Federal Surveys Undercount People with Disabilities as Defined by the Amendments to the Americans with Disabilities Act

Introduction

Disability can be defined and identified in multiple ways. Since 2008, several federal surveys, such as the American Community Survey (ACS), have used six yes-or-no questions (referred to as the ACS-6) to identify people having serious difficulty with hearing, vision, cognition, ambulation, self-care, or independent living. Respondents who answer “yes” to any ACS-6 are considered to have a disability (Bureau of Labor Statistics 2025). The ACS-6 set a minimum standard for survey questions on disability (Landes and others 2025) and are not designed to identify all people with disabilities (Bureau of Labor Statistics 2025).

The ACS-6 largely align with the definition of disability from the World Health Organization's International Classification of Functioning, Disability, and Health (ICF) (Brault 2009). At the same time, the ACS-6 aim to measure disability in accordance with the Americans with Disabilities Act of 1990 (ADA; Public Law 101-336), a federal civil rights law prohibiting discrimination against people with disabilities (Brault, Stern, and Raglin 2007). The ADA defines disability in three ways: (1) a physical or mental impairment that substantially limits one or more major life activities (MLAs), (2) a record of such an impairment, or (3) being regarded as having such an impairment.

The ADA does not require the federal government to measure disability in surveys, but the Census Bureau developed the ACS-6 to align with the ADA of 1990 because of an interest in assessing the effect of the ADA (Brault, Stern, and Raglin 2007). The ACS-6 have remained unchanged for almost two decades, even though the ADA was amended in 2008. The ADA Amendments Act of 2008 (ADAAA; Public Law 110-325) expands the definition of MLAs to include major bodily functions and provides a nonexhaustive list of MLAs (for example, seeing and working), including a nonexhaustive list of the operation of major bodily functions (for example, normal cell growth and endocrine functions).¹ Consequently, the ADA's definition of disability now includes physical or mental impairments that substantially limit the operation of one or more major bodily functions (hereafter, major bodily function limitations).

Because the ACS-6 have not adapted to the ADAAA, this article examines three fundamental questions about disability measurement:

1. To what degree do federal surveys identify people with major bodily function limitations as having a disability?
2. How does that compare with the rate at which the surveys identify people with impairments that substantially limit MLAs as defined before the ADAAA (hereafter, MLA limitations)?
3. By how much would disability prevalence estimates increase if the ACS-6 were expanded to include major bodily function limitations?

This article is the first to point out that the ACS-6 have not evolved to align with the ADAAA. It aims to inform a range of stakeholders of disability measurement (such as the disability community, policymakers, and researchers) about this misalignment and the extent to which it leads to undercounting people with ADA-defined disabilities.

I find that among people aged 18–64 who ever had major bodily function limitations—for the purpose of this article, people who ever had cancer, diabetes, epilepsy, or major depressive disorder (MDD)—less than 40 percent were identified by the ACS-6 as having a disability. This share was statistically significantly lower than the share of people with MLA limitations identified by the ACS-6 as having a disability. For these comparisons, I use data from the 2022 Medical Expenditure Panel Survey (MEPS), 2015 and 2017 National Health Interview Survey (NHIS), and 2020 National Survey on Drug Use and Health (NSDUH).

These findings show a disparity in disability identification in federal surveys: The ADA protects both people with major bodily function limitations and people with MLA limitations, but the former were identified as having a disability at a substantially lower rate than the latter. This disparity may prevent federal surveys, such as the ACS, from effectively serving as a resource for disabled people when resulting survey statistics are used to plan and fund government programs and services (Ross 2023). This disparity may adversely affect a federal agency's ability to publish disability regulations to protect people with disabilities because this disparity underestimates the number of people with disabilities based on the current ADA disability definition, which in turn could underassess the number of beneficiaries of a regulation and ultimately underestimate regulatory benefits. Disability statistics collected by federal surveys have been used to evaluate regulatory benefits and costs. To publish a regulation, a federal agency is required to show regulatory benefits outweigh costs. An example of disability regulations is the recently published final rule updating the regulation implementing Title II of the ADA to add more specific requirements about web and mobile application accessibility (Department of Justice 2024).

One way for federal surveys to address this disparity is to expand the scope of disability measurement to include major bodily function limitations. I quantify that if this scope were expanded to include cancer, diabetes, epilepsy, and MDD, the estimated disability prevalence among people aged 18–64 in the 2023 ACS would increase from 11 percent to roughly 27 percent, suggesting that 33 million additional people would be identified as having a disability. Comparable increases would be found in the MEPS, NHIS, and NSDUH.

It is important to note, though, what my study does not examine: the normative question of whether the ACS-6 should follow the expanded definition of major life activities in the ADAAA. Nevertheless, this article can serve as a technical reference in this examination. It is up to the Census Bureau, other federal agencies, and relevant stakeholders to consider whether or how the ACS-6 should be revised.

Literature Review

This article focuses on disability as defined by the ADA, although there are alternative definitions, including disability as defined by the Social Security Disability Insurance and Supplemental Security Income programs (Social Security Administration 2012), the Nagi model summarized in Burkhauser and others (2002), and the ICF framework summarized in Burkhauser, Houtenville, and Tennant (2012). The Nagi model considers disability as a process in which an individual's illness interacts with the socioeconomic environment. Under the ICF framework, disability refers to the presence of an impairment, activity limitation, and/or participation restriction, based on a health condition.

The ACS-6 largely align with the ICF framework (Brault 2009) and aim to follow MLAs as defined before the ADAAA. This is because the ACS-6 do not focus on the presence of specific conditions but on the realized effects of such conditions.

Prior studies have examined the strengths and weaknesses of the ACS-6 (for example, Burkhauser, Houtenville, and Tennant 2012; Burkhauser and others 2014; Hall and others 2022; Karpman and Morriss 2024; Landes and others 2025; Weeks and others 2021). The Census Bureau recently proposed, then paused, revising the ACS-6 (Santos 2024). This proposal used the Washington Group Short Set on Functioning (WG-SS) to replace the ACS-6 (Steinweg and others 2023). The WG-SS differs from the ACS-6 because it includes a question about communication difficulty and uses graded responses, as opposed to yes-or-no responses in the ACS-6. The WG-SS matches the ICF framework and follows MLAs as defined before the ADAAA but does not consider major bodily function limitations.

This article contributes to the literature on the share of people with physical or mental health conditions identified by the ACS-6 as having a disability. For example, it shares the perspective of Hermans, Morriss, and Popkin (2024) that it is valuable to measure disability in a manner that better matches the ADA. It implements this perspective before their paper was published. This article complements Burkhauser, Houtenville, and Tennant (2012), who found that 45 percent of people with a work-activity limitation answered “no” to all ACS-6. Burkhauser and others (2014) found that 34 percent of people receiving Social Security Disability Insurance, Supplemental Security Income, or both answered “no” to all ACS-6.

Hall and others (2022) and Karpman and Morriss (2024) found up to 32 percent of people aged 18–64 with physical or mental health conditions were not identified by the ACS-6 as having a disability, while Landes and others (2025) reported a share of up to 25 percent for people aged 18 or older. These studies used their findings to advocate expanding the ACS-6 to improve the accuracy of disability prevalence estimates. However, these studies did not relate this issue to the expanded legislative definition of MLAs.

By contrast, I use the ADAAA's expanded definition of MLAs to measure disability. I find a substantially higher share, more than 60 percent, of people with major bodily function limitations were not identified as having a disability. This is at least partly because my analysis sample is more consistent with the current ADA disability definition. First, my analysis sample includes people with major bodily function limitations without requiring that these limitations affect daily activities; conversely, their analyses were restricted to people with physical or mental health conditions that affected daily activities, required the use of assistive equipment or devices, or both. This restriction deviated from the current ADA disability definition because impairments substantially limiting the operation of bodily functions are disabilities themselves, regardless of whether they substantially limit MLAs as defined before the ADAAA, require the use of assistive equipment or devices, or both.

Second, those studies focused on people who currently have a physical or mental health condition, but I mirror the ADA's coverage of people who have a record of disability by including people who ever had a major bodily function limitation. Some of these people do not currently have a limitation and may be more likely to answer “no” to all ACS-6.² Still, they are considered as having an ADA-defined disability.

Data and Methods

The population of interest for my study includes working-age individuals aged 18–64 to facilitate comparison with the findings of Hall and others (2022) and Karpman and Morriss (2024). I analyze major bodily function limitations by examining four conditions—cancer, diabetes, epilepsy, and MDD—listed as examples of major bodily function limitations in the federal regulation implementing the ADAAA of 2008 (Department of Justice 2016). These four conditions are chosen because their presence can be observed in three recent federal surveys listed in Table 1: the 2022 MEPS, 2015 and 2017 NHIS, and 2020 NSDUH. This approach allows identification of people who ever had cancer, diabetes, epilepsy, or a major depressive episode (MDE), which leads them to be considered as having a disability because they meet the first or second way of the ADA disability definition.

Likewise, I analyze MLA limitations by examining six activities—caring for oneself, walking, seeing, hearing, working, and performing manual tasks—listed as examples of MLA limitations in the federal regulation implementing the ADA of 1990 (Department of Justice 1991). These six activities are chosen because their presence can be observed in the surveys mentioned above (in fact, the first four are captured by the ACS-6). One caveat is that the surveys do not identify people who ever had MLA limitations, only people who currently have such limitations, which lead them to meet only the first way of the ADA disability definition.

I analyze the degree to which people who ever had one of the four major bodily function limitations answered “yes” to any ACS-6. I use within-survey mean comparison to contrast that rate with the rate from people who currently have one of the six MLA limitations. People with the first four MLA limitations were, by definition, fully identified by the ACS-6 as having a disability—their rate is 100 percent—because the ACS-6 capture those limitations. See Appendix A for additional data description and a justification for using MDE as a proxy for MDD.

I conduct a robustness check to assess the extent to which my findings would change if my analysis sample were restricted to people who currently have, rather than ever had, major bodily function limitations. An advantage to this restriction is that it allows direct comparison of answers from people with current major bodily function limitations to the answers of people with current MLA limitations. However, this restriction deviates from the ADA disability definition by excluding people with a record of a major bodily function limitation.

I perform single imputation using logistic regression to estimate by how much disability prevalence estimated by the ACS-6 would increase if the scope of disability measurement were expanded to include cancer, diabetes, epilepsy, and MDE. This imputation combines data from the MEPS, NHIS, NSDUH, and ACS and accounts for their survey design parameters (sampling weights, stratification, and clustering). Similar data combinations have been used in previous studies. For example, Ingram and others (2003) combined NHIS and Census Bureau data to predict the single race that best described respondents who reported multiple races in the 2000 Census. Schenker, Raghunathan, and Bondarenko (2010) combined data from the NHIS and the National Health and Nutrition Examination Survey to improve the accuracy of self-reported disease occurrence.

The final variable of interest is a binary variable indicating whether a respondent answered “yes” to any ACS-6, cancer, diabetes, epilepsy, or MDE. To create this variable, I follow medical literature and use data on cancer and diabetes from the MEPS and NHIS, epilepsy from the NHIS, and MDE from the NSDUH to impute disease occurrence for respondents in the other surveys that otherwise would lack such information. The analysis sample for imputation includes adults aged 18–64 who answered “no” to all ACS-6. I exclude from imputation people answering “yes” to any ACS-6 because it is already known that the final variable of interest should be coded “yes” for them.

All four surveys are nationally representative, so differences in demographics—sex, age, and race and ethnicity—across surveys are assumed, and empirically verified, to be small. Nevertheless, I adjust for demographic differences by (1) dividing respondents into blocks based on their demographic characteristics and (2) performing within-block logistic regression that controls for demographics. See Appendix B for details on the imputation, an empirical assessment of demographics, and a sensitivity analysis.

Results

Table 2 presents disability rates, estimated by the ACS-6, of civilian noninstitutionalized adults aged 18–64 from four surveys. The estimate is the lowest in the ACS (11.1 percent), followed by 11.2 percent in the MEPS and 14.4 percent in the NHIS, and is the highest in the NSDUH (16.2 percent). Despite this variability, these rates closely align with those previously published by other studies (for example, Mitra and others 2022). The Bureau of Labor Statistics (2025) suggests that this variability may be explained by survey differences in some attributes, such as survey context.

Table 3 shows how the ACS-6 were answered by people with major bodily function limitations (Panel A) and MLA limitations (Panel B), as well as the prevalence of each limitation. Among adults aged 18–64, approximately 13 percent were ever told by a health professional that they had cancer, diabetes, or epilepsy. Seventeen percent experienced at least one MDE in their lifetime. Between 8 percent and 11 percent had difficulties caring for oneself, walking, seeing, or hearing. For difficulties working or performing manual tasks, the share ranged from 7 percent to 13 percent.

Among adults aged 18–64 who ever had cancer, diabetes, or epilepsy, 34 percent answered “yes” to any ACS-6. For people who ever had an MDE, the share was 38 percent. These findings indicate that less than 40 percent of people with major bodily function limitations were identified by the ACS-6 as having a disability.

As mentioned earlier, 100 percent of people with difficulties caring for oneself, walking, seeing, or hearing answered “yes” to any ACS-6 because the ACS-6 capture these difficulties. Approximately 70 percent of people with difficulties working or performing manual tasks answered “yes” to any ACS-6. These two estimates indicate that at least 70 percent of people with MLA limitations were identified by the ACS-6 as having a disability.

Collectively, these results show a disparity in disability identification in federal surveys: While the ADA protects both people with major bodily function limitations and MLA limitations, the former were identified as having a disability at a substantially lower rate compared with the latter (less than 40 percent versus at least 70 percent). This difference is statistically significant (see Table 3 notes).

A robustness check indicates that if the analysis sample were restricted to people who currently have (rather than ever had) major bodily function limitations, less than half would be identified as having a disability. This suggests that such a restriction would reduce—but not eliminate—the disparity in disability identification because the gap would still be greater than 20 percentage points (“at least 70 percent” minus “less than half”). The estimate of “less than half” is based on a separate analysis indicating that 49 percent of people who currently have an MDE answered “yes” to any ACS-6. This is an 11-percentage-point increase from individuals who ever had an MDE (49 percent versus 38 percent). I assume this 11-percentage-point increase holds for people with cancer, diabetes, or epilepsy, so among people currently with those conditions, 45 percent (34 percent + 11 percent) would answer “yes” to any ACS-6. The NSDUH identifies both people who ever had and currently have an MDE, but the MEPS and NHIS identify only people who ever had cancer, diabetes, or epilepsy.

Table 4 shows the extent to which disability prevalence estimates in federal surveys would increase if the scope of disability questions were expanded to include cancer, diabetes, epilepsy, and MDE. In Panel A, the benchmark includes the ACS-6 only, so its disability prevalence estimates are the same as those in Table 2. The additional rows in Panel A report disability estimates if the scope were expanded to include diseases whose information is available in the surveys. For example, NHIS data show that the disability prevalence estimate would increase to 23.3 percent if cancer, diabetes, and epilepsy were included.

Panel B shows imputed results for three scope expansions. For instance, the imputed 2023 ACS data show that if all four diseases were included, the disability prevalence estimate would be 27.1 percent, an increase of 16.0 percentage points from the benchmark (27.1 percent − 11.1 percent). The 2023 ACS estimates that the number of Americans aged 18–64 is 203 million (Census Bureau n.d.); therefore, the number of additional people who would be considered as having a disability would be 32.5 million (203 × 0.16). A subgroup analysis indicates that the increase in disability prevalence estimate would be larger for women, people aged 50–64, and people who identify as non-Hispanic White.

When all four diseases were included, the increase in disability prevalence estimate varies across surveys by less than 4 percentage points (from 16.0 to 19.6), suggesting the increase is robust across surveys. Nevertheless, the imputed disability prevalence estimates should be interpreted as rough estimates because of three potential caveats: (1) portability, (2) model misspecification and comorbidity, and (3) invalid confidence interval.

The imputation assumes portability—a logistic regression fitted to data in one survey applies to data in another. Some evidence supports this assumption (for example, all four surveys are nationally representative, and the MEPS uses the NHIS respondent pool as its sampling frame), but some issues may challenge this assumption. For instance, survey context varies across the four surveys: the MEPS, NHIS, and NSDUH are health surveys, while the ACS provides detailed population and housing information. This variation may result in higher disease prevalence estimates in the health surveys (the MEPS, NHIS, and NSDUH) compared with those in the non-health survey (the ACS). The Bureau of Labor Statistics (2025) suggests that respondents may be more likely to answer “yes” to disease occurrence questions that appear in health surveys than those in non-health surveys.

Additionally, the imputation may misspecify the relationship among the four diseases, leading to either an over- or underestimation of comorbidity—people who ever had cancer/diabetes/epilepsy and MDE—and subsequently to an under- or overestimation of disability prevalence when all four diseases were included. This is because the imputation aims to estimate the share of people in the union of three sets (cancer/diabetes, epilepsy, and MDE). Although the MEPS, NHIS, and NSDUH allow me to estimate the share in each set individually, no survey provides information on the extent to which these sets overlap. To mitigate this potential source of error, I use medical literature to guide model specification and divide respondents into blocks (see Appendix B for details).

In Table 4, when the scope includes the ACS-6 plus all four diseases, the ACS estimate is rounded down to 27.1 percent (from 27.146 percent), with a 95 percent confidence interval from 27.109 percent to 27.183 percent. This interval reflects the variability that would arise if the ACS drew a new sample on each replication, but the sample from the other three surveys (which influences the imputed values) remained fixed across replications. Some researchers may argue that we should be interested in replications where all surveys drew new samples on each replication, which would yield somewhat wider confidence intervals. Alternative methods are available for estimating the confidence interval, but they would likely yield qualitatively similar results.

Conclusion

This article is the first to highlight that the ACS-6 have not adapted to the ADAAA of 2008, which expands the definition of MLAs to include major bodily functions. I show that among people aged 18–64 who ever had major bodily function limitations, less than 40 percent were identified by the ACS-6 as having a disability, which is substantially lower than the share identified among people with MLA limitations. I further find that if the scope of the ACS-6 were expanded to include major bodily function limitations, the 2023 ACS disability prevalence estimate for people aged 18–64 would increase from 11 percent to roughly 27 percent.

Overall, these findings suggest that the ACS-6 only loosely align with the ADAAA. This is expected because the ACS-6 were developed before the ADAAA was enacted. The ACS-6 are not required to align with the ADA of 1990 or evolve to align with any amendments to the ADA.

If the ACS-6 were revised, they would be implemented in the context of the ADAAA, which remains one of the major laws protecting Americans with disabilities against discrimination. My findings prompt a question for stakeholders of disability measurement: If the ACS-6 were revised, what would be the socially optimal extent to which the revised ACS-6 measure disability using a framework aligning with the current ADA disability definition? The degree could range from zero (no alignment) to 100 percent (complete alignment). Full alignment is unlikely to be socially optimal because dozens of questions are required to identify all people with disabilities (Bureau of Labor Statistics 2025), but virtually no federal surveys can afford the time to administer that many questions. An ACS content test was conducted in 2022 to evaluate the recently proposed, then paused, ACS-6 revision (Santos 2024), and the content test's evaluation report (Steinweg and others 2023) makes no mention of measuring disability using a definition in keeping with the ADA. In fact, I could not find any discussion about the ADA in the 2023 report. It is up to the Census Bureau, other federal agencies, and their stakeholders to determine whether this change in the amount of discussion about the ADA is socially optimal.

Appendix A: Data Description

Table 1 lists the three nationally representative surveys used to analyze the degree to which people with major bodily function limitations—cancer, diabetes, epilepsy, or MDD—answered “yes” to any ACS-6, compared with people with MLA limitations—difficulties caring for oneself, walking, seeing, hearing, working, or performing manual tasks. The first four difficulties are captured by four questions in the ACS-6.

2022 MEPS

I use four data fields from the 2022 MEPS Household Component (HC): (1) the ACS-6, (2) cancer and diabetes, (3) difficulties working and performing manual tasks, and (4) demographics. I choose 2022 because, as of August 2025, it is the most recent year for which the MEPS HC full-year consolidated data file is available. Cancer and diabetes are MEPS priority conditions, so respondents with cancer or diabetes are identified by a direct, dichotomous question for each condition, asking respondents whether they were ever told by a health professional that they had cancer or diabetes. I therefore classify people with major bodily function limitations in a dichotomous manner. The MEPS does not ask respondents whether they currently have cancer or diabetes.

Because epilepsy and MDD are not MEPS priority conditions, no direct question is available for these conditions. The MEPS medical conditions file collects information on a variety of conditions, including epilepsy and MDD, from respondents who mention a condition and receive treatment in the survey reference year. However, this collection is unlikely to yield nationally representative estimates of disease prevalence because it imposes the requirement of receiving treatment. As a result, I do not use information about epilepsy or MDD captured in the MEPS medical conditions file.

People with difficulties working are identified by two questions. The first is a dichotomous question asking respondents whether they are limited in any way in their ability to work at a job, do housework, or go to school because of an impairment or a physical or mental health problem. Those answering “yes” receive a follow-up question asking them to specify all activities (working at a job, doing housework, or going to school) limited by an impairment or health problem. These two questions allow me to classify people with difficulties working in the same dichotomous manner as those with major bodily function limitations.

People with difficulties performing manual tasks are identified by a question asking respondents how much difficulty they have using fingers to grasp or handle something, such as picking up a glass from a table or using a pencil to write. This question has four response options: (1) completely unable to do it, (2) a lot of difficulty, (3) some difficulty, and (4) no difficulty. I classify people choosing one of the first three options as having difficulties performing manual tasks. An alternative approach, classifying only those choosing one of the first two options, results in a lower prevalence of such difficulties but a higher share of individuals with such difficulties answering “yes” to any ACS-6.

2015 and 2017 NHIS

I use four NHIS data fields: (1) the ACS-6; (2) cancer, diabetes, and epilepsy; (3) difficulties working and performing manual tasks; and (4) demographics. People who ever had cancer, diabetes, or epilepsy are identified by the same direct, dichotomous questions as in the MEPS. The NHIS does not ask respondents whether they currently have cancer, diabetes, or epilepsy.

People with difficulties working are identified by a dichotomous question about limitations in the kind or amount of work they can do because of a physical, mental, or emotional problem. People with difficulties performing manual tasks are identified by a question asking how difficult it is for the respondent to use fingers to grasp or handle small objects without using any special equipment. This question has six response options: (1) can't do at all, (2) very difficult, (3) somewhat difficult, (4) only a little difficult, (5) not at all difficult, and (6) do not do this activity. I classify people choosing one of the first three options as having difficulties performing manual tasks, consistent with the MEPS approach.

I choose 2015 and 2017 because they are the two most recent years for which the NHIS included both the ACS-6 and a direct question about epilepsy. The ACS-6 were removed from the NHIS after 2017, and the direct question about epilepsy was not available in 2016. Combining the 2015 and 2017 datasets increases sample size, consistent with the practice of Weeks and others (2021). I restrict the analysis sample to adult respondents selected to answer the ACS-6 and questions about medical conditions, including cancer, diabetes, and epilepsy.

2020 NSDUH

I use three NSDUH data fields: (1) the ACS-6, (2) MDE, and (3) demographics. I choose 2020 because it is the most recent year for which the NSDUH included the ACS-6. The NSDUH does not have a direct question about MDD, instead using a series of questions to measure whether a respondent had experienced an MDE in their lifetime or in the past year. I classify respondents who experienced an MDE in their lifetime as those who ever had MDD, and respondents who had experienced an MDE in the past year as those who currently have MDD.

I use the MDE measurement as a proxy for MDD because the federal regulation implementing the ADAAA of 2008 (Department of Justice 2016) lists MDD as an example of major bodily function limitations, but the NSDUH measures only MDE. The MDE measurement can serve as a proxy for MDD because, according to personal correspondence with the Substance Abuse and Mental Health Services Administration, the vast majority of MDE is MDD. It should be noted, though, that MDE includes episodes that occur as part of bipolar disorder, but MDD excludes bipolar depression. The NSDUH measures MDE, not MDD, because it cannot distinguish between MDE that occurs only in the context of depression and MDE that involves mania (bipolar depression). This is because the NSDUH does not ask about lifetime mania.

Information on difficulties working or performing manual tasks is not available in the NSDUH because it does not ask the former to all respondents aged 18–64 and does not have a question about the latter at all.

Demographics

Demographic variables common to all three surveys—sex, age, and race and ethnicity—are used to impute disease occurrence. Sex is a binary variable (women versus men) in all three surveys. Age is a continuous variable in the MEPS and NHIS, but in the NSDUH, it is a continuous variable only from 12 to 21 and a categorical variable from 22 to 64 (22–23, 24–25, 26–29, 30–34, 35–49, and 50–64). To reconcile this difference, I create three categorical variables for age: (1) a binary variable (18–49 and 50–64), (2) a three-category variable (18–25, 26–49, and 50–64), and (3) a six-category variable (18–21, 22–25, 26–29, 30–34, 35–49, and 50–64). As explained further in Appendix B, the first two variables are used to divide the analysis sample into demographic blocks, but the third is used as covariates in the logistic regression.

The MEPS and NHIS report race and ethnicity separately, but the NSDUH combines race and ethnicity into a seven-category variable. To adjust for this difference, I create two categorical variables for race and ethnicity: (1) a binary variable (non-Hispanic White and all other racial/ethnic identifications) and (2) a four-category variable (non-Hispanic White; non-Hispanic Black; non-Hispanic, all other races; and Hispanic, any race). Both variables are used to divide the analysis sample into demographic blocks, but the second is also used as covariates in the logistic regression.

Appendix B: Imputation

The imputation for my study comprises four stages:

1. Construct a range for the share of adults aged 18–64 who answered “no” to all ACS-6 but “yes” to cancer, diabetes, epilepsy, or MDE. Use this range to calculate a range for the mean of the final variable of interest, which serves as a reality check for the final imputed values.
2. Divide the analysis sample into demographic blocks and perform within-block regression controlling for demographic characteristics.
3. Use medical literature to inform the model specification of imputing disease occurrence.
4. Conduct a sensitivity analysis of four specifications with differing covariates to select the main specification.

Calculate a Range for the Mean of the Final Variable of Interest

No single survey collects information on all four diseases, so the collective prevalence of these four diseases is unknown. Nevertheless, I take three steps to calculate a range for the mean of the final variable of interest. First, I note that respondents whose final variable of interest is coded “yes” can be divided into two groups: (1) those answering “yes” to any ACS-6 and (2) those answering “no” to all ACS-6 but “yes” to disease occurrence questions—whether they were ever told by a health professional that they had cancer, diabetes, epilepsy, or MDE.

Second, I report in Table B-1 the share of people answering “no” to all ACS-6 but “yes” to disease occurrence questions. This table shows that 8.9 percent of civilian noninstitutionalized adults aged 18–64 answered “no” to all ACS-6 but “yes” to cancer, diabetes, or epilepsy. This share was 10.5 percent for MDE. I construct lower and upper bounds of the share of people who answered “no” to all ACS-6 but “yes” to cancer, diabetes, epilepsy, or MDE. The lower bound assumes a complete overlap between people belonging to the 8.9 percent and 10.5 percent, but the upper bound assumes no overlap. The lower bound is therefore 10.5 percent, but the upper bound is 19.4 percent (8.9 percent + 10.5 percent).

Third, I combine the estimated share of people answering “yes” to any ACS-6 with the lower and upper bounds of the share of people who answered “no” to all ACS-6 but “yes” to cancer, diabetes, epilepsy, or MDE. This combination creates a range for the mean of the final variable of interest. For example, in the 2023 ACS, 11.1 percent of people answered “yes” to any ACS-6, indicating that the mean of the final variable of interest in the 2023 ACS should be between 21.6 percent (11.1 percent + 10.5 percent) and 30.5 percent (11.1 percent + 19.4 percent).

Divide Analysis Sample Into Demographic Blocks

Table B-2 reports demographic characteristics of adults aged 18–64 answering “no” to all ACS-6. Their characteristics were similar across the ACS, MEPS, and NSDUH but differed slightly in the NHIS. For example, the NHIS reported a higher rate of people who identified as non-Hispanic White than the other three surveys.

Table B-3 shows the association between disease prevalence and demographic characteristics of adults aged 18–64 answering “no” to all ACS-6. Epilepsy was more common among people who identified as non-Hispanic White than among those who identified with other racial or ethnic groups. Cancer and diabetes were much more prevalent among people aged 50–64 than among younger age groups. MDE was more prevalent among women, adults aged 18–25, and people who identified as non-Hispanic White.

I divide the analysis sample into blocks based on demographics—sex, age, and race and ethnicity—and perform within-block logistic regressions controlling for demographics. This allows regression coefficients to vary by block, which can improve the adjustment for demographic differences across surveys. The number of blocks created for each disease depends on disease prevalence and how that prevalence varied by age. The number of blocks is as follows:

1. Four for epilepsy based on the binary age variable (18–49 and 50–64) and binary racial/ethnic variable (non-Hispanic White and all other racial/ethnic identifications).
2. Twenty each for cancer/diabetes and MDE based on the three-category age variable (18–25, 26–49, and 50–64), the binary sex variable (women and men), and the four-category racial/ethnic variable (non-Hispanic White; non-Hispanic Black; non-Hispanic, all other races; and Hispanic, any race). However, the binary racial/ethnic variable is used for two populations with low prevalence of specific disease (ages 18–25 with cancer/diabetes and ages 50–64 with MDE).

Use Medical Literature to Inform the Model Specification of Imputing Disease Occurrence

I impute three disease occurrence variables: cancer/diabetes, epilepsy, and MDE. I begin by imputing epilepsy occurrence because medical literature suggests that epilepsy may be a risk factor for cancer, diabetes, and MDE (for example, Adelöw and others 2006; Kanner and Balabanov 2002; Li and others 2021). This approach allows me to use epilepsy (both actual and imputed values) as a covariate when imputing the occurrences of cancer/diabetes and MDE.

Both the MEPS and NHIS provide information on cancer and diabetes, so the relationship between these two diseases is identified. I therefore impute a single disease occurrence variable—cancer/diabetes—indicating whether the respondent ever had cancer or diabetes, rather than imputing separate indicators, to avoid over- or underestimating their co-occurrence.

MDE data come from the NSDUH, so I impute a separate disease occurrence variable for MDE. The relationship between cancer/diabetes and MDE is not identified because information on cancer/diabetes and MDE is not provided by a single survey. Simultaneity may arise between cancer/diabetes and MDE because cancer and diabetes could be a risk factor for MDE, and vice versa (for example, Gillett and others 2024; Mallet and others 2018; Mössinger and Kostev 2023). However, I am not concerned about simultaneity here because my objective is just to impute missing data on disease occurrence, not to interpret regression coefficients.

I impute four specifications with differing covariates and show how the results vary as I change the covariates. I choose the specification that produces the most reasonable mean imputed values as the main specification and report those results in Table 4. The four specifications are:

1. Covariates include demographics³ only.
2. Covariates include demographics and epilepsy.⁴
3. Covariates include demographics and epilepsy. Cancer/diabetes is added as a covariate when imputing MDE.
4. Covariates include demographics and epilepsy. MDE is added as a covariate when imputing cancer/diabetes.

After performing a logistic regression, I estimate predicted probability of disease occurrence for respondents who otherwise would not have disease occurrence information. I use the predicted probability and a random number generator to impute a dichotomous disease occurrence outcome for these respondents. For example, for respondents with a predicted probability of 0.6, I use a random number generator to impute a dichotomous outcome of 0 or 1 so that they have a 60 percent chance of getting a 1 and a 40 percent chance of getting a 0.

Use Sensitivity Analysis to Choose the Main Specification

Table B-4 presents results from the four specifications. This presentation serves as a sensitivity analysis, showing how results vary across specifications and aiding selection of the most suitable specification. Under every specification, the mean of the imputed key outcome—the share of people who would answer “yes” to any ACS-6, cancer, diabetes, epilepsy, or MDE—falls within the range presented earlier. The mean of the imputed key outcome varies little across specifications, suggesting robust results. For example, as noted earlier, the range for the mean of the imputed key outcome in the 2023 ACS is from 21.6 percent to 30.5 percent. Across the four specifications, the mean of the imputed key outcome in the 2023 ACS varies from 27.1 percent (in specification A) to 28.2 percent (in specification C).

Because all four specifications yield mean estimates near 27 percent, the selection of a main specification from the four specifications may not be crucial. This is because the potential over- or underestimation of comorbidity discussed below may lead to an under- or overestimation of the mean of the imputed key outcome that exceeds the difference in the mean estimates across the four specifications.⁵ I therefore choose specification A, which produces the lowest estimate, as the main specification presented in Table 4.

As discussed in the Results section, the imputation may misspecify the relationship among the four diseases, resulting in an over- or underestimation of comorbidity—people who ever had cancer/diabetes/epilepsy and MDE—and consequently lead to an under- or overestimation of disability prevalence if the scope of disability measurement were expanded to include all four diseases.

On the one hand, it appears that in the ACS the imputation results in an overestimation of comorbidity and therefore leads to an underestimation of the key outcome's mean. The MEPS and ACS have nearly identical demographics (Table B-2), and their differences in the estimated disability prevalence in the first three rows of Table B-4 were less than 1 percentage point.⁶ However, in the fourth row, the mean of the key outcome in the ACS was lower than that in the MEPS by more than 3 percentage points (27.1 percent versus 30.8 percent). One explanation for this larger gap is that ACS respondents were more likely to have comorbidity than MEPS respondents. However, a more plausible explanation is that this larger gap results from the imputation's overestimating comorbidity in the ACS, because virtually all MEPS data on cancer and diabetes come from survey responses, whereas all ACS disease occurrence data come from imputation.

On the other hand, the imputation relies on a conditional independence assumption (for example, the occurrence of cancer/diabetes is independent of that of MDE, conditional on demographics). It is suitable to make this assumption because no survey covers all four diseases, resulting in an unidentified relationship between the four diseases. However, if this assumption does not hold, the imputation may underestimate comorbidity, resulting in an overestimation of disability prevalence in the ACS.

I use the imputation of two disease occurrence variables—cancer/diabetes and MDE—in specification A to illustrate the potential underestimation of comorbidity. Similar discussions can apply to epilepsy and the other specifications. In specification A, only demographics are included as covariates. For the sake of discussion, I impute the occurrence of cancer/diabetes before that of MDE. As a result, the imputed values for cancer/diabetes are just a function of demographics plus some random noise. Because MDE is imputed after cancer/diabetes, the imputation evenly distributes the NSDUH's MDE occurrence—a 1 in MDE value—between people in the other three surveys with a 1 in their cancer/diabetes values and those with a 0 in their cancer/diabetes values. This even distribution is accurate if the occurrences of MDE and cancer/diabetes are conditionally independent, given demographics.

If, however, the occurrences of cancer/diabetes and MDE are positively correlated—people with cancer/diabetes are more likely to have MDE than other people with the same demographics—the imputation will underestimate comorbidity because it does not consider this correlation. This lack of consideration results in distributing too much MDE occurrence to people with a 0 in their cancer/diabetes values but too little to those with a 1 in their cancer/diabetes values. This leads to an overestimation of people who receive a 1 in these two variables and finally an overestimation of disability prevalence if the scope of disability measurement were expanded to include cancer, diabetes, and MDE.