that these rates are measured through a sample rather than a census of establishments. Next, I examine data on quit rates predating the JOLTS program, to see whether the rates we have been observing recently are high relative to those in the second half of the 20th century. I then examine available JOLTS estimates to understand why they have risen of late. How much of the rise can be explained simply by the labor market tightening that began after the Great Recession and then, following a brief downturn during the pandemic recession in the first half of 2020, resumed its course? Have there been shifts in employment to industries in which quit rates tend to be higher, or have quit rates risen across industries?

Data on labor turnover

In a 1982 Monthly Labor Review article, Carol M. Utter detailed the BLS experience with labor turnover surveys (which primarily covered manufacturing) during the 20th century.^[3] From 1930 until budget cutbacks ended the BLS Labor Turnover Survey (LTS) in December 1981, information was collected on both accessions to an employer’s workforce (broken down, at times, into new hires, recalls, and transfers) and separations (divided into quits, layoffs, and discharges). However, the process by which these earlier estimates were produced was inconsistent over time, and it is inconsistent with the current process under JOLTS.

In 1926, the Metropolitan Life Insurance Company initiated a labor turnover survey (which would eventually develop into the LTS) in order to enable personnel managers in manufacturing plants to assess where their employees stood relative to national benchmarks. In 1929, this project was turned over to BLS, which began collecting monthly data in 1930. BLS started with a sample of 175 large establishments, accounting for 25 percent of manufacturing employment, and, over the next 10 years, expanded the sample size. According to the National Bureau of Economic Research, the calculation of quit rates did not cover all manufacturing workers until 1943, because, prior to that year, it included only production workers.^[4]

While BLS was increasing the sample size for the LTS, some state employment security agencies affiliated with the U.S. Employment Service of the Department of Labor began collecting labor turnover data for their own purposes. BLS began to enter cooperative agreements with state agencies for the joint collection of these data, starting with an agreement with Connecticut in 1954. By 1964, these agreements reached all 50 states and the District of Columbia, with the LTS sample totaling 40,000 reporting establishments in manufacturing and mining.^[5] Before the LTS was discontinued in 1981, there was a brief period during which the survey also collected job openings data, as well as information on some nonmanufacturing industries in certain metropolitan areas.

Nearly two decades later, BLS resumed the collection of information on job openings and labor turnover by initiating JOLTS. This survey, which first released official estimates in December 2000, currently has a sample of about 20,700 nonfarm business and government establishments covering the entire nonagricultural workforce, including workers employed in manufacturing, services, and government. Besides differing in scope, the LTS and JOLTS differ in their definitions of key concepts. The LTS defines quits as (1) terminations of employment initiated by the employee, (2) instances of failure to report for work after being hired (if the employee was previously counted as a new hire), and (3) unauthorized absences from work if, on the last business day of the month, the employee has been absent for more than 7 consecutive calendar days.^[6] The second and third categories likely are very small relative to the first, but precise data on their relative size do not exist. JOLTS defines quits as employees who left their job voluntarily, excluding those who retired or transferred to other locations.^[7] Quit rates are computed by first estimating the number of quits for the entire reference month and then dividing total quits in a sector by employment in that sector. The resulting ratios are multiplied by 100 and converted into percentages.

The JOLTS design is based on a random sample stratified by ownership, region, industry sector, and establishment size class. Surveyed establishments are drawn from a universe of more than 9.4 million establishments compiled by the BLS Quarterly Census of Employment and Wages program. This universe includes all employers subject to state unemployment insurance laws and all federal agencies subject to the Unemployment Compensation for Federal Employees program. Each month, employment estimates are benchmarked (or ratio adjusted) to the strike-adjusted employment estimates of the Current Employment Statistics survey.

To avoid mistaking random fluctuations for genuine trends, one should keep in mind how large a change in quit rates must be in order to be statistically significant. For this purpose, JOLTS publishes, for each year and each series, median standard errors, which are derived from the standard errors of not seasonally adjusted monthly estimates for the previous 5 years. Using these standard errors, one can form confidence intervals around two estimates being compared; if the intervals overlap, the difference between the estimates is not statistically significant.^[8] Alternatively, JOLTS also publishes, on a monthly basis and by series, the minimum month-to-month and year-to-year changes required for statistical significance.

Armed with standard errors, one can assess which recent month-to-month and year-to-year changes are statistically significant. The standard error for the economywide quit rate in 2021 was 0.050, while that for 2020 was 0.046. By creating confidence intervals around monthly quit-rate estimates during these 2 years and by checking whether these intervals overlap, one can easily see that changes of 0.2 percentage points or more are significant at the 5-percent level. The quit rate rose from 1.6 percent in April 2020 to 3.0 percent in November 2021, a gain of 1.4 percentage points; the 95-percent confidence intervals for these rates are more than 1.2 percentage points apart.

Historical data on quit rates

Thus far, the discussion has centered on seasonally adjusted (as opposed to not seasonally adjusted) quit rates. Over the course of a given year, the pattern of quits, just like that of many other labor market phenomena, is affected by seasonal trends. For instance, people enrolled in school may quit their summer jobs at the end of the season in order to return to full-time study. Likewise, individuals thinking of leaving a job may stay on past the end of the year so not to miss out on a holiday bonus. Separating the effects of these seasonal events from trends over time customarily involves presenting seasonally adjusted statistics. Because this section compares seasonally adjusted JOLTS data with historical data, performing a valid comparison requires the use of seasonally adjusted historical data as well.

LTS seasonally adjusted data are available only for the 1959–81 period, only for manufacturing. Although manufacturing was a larger part of the economy in that period than it has been under JOLTS, it still did not account for most of the economy. Yet, it remains possible to better understand current quit rates by comparing historical manufacturing quit rates with 21st-century quit rates in manufacturing and the overall economy (excluding agriculture).

For this comparison, it may be informative to provide some context on manufacturing’s place in the economy in the two eras. According to data from the U.S. Bureau of Economic Analysis, manufacturing was the largest major sector of the economy from 1959 to 1981, accounting for an average of about 24 percent of gross domestic product (GDP).^[9] In contrast, since the beginning of JOLTS in 2000, manufacturing has accounted for half that share, or about 12 percent of GDP.^[10] Because sectors with higher average pay tend to have lower quit rates—a relationship discussed in more detail below—it is also informative to see how manufacturing compares with other sectors in terms of compensation in both periods. From 1959 to 1981, manufacturing compensation per full-time equivalent employee—a broad measure of average pay—exceeded compensation in nonmanufacturing industries by about 20 percent, on average.^[11] Although manufacturing has shrunk as a share of GDP, it remains a relatively high-paying sector, employing workers who have averaged a 15-percent compensation premium since JOLTS began.^[12]

Because the seasonally adjusted quit rates from the LTS are available only for manufacturing, they cannot be used to assess the aforementioned general proposition positing a negative relationship between pay and quit rates across sectors.^[13] For the 2001–20 period, it is possible to use JOLTS data to calculate average annual quit rates for each of 18 sectors and then correlate these rates with the ratio of compensation to the number of full-time equivalent employees. Making this computation and averaging the correlation coefficients over the years in the period, one obtains a coefficient of −0.64 for the case in which the sectors are not weighted by their employment and a coefficient of −0.69 for the case in which the sectors are weighted. (A correlation coefficient of −1.00 indicates a perfect negative relationship.) Thus, there is evidence that higher pay deters quitting, although compensation is far from being the only factor.

Examining the relationship between quit rates in manufacturing and quit rates in the total economy during the JOLTS period can reveal what economywide quit rates would have been back in the 1960s and 1970s, had they been measured by the LTS. Chart 1 displays the quit rates for manufacturing and the total economy from December 2000 (just before the start of the 2001 recession) to December 2021. A few things are notable from the graph. First, quit rates generally rose during the period’s two long expansions, which lasted from November 2001 to December 2007 and from June 2009 to February 2020. Second, in 2021, the economywide quit rates rose to levels not seen in the earlier history of JOLTS, peaking at 3.0 percent. Third, manufacturing quit rates followed a pattern nearly identical to that of economywide quit rates, with a correlation coefficient of 0.90 between the two series (the maximum coefficient is 1.00). Finally, in every month over the period, the manufacturing quit rate was lower than the economywide quit rate.

These relationships, along with the fact that data for manufacturing exist for the 1959–81 period, suggest the following experiment. Assuming that the relationship between manufacturing and economywide quit rates was the same over this period as it has been since the turn of the 21st century, one can estimate what economywide quit rates would have been in the 1960s and 1970s. To do so, one can use linear regression and the following specification:

where QR_E,m is the economywide quit rate in month m, QR_M,m is the manufacturing quit rate in month m, b₀ is the constant, b₁ is the coefficient, and ε_m is the error term. The regression can be run by using the JOLTS data, which provide quit rates for both manufacturing and the total economy. Then, the estimated coefficients and the manufacturing quit rates from the LTS can be used to predict what the economywide quit rates might have been from 1959 to 1981. This calculation is based on the following equation:

Running regression (1) by using information from JOLTS for the period from December 2000 to December 2021, one arrives at the following relationship: 

Plugging the manufacturing quit rates for 1959–81 in equation (3), one can obtain an estimate of what the economywide quit rates would have been during that period. Chart 2 displays the results of this calculation, showing both actual manufacturing quit rates and predicted economywide quit rates. According to these estimates, the highest economywide quit rates were seen in 1973, toward the end of the business cycle expansion that took place between November 1970 and November 1973, with rates hitting as high as 3.3 percent, exceeding the JOLTS high by 0.3 percentage points, or roughly 10 percent. The estimated quit rates also exceeded 3.0 percent in the last year of the expansion from February 1961 to December 1969.

While these estimates suggest that quit rates in the past have reached levels as high as those seen recently, they come with some caveats. Besides the fact that the economywide estimates are based on a simple backward extrapolation of the current relationship between manufacturing and total quit rates, LTS microdata are unavailable, so it is not possible to assess statistical significance. In addition, as noted earlier, quits are defined slightly more broadly in the LTS than in JOLTS, which suggests that one might need to lower the economywide quit-rate estimate. On the other hand, as also noted previously, manufacturing provided higher pay in the 1960s and 1970s than in the 21st century, which indicates a bigger gap between the manufacturing quit rate and the economywide quit rate.

Why did quit rates rise in 2021?

This section evaluates whether a tightening of the labor market can explain the recent rise in quit rates. It also examines which sectors of the economy have contributed the most to that rise.

Measures of labor market slack and rates of resignation

Although, as noted, historical quit rates in the United States have likely reached levels as high as those observed today, it remains unclear why the rates seen over the last year are higher than those recorded in the preceding two decades. One likely reason for this difference is labor market tightness, and there are many possible explanations for why and how the labor market rebounded after the pandemic-induced recession in the first half of 2020. These explanations focus on developments such as the end of lockdowns, the stimulus coming from increased generosity of and enhanced eligibility for unemployment insurance benefits, relief payments to individuals, and increases in food assistance. Other developments that may have caused individuals to leave their jobs (and perhaps the labor force) include the desire of workers to protect themselves and their families from COVID-19, as well as challenges in providing childcare as a result of pandemic-related closures of childcare centers and the widespread use of remote schooling.^[14] Although it is beyond the scope of this article to test each of these explanations individually, they all suggest a tightening of the labor market. It is possible to assess whether the recent rise in quit rates is merely a function of a decline in labor market slack.

It has already been noted that quit rates tend to rise during expansions, when workers’ prospects for finding a better job brighten. With the unemployment rate having returned to low levels, is the recent increase in quit rates merely a function of a tightening labor market?

Addressing this question requires a measure of labor market slack. How this measure should be constructed is an active area of research, and currently no consensus exists on the best measurement approach. As discussed by Katharine G. Abraham, John C. Haltiwanger, and Lea E. Rendell, the unemployment rate has long been used as a measure of labor market tightness, given its evident correlation with the business cycle.^[15] The increased popularity among economists of search-and-matching models—which seek to explain a number of features of the macroeconomy—has increased the importance of considering job openings along with the unemployment rate. In these models, labor market tightness is measured by the ratio of job openings (the number of positions employers wish to fill) to the level of unemployment (the number of people who do not have a job and actively look for work). Thus, the higher the ratio, the tighter the labor market. Abraham, Haltiwanger, and Rendell develop what they term a “generalized measure of labor market tightness,” a measure whose denominator accounts for all potential jobseekers (not just the unemployed) and whose numerator is adjusted for firms’ recruiting intensity.^[16] Regis Barnichon and Adam Hale Shapiro recently looked at nine different measures of slack, testing their ability to predict inflation.^[17]

Again using linear regression, the analysis below examines whether the relationship between quit rates and two measures of labor market tightness or slack—the unemployment rate and the ratio of job openings to the number of unemployed people ^[18]—has changed since COVID-19 began.^[19] Because economic theory does not dictate the form of this relationship, two regression specifications are used. In the first specification, quit rates are a linear function of labor market slack:

Because quits are responding to the level of labor market tightness, the dependent variable in equation (4)—the quit rate—is specified as a function of the lagged value of the measure of labor market tightness, where the lag is 1 month. The second specification allows a more general relationship between slack and quit rates, positing quit rates as a quadratic function of slack:

For each measure, two regressions—one linear and one quadratic—are first run on the entire sample of JOLTS economywide quit rates, to get a sense of the relationship for the entire period from December 2000 to December 2021. The sample is then restricted to the prepandemic period (December 2000–February 2020), and the relationship found for this time span is projected forward, up through the end of 2021. If the regressions underpredict the quit rates, one can infer that the relationship between labor market tightness and quit rates has changed such that, at any given level of labor market tightness, there is now more quitting. Conversely, if the regressions overpredict the quit rates, one can infer that there is now less quitting at given levels of labor market tightness.

Table 1 summarizes the regression results. The table’s first data row (linear specification) in panel A shows that, as expected, when the unemployment rate goes up, the quit rate goes down. The quadratic specification, presented in the second data row, shows a similar relationship, although in this case the positive sign on the squared term indicates that the slope tends to be less negative at higher unemployment rates. The fit for the quadratic specification (R-squared of 0.58) is better than that for the linear specification (R-squared of 0.52). The next two rows of table 1 repeat these specifications, this time for the prepandemic period. The fit is now much better, with R-squared climbing above 0.9 for both the linear and quadratic specifications.

In the linear specification, the predicted quit rates are lower than the actual rates by an average of 0.74 percentage points for all of 2021 and by 0.82 percentage points for the second half of the year. In the quadratic specification, these differences are, respectively, 0.76 percentage points and 0.83 percentage points, showing a similar underprediction. Thus, both specifications suggest that, during the pandemic, the relationship between the unemployment rate and the quit rate has changed such that, at any level of the unemployment rate, the recent period exhibits a higher rate of quitting.

Because this finding could be sensitive to the choice of measure of labor market tightness, it is useful to try the alternative measure mentioned earlier, namely, the ratio of job openings to the number of unemployed people. As shown in panel B of table 1, the results based on this alternative measure for the entire JOLTS period indicate that the quadratic specification is a better fit (R-squared of 0.76) than the linear specification (R-squared of 0.70). As expected, both specifications suggest that when the labor market tightens, the quit rate goes up. When the regression is restricted to the prepandemic period from December 2000 through February 2020, the fit is much improved for the quadratic specification (R-squared of 0.88) and mildly improved for the linear specification (R-squared of 0.74). As with the regression using the unemployment rate as a measure of slack, the coefficients obtained for this shortened period are extrapolated to the period from March 2020 through December 2021, with the ratio of job openings to the number of unemployed people being used to make predictions.

In the quadratic specification, the model underpredicts the rate of quitting by an average of about 0.53 percentage points for all of 2021 and by 0.73 percentage points for the second half of the year. These levels of underprediction are somewhat lower than those obtained for the quadratic specification in the analysis using the unemployment rate as a measure of labor market tightness. The linear model also underpredicts the rate of quitting, although the levels of underprediction in this case average just 0.33 percentage points for all of 2021 and 0.26 percentage points for the second half of that year. Thus, while the levels of underprediction differ somewhat across specifications and measures, they consistently suggest that the relationship between labor market slack and quit rates has changed such that, conditional on the level of slack, there is a higher level of quitting in the recent period.

In sum, the results of the regression analysis suggest that if the relationship between quit rates and labor market slack had not changed, quit rates would still have risen from their April 2020 level of 1.6 percent, albeit not to the heights actually seen. For the linear and quadratic specifications using the unemployment rate as a measure of labor market tightness, the model predicts that quit rates would have risen to 2.2 percent. The linear model based on the ratio of job openings to the number of unemployed people predicts a peak of 2.3 percent. Only in the quadratic model for this measure do the predicted quit rates come close to the actual rates, with the predicted high being 2.9 percent.

Decomposition of the rise in quit rates

The preceding analysis indicates that the quit rates recorded recently are high for the 21st century, even after accounting for the degree of labor market tightness. Given this finding, it is useful to examine which sectors of the economy are contributing most to the rise in quit rates. One standard way to do this is through a decomposition analysis, whereby the change in the quit rate over a given period can be broken into three components: (1) a “within” component attributable to increases in quit rates within sectors, (2) a “between” component attributable to shifts in employment across sectors, and (3) a component attributable to an interaction of the “between” and “within” components. The equation for the decomposition is

where QR is the quit rate, π_s0 is the share of employment in sector s at the beginning of the period, π_s1 is the share of employment in sector s at the end of the period, and S is the number of sectors. The expression in the first set of square brackets is the “within” component, that in the second set of brackets is the “between” component, and that in the third set of brackets is the interaction term. The decomposition is performed for two periods: (1) from the end of the Great Recession in June 2009 to the peak in quit rates in November 2021, and (2) from April 2020 (after the quit rate fell early in the pandemic) to November 2021. Because neither of these periods is very long, one would not expect major shifts in employment across sectors and, thus, a large “between” component. Nonetheless, the decomposition technique can reveal which sectors contribute the most to the “within” component, possibly providing clues about why quit rates have been rising.

Panel A of table 2 displays the decomposition for the period from April 2020 to November 2021, during which the total nonfarm quit rate (see bottom of panel) nearly doubled, from 1.6 percent to 3.0 percent. Besides displaying this overall rate, the table shows quit rates for 19 industry sectors, indicating that, at the beginning of the period, these rates ranged from 0.6 percent for the federal government and durable goods manufacturing all the way up to 3.8 percent for accommodation and food services. As noted earlier, sectors with higher compensation tend to have lower quit rates. During the 19-month period, all sectors saw quit-rate increases, although these increases were quite small for the federal government, state and local government education, and educational services. In fact, the change for these sectors is not statistically significant, nor is it for the information sector.

Calculating the “within” component of the decomposition involves taking the change in quit rates in a sector and weighting it by the sector’s employment share. As seen in the table, the “within” component was responsible for more than 1.2 percentage points of the 1.4-percentage-point increase in the total nonfarm quit rate during the period, suggesting that the “between” component and the interaction term were unimportant. The sectors with the greatest contribution to this component were retail trade, professional and business services, accommodation and food services, and healthcare and social assistance. The “within” components of these four sectors were responsible for about 61 percent of the overall “within” component,^[20] both because these sectors had large increases in their quit rates and because all of them, except accommodation and food services, were relatively large.

Panel B of table 2 displays an identical decomposition for the period from June 2009 to November 2021, during which the total nonfarm quit rate (see bottom of panel) rose from 1.3 percent to 3.0 percent. Although this longer period allows for greater shifts in employment across sectors, the case remains that the “within” component of the decomposition was responsible for nearly the entire quit-rate increase. Over the period, all sectors, including the federal government and educational services, saw notable increases in their quit rates, and these increases were all statistically significant, with the exception of that for other services. The sectors with the greatest contribution to the “within” component (about 66 percent) were the same as those identified in the first decomposition, but their ordering is somewhat different, with accommodation and food services having the largest contribution, followed by retail trade, professional and business services, and healthcare and social assistance.^[21]

This finding is partly consistent with a recent study by the Pew Research Center, according to which most workers who quit their jobs in 2021 cited low pay, no advancement opportunities, and feeling disrespected as the main reasons for their resignations.^[22] Indeed, compensation-based factors are relevant to the low-paying sectors of retail trade and accommodation and food services. However, the other factors identified in the Pew study could apply to any industry sector, and the study did not examine the reasons for the change in quit rates. In the case of healthcare and social assistance, one can also speculate that the COVID-19 pandemic might have made some healthcare-related jobs more stressful, leading to worker burnout and higher quit rates.

Conclusion

This article has offered a broader perspective on the recent rise in quit rates, a phenomenon called the “Great Resignation.” The historical data examined in the article suggest that recent quit rates, while certainly high for the 21st century, are not the highest historically. Nonetheless, the pace of resignations seems to have risen more quickly than one would have expected from labor market tightening alone. Future research should assess alternative explanations for this development, taking into account pandemic-related factors such as increased stimulus payments, health concerns, childcare issues, and changing attitudes toward work. Examining which demographic groups have seen their quit rates rise most quickly might provide clues here.^[23] Future research should also pinpoint what is happening to workers who are resigning for the first time: are they leaving the labor force or moving on to better jobs?