The purpose of the current study is to examine whether individuals are differentially classified as having a disability under the Washington Group Short Set (WG-SS) depending on whether responses are provided by a proxy or self-reported. This study uses data from the National Health Interview Survey (NHIS) to evaluate whether proxy respondents are more or less likely to classify household members as having a disability under the WG-SS framework, even after accounting for sociodemographic characteristics; disability classification followed the recommendations for use provided by the Washington Group (Washington Group on Disability Statistics, 2020). This work aims to assess the response processes and person-centered accuracy of the WG-SS in the context of household surveys where proxy reporting is common.

We hypothesize that disability classification under the WG-SS is associated with reporting source, such that individuals reported on by a proxy will be classified differently—potentially more frequently—as having a disability compared to those who self-report. By distinguishing between the likelihood of disability identification, this study contributes to a more nuanced understanding of how proxy response influences disability classification and prevalence estimates in U.S. health surveillance.

Data from the 2010 through 2018 National Health Interview Survey (NHIS) were acquired from IPUMS (Blewett et al., 2019). The total pooled sample size was 284,809, including both children and adults. Because the WG-SS questions are only administered to adults (18+), 144,925 cases for individuals under age 18 were removed. Of the people who were administered the WG-SS, 5,554 did not answer any of the six questions and were also removed from the sample. After removing these people, the sample size was 134,330. Following IPUMS guidance (https://nhis.ipums.org/nhis/userNotes_variance.shtml), the sample weight variable (SAMPWEIGHT) was divided by the number of years pooled together (nine) to produce weights representing the average U.S. population across the 9-year period.

The dependent variable in the current study is whether a person is classified as having a disability based on their responses to the WG-SS items. Following the Washington Group's recommendations, if a person endorses “a lot of difficulty” or “cannot do at all,” they are considered as having a disability in the corresponding domain (Washington Group on Disability Statistics, 2020). People reporting at least one WG-SS disability are coded as 1; all others are coded as 0. Table 1 shows the six disability domains and the corresponding item on the survey.

The independent variable of interest in the current study is whether the NHIS interview was completed by a proxy or not. This variable was created using two NHIS variables: (1) PROXYSA, which indicates whether a person used a proxy due to a physical or mental condition; and (2) SAPROXYAVAIL, which indicates whether a knowledgeable proxy was available to complete the interview. Respondents who did not use a proxy were coded as 0, while those who needed a proxy, and had one available, were coded as 1. In our sample, all respondents who needed a proxy had one available.

The sociodemographic variables used in the analysis were chosen because previous literature suggests they are associated with disability status, proxy respondent, or both. The included variables are as follows: age (18–85; Lauer et al., 2019; Mactaggart et al., 2016; Todorov and Kirchner, 2000), sex (0 = male, 1 = female; Elkasabi, 2021; Mactaggart et al., 2016; Lauer et al., 2019), marital status (0 = married, 1 = not married; Saito et al., 2024), race/ethnicity (0 = White, 1 = non-White; Lauer et al., 2019), Hispanic ethnicity (0 = no, 1 = yes; Lauer et al., 2019), education (0 = some college or more, 1 = high school or less; Elkasabi, 2021; Hall et al., 2022a; Shandra, 2018), employment status (0 = employed, 1 = not employed; Amilon et al., 2021; Hendershot, 2004; Shandra, 2018), health (0 = poor to 4 = excellent; Amilon et al., 2021; Amilon and Christensen, 2025; Li et al., 2015), and insurance coverage (0 = has coverage, 1 = no coverage; Kaye, 2019).

A binary logistic regression was conducted to examine the relationship between proxy respondents (IV) and WG-SS disability status (DV) after controlling for sociodemographic characteristics. All analyses were conducted in Rstudio (Posit Team, 2024). The svyglm function in the survey package (Lumley, 2020) was used to conduct the logistic regression and to adjust for person weighting, stratification, and clustering due to the complex survey design of the NHIS. A likelihood-ratio test (LRT; Glover and Dixon, 2004), Wald test (Fox, 1997), and model accuracy (Baratloo et al., 2015), were used to assess model fit and performance. Listwise deletion was used for incomplete cases, resulting in a sample size of 133,025 people.

Using the WG-SS disability items, 10.9% of the sample reported having a disability. The majority of the sample were female, White/Caucasian, non-Hispanic, married, had some college education or more, were unemployed, and had health insurance coverage. The average age of the sample was 49 years (SD = 18.4), and the average self-reported health was 2.66 (SD = 1.08). A breakdown of disability domains and demographic characteristics by WG-SS disability indicator (described in Section 2.2) is presented in Tables 2, 3, respectively.

A binary logistic regression was conducted to examine whether individuals were more likely to be classified as having a WG-SS disability when a proxy respondent completed the NHIS survey, controlling for sociodemographic characteristics. A likelihood ratio test (LRT) showed the full model fit significantly better than the null model, χ²(df = 9) = 24635.61, p < 0.001. A Wald test indicated the overall effect of the variables are statistically significant, χ²(df = 23) = 4912.50, p < 0.001. Using a threshold of model fitted values >0.50, participants were classified as having a WG-SS disability, otherwise no WG-SS disability, and model accuracy was computed. Model accuracy was calculated as the number of correctly identified cases divided by total cases. The model's accuracy was 90.35% (120,915/133,025), meaning the model correctly classified 90.35% of the cases. Proxy respondent was statistically significant, exp(β) = 4.48, se = 0.08, p < 0.001, 95% CI [3.86, 5.21], indicating proxy respondents were 4.48 times more likely to endorse a disability than self-report respondents. See Table 4 for the full regression table.

This study has several limitations that should be considered when interpreting the findings. First, although we examined the association between proxy response and disability classification using the WG-SS, the analysis was limited to a binary disability indicator (i.e., presence vs. absence of disability) rather than assessing the severity of functional limitations. Consequently, we cannot determine whether proxies differentially report the severity of disability across functional domains, which may be relevant for more nuanced analyses of disability experience. Second, the reasons why a proxy was used in the NHIS (i.e., physical or mental condition prohibiting self-response) are not fully captured in the analysis. Although the NHIS designates proxies only under specific conditions, in practice, there may be inconsistencies in how interviewers determine the need for a proxy or in how household members decide who responds and these factors could play a role in disability classification. Additionally, individuals were removed prior to analyses because they did not complete any of the WG-SS items and this could affect the representativeness of the sample. Finally, the cross-sectional design of the study precludes causal inference. While we found an association between proxy response and disability classification, we cannot conclude that proxy reporting itself causes differences in disability prevalence estimates.

Future studies should build on these findings by using designs and methods that directly test the impact of proxy response on disability classification. Experimental designs that randomly assign self-respondents to conditions with or without proxy reporting could help disentangle the effects of proxy use from characteristics of respondents who typically require a proxy. Applying differential item functioning (DIF) techniques would provide a more nuanced understanding of which WG-SS items are most sensitive to proxy vs. self-report differences, and which respondent or contextual factors drive those differences. Additional research should also explore how characteristics of the proxy (e.g., relationship to the respondent, caregiving role, health literacy) shape reporting patterns, to inform survey protocols and improve the validity of disability measurement.

The differences in proxy reporting are not merely technical. They have real implications for equity, access, and policy accountability. If proxy reporting systematically increases (or decreases) the likelihood of disability classification, prevalence estimates—and the programs or protections tied to those numbers—may be distorted. This is especially problematic when proxies are more likely to report for populations already at risk of under- or overrepresentation in disability data, such as older adults, people with cognitive disabilities, and children and adolescents.

In surveys where proxy reporting is used, the source of the response may influence whether someone is counted as having a disability. This finding has important implications for survey design, cognitive testing, and the interpretation of disability data. We recommend that future validation studies of the WG-SS—and disability measurement tools more broadly—treat reporting source not as a procedural footnote, but as a meaningful source of variation with implication for data quality, equity, and the validity of disability statistics.