Disability and poverty in Nigeria

Are people in households with a severely disabled member more often poor? This page answers with the National Bureau of Statistics' Multidimensional Poverty Index survey 2021, in an analysis whose plan was registered publicly before any poverty figure was split by disability. Poverty here is NBS's multidimensional poverty: being deprived in at least 26% of a weighted list of health, education, living standards, and work and shocks.

The main result

People in households with a member aged 5 or older who has a severe difficulty are 11.7 percentage points more likely to be multidimensionally poor than people in comparable households (95% confidence interval 6.1 to 17.2 points).

In comparable households without such a member, 62.9% of people are multidimensionally poor. Standardised to the same mix of households, the figure with such a member would be 74.6%. The wording follows a rule fixed before the analysis: the whole 95% interval lies above zero, and wholly above the smallest difference of interest fixed in advance, +2 points.

All households (the pre-registered main estimate)

+11.7 percentage points

95% confidence interval 6.1 to 17.2 points. 51,422 households, 680 of them with such a member.

Standardised to the whole population of Nigeria in households with a member aged 5 or older.

Secondary: households where the comparison is better supported

+8.1 percentage points

95% confidence interval 2.1 to 14.1 points. 18,754 households, 508 of them with such a member.

Comparable households without such a member: 66.8% poor.

Few households have a member with a severe difficulty: 680 of 51,422 in the analysis. So for most households, the model's estimated chance of having such a member was below 1 in 100, the cut-off fixed in the pre-registration (0.01). That was the case for 61.7% of people. For them there are very few comparable households with such a member, so the main estimate leans on a small number of households. The second estimate uses only the households whose estimated chance was between 0.01 and 0.99 (38.3% of people), where the comparison is better supported. It describes those households, not all of Nigeria.

Of 52,022 households in the sample, 545 (1.1% of people) were left out for missing poverty data and 59 (0.2% of people) for missing disability answers.

How the estimated chance of having such a member was spread
The model's estimated chance that a household has a member with a severe difficulty (propensity, averaged over the repetitions, before the 0.01 cut-off), by group
QuantileWith such a memberWithout
1% of households below0.0020.001
5% of households below0.0030.002
25% of households below0.0100.004
50% of households below0.0220.007
75% of households below0.0480.014
95% of households below0.1120.044
99% of households below0.1800.095

Download this table: CSV, 1 KB · JSON, 2 KB

An adjusted association, not a causal effect

This does not show that disability makes households poorer. Two reasons:

  1. Poverty can come first. Earlier poverty can lead to disability, for example through poor nutrition, unsafe work or little health care, and earlier poverty also goes with poverty today. A survey at one point in time cannot tell which came first.
  2. Some shared factors were not measured. Religion, ethnicity, conflict, local health services and the household's earlier poverty are not in the survey, and a factor like these could account for part or all of the difference. How strong such a factor would need to be.

How severe difficulty was measured

Who counts as having a severe difficulty comes from MPI's own disability questions (not the Washington Group's DISABILITY3 measure). For seeing, hearing and walking, the survey asks how hard it is only of people who said yes to a difficulty and use glasses, a hearing aid or walking equipment. People who said yes but have no aid were not asked, and here they count as not severe. Aids cost money, so this probably puts some severely disabled people from poorer households in the comparison group; the pre-registration expected this to pull the estimate towards zero. Counting all of them as severe instead (secondary question S5, 4,086 households with such a member) gives 5.4 points (95% confidence interval 1.5 to 9.7). The yes/no screener does not separate some from a lot of difficulty, so this group probably includes milder cases; S5 came out lower than the main estimate, so it is not an upper bound.

The same comparison, five ways

The main estimate, the better-supported comparison, a familiar survey-weighted linear model and two secondary definitions (S2: NBS's broad definition; S3: the household head has the difficulty). They answer slightly different questions, so they are not ranked against each other: no comparison between them was planned.

How much more likely people in households with a severely disabled member are to be multidimensionally poor, five ways, with 95% confidence intervals

The same figures as a table:

How much more likely people in households with a severely disabled member are to be multidimensionally poor, five ways, with 95% confidence intervals. Source: NBS MPI 2021, Curb Cut estimates (pre-registered, OSF kes3p).
ComparisonDifference (percentage points)95% confidence interval, lower95% confidence interval, upper
All households (primary)11.76.117.2
Better-supported comparison (overlap)8.12.114.1
Survey-weighted linear model8.03.412.6
NBS's broad definition (S2)6.83.510.2
Household head has the difficulty (S3)10.24.116.4

Download this table: CSV, 1 KB · JSON, 2 KB

How we checked it

Every check below was fixed in the pre-registration before any poverty figure was split by disability. A failure would be reported here, not hidden.

First: reproducing NBS's own figures

NBS's 2022 report (Table 7) says that, on its broad definition, people in households with a person with a disability are more often poor. It is a raw comparison, with nothing taken into account. We rebuilt the poverty index from the survey and checked that it gives the same figures, within 2.0 points each. The check passed.

Reproducing NBS Table 7 (broad definition, unadjusted)
FigureOursNBS Table 7Within tolerance
People who are poor, households with a person with a disability (%)72.071.4yes
People who are poor, households without (%)61.861.8yes
Gap (percentage points)10.19.6yes

Download this table: CSV, 1 KB · JSON, 2 KB

A familiar benchmark

A survey-weighted linear model with the same factors, the kind of model most survey analysts use, gives 8.0 points (95% confidence interval 3.4 to 12.6), against 11.7 for the main estimate. It is a check, not a second answer.

Checks that should find nothing

Leave-one-zone-out figures
The main estimate with one zone left out at a time (one repetition), percentage points. Low and High: the 95% confidence interval
Left outDifferenceLowHigh
Without the North Central zone12.05.818.2
Without the North East zone10.22.517.9
Without the North West zone10.62.119.1
Without the South East zone11.76.117.3
Without the South South zone12.76.319.1
Without the South West zone9.34.314.4

Download this table: CSV, 1 KB · JSON, 2 KB

Other reasonable choices

Eight other choices fixed in advance (fewer factors, other methods, two other ways of counting deprivation) give estimates between 10.1 and 13.0 points.

The main estimate under other reasonable choices, percentage points. Low and High: the 95% confidence interval
ChoiceDifferenceLowHigh
Without the head's education13.07.218.2
Place and the head's sex and age only12.67.518.0
Place as state only (no senatorial district)11.55.717.4
Logistic regression for both models10.13.816.4
Random forest for both models12.25.818.5
Gradient boosting for both models10.34.715.9
Cooking fuel without charcoal13.07.317.9
Don't-know shock answers counted as not deprived11.86.317.4

Download this table: CSV, 1 KB · JSON, 2 KB

How the methods were chosen

The analysis uses double machine learning: one model predicts poverty, another predicts which households have a member with a severe difficulty, and the two are combined so that errors in either one matter less. For each model, three methods were scored on households they had not seen, before any difference was computed. The best one was chosen, unless a simpler one was within the margin of error. That gave gradient boosting for poverty and logistic regression for the exposure.

How well each method predicted, mean log loss over 5 cross-fitting folds (lower is better)
MethodPoverty modelExposure model
Logistic regression0.52450.0644
Random forest0.52800.0637
Gradient boosting0.51980.0761

Download this table: CSV, 1 KB · JSON, 2 KB

How strong would a hidden factor need to be?

Nothing that was not measured can be ruled out. What can be said is how strong such a factor would need to be to change the answer. No decision depended on these numbers; they are reported as they came out.

The worst case

Suppose a hidden factor pushed in exactly the direction that matters. If it accounted for 1.8% of what is left unexplained in both poverty and in which households have such a member, it could bring the main estimate to zero. At 1.0%, the 95% interval would include zero. These robustness values are small: the comparison is fragile to a factor of this kind. A factor of 3% strength would leave the estimate anywhere between -7.5 and 31.1 points.

Compared with factors we did measure

Measured factors give a sense of scale. A hidden factor as strong as each of these, and lined up with poverty and exposure the way each of them is, would move the estimate only a little:

Where a hidden factor as strong as an observed one would move the main estimate, percentage points (DoubleML bounds)
Hidden factorEstimate between95% interval
As strong as the household head's education10.7 to 12.75.9 to 17.3
As strong as place (state and senatorial district together)11.6 to 11.86.9 to 16.3
As strong as living in an urban or a rural area11.5 to 11.96.7 to 16.5

Download this table: CSV, 1 KB · JSON, 2 KB

The worst case and these comparisons answer different questions: the worst case assumes the hidden factor lines up perfectly, while the measured factors line up only weakly.

A second method

The same question asked of a weighted linear model (sensemakr; this model ignores the survey design, so its standard error is too small) gives 8.0 points before adjustment, a robustness value of 2.2%, and these adjusted figures:

Weighted linear model (ignoring the survey design), adjusted for hidden factors benchmarked on the head's education (sensemakr), percentage points. Low and High: the 95% confidence interval
Hidden factorDifferenceLowHigh
A hidden factor as strong as the head's education6.03.09.1
A hidden factor 2 times as strong as the head's education4.11.17.0
A hidden factor 3 times as strong as the head's education2.1-0.84.9

Download this table: CSV, 1 KB · JSON, 2 KB

The E-value

People in households with such a member are 1.19 times as likely to be poor (95% confidence interval 1.10 to 1.27). The E-value is 1.66: a hidden factor would need to make both poverty and having such a member about 1.66 times as likely to explain the whole difference away, and 1.42 times to bring the interval to 1. Two cautions:

Other questions we registered

Each is labelled secondary. S1 (the deprivation score instead of the poor or not-poor cut), S2 (NBS's broad definition) and S3 (the head has the difficulty) were registered as a family: All three of S1, S2 and S3 are supported after correcting for testing three questions (Holm's method). S4 to S6 are descriptive.

Secondary questions. S1 and S4 in points out of 100, the others in percentage points. Low and High: the 95% confidence interval; Holm: the p-value corrected across S1-S3
QuestionDifferenceLowHighHolmWording
S1: deprivation score (0-100)4.52.96.0<0.0001supported
S2: NBS's broad definition (some difficulty or worse, or a skin condition)6.83.510.20.0001supported
S3: the household head has a severe difficulty10.24.116.40.0011supported
S4: Health (points out of 100)2.00.93.2descriptive
S4: Education (points out of 100)0.4-0.21.0descriptive
S4: Living standards (points out of 100)0.3-0.51.0descriptive
S4: Work and shocks (points out of 100)1.30.52.2descriptive
S5: people not asked how severe counted as severe5.41.59.7descriptive
S6: self-care, remembering or communicating only12.46.318.9descriptive

Download this table: CSV, 1 KB · JSON, 3 KB

Where the difference sits (S4)

The poverty index has four dimensions. This shows how much of the difference in the deprivation score (S1) sits in each. It is descriptive: it shows where, not why. The four parts are separate estimates and do not add exactly to S1.

Where the difference sits: each MPI dimension's part of the deprivation score, with 95% confidence intervals

The same figures as a table:

Where the difference sits: each MPI dimension's part of the deprivation score, with 95% confidence intervals. Source: NBS MPI 2021, Curb Cut estimates (pre-registered, OSF kes3p).
ComparisonDifference (points out of 100)95% confidence interval, lower95% confidence interval, upper
Health2.00.93.2
Education0.4-0.21.0
Living standards0.3-0.51.0
Work and shocks1.30.52.2

Download this table: CSV, 1 KB · JSON, 2 KB

Within urban and rural areas and within each zone (exploratory)

Exploratory and noisy: each group has few households with such a member, so these intervals are wide. They are descriptions, not findings, and are not compared with each other.

The main comparison within groups (exploratory X1), percentage points. Low and High: the 95% confidence interval
GroupDifferenceLowHighHouseholdsWith such a member
Rural areas6.70.812.036,886504
Urban areas29.418.240.514,536176
North Central zone0.2-14.615.010,00098
North East zone6.4-5.017.88,58187
North West zone16.910.323.19,844136
South East zone11.8-3.026.76,820126
South South zone7.9-2.918.78,271173
South West zone17.2-8.743.27,90660

Download this table: CSV, 1 KB · JSON, 3 KB

Exploratory: the same people five years apart

Exploratory, not confirmatory

This part compares only two survey rounds, five years apart. It was registered as exploratory: two outcomes, intervals only, no claim that a difference is supported.

The General Household Survey-Panel visited the same households in 2018/19 and 2023/24, and both rounds asked the Washington Group's six questions. Of 15,229 people aged 15 and over in 2018/19, 13,266 were found again (87.1%). After keeping those with no severe difficulty in 2018/19 and complete answers, 10,933 could be used. 364 of them had a severe difficulty by 2023/24.

A rule fixed in advance said each outcome is analysed only if the survey could detect a difference of at most half a standard deviation of the change. For consumption, the smallest detectable difference was 0.49 standard deviations, so it was analysed. For work, the smallest detectable difference was 0.28 standard deviations, so it was analysed.

Exploratory: change between 2018/19 and 2023/24 for people whose severe difficulty began between the rounds, against comparable people. Low and High: the 95% confidence interval
OutcomeDifferenceLowHighPeopleWith onset
Log household consumption per person (log points)0.035-0.0450.11410,933364
Worked in the last 7 days (percentage points)-26.4-33.2-19.710,628360

Download this table: CSV, 1 KB · JSON, 2 KB

R (fixest) gave the same estimates and intervals to at least six decimal places. Both models take into account age band, sex, zone and urban or rural, and use the survey's weight for people followed over time.

What cannot be tested. The comparison assumes that, without the onset, people with onset would have changed like the comparison group. That assumption cannot be checked with two rounds. Onset is only seen at the second round, so the order of change (the difficulty first, or the change in work first) is unknown.

Questions that differ between the rounds. In 2023/24 the communication question adds "using your usual language". The work questions cover the same three kinds of work in both rounds (for someone outside the household, on the household's farm, and in the person's own business), but the wording differs: in 2018/19 the business question says "on their own account or in a business", in 2023/24 "a non-farm household business". The 2023/24 questions on helping in another member's business or farm are not used, because 2018/19 has no separate helping question; the 2018/19 farm question probably includes helpers, which may make work in 2023/24 look slightly lower.

Who was found again (attrition)

People not found again were younger on average. No outcome is involved in this comparison.

People found again in 2023/24 (stayers) and not found (leavers), on 2018/19 characteristics
CharacteristicStayersLeaversStandardised difference
Age (years)37.3731.860.32
Women (share)0.510.53-0.03
Urban (share)0.290.280.03
Severe difficulty in 2018/19 (share)0.020.03-0.08

Download this table: CSV, 1 KB · JSON, 2 KB

How to read survey figures

Four short examples, using this page's own numbers.

What a 95% confidence interval means

The main estimate is 11.7 points, with a 95% confidence interval of 6.1 to 17.2. A survey talks to a sample of households, not to everyone, so a different sample would give a slightly different figure. The interval shows that uncertainty: if the survey were repeated many times and an interval made the same way each time, about 95 in 100 of them would contain the true value. This interval does not include zero, so the survey can show a difference. It does not cover other kinds of uncertainty, such as factors that were not measured.

Survey weights

Each household in the survey stands for many people in Nigeria, and some kinds of household were sampled more than others. The weights used here are population weights (the household's weight times its number of members), so that the 51,422 households in the analysis represent the whole population. The price is precision: with these unequal weights, the households carry about as much information as 26,009 households with equal weights would.

An adjusted association, not an effect

On rainy days, more people carry umbrellas and more pavements are wet. Umbrellas do not wet pavements: the rain does both. Comparing only days with the same weather takes the rain out of the comparison. That is what adjusting does. Here we compared households in the same senatorial district, urban or rural, with the same mix of ages, and with a head of the same sex and similar age and education. What remains is an adjusted association. It could still come from a factor nobody measured, or from poverty coming first, so it is not proof that one leads to the other.

Percentage points and percent

Among comparable households without such a member, 62.9% of people are poor; standardised to the same mix of households, the figure with such a member would be 74.6%. The difference is 11.7 percentage points: points are a simple subtraction of two percentages. Said another way, the share is 19% higher than the comparison households' share: percent compares the difference with the starting figure. Both are correct; points are used on this page because they do not depend on the starting figure.

Registration, data and citations

Registered before the analysis

Registration
https://osf.io/kes3p/ (Open Science Framework, Secondary Data Preregistration template)
Registered
2026-11-01 05:49:21 UTC
Plan and code frozen
tag p5-freeze, commit 22b8acee10b304f25201627d5cf4da47c2544a19
Deviations from the plan
none so far (deviations log)

The pre-registration as frozen (on GitHub, at the freeze tag). The published estimates and checks are on GitHub (p5_mpi_estimates.csv, p5_mpi_summary.csv, p5_ghs_estimates.csv, p5_ghs_summary.csv).

The policy brief

Download the two-page policy brief (PDF), or read the same brief as text on GitHub.

Data sources and citations