Appendix to Don’t Choose Your Own Adventure: Understanding Middle-Class Earnings Trends
Nonworkers, Part-Year Workers, and Part-Time Workers
This is a technical appendix to my new paper for Civitas Outlook, “Don’t Choose Your Own Adventure: Understanding Middle-Class Earnings Trends” (also available on First World Problems). It runs through the experimentation I conducted to choose full-year workers as the most informative subsample of men and women when looking at long-run earnings trends.
One way to address the question of how to deal with nonworkers, part-year workers, and part-time workers in estimating meaningful median earnings trends is to exploit a little-used feature of the CPS—the fact that it collects income information two different times for most households. The way the CPS works is that everyone is surveyed in four consecutive months, before taking a year off and then being interviewed again in the same four months the following year. Therefore, if someone is observed as a nonworker, part-year worker, or part-time worker in one year, we may be able to leverage information from their earnings one year earlier or one year later.
For example, in my sample of people at least 15 years old, I am able to match 71 percent of them to another earnings record in 1973 and 64 percent in 2022.1 (I stop at 2022 instead of 2023 because 2024 data was unavailable at the time of writing, so I could only check if men without earnings in 2023 had them in 2022, not 2024.) Among people who have no earnings, I can match 75 percent of them to another earnings record in 1973 and 67 percent in 2022. Most of those with no earnings who can be matched also have no earnings in the other year they are observed; just 12 percent in 1973 and 9 percent in 2022 can be matched to a positive earnings amount.2 (Note that many of these people with no earnings in consecutive years are retired.)
Let’s focus, first, on men, since their earnings trends look much worse than those of women and assessing their trends is more sensitive to measurement issues. For those who we can match across surveys, we can use their linked earnings information to get a better sense of what earnings trends would look like if we could observe everyone’s earnings. If someone does not have earnings in one year but does in an adjacent year, we can swap in the nonzero amount.
Of course, some men have no earnings in a given year because of subpar economic conditions. It’s not informative to ask, “What if the economy were great?” to the extent that most nonworking men only wish they had a job. But we know from the CPS why someone went an entire year without working, why they only worked part of the year instead of the whole year, and why they worked part-time instead of full-time. Since we can identify men who say the economy was to blame for their working less than year-round or less than full-time, we can keep them in the data as below-median earners and let the strength of the economy affect the results. That’s what I do in these experiments; values are replaced with earnings from a linked survey only if someone had a low initial value for some reason other than economic problems.3
Importantly, the sample that can be linked has similar earnings growth to the full sample of men if all non-positive earnings are dropped. Among men with positive earnings, the median increased 48 percent from 1973 to 2022 for the full sample and 47 percent for the linked sample.4 (Note that the analyses in this appendix use the MACPI, described in the paper, to adjust for inflation. Using the PCEPI instead would affect the magnitudes but not the relative comparisons of different measures.)
However, when including men with no earnings or negative earnings, growth is worse in the linked sample than the full sample. Among all men, the median fell by 1 percent in the full sample but by 15 percent in the linked sample.5
So, staying with the linked sample, which is closer to The Truth: a 47 percent increase, or a 15 percent decline? For starters, if we replace earnings of $0 with positive values from the year before or after—but only for men who were nonworkers for reasons other than difficulty finding a job—median earnings falls by 1 percent rather than 15 percent.6 That suggests that if we could link more nonworkers to past or future earnings, the earnings trend would improve substantially relative to the trend when all nonworkers are included.
Some men who don’t work clearly could but are doing other things instead. If we drop the remaining men with non-positive earnings (except for those who had no earnings due to economic conditions), the median increases 40 percent from 1973 to 2022.7 That’s a bit less than the 47 percent we got before we added people whose earnings we could pull from a second interview (or even the 44 percent we’d have obtained from that sample if we’d left in those with no earnings because of economic conditions, not shown). That suggests that the rising share of men with no earnings causes men’s trends to be too strong relative to what it would be if everyone’s earnings could be observed in a given year.
Addressing people with negative or no earnings isn’t the only complication to address. Some people work only part of the year for voluntary reasons having nothing to do with the state of the economy. For instance, graduating seniors may only work for half the calendar year. Parents may take time off after the birth of a child. Some part-year workers are older and retired midway through the year. These part-year jobless spells are often just continuations of the full-year jobless spell that we are trying to address.
Replacing a value of $0 with earnings from the year before or after someone is observed may only replace it with a value that itself still reflects (voluntary) time out of the labor force, understating what the person would make as a full-year worker. If we only replace values of $0 using full-year earnings from the year before or after, dropping all other non-positive earnings (unless due to economic problems), the increase in median earnings bumps up to 45 percent from the 40 percent in the last step.8
Further, we can also check what happens when we replace earnings of men working only part of the year with their observed full-year earnings from the year before or after (where possible). If we make that replacement too—again, only for men whose part-year employment had nothing to do with economic conditions—on top of replacing non-positive earnings with positive earnings from the year before or after, the increase in median earnings is 43 percent.9
If we also drop the remaining workers who worked only part of the year for reasons other than economic conditions (those who weren’t full-year workers the other year they were observed), median earnings rise by 42 percent.10 Note that this estimate still includes non-working men and part-year men if economic conditions were the reason they didn’t work or only worked part of the year.11
These estimates have dropped all non-workers and part-year earners for whom we can’t pull full-year earnings from an adjacent year, unless their status is due to economic conditions. That seems reasonable in the case of men who consistently aren’t working because they are in school, taking care of home or family, or retired. It also seems reasonable for unambiguously disabled adults.
However, over the long run, while disability rates of men have been fairly constant, the likelihood of working conditional on reporting a disability has fallen.12 At the same time, federal disability benefit rolls have increased substantially.13 Disability benefit receipt increases predictably during economic downturns, suggesting that some men experiencing economic problems respond by turning to disability benefits. If the increased rate of disability benefit receipt were suppressing a bigger rise in the number of men saying they don’t work because of the economy, then dropping disabled non-workers would overstate the gains that able-bodied men have made over time.
One reasonable way of addressing this issue is to add back to the analysis “excess” non-workers and all part-year workers who cite illness or disability as the reason for not working year-round. I determine the share of adults in 1973, by age group, who were nonworking and disabled, then I add back in the excess share of non-working disabled adults (the number above the 1973 share) in every year.14 Then I add back all part-year disabled workers. (I ignore any changes in part-time work, since that information is unavailable for 1973.) This final sample is 27 percent smaller in 1973 and 35 percent smaller in 2022 than the full sample of men ages 15 and older and linkable to a second earnings record. It is 9 percent smaller than the sample of such men with positive earnings in 1973 and 1 percent smaller in 2022.15
After making this final adjustment, median male earnings rises by 32 percent from 1973 to 2022—significantly lower than the 42 percent increase before adding back these sick and disabled men.16
To recap where we have arrived: some men either don’t work or work only part of the year for reasons having nothing to do with economic conditions. In the data, for a subset of men, we can observe earnings from one year earlier or one year later. Some of the men who didn’t work year-round in one year did so in an adjacent year, and some who worked part-year worked year-round in an adjacent year. We can (1) swap in their full-year earnings and (2) exclude the remaining non-workers and part-year earners whose status had nothing to do with economic conditions nor with their health, and (3) exclude a number of disabled non-workers who would have been non-workers even in 1973.
The resulting 32 percent increase in median male earnings is closer to the 47 percent rise in the median when we looked at all men with positive earnings than to the 15 percent fall when we looked at all men with or without earnings. If we just look at full-year workers (without swapping in any full-year values from adjacent years for non-workers and part-year workers), median earnings increase by 36 percent.17 That suggests that the simple trend for full-year earners is a solid basis for assessing male earnings trends, at least when comparing similar points in the business cycle. (There will be more part-year workers and non-workers when times are bad than when times are good, but the checks here suggest that taking that into account wouldn’t change the 50-year trend comparing two strong economic years.)
Recall that the median earnings trend for the CPS sample linkable across years was somewhat different than the trend for the full CPS sample. If we go back to the full CPS sample and look at median earnings among men who worked all year, the increase from 1973 to 2022 was 41 percent, and the increase from 1973 to 2023 was 40 percent, or $17,900. From 1973 to 1989, the median rose by 11 percent, or $5,000. It rose by 26 percent from 1989 to 2023, or $12,900.18 The 1973-to-2023 increase using the PCEPI instead of the MACPI was 13 percent.
When we conduct the same exercise for women, we get similar results. The median rises by 117 percent from 1973 to 2022 among full-year workers in the linked sample and by 124 percent after we make all the adjustments, including “excess” disabled adults.19 If anything, using the year-round-worker trend understates gains for women. In the full sample (linkable or not), median earnings among year-round workers increases 115 percent from 1973 to 2022 and 115 percent from 1973 to 2023.20 Using the PCEPI, the 1973-to-2023 increase is 75 percent.
Thus, the trend for year-round workers is an appropriate approximation of the increase in median earnings among both men and women using more careful adjustments to account for changes in non-work, part-year work, and part-time work.
Linking people in the data to their second interview is not as straightforward as would be ideal. Households that move from an address are not followed; the new household living at the old address is interviewed instead. Not only do households move, but household membership changes. Within households, there are sometimes data inconsistencies requiring that demographic variables be checked to make sure one is truly linking the same person. There are both people incorrectly matched to someone else and people who fail to be matched. However, researchers have developed sophisticated approaches to conducting these linkages. See Scott Winship, “Economic Instability Trends and Levels across Household Surveys,” Final Report Submitted for the National Poverty Center Survey of Income and Program Participation (SIPP) Small Grants Competition, January 2011, https://web.archive.org/web/20131104184329/http://npc.umich.edu/news/events/census_sipp_conf/winship.pdf.
Matching is easier using the IPUMS CPS data, which I use from 1989 to 2023, since the IPUMS team has matched records and included linking variables. My linking of 1973 data to the 1972 and 1974 estimates is based on the approach in Winship, “Economic Instability Trends and Levels across Household Surveys.” The latter uses methods similar to those used by the IPUMS team.
Nonworkers are asked whether the reason was illness or disability, taking care of home or family, going to school, being retired, (through 1983) being in the Armed Forces, being unable to find work, or “other.”
For this sample of men, I can link 70 percent to another earnings record in 1973 and 62 percent in 2022.
I can link 70 percent of men in 1973 and 64 percent in 2022.
For reference, this recoding affects 3.6 percent of the unweighted sample of men age 15 or older linked to a second earnings record in 1973 and 4.5 percent in 2022. It replaces 19.0 percent of the zeroes in 1973 and 13.2 percent in 2022. It replaces 19.4 percent of the zeroes not due to difficulty finding work in 1973 and 13.5 percent in 2022. Among men ages 25 to 54, it replaces 24.0 percent of the zeroes not due to difficulty finding work in 1973 and 26.7 percent in 2022.
This drops 15 percent of cases from the unweighted sample of men age 15 or older linked to a second earnings record in 1973. Without the recoding of zeroes to a lagged or lead value, this step would drop 19 percent of cases. The figures for 2022 are 29 percent and 33 percent. The recoding by use of second observations reduces the number of men dropped when excluding zeroes by roughly 10 to 20 percent. For men ages 25 to 54, relatively few cases are dropped—just 2.8 percent in 1973 and 6.5 percent in 2022.
This recoding affects 1.0 percent of the unweighted sample in 1973 and 2.0 percent in 2022. It affects 5.3 percent of the zeroes not due to difficulty finding work in 1973 and 6.0 percent in 2022. After this recoding, dropping the remaining zeroes removes 17.8 percent of cases in 1973 and 31.4 percent in 2022. Among men ages 25 to 54, the recoding affects 9.2 percent of zeroes not due to difficulty finding work in 1973 and 15.9 percent in 2022. After the recoding, dropping remaining zeroes removes 3.3 percent of cases in 1973 and 7.5 percent in 2022.
“Economic reasons” for part-year work include having spent time unemployed (looking for work) and indicating that part of the year was spent not working or looking due to no work being available. This recoding affects 4.4 percent of the unweighted sample in 1973 and 2.7 percent in 2022.
This drops an additional 10.2 percent of cases from the unweighted sample in 1973. For 2022, it drops an additional 4.0 percent. For men ages 25 to 54, the percentages are 2.9 in 1973 and 2.0 in 2022.
It also includes part-time men. The CPS lacks a variable for 1973 that indicates why someone worked part-time rather than full-time. From 1989 to 2022, however, the increase in men’s earnings at the previous stage (30 percent) is only slightly lower (28 percent) if full-year, part-time earnings are replaced with full-year, full-time earnings from another year and if part-year, part-time earnings are replaced with part-year, full-time earnings. (In all instances, the original values are retained if a worker was part-time due to economic problems.) This recoding affects 2.2 percent of the unweighted sample in 1989 and 1.9 percent in 2022. Dropping the remaining part-time workers (except those working part-time due to economic problems) lowers the increase in the median to 23 percent. Doing so removes an additional 7.7 percent of cases in 1989 and 5.2 percent of cases in 2022. This last adjustment appears sensitive to the end year chosen—using 2021 instead of 2022, the median rises by 27 percent, which is only a bit less than the 29 percent before dropping part-time workers.
Winship, “America Is Still Working.”
Ibid. Scott Winship, “What’s behind Declining Male Labor Force Participation: Fewer Good Jobs or Fewer Men Seeking Them?” Mercatus Center, 2017, https://www.mercatus.org/research/research-papers/whats-behind-declining-male-labor-force-participation.
I establish these 1973 benchmarks separately for men ages 15-44, and 45-64. I don’t add any disabled workers back in if they are age 65 or older. I retain first the non-working disabled men whose earnings have been replaced by full-year earnings from another year, followed by the remaining “excess” non-working disabled men (who have earnings of $0).
For men ages 25 to 54, the final sample is smaller than the full sample of linkable men 25-54 only by 5 percent in 1973 and 9 percent in 2022. It is smaller than the full sample of linkable men with positive earnings by 1 percent in 1973 but larger by 1 percent in 2022.
The approach I use assumes that all “excess” disabled non-workers would have below-median earnings if they worked, so it likely understates the true counterfactual if disability had not grown more common among working-age men. Adding excess disabled men does not change the unweighted sample in 1973 (by design) and increases it by 1.7 percent in 2022. Adding disabled part-year workers increases the 1973 sample by 1.4 percent and the 2022 sample by 0.5 percent.
The median for year-round full-time workers increased by a bit more—41 percent.
For year-round full-time workers, the median rose 41 percent from 1973 to 2023, 9 percent from 1973 to 1989, and 29 percent from 1989 to 2023.
The median for year-round full-time workers increased 112 percent.
The increase from 1973 to 2023 among year-round full-time workers was 113 percent.