An official website of the United States government
Labor is an important input to the production process, and the correct measurement of hours worked is critical for estimating productivity growth. In November 2022, the U.S. Bureau of Labor Statistics (BLS) will introduce a new method for measuring hours worked by employees for its major-sector productivity data.1 This new method uses all-employee hours from the BLS Current Employment Statistics (CES) survey, also known as the establishment survey, as its main data source. The CES all-employee hours series was first introduced in March 2006 as a research series and became the official BLS hours series in 2010. The new method for estimating hours worked improves on the current method, which uses the CES production-employee data and relies on several assumptions that no longer hold.
The BLS productivity program uses the CES survey as its primary source of hours data rather than the Current Population Survey (CPS), also known as the household survey, because (1) output data come from establishments and thus hours data from establishments are more likely to be consistent with the output data, (2) industry coding in the establishment survey is more accurate and more consistent with that of the output data, (3) the larger sample in the establishment survey provides better industry coverage and reduces the variability of industry-level estimates, and (4) employment estimates from the establishment survey are benchmarked annually to the employment universe by industry.2
A drawback of using the CES hours data is that the CES survey collects data on hours paid, whereas hours worked is the appropriate concept for measuring productivity. The CES hours-paid data include paid leave and exclude off-the-clock hours. For salaried and commission-only employees (henceforth referred to as salaried employees), the CES survey questionnaire explicitly instructs respondents to report hours based on their standard workweeks. Therefore, it is necessary to adjust the CES hours-paid data to estimate the number of hours worked by removing paid time off and adding in off-the-clock hours. The new method uses data from two other BLS surveys—the National Compensation Survey (NCS) and the CPS—to make these adjustments.
In the section that follows, we describe the current method for estimating hours worked. The next section explains why it is important to make the change to the new method. The section after that describes the new method that will be used to adjust the CES all-employee hours series to account for paid time off and off-the-clock work. The next section compares estimates of hours worked and labor productivity that were constructed using the current method with research estimates that use the new method over the 2006–21 period. The final section concludes.
The BLS productivity program last changed its method for estimating hours worked in 2004 and implemented that method beginning with data from 1979.3 The current method uses three sources of data to estimate hours worked by wage and salary employees. The main source is the CES survey, which is a monthly payroll survey that covers about 689,000 establishments. The CES survey collects data on employment and total hours paid for all employees and for production employees. Employment includes all employees who were paid (worked or were on paid leave) during the pay period that includes the 12th of the month. Total pay-period hours are converted to total weekly hours by using conversion factors that vary depending on the number of days in the pay period.4 The CES program calculates average weekly hours paid as total weekly hours paid divided by employment.
Until 2006, the CES survey collected hours data only for production and nonsupervisory employees (henceforth referred to as production employees).5 Therefore, in the current method, it is necessary to estimate hours of production and nonproduction employees separately. The first step is to convert CES production-employee hours to an hours-worked basis by using production-employee hours-worked-to-hours-paid (HWHP) ratios from the NCS, which is an establishment survey that collects data on all forms of compensation, including paid leave.6 The NCS HWHP ratios adjust hours paid for paid time off (annual leave earned, including paid holidays, and average annual sick leave taken). More formally, production-employee hours worked are calculated as follows:
where AWHPCES is quarterly average weekly hours paid for production employees from the CES survey, HWHPPNCS is the HWHP ratio for production employees from the NCS, and EMPPCES is quarterly average production-employee employment from the CES survey.7 Multiplying the weekly hours estimate by 52 “annualizes” the data so they are comparable with annual hours estimates. BLS applies these HWHP ratios at the three-digit industry level of the North American Industry Classification System (NAICS).8
To estimate nonproduction-employee hours worked, BLS uses additional data from the CPS, which is a monthly survey of about 60,000 households that collects demographic and job-related data for civilians ages 15 years and older. The CPS collects information about actual hours worked during the week that includes the 12th of the month.9
The first step is to classify employees as either production or nonproduction by using industry and occupation codes. Next, data on actual hours worked are used to calculate the ratio of nonproduction-employee average weekly hours to production-employee average weekly hours (henceforth referred to as the NPPCPS ratio).10 The NPPCPS ratio is multiplied by production-employee average weekly hours. Thus, nonproduction-employee hours are calculated as follows:
where EMPallCES is CES employment for all employees, and (EMPallCES – EMPPCES) is nonproduction-employee employment from the CES survey.
The NPPCPS ratio captures some of the variation in off-the-clock hours for nonproduction employees because it uses actual hours worked. But the current method misses all the off-the-clock hours worked by salaried production employees. This is an important omission because production employees make up 80 percent of wage and salary employment, and 30 percent are salaried, which translates into about 24 percent of total wage and salary employment. In contrast, salaried nonproduction employees account for only 14 percent of wage and salary employment (about 20 percent of wage and salary employment times about 70 percent salaried).11
Hours worked for all employees are calculated as the sum of production- and nonproduction-employee hours worked:
There are three compelling reasons to switch to the CES all-employee series as the main source of hours data: (1) the all-employee data are of higher quality than the production-employee data, (2) the ability to replicate the CES production-employee concept using CPS data has diminished, and (3) there is new evidence of bias in the NPPCPS ratio. We discuss the three reasons in more detail in the paragraphs that follow.
First, the CES all-employee data are more reliable than the CES production-employee data because the production-employee employment estimate is more susceptible to sampling error and nonresponse bias. Unlike total employment, production-employee employment is not estimated directly, nor is it benchmarked to the Quarterly Census of Employment and Wages (QCEW).12 CES survey respondents report the number of production employees in goods-producing industries and the number of nonsupervisory employees in service-providing industries. Within each estimating cell (defined by industry), production-employee employment is calculated as the product of total employment and the ratio of production-employee employment to total employment for establishments in that cell. Both are estimated from the sample, but only total employment is benchmarked to the QCEW. In addition, nonresponse is significantly higher for production-employee employment than for all-employee employment. Thus, estimates of production-employee employment could be biased if the production-employee ratio is different for responding and nonresponding establishments—even though all-employee employment totals are correct (after benchmarking).13 There is also evidence that some respondents have difficulty classifying employees as production or nonproduction employees.14 Although there have been no recent studies to determine whether this classification issue has gotten worse over time, it was deemed an important factor by the CES program in its move to collect all-employee hours.15
The second reason for the method change is that the ability of the NPPCPS ratio to replicate the CES production-employee concept by using industry and occupation data has diminished. Chart 1 shows the percentage of wage and salary employees who are classified as nonproduction employees in both the CES survey and the CPS. From 1994 through the mid-2000s, the two series track each other closely. In both series, the percentage of nonproduction-employees ranged from 18.2 to 20.4 percent. The largest differences were in the 1.7- to 1.9-percent range, and there was no apparent trend. However, beginning around 2006, the two series began to diverge. The CES series reached its peak of 19.2 percent in early 2004, but then it declined by 1.7 percentage points, to 17.5 percent in the fourth quarter of 2007, with little movement the ratio since then. In contrast, the percentage of nonproduction employees in the CPS started increasing in 2006—from about 18.5 percent to about 21.0 percent by the end of 2019. Thus, the two series went from being within 2 percentage points of each other in the pre-2006 period and tracking each other fairly closely to diverging and differing by more than 3 percentage points in the post-2006 period.
There have been changes to both surveys that could have contributed to this divergence. The CES program introduced probability sampling by industry and changed to a state-based design between 2000 and 2003, with both changes making the sample more representative and improving data quality.16 In January 2003, the CPS adopted the 2002 U.S. Census Bureau occupational classification system that included additional “first-line supervisor” codes.17 There is no way to determine whether either of these changes is responsible for the divergence, but it is clear that the two series no longer track each other closely.
The third reason for adopting the CES all-employee hours data is that there is new evidence of bias in the NPPCPS ratio. The current method’s reliance on the NPPCPS ratio implicitly assumes that any CPS reporting bias in average weekly hours worked is the same for both groups.
After the NPPCPS ratio method was developed, BLS introduced the American Time Use Survey (ATUS), which provides an avenue for testing bias in hours reports in the CPS.18 In several different studies, Harley Frazis and Jay Stewart compared CPS hours data with hours data from the ATUS and found that reported hours in the CPS data were accurate, on average.19 However, in their 2004 study, Frazis and Stewart also found variation in reporting accuracy across demographic groups. In particular, they found that workers with more education overstated their hours, while those with less education understated their hours. This suggests that the NPP ratio could be biased upward because, on average, production employees have less formal education than nonproduction employees. In addition, Lucy P. Eldridge and Sabrina Wulff Pabilonia found that nonproduction employees are more likely to bring work home from their workplaces, which could make it more difficult for them to recall their hours worked.20
To examine the bias in the NPPCPS ratio, we compare hours ratios calculated from CPS and ATUS data. We consider the ATUS estimates to be the more accurate hours measure, given that responses are less likely to be subject to recall bias and aggregation bias.21 Table 1 shows that there is an upward bias in the NPPCPS ratio in most years and, more importantly, that the bias varies over time and over the business cycle. The new method, which uses the CES all-employee hours data and therefore does not need the NPP adjustment, addresses these three issues.
|Year||Current Population Survey|
|American Time Use Survey|
Note: Ratios of nonproduction-employee average weekly hours to production-employee average weekly hours (NPP) compare average weekly hours of nonproduction employees to production employees. The CPS-based NPP ratios were calculated by using the basic monthly CPS. The ATUS-based NPP ratios were calculated by using diary days that fell within CPS reference weeks and are corrected for sample composition and rotation-group bias. All ratios are calculated using the main job only. The calculations of hours for the ratios follow the method used by Harley Frazis and Jay Stewart in “Comparing hours worked per job in the CPS and the ATUS,” Social Indicators Research, vol. 93, no. 1 (August 2009), pp. 191–95; and “Why do BLS hours series tell different stories about trends in hours worked?” in Katharine G. Abraham, James R. Spletzer, and Michael J. Harper, eds., Labor in the New Economy, National Bureau of Economic Research: Studies in Income and Wealth, vol. 71 (Chicago, IL: University of Chicago Press, 2010), pp. 343–72. CPS = Current Population Survey; ATUS = American Time Use Survey.
Source: U.S. Bureau of Labor Statistics.
The new method uses the CES all-employee hours data and therefore eliminates the need to separately estimate hours worked by production and nonproduction employees.22 However, like the current method, the new method requires the adjustment of CES hours-paid data to an hours-worked concept.23 The new method can be expressed as follows:
The ideal hours-worked-to-hours-paid adjustment ratio, HWHP*, compares hours worked to hours paid, where hours worked include paid hours worked and hours of off-the-clock (OTC) work, and hours paid include paid hours worked and hours of paid time off (PTO).24 The ratio can be written as
Equation (5) can be decomposed into separate PTO and OTC adjustment ratios:
The first term is the PTO adjustment, or PTO ratio. It is a ratio of paid hours worked to hours paid and is conceptually the same as the NCS HWHP ratio used in the current method. The NCS calculates paid hours worked as hours paid minus hours of PTO. The second term is the OTC adjustment, or OTC ratio, which is a ratio of hours worked (paid hours worked plus OTC hours) to paid hours worked.
In the new method, the PTO ratio is constructed from NCS data as is currently done, except that it is calculated for all employees rather than for production employees. Because the NCS PTO ratio is calculated from establishment data, we assume that estimates of the level of PTO hours are more accurate than those that would be obtained from CPS data. And because the NCS PTO data are based on annual leave earned and average sick leave taken, the NCS PTO ratio is free of seasonal variation.25
The OTC ratio is constructed from CPS data. Note that, even though we estimate the PTO ratio from the NCS and the OTC ratio from the CPS, the ratios are constructed to be consistent with each other. That is, the denominator in the OTC ratio is the same (at least conceptually) as the numerator in the PTO ratio. This gives us26
To estimate the new HWHP ratio for all employees, the new method focuses on the estimation of the OTC ratio using CPS data.
Before turning to the details of the calculations, we describe the data issues and provide a brief outline of our approach.27 The numerator of the OTC ratio is hours worked. The CPS collects data on actual hours worked directly, but it does not collect information on paid hours worked, the variable in the denominator of equation (7). This variable can be estimated by combining other information from the CPS with a few reasonable assumptions.
First, we assume that hourly paid workers are paid for all of the hours they work (i.e., they do not work off the clock).28 Because the CPS collects data on hourly status only for main jobs, our second assumption is that workers on their second jobs are paid on an hourly basis (i.e., no OTC work).29 Thus, for hourly employees, paid hours worked are assumed to be equal to actual hours worked.30 The CPS collects hourly status data as part of its earner-study questions, which are asked only in the outgoing rotations (months in sample 4 and 8). Thus, it is necessary to impute hourly status for the other six rotation groups.31
For main jobs in which the employee is both full time and salaried, we estimate paid hours worked as paid hours minus hours of PTO, which is consistent with the NCS concept.32 The CPS collects information about usual hours worked on each job, but this may not correspond to hours paid for salaried employees. For example, a person may usually work 45 hours per week even though she or he is paid for 40 hours. We calculate time off relative to hours paid. If this employee worked 36 hours, time off would be 4 hours rather than 9. As to whether the employee was paid for the time off, the CPS collects this information only if the employee did not work during the reference week. Thus, in situations in which the employee both worked and took time off, there is no information about whether the time off was paid.
The new method addresses these shortcomings by using variables from the CPS to impute the missing data. We start by assuming in most cases that employees who are full time and salaried on their main job are not paid for more than 40 hours per week, a standard workweek.33 Next, we compare actual hours worked with usual hours paid (usual hours worked topcoded at 40) to determine whether the person took time off. We then estimate the probability that time off was paid and use these predicted probabilities, prob(PTO), to adjust time off to PTO. Once we have estimates of hours paid and hours of PTO, calculating paid hours worked and the OTC ratio is a straightforward process. In the subsections that follow, we describe the method for imputing hourly-paid status and estimating paid hours worked (hours paid minus hours of PTO) for full-time, salaried employees.
Given our assumption about hourly employees, we need to know whether each CPS respondent is paid hourly.34 Because this variable is available only in the outgoing rotation groups of the CPS (one-fourth of the sample and only for main jobs), we must estimate the probability that the worker is paid hourly (prob(hourly)) for the rest of the sample.
To estimate prob(hourly), we use a random-forest algorithm.35 This algorithm is applied to a “training dataset” that has no missing values for the dependent variable. We used the outgoing-rotation-group data, which has information about hourly/salaried status, as the training dataset to generate the predicted probabilities. Once the algorithm has been trained, the results are used to generate predicted values in the target dataset (the six rotation groups for which data on hourly/salaried status are not collected).
The random forest is composed of repeated decision trees that use random subsets of data generated by the algorithm. Each node of a given tree uses one predictor of a random subset of the predictor variables (known as “features” in machine learning language).36 For each observation, each decision tree will make a prediction. The algorithm then combines these individual decision trees into a “forest.” A key feature of the random-forest algorithm is that, although the predictions from individual trees tend to be poor, the prediction from the forest is quite accurate.
The final prediction for each observation in our training dataset is the average of these predicted probabilities. Predicted probabilities (prob(hourly)) for observations in the target dataset are generated from similar observations in the training dataset. The simplest way to interpret prob(hourly) is that it is equal to the fraction of people represented by the observation that were paid on an hourly basis. To illustrate, an observation that has a sample weight of 2,400 represents 2,400 workers. If prob(hourly) = 0.75, 1,800 of the people represented by the observation were paid hourly, and the other 600 were not.
For full-time, salaried employees, we estimate hours paid and hours of PTO, which are needed to calculate the denominator of the OTC ratio (paid hours worked). To estimate hours paid, we start by topcoding usual hours at 40, which is the standard workweek in most industries.37 For respondents who report “hours vary” on their main job, we assume usual hours worked and paid are equal to 40 if the respondents report that they usually work full time.38 We then assume that hours paid are equal to the topcoded values of usual hours worked. Next, for each job, we calculate time off as the difference between usual hours paid and actual hours worked:
If actual hours worked are greater than usual hours paid, then time off is set to zero. Determining whether the respondent was paid for the time off is more complicated. If respondents report that they were employed but not at work, the CPS asks if they were paid for the time off. If the respondents took time off but worked for part of the week, they are not asked if they were paid for their time off. For these observations, we must estimate the proportion of time off that was paid. As was done for prob(hourly), we use the random-forest algorithm to generate predicted probabilities. The algorithm is trained on the subset of CPS respondents who are asked if their absence from work was paid. Thus, hours of PTO for an observation are equal to
where prob(PTO) is estimated for those not asked about paid absences and is equal to either 0 or 1 for those who were asked. As was done for prob(hourly), prob(PTO) is estimated using the random-forest algorithm and is equal to the fraction of people represented by the observation that was paid for time off.39
The denominator of the OTC ratio is paid hours worked. By assumption, paid hours worked are equal to actual hours worked for all part-time employees and full-time employees who are paid hourly. For full-time employees who are paid a salary, paid hours worked equals hours paid minus hours of PTO.40 Thus, for all full-time employees, paid hours worked can be written as
The OTC ratio is calculated by (1) summing actual hours worked over all observations, where each observation is a job; (2) summing paid hours worked over all observations; and (3) dividing (1) by (2):
This OTC ratio is seasonally adjusted and then multiplied by the NCS PTO ratio to give us the new HWHP ratio for all employees.
The final step is to multiply the new HWHP ratios by the CES all-employee hours paid data as in equation (4). Hours worked for employees are calculated at the three-digit NAICS level at a quarterly frequency.41 Because the CPS sample size varies from month to month, we tested whether constructing OTC ratios from 3 months of data produces reliable results. To do this, we estimated the sample mean percent margin of error corresponding to a 90-percent confidence interval by resampling the microdata underlying both the ratio’s numerator and its denominator with a nonparametric bootstrapping algorithm and verified that this value fell below 10 percent in both cases.42 Thus, we are confident in the reliability of the quarterly estimates.
In this section, we examine how the research hours series constructed using the new all-employee method compares with the series constructed using the current method.43 We first compare hours worked for all employees in the private nonfarm sector to show the impact of adjusting CES all-employee hours paid for PTO and OTC hours.44 Next, we discuss the impact on measures of hours worked by employees in 14 major industries. Finally, we show the impact of the new hours method on measured productivity in the nonfarm business sector.45
Chart 2 shows the impact of the different adjustments on total hours worked for employees in the private nonfarm sector.46 All series have the same general trend, which is not surprising because changes in hours are driven primarily by changes in employment and all series use the same CES employment data. However, the series differ in levels. The hours-worked levels for the new series are about 3.3 percent higher than those calculated with the current method over the 2006–21 period. This difference declines from about 3.3 percent in 2006 to about 2.7 percent in 2021, although there is quarter-to-quarter variation around that downward trend.
The PTO ratio adjustment removes paid time off from CES hours paid and is seen in chart 2 as the difference between the graphs of “Hours paid, CES,” and “Paid hours worked.” The “Paid hours worked” series is, on average, about 7.2 percent lower than “Hours paid, CES,” series, which implies that the PTO ratio is about 0.928. Over our sample period, the PTO ratio declined gradually from 0.929 in 2006 to 0.925 in the fourth quarter of 2019. The PTO ratio decreased sharply in the first quarter of 2020, to 0.917, and then increased to 0.922 in the second quarter of 2020. The sharp decline in the first quarter was likely due to the massive job losses in the leisure and hospitality sector and other low-wage sectors in which paid leave is less prevalent.
The impact of the OTC adjustment is seen in chart 2 as the difference between “Paid hours worked” and “Hours worked, new.” The OTC ratio averaged 1.043 over the entire sample period and exhibited a fair amount of quarter-to-quarter variation. It fell from 1.049 in 2006 to 1.040 in the fourth quarter of 2019. Between the fourth quarter of 2019 and the second quarter of 2020, it fell from 1.040 to 1.031. The ratio remained in the range of 1.034 to 1.037 through the end of 2021. The sharp decline in the OTC ratio coincided with the coronavirus disease 2019 (COVID-19)-related shutdowns, which we would expect to decrease the need to work extra hours.
The average OTC ratio of 1.043 may be higher than one would expect given that salaried employees are only 39 percent of all employees and would seem to imply that salaried employees work a lot of OTC hours. But it is important to recognize that the ratio is calculated relative to paid hours worked—not paid hours. A simple numerical example will provide some intuition. Suppose that full-time salaried employees are paid for 40 hours per week, that they work 41 hours on average, and that they represent 39 percent of wage and salary employees.47 If OTC hours are calculated relative to the standard 40-hour workweek, then the aggregate OTC ratio would be about 1.01 = (1 + ((41–40)/40) × 0.39). However, the OTC ratio we construct is calculated relative to paid hours worked to be conceptually consistent with the PTO ratio calculated by using NCS data. Under the same assumptions, but using paid hours worked (37.1 = 40 × 0.928) in the denominator, the aggregate OTC ratio is 1.041 = (1 + ((41–37.1)/37.1) × 0.39).
Next, we consider differences in growth rates. Chart 3 shows the long-term growth in aggregate hours worked for employees, as measured by the current series and the new series. Both series are expressed as indexes with a base period of second quarter 2006. The two series are nearly identical over the 2006–19 period, indicating that long-run growth rates are similar. For the business cycle spanning from the fourth quarter of 2007 to the fourth quarter of 2019, the current hours series had an annual average growth rate of 0.8 percent, while the comparable rate for the new series was 0.7 percent. (See the first line of table 3, below.) Since the trough of the COVID-19 recession in the second quarter of 2020, hours worked calculated with the new method have grown slightly slower than hours worked calculated with the old method, which reflects the impact of the new OTC ratio.
Chart 4 shows the annualized quarterly growth rates for the two hours-worked series from the third quarter of 2006 to the fourth quarter of 2019. We exclude the 2020–21 data so we can more easily see differences in the quarterly growth rates that would otherwise be obscured by the enormous declines that occurred during the early stages of the COVID-19 pandemic. Even though the growth rates of the two series track each other closely, in most cases, the quarterly growth rates differ in any given quarter. As we can see in chart 5, most of the differences in the quarterly growth rates between the new and old hours series are less than 1 percentage point (annualized). Only six quarters differ by more than 1 percentage point. Some of the largest differences occurred around the Great Recession (fourth quarter 2007 to second quarter 2009) and the COVID-19 recession (fourth quarter 2019 to second quarter 2020). For example, in the first quarter of 2010, in the aftermath of the Great Recession, the growth rate of the new hours series was 1.3 percentage points higher than that of the current hours series. Similarly, in the second quarter of 2020, during the COVID-19 recession, the growth rate of the old hours series was 2.3 percentage points higher than that of the new hours series, while after the pandemic recession in the fourth quarter of 2021, the growth rate of the new hours series was 2.1 percentage points higher than that of the old hours series. Thus, the new series tells a somewhat different story about quarter-to-quarter changes, especially around turning points in the business cycle.
Thus far, our comparison has focused on the aggregate private nonfarm sector. However, examining industry detail is also important for two reasons: (1) BLS builds its aggregate hours estimates from industry data, and (2) BLS publishes productivity estimates at the industry level. Charts 6a and 6b show the OTC ratios for major industries in the goods-producing and service-providing sectors, respectively. In all major industries, the ratios are either falling slightly or steady over the period. Except for nonfarm natural resources, OTC ratios are in the 1.04–1.05 range in 2006. They fall slightly to the 1.03–1.04 range by 2021, although the range widens during and after the Great Recession. The OTC ratio for nonfarm natural resources is substantially higher, falling from 1.11 in 2006 to 1.08 in 2021. In the service-providing industries, wholesale trade and transportation and warehousing have the highest OTC ratios for most of the period, while retail trade, education and health services, and leisure and hospitality have the lowest OTC ratios.
In charts 7a to 7n, we compare the different employee hours series for the 14 major industries. To make it easier to compare the differences across industries and determine which industries contribute the most to the aggregate difference in levels, the range of the scales shown on the vertical axes in these charts are scaled so that the difference between the maximum and minimum values equals 14. We see that the hours worked levels calculated with the new method are noticeably higher than those calculated with the old method in most of the service industries. The two hours measures are about the same in durable and nondurable goods manufacturing, as well as in utilities. The new measure is only slightly higher in natural resources, information, and construction. These differences show the substantial number of off-the-clock hours that were previously unaccounted for.
Table 2 shows the percent difference between the two series for the fourth quarter of 2021 and reveals stark differences across industries. The smallest differences occurred in durable and nondurable goods manufacturing, both of which were less than 0.5 percent in absolute value.48 By contrast, the differences in wholesale trade, transportation and warehousing, utilities, information, and other private services (excluding households) were all 4 percent or more.49 The new hours series is 1.3 percent higher than the old hours series in construction, in which approximately 42 percent of employees are nonhourly workers.
|Industry sector||Percent difference|
Nonfarm natural resources
Durable goods manufacturing
Nondurable goods manufacturing
Transportation and warehousing
Professional and business services
Education and health services
Leisure and hospitality
Other private services (excluding households)
Source: U.S. Bureau of Labor Statistics.
As can be seen in table 3, the growth rates over the 2007–19 period show that, in all major industry sectors, the difference in the long-run growth in hours between the old and new estimation methods was less than 0.9 percentage point in absolute value.
|Industry sector||Business cycle, fourth quarter 2007 to fourth quarter 2019||Recession period, fourth quarter 2007 to second quarter 2009||Post-recession period, second quarter 2009 to fourth quarter 2019|
Nonfarm natural resources
Durable goods manufacturing
Nondurable goods manufacturing
Transportation and warehousing
Professional and business services
Education and health services
Leisure and hospitality
Other private services (excluding households)
Note: The National Bureau of Economic Research determined that a peak in economic activity occurred in the fourth quarter of 2007, a trough occurred in the fourth quarter of 2009, and a peak occurred in the fourth quarter of 2019, representing a full peak-to-peak business cycle.
Source: U.S. Bureau of Labor Statistics.
Finally, we look at the impact of using the new hours method on labor productivity in the nonfarm business sector. (See chart 8a.) Again, long-run growth is similar for the two series over the 2006–21 period. However, during the COVID-19 recession, productivity grew at a faster pace when calculated with the new hours method. Charts 8b and 8c show the impact of the new method for durable and nondurable goods manufacturing. Long-run productivity growth is similar, with the lines representing the new and old methods being nearly identical throughout the period, although we again observe some differences in quarter-to-quarter growth.
In November 2022, the BLS productivity program will introduce a new method for estimating hours worked that uses the CES all-employee hours data, which were first published in March 2006. A key advantage to using the all-employee data is that it is no longer necessary to estimate the hours of production or nonproduction workers separately. In addition, the new method does a better job of adjusting CES hours-paid data to an hours-worked concept. The new method still uses NCS data to adjust for paid time off, but it uses CPS data to make an additional adjustment for off-the-clock hours. This latter adjustment accounts for hours worked by full-time, salaried employees beyond their paid work hours.
The new method for measuring hours is a notable improvement over the current method, especially in its incorporation of off-the-clock hours, but it does not result in substantial revisions to productivity estimates. The new method produces results that are similar, but not identical, to those produced with the current method. Both series have nearly identical long-run growth rates, but total hours are about 3.3 percent higher, on average, with the new method. The main difference between the new and old data series is that they tell somewhat different stories about quarter-to-quarter growth of hours worked.
Lucy P. Eldridge, Sabrina Wulff Pabilonia, Drake Palmer, Jay Stewart, and Jerin Varghese, "Improving estimates of hours worked for U.S. productivity measurement," Monthly Labor Review, U.S. Bureau of Labor Statistics, October 2022, https://doi.org/10.21916/mlr.2022.27
1 There will be no change to the measurement of hours worked by the self-employed and unpaid family workers.
2 The Current Employment Statistics (CES) survey sample is 6 times larger than the Current Population Survey (CPS) sample and is stratified by industry; the CPS sample is not stratified by industry. CPS sample weights are benchmarked to match the annual census population controls, which are based on the most recent decennial census population count supplemented with birth and death data and estimates of net international migration.
3 This method is linked to historical data. Prior to 2004, the U.S. Bureau of Labor Statistics (BLS) adjusted the hours of production employees in the manufacturing sector by using the ratio of office to nonoffice worker hours extrapolated from a survey that was last conducted in 1978. In the nonmanufacturing sector, OPT assumed that supervisory employees worked the same hours as nonsupervisory employees. For details, see Lucy P. Eldridge, Marilyn E. Manser, and Phyllis F. Otto, “Alternative measures of supervisory employee hours and productivity growth,” Monthly Labor Review, April 2004, pp. 9–28, https://www.bls.gov/opub/mlr/2004/04/art2full.pdf.
4 For monthly and semimonthly pay periods, this will depend on the number of days in the month.
5 The CES survey instructs respondents to classify employees as production or nonproduction employees in goods-producing industries and as nonsupervisory or supervisory employees in service-providing industries. This results in inconsistencies in how some employees are classified. For example, an accountant who does not supervise other employees would be classified as a nonproduction employee if employed in a goods-producing industry but as a nonsupervisory employee in a service-providing industry. Moreover, a record analysis survey conducted in 1981 showed that respondents do not always follow CES instructions when classifying their employees. See Employer Records Analysis Survey of 1981: Final Report (U.S. Bureau of Labor Statistics, 1983).
6 One of the primary uses of the National Compensation Survey (NCS) data is to construct the Employment Cost Index. Prior to 2000, the hours-worked-to-hours-paid (HWHP) ratio was estimated using data from the now-discontinued Hours at Work Survey. For more information, see the NCS page at https://www.bls.gov/ncs/home.htm.
7 The seasonally adjusted CES data are published monthly. The quarterly average weekly hours estimate is constructed by averaging total hours for the 3 months of the quarter and dividing that average by the corresponding average number of employed production employees. Quarterly employment is an average of the 3 months of the quarter. The NCS HWHP ratio is a 3-year moving average of annual ratios. Quarterly values of the NCS HWHP ratios are interpolated using the Denton procedure.
8 For a few industries that have data gaps, BLS uses a higher aggregate HWHP ratio for a three-digit NAICS industry level in the North American Industry Classification System (NAICS). Because leave policies do not change quickly over time, a 3-year moving average is used to limit the volatility of the series.
9 Reference weeks in November and December are sometimes moved up to avoid conflicts with the Thanksgiving and Christmas holidays. The CPS goes to great lengths to collect data on the number of actual hours worked. CPS respondents are asked to report their usual hours on their main jobs as well as other jobs they may hold. For their main job, they are then asked if they took time off or worked extra hours during the previous week; they are then asked to report their actual hours worked. For other jobs, respondents are only asked to report their actual hours worked during the previous week.
10 Hours from the CPS are not used to directly estimate nonproduction-employee hours because the CPS sample is not stratified by industry, and therefore employment totals will not necessarily match CES totals by industry.
11 The nonhourly group includes those who receive a salary, work for commission, or are paid in kind from a private employer. Throughout this article, we refer to these employees as salaried.
12 See “Employment, hours, and earnings from the establishment survey,” in Handbook of Methods (U.S. Bureau of Labor Statistics), p. 5, https://www.bls.gov/opub/hom/pdf/ces-20110307.pdf.
13 It is worth noting that the response rate for both production-employee hours and all-employee hours is substantially lower than the response rate for total employment. The response rate for reporting all-employee counts is about 53 percent, on average, but it varies by establishment size. Conditional on reporting an all-employee count, the response rate for hours is about 57 percent, which translates to an unconditional response rate of about 30 percent. BLS has studied the impact of nonresponse on CES earnings using data from the QCEW and found a downward bias resulting from nonresponse. See Jeffrey Groen, Kerrie Leslie, Julie Gershunskaya, Patrick Hu, Tran Kratzke, Michael McCall, Edward Park, and Anne Polivka, “An investigation into nonresponse bias in CES hours and earnings,” internal report (U.S. Bureau of Labor Statistics, 2013). The QCEW does not collect data on hours, so the Groen et al. study compared CES and CPS hours data. See also Harley Frazis and Jay Stewart, “Why do BLS hours series tell different stories about trends in hours worked?” in Katharine G. Abraham, James R. Spletzer, and Michael J. Harper, eds., Labor in the New Economy, National Bureau of Economic Research: Studies in Income and Wealth, vol. 71 (Chicago, IL: University of Chicago Press, 2010), pp.343–72, https://www.nber.org/system/files/chapters/c10828/c10828.pdf.
14 The CES program has attempted to address this issue, but employers tend to report data they use for their business, which may not conform to the CES concepts. See Employer Records Analysis Survey of 1981: Final Report.
15 See Patricia Getz, “CES program: changes planned for hours and earnings series,” Monthly Labor Review, October 2003, pp. 38–39, https://www.bls.gov/opub/mlr/2003/10/ressum1.pdf. Getz noted, “Another impetus for the transition to an all-employee definition is the increasing difficulty that many respondents have compiling information for the production and nonsupervisory worker definitions presently used by the CES program. Most payroll recordkeeping does not allow for the easy identification of workers defined by CES categories of production and nonsupervisory workers.” See also Angie Clinton, John Coughlan, and Brian Dahlin, “New all-employee hours and earnings from the CES survey,” Monthly Labor Review, March 2010, pp. 34–40, https://www.bls.gov/opub/mlr/2010/03/art3full.pdf.
16 See Teresa Morisi, “Recent changes in the national Current Employment Statistics survey,” Monthly Labor Review, June 2003, pp. 3–13, https://www.bls.gov/opub/mlr/2003/06/art1full.pdf.
17 For more information on this and related changes, see The Employment Situation: January 2003, USDL-03-46 (U.S. Department of Labor, February 7, 2003), p. 6, https://www.bls.gov/news.release/archives/empsit_02072003.pdf.
18 Misreporting in the CPS was a concern at the time of implementation because research by John P. Robinson and Ann Bostrom cast doubt on the accuracy of hours-worked data from household surveys. Comparing responses to retrospective questions to time-diary data, which are generally considered to be more accurate, the authors found that respondents answering retrospective questions tend to report longer work hours and that the difference between responses to retrospective and time-diary estimates had increased over time. The method for the ratio of nonproduction-employee average weekly hours to production-employee average weekly hours (NPP ratio) assumes that any bias would be similar for production and nonproduction employees. However, there are a number of issues with the Robinson-Bostrom study. In particular, the time-diary surveys that they used collected hours data, but the questions differed across surveys and time periods. In a 2004 study, Harley Frazis and Jay Stewart showed that responses to “usual hours” and “actual hours” questions differ significantly. The exact question wording also matters. In addition, it is not clear what sample weights were used in the Robinson-Bostrom study; nor is it clear how work was defined in the time-diary data. Still, the main finding in the 2004 Frazis-Stewart study (and confirmed in subsequent Frazis-Stewart studies) is that average weekly hours calculated from the CPS data are accurate, on average, although some groups overreport hours while others underreport hours. For more information, see Robinson and Bostrom, “The overestimated workweek? What time diary measures suggest,” Monthly Labor Review, August 1994, pp. 11–23, 1994, https://www.bls.gov/opub/mlr/1994/08/art2full.pdf; see also Frazis and Stewart, “What can time-use data tell us about hours of work?” Monthly Labor Review, December 2004, pp. 3–9, https://www.bls.gov/opub/mlr/2004/12/art1full.pdf; Frazis and Stewart, “Where does the time go? Concepts and measurement in the American Time-Use Survey,” in Ernst Berndt and Charles Hulten, eds., Hard to Measure Goods and Services: Essays in Memory of Zvi Griliches, National Bureau of Economic Research: Studies in Income and Wealth, vol. 67 (Chicago, IL: University of Chicago Press, 2007), pp. 73–97, https://www.nber.org/system/files/chapters/c0874/c0874.pdf; Frazis and Stewart, “Comparing hours worked per job in the CPS and the ATUS,” Social Indicators Research, vol. 93, no. 1 (August 2009), pp. 191–95, https://doi.org/10.1007/s11205-008-9380-y; and Frazis and Stewart, “Why do BLS hours series tell different stories about trends in hours worked?”
19 See Frazis and Stewart, “What can time-use data tell us about hours of work?”; Frazis and Stewart, “Where does the time go? Concepts and measurement in the American Time-Use Survey”; Frazis and Stewart, “Comparing hours worked per job in the CPS and the ATUS”; and Frazis and Stewart, “Why do BLS hours series tell different stories about trends in hours worked?”
20 Lucy P. Eldridge and Sabrina Wulff Pabilonia, “Bringing work home: implications for BLS productivity measures,” Monthly Labor Review, December 2010, pp. 18–35, https://www.bls.gov/opub/mlr/2010/12/art2full.pdf.
21 See F. Thomas Juster, “The validity and quality of time use estimates obtained from recall diaries,” in F. Thomas Juster, and Frank P. Stafford, eds., Time, Goods, and Well-Being (Ann Arbor, MI: University of Michigan Press, 1985), pp .63–91.
22 For quarterly series prior to the second quarter of 2006 and annual series prior to 2007, the new hours-worked time series is constructed by linking the level of the new hours-worked series to the movements of the previous series. This approach prevents a break in the series.
23 The importance of off-the-clock (OTC) hours has been recognized in previous research. For example, Stephanie Aaronson and Andrew Figura examine the cyclicality of OTC hours in their article, “How biased are measures of cyclical movements in productivity and hours?,” Review of Income and Wealth, vol. 56, no. 3 (September 2010), pp. 539–58. In addition, Harley Frazis and Jay Stewart consider whether the difference in hours concepts (worked versus paid) can explain the divergent trends in CES and CPS hours measures in their article “Why do BLS hours series tell different stories about trends in hours worked?,” and Lucy P. Eldridge and Sabrina Wulff Pabilonia examine whether increasing amounts of unpaid overtime work brought home from the office may bias estimates of productivity growth in their article “Bringing work home: implications for BLS productivity measures.”
24 The subscript denoting all employees is dropped from this point forward.
25 Because the CPS does a better job of capturing variation in actual leave taken, we experimented with constructing a hybrid paid-time-off (PTO) ratio that combines estimates of variation in PTO around the trend using CPS data and with NCS levels. However, because nearly all of the variation in actual leave taken is seasonal, the gain to using actual variation was minimal. A notable exception occurred during the coronavirus disease 2019 (COVID-19) pandemic. Research in progress will use a hybrid PTO adjustment to examine the impact of quarterly variation in paid leave during the pandemic on hours of work.
26 This measure of OTC hours differs from definitions used by previous researchers who analyzed OTC work. Both Aaronson and Figura (2010) and Frazis and Stewart (2010) calculated OTC hours as the difference between hours worked and (simulated) hours paid directly from the CPS, without adjusting hours paid for PTO. See Aaronson and Figura, “How biased are measures of cyclical movements in productivity and hours?” and Frazis and Stewart, “Why do BLS hours series tell different stories about trends in hours worked?”
27 Our approach is similar to that of Frazis and Stewart in their 2010 article “Why do BLS hours series tell different stories about trends in hours worked?,” except that we focus on all employees rather than on production employees. Our approach improves on the treatment of PTO and uses data-science tools to perform imputations.
28 Unlike full-time, salaried work, for which the standard workweek is usually 40 hours, there is no standard workweek for part-time, salaried work. Because only 3 to 4 percent of wage and salary workers are both part time and salaried, this assumption should have a minimal impact on estimates of aggregate hours.
29 We think that this is a reasonable assumption because usual hours worked on second jobs are, on average, about 14 hours per week, and those who work part time are more likely to be paid hourly. In addition, only about 5.5 percent of employed people in the CPS hold more than one job; thus, the impact of these assumptions is relatively minor.
30 Note that the CPS data are on a person basis, whereas the CES data are on a job basis. People who work at two jobs are counted once in the CPS but twice in the CES survey. Because the CPS OTC ratio will be applied to the CES data, we converted the CPS data to a job basis by creating separate observations for second jobs.
31 Sampled households are in the CPS for 4 consecutive months, out of the sample for 8 months, and back in the sample for another 4 months. The questions on earnings, which include whether the worker is paid hourly, and the additional questions on second jobs are asked in months in sample 4 and 8—the outgoing rotations.
32 Between 37 and 42 percent of employees worked full time and were salaried on their main job from 2006 to 2021.
33 In some industries, where CES average weekly hours paid for all-employees exceed 40 hours and it is possible that some employees regularly receive overtime pay, we assign the topcode as the 3-year moving average in that industry.
34 Hourly/nonhourly status is collected as part of the CPS earner-study questions. Respondents are asked about the easiest way to report their earnings. If respondents indicate “hourly,” they are classified as hourly. If they indicate that it is easier to report their earnings at some other periodicity (weekly, biweekly, etc.), they are asked if they are paid hourly.
35 Another variation of the random-forest model is to assign integer values (0 or 1) to each observation rather than computing probabilities. The two approaches generate similar results.
36 The variables (features) used to train the algorithm for prob(hourly) are age, sex, education, marital status, family income, weekly earnings, major industry, major occupation, and usual hours worked.
37 Usual hours paid is available for a subset of workers, and we use that measure when it is available. The value of the cap on weekly hours has virtually no effect on the growth rate of hours and labor productivity. It does, however, have a small impact on hours levels.
38 We also topcode total actual and usual hours worked per worker at 84 hours per week (7 days per week × 12 hours per day).
39 We use information from the subset of CPS respondents who are asked if their absence from work was paid as the training dataset to generate the predicted probabilities. The variables (features) used to train the algorithm for prob(PTO) are age, sex, education, marital status, number of children, family income, major industry, major occupation, class of worker, and usual hours worked. Because there was a large increase in unpaid absences due to “other reasons” reported by CPS respondents during the early stages of the pandemic that BLS believes should have been mostly classified as unemployed on temporary layoff and this should not be relevant to partial week workers, we use monthly data from March through August 2019 as the training datasets for monthly PTO predicted probabilities in the same months of 2020 as in 2019. For more information on this, see The Employment Situation: March 2020, USDL-20-0521 (U.S. Department of Labor, April 3, 2020), p. 5, https://www.bls.gov/news.release/archives/empsit_04032020.htm.
40 In a small number of cases, CPS respondents report their usual hours worked are zero. In these cases, we set paid hours worked for full-time salaried employees equal to actual hours worked topcoded at 40.
41 Under the old method, it was not possible to apply the NPPCPS ratio to three-digit NAICS industries, because the ratio was calculated by using 20 percent of the sample (nonproduction employees only) in the numerator and 80 percent of the sample (production employees only) in the denominator. The new method uses all observations in both the numerator and the denominator.
42 We also tested OTC ratios constructed from a single month of data and found that the results for some industries were not reliable.
43 The official series will track the research series closely, but there may be slight differences.
44 This sector corresponds to data collected in the CES survey and includes government enterprises and nonprofit institutions but excludes general government. This differs from the definition of the private nonfarm sector used in the BLS total factor productivity data, which also removes government enterprises.
45 The nonfarm business sector includes private nonfarm industries and government enterprises and excludes nonprofits and general government. The nonfarm business productivity estimates are calculated for all workers, not just employees, including hours worked by the self-employed and unpaid family workers.
46 The data in charts 2 through 8c are consistent with official BLS productivity data released on August 9, 2022.
47 This is approximately correct, as the vast majority of salaried employees work full time.
48 Approximately 42 percent of employees in durable goods manufacturing and 40 percent of employees in nondurable goods manufacturing were paid on a nonhourly basis in 2021.
49 Approximately 51 percent of employees in wholesale trade, 40 percent of employees in transportation and warehousing, 48 percent of employees in utilities, 62 percent of employees in professional and business services, 41 percent of employees in education and health services, and 50 percent of employees in other services were nonhourly employees in 2021.