Independent Validation

1. Validation approach

Pan Africa Data's national-level output is calibrated against the World Bank Macro Poverty Outlook (WB MPO) by design — anchored at the $3.65/day (2017 PPP) line for lower-middle-income countries and at the $6.85/day line for upper-middle-income countries. A direct comparison at the calibration threshold is therefore a fidelity check, not empirical validation. Presenting it as validation would be dishonest, and this page does not do so.

Instead the validation programme uses a three-way triangle. Three vertices: (a) Pan Africa Data's modelled marginalised class share, (b) the World Bank PIP's independently-published headcount at the same $3.65 line, and (c) each country's own national statistics office headcount at the office's own national poverty line. The empirically informative edges are the two that are not the PAD-to-WB calibration edge.

Why these six countries in v1

The panel spans both calibration paths (South Africa is upper-middle-income; the other five are lower-middle-income) and represents the largest African economies by GDP or population with a recent, publicly-published national household survey. Tanzania was considered and deferred to v2 pending publication of the Integrated Household Budget Survey (IHBS) 2024/25 — the currently-available survey is eight years old.

2. Master validation matrix

All figures below are percentage of population below the $3.65/day (2017 PPP) line for the 2023 reference year, except the NSO column where the reference year and threshold construct vary by country and are reported alongside the value. The NSO threshold construct is not $3.65/day — it is each country's own national poverty line — which is what makes the NSO column a genuinely independent measurement.

The zero gap for the five LMIC rows is a calibration match by design: Pan Africa Data anchors LMIC countries to the World Bank Macro Poverty Outlook at the $3.65/day line at ingest time. Only the South Africa gap is a genuine empirical distance versus the World Bank. For South Africa the calibration match is at the $6.85 line, where PAD's cumulative reads 59.81 % and WB PIP publishes 59.81 % — reported here rather than as a second column so the main matrix stays focused on a single comparison threshold.

3. South Africa — the only genuinely out-of-sample check in the panel

South Africa is the sole upper-middle-income country in the panel and provides the only genuinely out-of-sample check of Pan Africa Data against a benchmark at a non-calibration threshold in v1. South Africa is calibrated at the $6.85/day line as an upper-middle-income country. PAD's marginalised class share for 2023 is 35.75 %, which is a genuinely computed value at the $3.65 threshold and sits 5.36 percentage points above the World Bank's own $3.65 figure of 30.39 %.

Statistics South Africa's Lower-Bound Poverty Line of R1,300 per person per month for 2023 works out to approximately $4.81/day 2017 PPP, using the World Bank private consumption PPP factor for South Africa in 2017 (6.66 rand per international dollar) with an intervening CPI adjustment. At that higher-than-$3.65 threshold Statistics South Africa reports 37.9 %. Interpolating linearly between Stats SA's Food Poverty Line (17.6 % at approximately $2.87/day) and its Lower-Bound Poverty Line (37.9 % at approximately $4.81/day), the Stats SA reading at exactly $3.65 comes out to approximately 26 %. PAD's 35.75 % therefore sits about 10 percentage points above what Stats SA implies at that threshold and about 5 points above WB PIP.

4. Trajectory tracking, 2019–2023

The master matrix reads a single reference year. The table below shows Pan Africa Data's marginalised class share through five years of successive economic shocks — COVID (2020), the recovery (2021), Ghana's cedi crisis (2022), Nigeria's naira devaluation (2023) and Egypt's continuing inflation shock (2022–2023). Every year in which the World Bank has subsequently published a $3.65 figure, PAD lands on it.

Nigeria — the panel's clearest case of macroeconomic shock

PAD's marginalised class share rises monotonically from 56.85 % (2019) to 64.02 % (2023), a 7.17 percentage-point deterioration across the period spanning the June 2023 naira devaluation. When the World Bank Macro Poverty Outlook subsequently published its $3.65 figure for Nigeria in 2023, it landed at 64.02 % — the same value PAD's modelled trajectory had projected. The direction and terminal magnitude both match.

Kenya — spike and recover

63.89 % in 2019, rising to 65.65 % (2020) and peaking at 67.86 % in 2021 — capturing the COVID poverty spike and its persistence into the second wave year — then recovering to 65.49 % (2022) and 64.67 % (2023) as the Kenyan economy rebalanced.

Ghana, Egypt and Morocco

Ghana shows slow deterioration from 56.08 % (2019) to 57.91 % (2023) with 2022 as an inflection point, consistent with the cedi crisis of that year. Egypt shows the most volatile trajectory in the panel: 10.23 % (2019) declining to 7.09 % (2021–2022) before returning to 10.17 % (2023) as the inflation crisis hit household budgets. Morocco shows slow secular improvement from 9.63 % (2019) to 8.11 % (2023) — the only country in the panel that avoided a shock-driven poverty deterioration during the period.

The trajectory table is not a forecast back-test. Producing a true forecast back-test requires re-running the calibration from a frozen 2019 baseline using only 2019-vintage information, then comparing the four-year projection to what the World Bank subsequently published. That exercise is committed to as a separate publication in the next-steps section below.

5. How to read the numbers

Calibration fidelity versus empirical validation

For the five LMIC countries in the panel, the exact PAD-versus-WB-PIP match at the $3.65 line is a calibration fidelity check by construction. It confirms that (a) the calibration is being applied consistently across five different countries with materially different Gini values, consumption levels and CPI trajectories, and (b) the ingest is drawing from the correct World Bank source. It is not evidence of the model's predictive accuracy.

The independent signal is the NSO agreement with WB PIP

Each country's national statistics office publishes at a threshold specific to that country's own poverty-line construct — not at the $3.65/day World Bank line — so the NSO figure is not directly comparable to PAD or to WB PIP at $3.65 without interpretation of where the national threshold sits on the international scale. In every case in the panel the direction of the NSO reading is consistent with WB PIP once the threshold construct is accounted for. Because Pan Africa Data tracks WB PIP by design, PAD inherits this directional consistency with the six independent national sources.

South Africa is where the empirical content lives

The South Africa row is the only place in the v1 panel where the PAD-versus-WB-PIP comparison is a genuine empirical distance. The 5.36 pp gap is real and the reasons are stated openly.

6. Limitations disclosed

The validation programme is a v1 exercise; its limitations are stated in full here so an institutional reviewer can weigh them without having to search for them.

7. Sources and data provenance

Pan Africa Data

National-level modelled income distribution for six African countries, 2023 reference year unless otherwise noted, published on the 2021 purchasing-power-parity basis. The marginalised class boundary at $3.65/day (2017 PPP) corresponds directly to the World Bank $3.65 lower-middle-income poverty line. Model version PAD-v1.2-2026.

World Bank Poverty and Inequality Platform

Poverty headcount ratios sourced from the World Bank Macro Poverty Outlook series, March 2026 update. All values on the 2017 PPP basis with the June 2025 rebase applied.

National statistics offices — six independent third measurements

Purchasing-power-parity conversion for South Africa

Local-currency thresholds converted to 2017 PPP dollars using the World Bank private consumption PPP factor for South Africa in 2017 (6.66 rand per international dollar) and the ratio of the South African consumer price index in 2017 to 2023 (146.08 in 2017 versus 194.90 in 2023). Formula: $ 2017 PPP/day = ZAR/day × (CPI 2017 / CPI ref year) ÷ PPP 2017. This is the standard World Bank methodology for placing local-currency thresholds on the 2017 PPP scale and matches the basis used by WB PIP itself.

8. Next steps

Three extensions of the validation programme are planned.

Version 2 — extended panel

The panel will be extended to twelve countries by adding Tanzania (once IHBS 2024/25 is published), Rwanda (using EICV 5), Zambia (using LCMS 2022/23), Uganda (using UNHS 2019/20 or newer), Ethiopia (using WMS or HCE) and one further North African country. The refresh of the six existing rows will use newer NSO surveys where available.

Refined UMIC calibration

The current calibration approach for upper-middle-income countries is being revised in the next model version. The revision is expected to close most of the observed South Africa gap versus WB PIP at the $3.65 line.

Full historical back-test

A full historical back-test running the current model as if it were 2015 and measuring how accurately it would have predicted 2016–2025 for each of the 54 African countries at each of the three WB PIP thresholds is under development. That exercise is analytically substantial and will be published as its own paper.