Methodology: definitions, sources and limits

This page explains how the measured figures on the main page were made, what they can and cannot show, and where each choice is recorded. The project's full check log is docs/verification.md; references such as "D46" below are rows in it.

What is new here, and what is not

The Nigeria DHS 2023-24 final report already publishes disability for people aged 15 and over in every state, separately for women and men, without confidence intervals (Tables 14.2.1 and 14.2.2; D46). This site builds on those figures rather than replacing them.

What this site adds:

The definition: Washington Group Short Set

The DHS 2024 household questionnaire asks the Washington Group Short Set (WG-SS) about every household member aged 5 and over: how much difficulty they have seeing (even with glasses), hearing (even with a hearing aid), walking or climbing steps, remembering or concentrating, with self-care, and communicating. Each answer is "no difficulty", "some difficulty", "a lot of difficulty" or "cannot do at all".

DISABILITY3 (the disability figure on this site)
"A lot of difficulty" or "cannot do at all" in at least one of the six activities. This is the Washington Group's recommended cut-off (WG Document 7A; D23). Someone with no answer at that level and a "don't know" in any activity is left out of the denominator, as Document 7A says; the DHS report counts them as having no disability. "Don't know" is 0.0% in every row of the report's Table 14.1, so the difference is negligible (D48).
"Any difficulty"
At least "some difficulty" in at least one activity. This is a much wider group: in the NDHS 2024 report, 15.5% of people aged 15 and over, against 2.8% at the cut-off (Table 14.1, computed as 100% minus "no difficulty" minus "don't know").

Figures of about 15% and about 3% therefore describe different things. JONAPWD's 15% applies the global disability rate to population projections; nobody was counted. WHO's fact sheet says "An estimated 1.3 billion people experience significant disability", 16% of the world's population, but it does not define "significant disability" or cite a source report for the figure (D59), so it cannot be matched to the Washington Group cut-off.

Assumed rate vs survey: try other assumed rates against each state's survey interval.

Who is counted

The figures cover the de facto household population: people who slept in the household the night before the interview, as in the DHS report (D48). People living in institutions, and people without a household, are not in the survey. The headline is adults aged 18 and over; children 5–17 are reported separately. 49 people whose age is recorded as "don't know" belong to no age group, so they are outside the 18+ and 5–17 figures; the report counts them as 15+. Our pipeline reproduces every transcribed cell of the report's Tables 14.1 and 14.2 to the published decimal when it uses the report's own conventions (D48).

Privacy: only checked aggregates are published

The survey microdata are licensed and never leave the author's computer. Only weighted aggregates leave it, through one disclosure check: any cell based on fewer than 30 respondents is suppressed. Because published cells add up (states to zones, women and men to both sexes, wealth groups to their state), a lone suppressed cell could be worked out by subtraction, so wherever a set of cells that add up would hold exactly one suppressed cell, a second one is suppressed too (complementary suppression; D49). The number of people with a difficulty in a cell is never published.

Precision flags

Every estimate carries a flag based on the US National Center for Health Statistics data presentation standards for proportions (Parker et al. 2017; D42), with 95% Korn–Graubard confidence intervals:

Reliable
Meets the standards: enough respondents and effective sample size, and an interval that is not too wide for the estimate. The project adds a stricter rule: at least 10 people with the difficulty in the sample.
Review
Shown, but read with care: it rests on fewer than 10 people with the difficulty (or none), or on few degrees of freedom. Many single-activity state figures are in this group.
Unreliable
Too few respondents, or an interval too wide to be useful.

A flag is a screening label, not a guarantee: it says nothing about measurement error, non-response or a wrong design assumption. All 37 state headline estimates (adults 18+) are flagged reliable.

Small-area estimates were tried and change little

A Fay-Herriot model borrows strength from state characteristics to make small-sample state estimates more precise. Five candidate characteristics, and the rule for choosing among them, were written down and frozen before any model was fitted (share aged 60 and over, urban share, poverty, women's literacy, population density; the pre-declared plan and its result).

The result (D56): no model passed its pre-declared checks, and the characteristics explained too little to be chosen. More importantly, the states differ far more than the survey's sampling error, so most estimates would move by less than about 4%, and none by more than 11%; the precision gain is about 4% at the median. The survey's direct estimates are already reliable in all 37 states, so they are what this site shows. The full record is in data/public/sae on GitHub.

Spatial methods and their limits

All of these use the 37 direct state estimates, never individuals (D58).

No spatial analysis is made of JONAPWD's figures: they are 15% of each state's projected population (correlation 0.998 with population), so any map or spatial test would only show where people live (D17).

Disability and poverty: a pre-registered analysis

The disability and poverty page reports a separate analysis of the NBS Multidimensional Poverty Index survey 2021, whose plan was registered on the Open Science Framework before any poverty figure was split by disability. It uses MPI's own disability questions, not the Washington Group measure above, and reports an adjusted association, not a causal effect. The page explains its checks, how strong a hidden factor would need to be, and how to read survey figures, and links a two-page policy brief.

Data sources and citations

Disability (the headline)
The DHS Program, Nigeria Demographic and Health Survey 2024 (National Population Commission and ICF), household member recode (NGPR8BFL). Used only for the registered project "Curb Cut: disability prevalence and poverty in Nigeria". Report: NDHS 2023-24 final report (FR395), chapter 14.
WHO comparison figure
World Health Organization, fact sheet "Disability", 7 March 2023.
Population
National Bureau of Statistics (NBS), Demographic Bulletin 2023, Annex 3 (source: National Population Commission); cross-checked with WorldPop v3.0 (CC BY 4.0).
Small-area candidate characteristics
WorldPop 2020 age structures (Bondarenko et al. 2020, doi:10.5258/SOTON/WP00699, CC BY 4.0); JRC GHSL GHS-SMOD and GHS-POP R2023A (Schiavina et al. 2023); NBS, Nigeria Multidimensional Poverty Index 2022, Table D6; NBS and UNICEF, Nigeria MICS 2021 (MICS6), Table SR.6.1W. Details: covariates README.
Health facilities
CIESIN, Columbia University (2024), GRID3 NGA - Health Facilities v2.0, CC BY 4.0, doi:10.7916/kv1n-0743 (D57). Non-exhaustive and not validated by government.
Boundaries
Runfola, D. et al. (2020), geoBoundaries: A global database of political administrative boundaries, PLoS ONE 15(4): e0231866 (CC BY 4.0; source GRID3, 2022).
Assumed figures
Joint National Association of Persons with Disabilities (JONAPWD), 2025 Assumptive Disability Disaggregated Data.

The scripts that turn the microdata into the published aggregates are public, so anyone with their own DHS access can reproduce every figure: docs/reproduce.md.

Data API (optional)

The same public files are also served read-only by a small API at curb-cut-api.onrender.com (states, prevalence, the comparison and the poverty result, as JSON; at most 60 requests a minute per client). It runs on a free plan that sleeps when unused, so a request after a quiet spell can take about a minute to answer. This site never calls it: every page works without it. Details: docs/api.md.

Visit counts

Each page asks GoatCounter for a one-pixel image, so the site can say how many people use it. That is the only request a page makes to another site, and it runs no script. GoatCounter records the page and the time, and works out the browser, operating system, language and country from the request; it keeps only totals per hour or day. It stores no IP addresses and no full browser identifier, and sets no cookies or other storage in your browser. To tell visits apart, it keeps a random ID in memory for up to eight hours. The data are stored in Finland and Germany and are not shared with anyone else (GoatCounter's privacy policy, 8 June 2025). Blocking images from goatcounter.com stops the count and changes nothing else. The totals are public: Curb Cut's visit counts on GoatCounter.