Since the onset of the coronavirus pandemic and economic shutdown, potential misclassification errors in the reported unemployment rate have occurred. In particular, one issue that emerged in the March Employment Situation report and has remained an issue through May, is whether some people who responded to the survey that they were employed but not working should have been classified as unemployed.
The potential misclassification issue has arisen because the number of workers who are “absent from work due to other reasons” has spiked in an unusual way since March 2020. Observers have noted that many of those people should probably have been recorded as “on temporary layoff” and thus be counted as among the unemployed. To the degree that is the case, a more accurate measure of the unemployment rate is higher than the official measure.
The Bureau of Labor Statistics (BLS) categorizes people as employed or unemployed based on how they answer questions about their work status during a single reference week. The agency counts people as employed if they worked during the week or were employed but absent from work due to vacations or illness or “other reasons.” In a typical month, a small fraction of people report being “absent from work due to other reasons.” The misclassification issue has arisen as the survey instruments deployed by BLS to collect data on labor market conditions have largely remained unchanged.
To improve the quality of the data, BLS has taken steps in recent months to improve the accuracy of recorded responses. However, it is important to note that historically, BLS has not edited responses post hoc and has always declined to reclassify respondents, which would amount to interference with the data. Furthermore, BLS has presented the data as collected, provided microdata to researchers, and in each report since March provided details about the potential misclassification.
To explore this issue, we use the monthly Current Population Survey microdata—which is released by the U.S. Census Bureau and BLS and is the input into the Employment Situation reports—to assess how those who were potentially misclassified changed their reported employment status over time. As a first step, we reproduce the numbers released by BLS in the Employment Situation report regarding the official unemployment rate and the share of labor force participants who are “employed and absent from work due to other reasons” (i.e., workers who were potentially misclassified).
We find that many of those individuals who were potentially misclassified in April 2020 had a different reported employment status in May. For example, over those two months, 22 percent of the potentially misclassified workers in April had different responses in May and were then counted as officially unemployed. That we see such a large fraction classified as unemployed in the subsequent month suggests many of those individuals were initially misclassified. The change over time in how individuals are classified also suggests that BLS has indeed taken steps to increase the quality of the data.
Quantifying Potential Misclassification
Following the release of the May Employment Situation report, in which the unemployment rate unexpectedly fell by 1.4 percentage points, some have unfoundedly suggested that BLS deliberately lowered the unemployment rate in May. They have raised the potential misclassification issue to debate not just the level and trend in the unemployment rate, but also the veracity of the Employment Situation reports. This argument is incorrect.
First, BLS has been transparent about this potential misclassification. Since March, when the shutdowns began, BLS flagged the issue and provided an estimate of its size. As BLS wrote in May:
“However, there was also a large number of workers who were classified as employed but absent from work. As was the case in March and April, household survey interviewers were instructed to classify employed persons absent from work due to coronavirus-related business closures as unemployed on temporary layoff. However, it is apparent that not all such workers were so classified. BLS and the Census Bureau are investigating why this misclassification error continues to occur and are taking additional steps to address the issue.
If the workers who were recorded as employed but absent from work due to “other reasons” (over and above the number absent for other reasons in a typical May) had been classified as unemployed on temporary layoff, the overall unemployment rate would have been about 3 percentage points higher than reported (on a not seasonally adjusted basis). However, according to usual practice, the data from the household survey are accepted as recorded. To maintain data integrity, no ad hoc actions are taken to reclassify survey responses.”
Second, regardless of misclassification, the unemployment rate did fall from April to May. According to BLS, the potential misclassification issue could have led the actual unemployment rate in April to be roughly 5 percentage points higher than officially reported—larger than the effect in May. In other words, reclassifying such individuals as unemployed in both months would suggest an even larger decline in the unemployment rate in May than was officially reported.
Third, BLS makes the detailed survey data available to researchers, making it very hard to manipulate the statistics and very easy to replicate the results. Figure 1 shows the monthly share of the U.S. labor force that was employed but absent from work for a specified reason (e.g., vacation, illness, child care, other family obligations, parental leave, labor dispute, weather affecting their job, school/training, or civic/military duty), and the share that was employed but absent from work for “other reasons.”
It is the last category of respondents—those “employed but absent from work for ‘other’ reasons”—who are potentially misclassified. In the first two months of the year, when the economy was still expanding, a small portion of the labor force were absent from work for other reasons. In March, 1.3 percent of the labor force was categorized as employed and absent from work for other reasons. That percentage rose in April to 5.2 percent but then in May fell partway back to 3.4 percent.
Our analysis affirms that BLS flagged the potential misclassification issue in the first month in which its magnitude became significant (March). We also replicate the number and distribution of unemployed, employed absent from work for other reasons, and employed absent from work with a specified reason as published in the monthly Employment Situation reports. We find no basis for ill-considered claims that BLS has improperly handled either the underlying data or the subsequently produced statistics.
Employer-Employee Relationships during the Shutdown
Following individuals for consecutive months, we show the employment status in the subsequent month for those who were employed and absent from work for other reasons in the prior month (see figure 2). For example, the bottom horizontal stacked bar shows the distribution of employment statuses in February for those who said that they were employed but not at work for other reasons in January. Among those who reported they were employed but not at work for other reasons in January, only 4.7 percent reported they were unemployed in February.
Among those who were potentially misclassified in April and who are observed in the May data:
35 percent were working for pay in May; these respondents were classified as employed in both months and this evidence suggests that they maintained their employer relationship.
22 percent were classified as unemployed in May; that change in classification could have reflected a change in employment status or improved survey implementation geared toward minimizing misclassification.
29 percent were still labeled as absent for “other reasons” in May.
The remaining 15 percent were in other categories, mostly out of the labor force.
Compared to pre-pandemic months, the share switching status to employed was lower in May, the share switching to unemployed was higher, and the share remaining absent for other reasons is quite elevated. This suggests that BLS has been correct to point to potential misclassification of these respondents. Nonetheless, the evidence is inconclusive on how to systematically view those characterized as employed but absent from work for other reasons in a given month. That we see such a large fraction classified as unemployed in the subsequent month suggests many of those individuals were initially misclassified. But a notable fraction reported being employed and working in the subsequent month, suggesting that if those individuals were indeed unemployed in the prior month, then that was a temporary state.
We compare transitions from April to May for the employed, the unemployed, and those potentially misclassified in table 1. Despite demographic differences, the potentially misclassified appear to resemble those who are classified as unemployed on a temporary layoff. Both groups show similar changes in their employment status over the two months. For example:
Among those potentially misclassified in April:
35 percent were classified as employed in May; and,
51 percent were classified as either remaining “employed but absent for other reasons” or unemployed in May.
Among those classified as on temporary layoff in April:
31 percent were classified as employed in May; and,
56 percent were classified as either remaining “employed but absent for other reasons” or unemployed in May.
Nonetheless, some of those that are potentially misclassified are likely correctly classified as employed but not at work; the distribution of employment status changes for the potentially misclassified are also comparable to those who were not at work but for a listed reason.
Given the abrupt changes in the labor market, the number of workers employed but not at work for other reasons may indeed have risen. Knowing how large this group has been is difficult as the list of reasons people may be absent did not expand to include new options unique to the COVID-19 policy response. For example, individuals may be having a difficult time describing their status if their employer has reduced their hours to zero but is paying them, either by using a loan through the Paycheck Protection Program or by benefiting from the employee retention tax credit. These results help show why BLS flagged the issue of potential misclassification.
Another way to look at the paired months is to see what employment status a respondent had before becoming potentially misclassified. Figure 3 shows the employment status in April of those who were potentially misclassified in May. In both months, 45 percent were consistently coded as employed and absent from work for other reasons. In April the remainder of those potentially misclassified in May were split across employed (27 percent) unemployed (17 percent) and not in the labor force (11 percent). (We note that rates of misclassification are the same among those who do not have a paired observation, because it is their first month in the survey for 2020, as those who do.)
Evidence from paired months highlights the challenge of correctly classifying people on temporary layoff in a single month. This is especially true given the extraordinary month-to-month churn in the labor market currently, the unique circumstances of stay-at-home orders that have disrupted employer-employee relationships, and the challenges in conducting a survey during this period while maintaining fidelity to data collection and dissemination. In the end, a more accurate measure of the unemployment rate would be higher than the reported rate by close to the share (but not all of the share) that are absent from work for other reasons. Other research on this potential misclassification issue has reached this same conclusion.
The fact that the unemployment rate is being reported with error is obviously not ideal. The unemployment rate is one of the most important statistics the government produces, and it is typically considered a highly accurate and little-revised data point. That being said, BLS has been quite transparent about the issue, first raising it in March, and prominently giving detailed data about the degree of misclassification on the day of the Employment Situation report. In addition, the broad economic story has been correct. Whichever way one interprets the potentially misclassified individuals, the unemployment rate rose in March, spiked in April, and fell slightly in May. Further, the level of detail available in the microdata can help researchers better understand the nature of the misclassification. The challenges around accurately reporting the data highlight the importance of adequate funding and independence of the federal statistical agencies.
In 2017, The Hamilton Project and the American Enterprise Institute jointly published “In Order That They Might Rest Their Arguments on Facts: The Vital Role of Government-Collected Data” which provides evidence that the objective and impartial data collected by the federal statistical agencies provides value to businesses, policymakers, and families. This analysis in this blog affirms that in times of changing economic fortunes and under difficult circumstances for data collection, the federal statistical agencies, including BLS, continue to transparently and consistently produce vital statistics and to course-correct without malpractice or malfeasance.
The authors would like to thank Stephanie Aaronson, Martha Gimbel, Erica Groshen, and Ernie Tedeschi for thoughtful comments as well as Emily Moss and Jenn Umanzor for excellent research assistance.