Aerial view of two people shaking hands with office supplies in the background
State of the Economy

Inequality in labor force participation has diminished

Ryan Nunn
September 29, 2026
Printable PDF

Key Takeaways

  1. Labor market inequality can be assessed comprehensively with a basket of demographic variables, rather than by comparing workers across a single dimension

  2. Changes in outcomes for disadvantaged groups can be a leading indicator of overall labor market trends

  3. In the late 2010s, inequality in labor force participation and unemployment fell, then spiked with the 2020 pandemic

  4. The recovery from the pandemic saw inequality return to its declining trend

  5. Over the last year, the lowest-participating group gained about one percentage point of ground relative to the top-participating group

  6. However, the highest-unemployment group has seen its unemployment rise slightly relative to the lowest

The jobs report for September will provide headline averages of key labor market variables, but to dig deeper typically requires looking at the microdata, and sometimes requires new methods. We know that the unemployment rate (as of August) was 4.1% and the labor force participation rate was 61.6% (83.4% for people aged 25 to 54), but this doesn’t tell us about how widely varying experiences are for different groups of people. Gaps across groups are interesting in their own right, but they can also tell us something about how strong the labor market really is, because a labor market in which the bottom of the distribution is struggling tends to be a weakening labor market.   

One common way that labor analysts use the jobs report’s data to investigate this type of question is to separate key variables, like labor force participation, by educational attainment. Figure 1 does so for 25–54 year olds, which has the dual benefit of (for most people) capturing final educational level and keeping the focus on people of working age.1 The figure shows narrowing inequality—about two thirds of a percentage point— in participation in the late 2010s up until the pandemic.2 When the pandemic arrives, inequality skyrockets as less-educated workers lose their jobs and temporarily leave the labor force. During the subsequent recovery and expansion, that inequality diminished, leaving the gap between high school and college graduates almost exactly where it was in January 2020.

A different approach

While informative, cuts of data by education are limited; education is only one variable that correlates with labor market advantage. Other factors, from location to gender to race, all matter as well, and could interact as they predict labor market outcomes like participation.3 

In a Minneapolis Fed working paper, Narayan and Nunn propose a different approach that, in essence, generalizes the idea behind cutting labor market outcomes by education. Rather than use that single variable (or any other) on its own, they show how to map a large collection of variables to the predicted level of employment for any given worker. Using a random forest machine learning method, they classify all 25–54 year old workers into one of five quintiles, from the lowest probability of employment to the highest, based on their demographic characteristics. Starting in a pre-pandemic base period of 2017 and 2018, they then track the average employment rate of each group over time. 

Figure 2 implements this approach for 25–54 year old labor force participation up through August 2026.4 (For ease of interpretation, we show only the bottom, middle, and top quintiles.) Like Figure 1, it shows considerable narrowing of gaps between the top and bottom quintiles—though to a much greater extent than in Figure 1. In other words, inequality fell as individuals with low tendency to participate became more likely to do so, relative to those with a high tendency. What is different from Figure 1 is that we see inequality falling during the post-pandemic era by about as much as it did in the pre-pandemic era. Labor force participation by the bottom quintile rose from 63.7% in the immediate pre-pandemic period (calculated here as the 12 months ending with January 2020) to 65.8% in the most recent period (the 12 months ending with August 2026). By contrast, the top quintile ended essentially where it started, ticking up from 95.1% to 95.2% over the same time.

Appendix Table 1 gives a sense of who is in all five of the quintiles, displaying the share of each who are college graduates, are native-born, and so forth.5 Notably, higher quintiles tend to include more men, college graduates, and (in the top quintile) married individuals.    

The same method can be applied to other labor market outcomes like unemployment.6 The mapping from demographic characteristics to any particular outcome is different in each case: for example, women are much more likely to be found in the bottom quintiles of participation, but roughly evenly distributed across quintiles of unemployment.  

When applied to 25–54 year old unemployment, the trends in inequality are different but in a way that could be expected, given how unemployment rates tend to evolve across the business cycle. As the labor market tightens, unemployment rates typically fall for all groups. Because gaps in ratio terms tend to be stable—e.g., the Black unemployment rate is usually about twice the White rate—percentage point gaps tend to shrink. Figure 3 shows a somewhat similar pattern, with diminishing gaps during the strengthening labor markets of the late 2010s and early 2020s. By contrast, gaps have widened recently; the highest-unemployment quintile saw its unemployment rate rise by about 0.4 percentage points more from August 2025 to August 2026 than the lowest quintile.7

The importance of tracking labor market inequality

Outcomes like labor force participation and unemployment are persistently different across groups, as has been extensively documented by race, gender, and education, just to choose a few examples. Policies and longer-run structural factors matter, of course, but so does the strength of the economy. When labor markets are tighter, gaps between groups tend to be smaller.  

A comprehensive assessment of labor market strength is better if it pays attention to how those gaps shift over time. Declining inequality in participation is a welcome trend, albeit one that is counterbalanced by a moderate increase in unemployment inequality.  

The author is grateful to Ayushi Narayan and Kevin Rinz for insightful feedback on an earlier draft.  

Appendix

The table below shows demographic characteristics of the five quintiles by predicted labor force participation rate. Quintile 1 has the lowest participation rate and quintile 5 has the highest.  

The main approach in Narayan and Nunn (2024), and in this article, is to calculate quintiles in a base period (selected to be 2017 and 2018) and hold fixed the mapping between demographic characteristics and quintile. For instance, a 43 year old married native-born man with a college degree (and other characteristics predicting high participation) would be in the highest participation quintile in every month, regardless of how the empirical relationship between demographic characteristics and participation might change after 2018. This has the advantage of making it visually clear whether any “overfitting” occurs: if it did, we would expect to see large shifts just before and after the base period. The disadvantage is that, over time, the mapping between demographic characteristics and the labor outcome can become somewhat less accurate, and group shares can drift away from their original values.  

To show that this distinction makes little difference in practice, we show alternative figures for participation (Appendix Figure 1) and unemployment (Appendix Figure 2) that contemporaneously reapply the random forest method in every quarter. The patterns are qualitatively similar.

Footnotes

  • 1

    Figure 1 and subsequent figures show 12-month trailing averages that include the current period. When October 2025 is included in the 11 prior months, the average is taken over the non-missing values. 

  • 2

    This calculation is for January 2016 through January 2020, using (as elsewhere in the article) the trailing 12-month average in those months. 

  • 3

    When examining wages, analysts often study the worker-level distribution of wages, rather than average wages by education. In this way, they are able to look directly at the object of interest: inequality (e.g., the ratio of the 90th and 10th percentiles of wages) at the worker level. For “extensive margin” outcomes like employment, participation, and unemployment, there is no analogous distribution. The approach of Narayan and Nunn (2024) and this article is to approximate such a distribution with a rich set of observable covariates.  

  • 4

    There are some slight differences with Narayan and Nunn (2024). The most important is that, rather than show three-month trailing averages of the outcome, it shows 12-month trailing averages. For periods of time that include the October 2025 missing CPS data (other than October 2025 itself), the figure shows averages for the remaining 11 months.   

  • 5

    State of residence is also used as a predictor but not shown in the table.  

  • 6

    Narayan and Nunn (2024) focuses exclusively on the employment-to-population ratio. 

  • 7

    As elsewhere, note that the values for these periods represent the trailing 12-month average ending in the given month.