Each month the U.S. Bureau of Labor Statistics releases its Employment Situation report. This report includes about two dozen distinct datasets that can help economists, journalists and the public understand the health of the nation’s economy. It’s where to find monthly changes to the number of people employed and the unemployment rate, and it’s a report that’s widely covered in the media, but isn’t always covered with much context.
This is the bullhorn May employment data running across news chyrons and online headlines:
Two different surveys make up the monthly BLS report. There’s the Current Population Survey — also called the household survey — which surveys 60,000 households. And, there’s the Current Employment Statistics survey — also called the payroll survey — which surveys 142,000 businesses and government agencies.
Changes to employment numbers come from the payroll survey, while the unemployment rate comes from the household survey.
2. Sampling error, confidence intervals and statistical significance can change how we interpret the monthly employment numbers.
All surveys that sample a population come with sampling error. Again, BLS surveys are samples, not counts, of characteristics of the U.S. workforce. BLS aims to sample households and businesses that represent the variety of the general population, but a survey that is small compared to a total population will be less representative, have more error, and ultimately produce less precise estimates.
Because BLS monthly employment reports are not raw counts, the true numbers would probably not hit the headline numbers on the nose, and might not even come close. The household survey is smaller than the payroll survey, so it comes with a wider range of error. The less data, the less precise the estimate. Quarterly and annual BLS estimates are more precise because they use more data. Technical notes linked to BLS monthly economic situation releases include discussions of sampling error.
Confidence intervals represent the range BLS thinks the data would fall in if it counted up every job added or lost, rather than surveying a sample of businesses. A small confidence interval means the reported result is more precise — it is more likely to be closer to the true count. The confidence interval for job numbers from the payroll survey is usually around +/- 100,000. For May, the confidence interval is +/- 110,000, but for this example let’s stick with the +/- 100,000 number.
If the economy added 150,000 jobs in a given month, with a confidence interval of +/- 100,000, BLS says in its technical notes that it is giving a 90% chance that the actual number of jobs added is between 50,000 and 250,000, a pretty wide swing. (Other federal agencies explain confidence intervals in a slightly different way.)
If BLS reports that 50,000 jobs were added in a given month, they can’t say for sure whether the economy actually added any jobs — when factoring in sampling error, it’s possible the economy lost 50,000 jobs. This is exactly what happened in May. We can’t say for sure whether the economy actually added jobs. In fact, it may have lost jobs. This is how BLS puts it in their technical notes:
“Since this range includes values of less than zero, we could not say with confidence that nonfarm employment had, in fact, increased that month.”
The above bears repeating: when you factor in the confidence interval and find that the estimated change in the number of jobs could be more or less than zero, it is impossible to say with confidence that the number of jobs went up or down. Confidence intervals can change slightly each month, and estimates may not even be statistically significant in a given month, so it’s important to report on the technical notes.
The May employment numbers come with a confidence interval of +/-110,000 at 90% confidence, the BLS reports. In other words, the agency is 90% sure that if it directly counted the change in the number of jobs from April to May, it would find a gain of 185,000 or a loss of 35,000 jobs. Those 75,000 jobs added? That’s the middle of the confidence interval. The changes in jobs numbers shouldn’t be taken literally, as in, “the economy definitely gained 75,000 jobs.” The economy could have also lost 35,000 jobs. Or, again, gained 185,000. The May job increase is something of a Schrodinger’s cat of economics.
Changes in unemployment, estimated from the household survey, are another headline grabber. In April 2019, for example, unemployment went down compared to March. BLS reported 387,000 fewer people were unemployed in April. Was this a meaningful number? For BLS to say that it was confident in the estimate at a 90% level, the estimate had to be outside of +/- 271,000. Because their survey estimates showed a drop of 387,000 in the number of unemployed people, the estimate for April was statistically significant. BLS was signaling that this was a change that mattered.
In May, the number of unemployed people grew by 64,000. For BLS to say that it was 90% confident in this number, the estimate had to be outside of +/- 268,000. Because the growth in the number of unemployed is within that range, the result for May is not statistically significant. BLS is signaling they can’t confidently say whether the change this month was due to errors in the survey or something else, or because there was an actual change in the number of unemployed.
That may not matter much for a small change. An increase of 64,000 unemployed people is so small compared to the size of the labor force that it doesn’t affect the reported unemployment rate.
But say the increase in unemployed had been 250,000. That would have yielded an unemployment rate of 3.7%, and every small change in unemployment is newsworthy. But in that hypothetical, BLS still wouldn’t have been 90% confident that the uptick in the unemployment rate wasn’t due to survey error, or something else not accounted for. The uptick in unemployment would have grabbed headlines, but it might not have been signaling anything meaningful about the state of the economy.
BLS reports the number needed to achieve statistical significance for unemployment changes on this page, which it overwrites each month. The “needed” column indicates the number needed for the estimates to be statistically significant:
Historical versions of that table don’t seem to be available on the BLS website, but we reached out to BLS and they provided monthly records going back to January 2015, when they first started producing the table. You can download those historical tables here in PDF and Excel formats:
There’s also nonsampling error. Nonsampling error happens when, for example, an establishment doesn’t respond to its survey in a timely or complete way. Survey responses for May will continue to trickle in even after preliminary results are reported. This is why BLS makes revisions.
3. Some of the data are preliminary.
Final employment numbers for May won’t be out until July. BLS revises its payroll survey data in each of the two months following an initial release. So the estimated 75,000 jobs added in May will be revised in June and finalized in July. BLS revises the data using additional survey responses that come in late, and by making further seasonal adjustments.
Sometimes there is a small difference in preliminary versus final numbers. In February 2017, BLS reported the economy added 235,000 jobs. That April, the final numbers for February came in at 232,000 jobs added — off by just 3,000 from the preliminary estimate. In December 2018, BLS reported that the economy added an impressive 312,000 jobs. Two months later that number was revised down 85,000 jobs to 227,000 — strong, but not as strong. And in September 2017, BLS reported that the economy lost 33,000 jobs. But in November when BLS released its final revision, the economy had actually added 38,000 jobs — off by 71,000 from the preliminary.
Revisions can shift narratives on how the economy is doing. They provide an opportunity to reframe coverage of preliminary data and to offer more accurate information.
4. There are major and persistent differences in employment across racial and educational boundaries.
Two things persist month-over-month, even when the economy is strong: the unemployment rate for black people is usually about 40% higher than the national rate, and people with less education face worse employment prospects.
Black unemployment has been trending downward for a decade, and the unemployment rate for black people is now lower than before the Great Recession, which lasted from 2007 to 2009. Still, black unemployment remains about twice that of white unemployment and triple the rate of Asian unemployment. Unemployment rates for Hispanic or Latino people are also persistently higher than national unemployment rates. The last time white unemployment was at the current black unemployment rate of 6.2% was more than half a decade ago in October 2013 — a time when black unemployment was at about 13%.
Black unemployment has been twice that of whites since at least the 1950s, but economists still struggle to explain why the gap persists. “Blacks have substantially higher and more cyclical unemployment rates than whites, and observable characteristics can explain very little of this differential, which is importantly driven by a comparatively higher risk of job loss,” Federal Reserve economists wrote in a 2017 paper.
That paper also found that blacks who want to work full-time are more likely than whites to work part-time. After the Great Recession, many black men who had worked full-time did not transition back to full-time work, while white men had a much easier time returning to full-time work, according to the authors.
The racial unemployment gap is sometimes boiled down to the so-called “first fired, last hired” phenomenon cited in the Economic Report of the President sent to Congress in March 2019. A comprehensive study in Demography from February 2010 took BLS household survey data from 1989 to 2004 and found “considerable evidence” that blacks are the first fired as the economy sours, but no evidence that they are the last hired when the economy rebounds. Whatever the reasons, the gap remains no matter how the economy is doing.
Like the racial unemployment gap, the data for each month show an educational attainment gap.
More education usually means better prospects for employment. This is true in developed North American and European economies across the world, according to research from the Organization for Economic Cooperation and Development. More education also means higher pay. The average college graduate makes $78,000 per year compared to $45,000 for workers who only graduated high school, a June 2019 analysis from the New York Fed found.
Employment doesn’t just improve quality of life and increase lifetime wealth — it can also affect how long people live. The authors of a 2018 paper in the Journal of Racial and Ethnic Health Disparities analyzed data from a longitudinal study that followed employment over 25 years for 2,025 whites and 1,156 blacks from 1986 to 2011. They found that employment can improve life expectancy, but the effects are less for blacks, women, and people with lower levels of education.
“We do not argue that women or blacks, particularly black men, are unable to translate employment to health,” the authors write. “Instead, we argue that how American society operates minimizes the health gain of certain groups, even when they secure a job.”
5. Wages aren’t rising at the rate they appear to be.
“We do not foresee a sharp pickup in wage growth nationally if the labor market continues to tighten as many anticipate,” wrote San Francisco Fed economists Sylvain Leduc, Chitra Marti, and Daniel J. Wilson in January 2019.
There may also be side effects beyond workers having to stretch their earnings farther. One study, in the American Journal of Preventive Medicine, found an association between increases in state minimum wages and slower growth in suicide rates.
6. The unemployment rate doesn’t include people who are not part of the labor force, such as people who want to work but have given up looking.
The unemployment rate comes from some pretty simple math. It’s the number of unemployed people divided by the size of the labor force. The resulting fraction is converted to a percentage, and there you have the unemployment rate.
The labor force is made up of people who are employed and most but not all people who don’t have a job. Unemployment numbers don’t include people in government-run institutions — like prisons. The labor force also doesn’t include people who haven’t looked for work recently, are discouraged over job prospects, or are not looking for work for other reasons. BLS calculates the labor force participation rate by comparing the labor force to the civilian non-institutional population.
The labor force participation rate has been declining since the Great Recession. Since early 2013 the rate has hovered each month around 63%, down from a high of about 67% in 2000 and about 66% in 2007 before the recession. Today’s participation rate matches rates last seen in 1977, when the participation rate for women was 48%. The participation rate for women today is about 57%.
Economists often chalk up declining labor participation rates to demographic changes. Baby boomers are retiring (and living longer), while the participation rate for men age 25 to 54 is declining, according to a September 2016 analysis from BLS economist Steven F. Hipple. Princeton economist Alan B. Kreuger comes to largely the same conclusions in a fall 2017 paper published in Brookings Papers on Economic Activity. Labor force participation among prime working-age men is “notably low” in the U.S. compared to other countries with similarly advanced economies, Kreuger finds. Health problems may be playing a role.
“Labor force participation has been declining for prime age men for decades, and about half of prime age men who are not in the labor force may have a serious health condition that is a barrier to working,” Kreuger writes. “Nearly half of prime age men who are not in the labor force take pain medication on any given day; and in nearly two-thirds of these cases, they take prescription pain medication.
Lastly, the data don’t include people who do farm work because BLS draws data in part from unemployment insurance records, and many jobs in agriculture are exempt from unemployment insurance, “making the sample frame for agriculture insufficient for calculating statistically sound estimates,” according to this BLS FAQ. (The Census of Agriculture from the U.S. Department of Agriculture offers more detail on farm labor.)
Bottom line, there are people not working who are not counted in the official unemployment rate.
7. Alternative measures of unemployment are always higher than the official rate.
These measures tend to follow similar trends to the official unemployment rate, fluctuating up and down based on how the economy is doing. The broadest measure includes all people not in the labor force and underemployed people — those working part-time who want to work full-time. This yields an unemployment rate roughly double the official figure. Again, none of these data include the 6.6 million people who are incarcerated, representing 4 percent of the current labor force.
These data may still underrepresent the underemployed. A paper published in July 2012 in Business Economics by University of Alabama economist Samuel N. Addy, and Michaël Bonnal and Cristina Lira, states that BLS labor underutilization data “do not consider underemployment among full-time workers and thus provide incomplete information on the full extent of underemployment.”
Under pressure: the employment situation as barometer
The monthly unemployment rate and job numbers are certainly worth reporting as a barometer of economic health. The Federal Reserve’s Federal Open Market Committee regularlylooks to unemployment as a leading economic indicator during its quarterly meetings, so it’s a number that matters to top economy watchers.
High unemployment is bad on its face because it means a lot of people are not working who want to be. Low unemployment is usually good, but not always. Inflation can rise if employers offer higher wages to compete in a tight labor market. If people are making more money, the theory is that goods producers respond by raising prices. Full employment doesn’t mean everyone has a job, but rather that “unemployment has fallen to the lowest possible level that won’t cause inflation,” as Clive Crook wrote last year in Bloomberg.
The U.S. Federal Reserve in March 2019 pinned 4.0% to 4.6% as the range for a “normal rate of unemployment.” This is well above the current unemployment rate. What gives? Dylan Matthews at Vox goes into some depth on why the U.S. economy may not actually be at full employment. The idea is that as long as jobs are being added and filled and inflation remains low, we’re not at full employment yet. That means there may still be room for the unemployment rate to fall.
Seasonal adjustment is another concept to understand. The headline numbers represent data that is seasonally adjusted, meaning BLS economists control for predictable seasonal fluctuations in employment. For example, major holidays tend to spike employment as stores hire workers to keep up with increased demand. Any major fluctuations in the widely reported monthly data are not because of seasonal changes.
Key links to BLS data
BLS puts out a lot of data each month, but it can be hard to know where to start. Here are some quick links:
BLS revises its payroll estimates twice. Once in the month after the preliminary release, and again in the month after that. Final numbers are sometimes close to the preliminary release, or they can drastically change and affect the perception of how the U.S. economy is doing.
BLS reports the number needed to achieve statistical significance for unemployment changes on this page, which it overwrites each month. We reached out to BLS and they provided monthly tables going back to January 2015, when they first started producing them. You can download those tables here in PDF and Excel:
Citation: Gertner, Alex K.; Rotter, Jason S.; Shafer, Paul R. “Association Between State Minimum Wages and Suicide Rates in the U.S.” American Journal of Preventive Medicine, May 2019, doi.org/10.1016/j.amepre.2018.12.008.
Leduc, Sylvain; Marti, Chitra; Wilson, Daniel J. “Does Ultra-Low Unemployment Spur Rapid Wage Growth?” FRBSF Economic Letter, January 2019, frbsf.org/economic-research/files/el2019-02.pdf.
Assari, Shervin. “Life Expectancy Gain Due to Employment Status Depends on Race, Gender, Education, and Their Intersections,” Journal of Racial and Ethnic Health Disparities, April 2018, doi: 10.1007/s40615-017-0381-x.
Cajner, Tomaz; Radler, Tyler; Ratner, David; Vidangos, Ivan. “Racial Gaps in Labor Market Outcomes in the Last Four Decades and over the Business Cycle,” Board of Governors of the Federal Reserve System, June 2017, doi.org/10.17016/FEDS.2017.071.
Leduc, Sylvain; Wilson, Daniel J. “Has the Wage Phillips Curve Gone Dormant?” FRBSF Economic Letter, October 2017, https://www.frbsf.org/economic-research/files/el2017-30.pdf.
Krueger, Alan B. “Where Have All the Workers Gone?: An Inquiry into the Decline of the U.S. Labor Force Participation Rate,” Brookings Papers on Economic Activity, Fall 2017, doi:10.1353/eca.2017.0012.
Addy, Samuel N.; Bonnal, Michael; Lira, Cristina. “Toward a More Comprehensive Measure of Labor Underutilization: The Alabama Case,” Business Economics, July 2012, doi.org/10.1057/be.2012.16.
Couch, Kenneth A.; Fairlie, Robert. “Last hired, first fired? Black-white unemployment and the business cycle,” Demography, February 2010, doi.org/10.1353/dem.0.0086.