Before the numbers speak, define who is being counted
To understand what work from home changed for people with disabilities, we first need to understand what the labor-market numbers are counting. This chapter focuses on the definitions, populations, and comparison periods behind the BLS snapshot and the IZA study, so readers can tell a real change in employment from a change in the denominator. The practical payoff is simple: a statistic about disability employment becomes useful only when its population, outcome, and time window are clear.
The CPS records functional difficulty and labor-force state
The Current Population Survey does not begin with a medical diagnosis. It asks whether household members have serious difficulty hearing, seeing, concentrating, remembering or making decisions, walking or climbing stairs, dressing or bathing, or doing errands independently. A “yes” to any of these six questions places a person in the CPS disability classification. The questions are designed to identify functional difficulty relevant to population statistics. They do not describe the cause or severity of a condition, whether someone wants a job, or which accommodation would work best.
That distinction prevents a common interpretive error. If a person is classified as having a disability and is not employed, the data do not tell us that they lack the capacity to work. They tell us that, during the survey reference period, the person did not meet the survey's definition of employment. The reasons may include health, retirement, caregiving, a preference not to work, discouragement, the absence of a suitable job, or a workplace arrangement that makes available work impractical. The survey classification is a starting point for asking which conditions shape participation, not an explanation of the outcome.
The labor-force categories make the same point from another angle. A person is employed if they worked for pay, worked in their own business or farm, worked at least 15 hours unpaid in a family business, or were temporarily absent from a job. A person is unemployed only if they had no employment, were available for work, and had made specific efforts to find a job in the preceding four weeks, subject to the survey's recall exception. Everyone else is not in the labor force. The civilian labor force is the sum of employed and unemployed people.
These are not interchangeable descriptions of nonemployment. Someone outside the labor force may want a job, but may not have searched recently; someone unemployed is actively in the market under the survey definition. In the BLS 2025 release, about 75% of people with a disability were not in the labor force, compared with about 32% of people with no disability. Yet among people outside the labor force, only about 3% of those with a disability said they currently wanted a job, compared with nearly 7% of those without a disability. The figures do not imply that most disabled nonparticipants are silently waiting for a remote job. They show that the category includes many different situations and must be interpreted carefully.
The denominators are just as important as the labels. The employment-population ratio divides employed people by the entire civilian noninstitutional population. The unemployment rate divides unemployed people by the labor force, not by the whole population. In the BLS table, 8.029 million people with a disability were employed and 725,000 were unemployed, for a labor force of 8.755 million. That produces an unemployment rate of 8.3%. The 22.8% employment-population ratio instead compares the 8.029 million employed people with the full disability population of 35.286 million. These two percentages answer different questions: how many people in the population are working, and how many active labor-force participants are jobless. A change in participation can move the first measure without appearing as a corresponding change in the second.
This distinction is central to the WFH question. If a remote arrangement makes work feasible for someone who had stopped searching, that person may move from “not in the labor force” to employed without ever appearing in the unemployment rate. If it helps an active jobseeker find a job, unemployment may fall. A study that only reports the employment-population ratio cannot tell which path occurred. The categories do not identify the mechanism, but they tell us where to look for one.
The headline employment gap is nevertheless substantial. In 2025, 22.8% of people with a disability age 16 and over were employed, compared with 65.2% of people with no disability. The unemployment rates were 8.3% and 4.1%, respectively. Age is part of the context: half of the population with a disability was age 65 or older, compared with about 18% of the population with no disability. When the comparison is restricted to ages 16 to 64, the disability employment-population ratio rises to 38.1%, but the no-disability ratio is still 74.8%.
The age adjustment illustrates why a fair comparison is not the same as a convenient one. The overall 16-and-over rate includes many older people with and without disabilities, but the two populations have very different age profiles. Restricting the comparison to working age removes some of that compositional difference; it does not remove the disability employment gap. At the same time, an age-restricted result cannot be used to describe older people whose employment decisions may involve retirement and different labor-market constraints. A number becomes more interpretable when the adjustment is named, but it does not become universal.
The BLS figures therefore establish two things at once. First, people with disabilities occupy a much less favorable labor-market position. Second, the position cannot be interpreted through a single number or a single presumed cause. The relevant question for WFH is not “How many disabled people are employed?” alone. It is whether changes in the conditions of work alter the probability that people who could work under one set of conditions can participate under another.
There is also a time-series warning attached to the latest BLS snapshot. October 2025 CPS data were not collected because of a federal government shutdown, so the annual estimates for 2025 are averages over 11 months. BLS says they are not strictly comparable with annual averages for other years. The release also notes that industry figures use a new classification system beginning in January 2025 without historical data being revised. These breaks do not make the snapshot useless; they tell us what role it can play. The release is strong evidence about the current measured level and the structure of the population, but it is not a clean additional point in an ordinary annual trend.
The research sample narrows the question to protect the comparison
The IZA paper asks a different question from the BLS release. Rather than describing everyone age 16 and over with any reported disability in a single calendar year, it studies people ages 18 to 64 and compares two periods: January 2018 through December 2019, before the pandemic, and July 2022 through June 2024, after the pandemic disruptions had ended and overall employment had returned to pre-pandemic levels. It omits 2020 and 2021, when the pandemic itself affected both work and survey conditions.
The paper's baseline also excludes cognitive difficulty and retains five categories: hearing, vision, ambulatory, self-care, and independent-living difficulties. The combined population in those categories remained broadly stable before and after COVID, at around 9 million in the paper's account. The authors call this group “physical disabilities,” but they explicitly recognize that self-care and independent-living limitations can involve more than purely physical restrictions. The label is an analytical convenience for a more stable comparison, not a universal definition of disability.
Why does stability matter? Suppose the number of people classified in a group rises after an event, and the newly included people happen to have higher employment prospects. The group's employment rate could rise even if the event did nothing to improve access for people who were already in the group. That is the composition problem the paper is trying to reduce. The number of people reporting cognitive difficulty rose from around 6 million to more than 7 million afterward in the paper's figures. Including them is not wrong, but it changes the question: the result would combine a possible work-organization effect with a shift in who appears in the measured population.
This is a general lesson about before-and-after evidence. A stable population is not automatically a causal control group, and an unstable population is not automatically unusable. The researchers are trying to ensure that a change in the measured outcome is less likely to be produced simply by a change in membership. They then use occupation-level differences in WFH to investigate whether the remaining employment change lines up with where remote work expanded. The sample restriction is therefore a prerequisite for the later design, not the conclusion itself.
The paper's narrower sample shows an employment rate for people with a physical disability rising from 36.2% in 2019 to 41.3% in the first half of 2024. In its broader trend description, the paper reports that employment for this group rose 12.4% after the pandemic while employment for people without a disability remained roughly constant. Those numbers belong to the paper's defined sample and periods. They should not be set beside the BLS 2025 rate as if the three figures were observations from one common series: the age ranges, disability categories, periods, and annual-versus-monthly constructions differ.
The choice of survey also matters. The paper uses CPS disability and employment information for its outcome. It reports that the American Community Survey classified roughly 40% more working-age people as disabled in 2022–2023 than the CPS and that the correlation of the main outcome across the surveys was only 0.33. The authors therefore use CPS disability status for the employment analysis, while using an ACS transportation-to-work question answered by workers without a disability to measure occupation-level WFH opportunities. A survey difference of this size does not prove that one survey is universally correct; it shows that definitions and collection contexts can materially change the apparent population and outcome.
The two-survey arrangement may seem inconsistent until the questions are separated. The CPS supplies the linked disability and labor-force statuses needed to construct the employment outcome. The ACS question supplies a way to observe whether a person usually worked from home, which is used to compare WFH opportunity across occupations. The researchers do not need the ACS and CPS to produce identical disability counts for that particular use of the ACS measure. They do need to avoid treating an ACS disability estimate as interchangeable with a CPS disability estimate. Measurement can be fit for one variable and unsuitable for another.
The measurement path can be summarized like this:
The diagram is not a claim that the research sample is the only valid population. It shows how a descriptive statistical system becomes a research comparison through successive choices. First, the CPS defines functional disability. Next, it assigns a labor-force state. The paper then restricts age, categories, and periods to make the before-and-after comparison less vulnerable to composition changes. Each filter improves one kind of interpretability while reducing the range of people to whom the result directly applies.
That trade-off is essential for reading the later evidence. The BLS snapshot tells us why accessibility in employment remains a consequential issue, including for people outside the working-age range. The IZA study gives a more controlled test for a particular working-age population whose measured composition is more stable. Neither source should be forced to answer the other's question. The research estimate can be persuasive within its defined sample without automatically describing people with cognitive disabilities, older retirees, workers in every occupation, or people whose home conditions make remote work difficult.
A precise population does not settle the causal question, but it makes the question possible. Now that the categories, denominators, and comparison windows are visible, we can examine the conditions that WFH might change: the cost of reaching work, the ability to control the work environment, and the occupational tasks that determine whether location can vary at all.