Disability Data Events

Why Do Disability Prevalence Estimates Vary? A Summary of the WG-CIP-DDI Webinar

Summary

Two surveys, same country, same questions — yet completely different disability prevalence estimates. A recent webinar unpacked why this happens, and why understanding the variation matters more than looking for a single "correct" number.

Organized by: Washington Group on Disability Statistics (WG), Center for Inclusive Policy (CIP) and the Disability Data Initiative (DDI)
Hosted by: Center for Inclusive Policy (CIP)
Moderated by: Jennifer Madans, WG and CIP

About the Webinar

On May 5, 2026, the Washington Group on Disability Statistics (WG), the Center for Inclusive Policy (CIP) and the Disability Data Initiative (DDI) co-organized a webinar exploring a common and persistent puzzle for disability data: why do disability prevalence estimates vary across data sources? And what do these differences mean for interpreting and using disability data?

The webinar was moderated by Jennifer Madans, Chair of the Washington Group on Disability Statistics and Senior Advisor at the Center for Inclusive Policy. Presentations were delivered by Daniel Mont, Co-founder and CEO of the Center for Inclusive Policy, and Sophie Mitra, Professor at Fordham University and Founder of the Disability Data Initiative (DDI). Daniel Mont and Sophie Mitra were joined by Elizabeth Lockwood, Representative at the United Nations for CBM Global Disability Inclusion, for a panel discussion.

Why Do Disability Prevalence Estimates Differ?

Daniel Mont opened by addressing a question that often puzzles policymakers and researchers: why do data sources from the same country often produce different estimates of disability prevalence? There is no single reason, and variation in estimates is often due to a combination of several factors. Disability is a complex phenomenon, with different analytic purposes requiring different conceptualizations of disability. This situation has resulted in different definitions of disability. Some definitions ask directly whether a person has a disability; others rely on medical diagnoses or program eligibility; and others, like the Washington Group Short Set on Functioning (WG-SS), focus on functional limitations, measuring how much difficulty a person has with specific daily activities. Because these approaches identify different people, they will also produce different prevalence estimates. In these cases, differences in disability prevalence estimates are not a flaw but instead a reflection of genuine differences in the analytic purposes of different disability measures.

Even within the same conceptual approach, differences in data collection methods, such as sampling frame, sampling methods, response rates and their effect on sample composition, placement of the questions in the questionnaire, questionnaire scope (multipurpose, single topic and type of topic), mode of question administration (in-person, phone, web), respondent selection (household respondent, self-response), question wording, the number of domains covered, the response scale and the cutoff used, and implementation factors like translation quality or interviewer training, can all shift who gets counted.

While these factors contribute to differences in disability prevalence estimates within countries, they can also lead to between-country differences in disability estimates. Structural differences across countries add another layer that can create additional differences in prevalence rates across countries: for example, countries with older populations will generally have higher prevalence rates, since disability rises with age. Different survival rates associated with various causes of disability also play a role.

Mont’s core message was that variation in estimates of disability prevalence is expected and often valid. Understanding why estimates differ is far more useful than assuming one figure is right and the rest are wrong.

Comparing Disability Estimates Across Data Sources

In the next section of the webinar, Sophie Mitra used data from the Disability Data Initiative (DDI), which has reviewed disability questions across more than 3,600 censuses and household surveys in 206 countries, to illustrate variation in estimates of disability prevalence. A review of the data shows that even when datasets include content covering the same six functional domains as the WG-SS, differences in administration, such as changes in question wording or in response options, result in variations in prevalence. In the 2018 Malawi census, for instance, the cognitive domain asked about difficulty “learning new things or solving problems or remembering” rather than using the original WG-SS phrasing, and the communication domain was simplified to ask only about difficulty speaking. Small as they may seem, these variations shift who gets identified as having a disability and therefore will be reflected in prevalence estimates.

However, prevalence estimates can differ even when the exact same question wording is used. The WG-SS was included in Senegal’s 2013 census and the 2018 Demographic and Health Survey (DHS), yet the prevalence estimate for adults (using the recommended cutoff of “a lot of difficulty” or “cannot do at all” in one or more of the functional domains) was 2.1% in the census and 4.7% in the DHS. The direction of the difference held consistently across men, women, rural and urban populations, and all age groups. The gap was largest among adults 65 and older: according to the census, an estimated 12.1% of adults 65 and older had disabilities, compared to an estimated 24.8% in the DHS. In general, surveys tend to produce higher disability prevalence estimates than censuses, while health surveys tend to have higher disability rates than surveys that focus on other topics, such as living standards or employment.

Estimates can also vary across datasets due to differences in sample composition. The surveys analyzed were based on samples with higher percentages of women and older persons than the populations reached by the censuses, for example. Because disability rates are higher for women and older adults, a sample composition with an overrepresentation of these groups will affect the overall disability prevalence for the country. That said, Mitra showed that the disability gaps in key outcomes such as educational attainment remain consistent across data sources, even when disability prevalence estimates differ. Importantly, variation in estimates across datasets is not unique to disability. Mitra showed that, for the overall population, estimates for indicators like education, assets, and housing also differ across the surveys and censuses using similar questions, yet these estimates are routinely produced and widely used.

Panel Discussion: Navigating Differences in Disability Data

The final section of the webinar included a panel discussion. Jennifer Madans introduced this section by providing a brief review of the reasons that disability prevalences vary. When presented with different prevalence estimates many come to the conclusion that all but one of the estimates are wrong. However, the variation reflects valid conceptual differences in how disability is defined and in approaches to identifying the population with disabilities. These differences reflect the complexity of disability and the variety of analytic or policy objectives for which data are collected. Even when using the same conceptual framework, data collection methods, such as different target populations, using self-response vs. a proxy/household respondent, different modes of data collection (interview, web, paper), and variation in field procedures (interviewer training, implementation, monitoring), can influence results. Differences in the demographic characteristics of the survey sample can also affect prevalence estimates. Since disability affects all aspects of life and the need for data for policy, program and research purposes varies, it will be necessary to conduct multiple data collections to obtain the required data. The different designs employed by these collections will produce different prevalence estimates.

Given these factors it should not be surprising that there is an apparent lack of consistency in prevalence estimates and an inherently complex and messy disability data landscape. While variation in disability prevalence estimates has been a major concern, variation is not unique to disability. Rather, variation is common across many population indicators, including as was shown for education and economic indicators such as estimates of assets. There is less concern about differences in these other topic areas and much to be learned about how variations in results are handled.

The task for the panel was to address how to effectively interpret and use data from these varied data sources.

As a data user, I am interested in the prevalence of disability. How do I determine which data source to use?

For national-level prevalence estimates, Daniel Mont indicated that the census is the starting point. It aims to reach the entire population and provides the geographic granularity, including subnational breakdowns, that surveys generally cannot provide. However, there are limitations: the limited space available in census questionnaires results in constraints on content. For example, persons with psychosocial disabilities and children with developmental disabilities tend to be undercounted. The census provides the broadest picture, but not the complete one. Censuses are also major undertakings for a country, which affects how enumerators are trained and supervised. Reductions in training, supervision and monitoring of field work can affect the quality of the disability data produced. Surveys allow for more control over field operations and generally produce higher quality data. Surveys are useful as a complement, offering the opportunity to administer more detailed question sets and better implementation quality for specific topics or populations.

Given that estimates are sensitive to the questions asked and the survey in which they are asked, do you think it would be useful for users if the source of the prevalence estimate was always mentioned?

In his response, Mont highlighted the importance of the source and the method involved in producing the estimate. He emphasized the importance of transparency: whenever a prevalence figure is published, its source, the questions used, the cutoff applied, and any implementation details that might affect interpretation should be reported. A number without that context is difficult to interpret or use responsibly.

In terms of disaggregating outcome indicators by disability, how do we choose which data source to use?

Sophie Mitra emphasized that the answer depends on what you’re trying to measure, the scope of your work and topic of interest. For cross-country analyses, instruments with internationally validated question sets, particularly the WG-SS, are preferable. There is more flexibility around disability questions for single-country work. In addition, the choice of a data source should be driven by the indicator of interest. For reproductive health outcomes, the DHS (Demographic and Health Survey) is the natural choice. For demographic characteristics like household composition, the census, with its much larger sample size, is needed. For employment indicators, the ideal source would be a labor force survey, but where those do not include disability questions, the census is often the best available alternative.

How can we communicate to data users, as well as the public and other stakeholders, about variation in disability prevalence estimates? How should NSOs (national statistical offices) approach this versus NGOs, OPDs?

Elizabeth Lockwood addressed the question of communication from the perspective of civil society and organizations of persons with disabilities. She emphasized that good data communication goes beyond reporting numbers: it means making gaps visible, showing not just what the data says, but which groups or disability types may be underrepresented or missing altogether. Findings should be shared in accessible formats and languages, including national sign languages, tested with persons with disabilities before publication, and written in simple language free of jargon, ensuring compatibility with assistive technologies. She also flagged the importance of safety and privacy, particularly in smaller communities where individuals may be identifiable, and called for using data storytelling to make findings meaningful and actionable for the communities the data is meant to serve.

Finally, Lockwood shared an example of effective advocacy. An OPD in Las Piñas, Philippines collected local data on unemployment among persons with disabilities, presented it to local authorities, and secured 400 employment placements as a direct result. The data was simple and clear; the impact was immediate and concrete.

Q&A Highlights

The questions below were submitted by attendees during the live Q&A session. Because these reflect common challenges faced by researchers, policymakers, and data users, this section is also designed to function as a reference FAQ that can be consulted beyond the webinar context.

How can we better ensure older census data still fully captures persons with disabilities?

Panelists noted that the situation is improving: the 2020 census round has seen substantially more countries adopt the WG-SS than previous rounds. The UNICEF-supported Multiple Indicator Cluster Survey (MICS), which uses the Child Functioning Module (CFM), now provides relatively consistent childhood disability estimates across many countries. Panelists also pointed out that overall prevalence, often obtained from the census, changes slowly, so the usual 10- year gap between censuses, while not ideal, may not be a major problem. Changes in prevalence can also be monitored by surveys, which may be collected more frequently. Disaggregated outcome indicators, such as employment or education gaps, can shift more quickly in response to policy changes, making newer data collected from surveys particularly valuable for monitoring program impact.

Do data systems measure the need for care and support or the impact on caregivers who cannot work outside the home?

Some data sources do address these topics. Household income and expenditure surveys collect information on the work behavior of all household members. Time use surveys, available in some countries, capture time spent on caregiving.

CIP and its collaborators have recently produced a review for the International Labour Organization examining what questions can be used to measure care-related needs and their economic consequences that can be added to labor force surveys. This resource should be available in the near future

What is the best way to interpret data when one person reports multiple disabilities?

Having multiple disabilities does not affect how the population with disabilities is identified in most cases. For example, using the WG-SS, a person is identified as having disabilities if they meet the threshold cutoff in at least one domain. However, the WG-SS can be used to identify many different populations with disabilities depending on how responses from the six questions are put together. Groups with difficulties in more than one domain can be identified. These more detailed characterizations of difficulties in functioning are particularly important for understanding goods and services people need. The WG-SS also allows for analyzing variation in the type and severity of functional difficulties, rather than treating disability as a single binary category, another strength of the question set. Instruments like the Model Disability Survey and Kenya’s recent national survey on care and support needs are beginning to capture this complexity in ways that can inform program design.

What does this mean for interpreting the commonly cited "16% disability prevalence"?

The 16% figure, drawn from the World Health Organization’s World Report on Disability, is based on a specific conceptual approach, specific set of questions, and a specific cutoff point. It is a global average, and individual countries will legitimately fall above or below it depending on their demographics, the causes of disability present, and the methodology of their data collections. Mont was explicit: a country whose own data collection is producing an estimate of 9% is not necessarily doing something wrong. The 16% is a useful reference point, but like any number, it needs to be understood in context. As he put it, a country that uses a very broad definition of disability may find that 80% of its population has some form of functional difficulty, which is not a meaningless figure if you’re designing accessible public infrastructure, though obviously this is not the right cutoff for a targeted disability support program.

What is the difference between disability prevalence calculated directly from the raw data and the adjusted prevalence of disability?

The unadjusted (crude) disability prevalence estimate tells us what percentage of persons in the population have a disability. The actual size of the population is useful when determining the need for assistive devices or other accommodations. However, when making international comparisons or comparisons across groups within a country, it is very useful to adjust by age. As prior research has shown, disability has a strong relationship with age – usually, disability rates are higher at older ages. So, if we do not age-adjust, countries that have a large old-age population may appear to have higher disability rates, compared to countries that have a large young-age population. Similarly, groups that are older, on average, than the general population may appear to have higher disability prevalence. By age-adjusting disability estimates, we can better compare disability rates across countries that have different age structures or across groups with different age distributions. Comparing age-adjusted disability estimates allows us to examine differences in the risk of disability – is an individual in group A at a higher risk of disability than those in group B?

What do you recommend when one person has multiple types of disabilities registered?

For many definitions of disability, persons with difficulty in one domain are treated the same as those with difficulties in more than one domain. As long as a person reports difficulties at the stated level in one domain, they are included in the population with disabilities. However, it could be of interest to look at people with difficulties in multiple domains because it is likely that they face even higher barriers to inclusion. If the sample is big enough, it is possible to look at the distribution by type of disability and by whether people have difficulties in multiple domains of functioning as well as how much difficulty they have in any of the domains. For example, using the WG-SS, an alternative way to define the population with disabilities is to include those with some difficulty in at least two domains, along with those with at least “a lot of difficulty” in at least one domain. It is also possible to disaggregate outcomes across three categories of disability – no difficulty or some difficulty in only one domain, at least some difficulty in two domains but no domain with a higher level of difficulty, and at least “a lot of difficulty” in at least one functional domain.

Which level of difficulty is recommended as the cutoff when measuring disability prevalence? For example, in Tanzania we use “some difficulty”, “a lot of difficulty” and “a lot / unable”.

The Washington Group has recommended a cutoff for use in international comparisons and official reporting of disability estimates of all persons with “a lot of difficulty” or “cannot do at all” in at least one functioning domain. This cutoff was chosen because it identifies the population most at risk of exclusion given the severity of functional difficulties. In addition, in cognitive and field testing of the WG-SS questions, those responding “a lot” or “cannot do at all” were very similar in their functioning characteristics across countries. Choosing this cutoff ensures that the group identified as “with disabilities” is homogenous with respect to their functioning and other characteristics.

However, other cutoffs can be used with the Washington Group questions depending on the purpose for which the data are being used. Empirically, it is possible to see if people with “some difficulty” in multiple domains or even people with “some difficulty” in only one domain, have worse outcomes than the general population, which would be evidence of disability in that country context. As is always the case, it is important to clearly state the cutoff used and the data collection that is the source of any estimate.

In many societies, including parts of Ethiopia, the main problem is that disability may be hidden due to stigma. Families may not report children with disabilities, and some conditions are not recognized as disabilities. How could we approach this?

This is one reason why the Washington Group questions do not mention the word disability – to avoid that stigma. We found in testing that people were more likely to report difficulties in functioning than they were to identify as having a disability. Of course, this probably does not totally eliminate the stigma issue, but we believe it lessens it.

Related Resources

You might also like