When care is free
Quasi-experimental evidence from Africa's user-fee abolition wave, 2001–2016
For decades, most public clinics in sub-Saharan Africa charged patients at the door. A delivery might cost a week's income; a child's malaria treatment, a day's income. Between 2001 and 2016, nine countries decided to stop charging mothers and young children. They did it in different years, for different reasons, which is exactly what makes the question answerable: did removing fees save lives?
This analysis treats the staggered rollout as a natural experiment. Countries that abolished fees are compared against 26 sub-Saharan countries that kept charging, using two designs that answer the question from different angles: local-projections difference-in-differences (LP-DiD) across all nine adopters, and a synthetic control case study of Sierra Leone's 2010 Free Health Care Initiative.
Under-5 mortality, pooled effect
LP-DiD across 9 adopting countries, 95% CI
Sierra Leone maternal mortality, 2010–13
vs. synthetic Sierra Leone (ridge-augmented), before Ebola. Placebo rank test p =
Design
9 + 26countries: 9 staggered adopters, 26 never-treated controls, 1995–2023, World Bank / UN IGME public data
Every number on this page is computed live from the result files produced by the analysis pipeline. Nothing is pasted in, so the story and the statistics cannot drift apart.
The policy wave
The staggered timing matters methodologically. A single before/after comparison confounds the policy with everything else happening at the time. Nine different adoption years, each compared against countries that never adopted, is a much harder pattern for a confounder to mimic.
Nine countries, fewer child deaths
The event study below traces outcomes relative to each country's abolition year. Estimates left of zero test the identifying assumption: if treated and control countries were on different paths before the policy, the design is in trouble. Estimates right of zero measure the effect.
Under-5 mortality falls by about
Four things to know before believing this chart:
- The pre-period is not perfectly flat. At t−6 the point estimate is positive and marginally significant, meaning treated countries were improving somewhat faster than controls before abolishing fees. Governments do not flip a coin to decide health policy. The effect estimate survives this (the post-period drop is larger and sharper than the pre-trend), but a reader should know it is there. Toggle to the classic TWFE estimator to see why estimator choice matters under staggered adoption.
- Mortality series are modeled. The UN IGME estimates smooth over sharp year-to-year changes, which biases against finding sudden policy effects. The true short-run effect is likely larger than what smoothed data can show.
- This is an intent-to-treat estimate of a policy announcement. Implementation quality varied enormously; Uganda's abolition was underfunded. The average mixes strong and weak implementations.
- The p-value depends on how strict you are. Cluster-robust inference gives p =
; reassigning the nine adoption years to random countries 1,000 times gives a design-based p = (one-sided). The estimate holds its sign and size across every specification tested (timing shifts, broad treatment coding, dropping Ebola countries, dropping any cohort), but under the strictest test it sits at the edge of significance, not comfortably past it. The methods page shows all of it.
Sierra Leone: the sharpest test
In April 2010, Sierra Leone abolished all fees for pregnant women, lactating mothers, and children under five in one stroke: the Free Health Care Initiative. It was the boldest version of the policy anywhere, in one of the world's most dangerous places to give birth.
The synthetic control method builds a "counterfactual Sierra Leone" from a weighted mix of countries that kept charging fees, matched on pre-2010 trajectories and structural characteristics. Sierra Leone's mortality was near the top of the donor pool, so the classic estimator struggles to match its level; the ridge-augmented estimator (AugSynth) fits the pre-period almost exactly. Both are shown.
The two outcomes tell different stories:
Maternal mortality diverges. Across 2010–13, before Ebola struck, actual maternal mortality averaged about
Under-5 mortality shows nothing. The actual and synthetic paths overlap almost perfectly. On smoothed mortality data, for a single country, the child-survival effect of the FHCI is not detectable, even though the pooled nine-country design finds one.
Placebo tests
Synthetic control has no standard errors, so inference works by pretending each donor country passed the policy and measuring the fake "effects." If Sierra Leone's gap is not unusual against that distribution, the result is noise.
For maternal mortality, Sierra Leone's 2010–13 gap is larger than 18 of the 20 placebos (rank test p ≈
What I would tell a policymaker
Removing point-of-care fees for mothers and children is followed by an acceleration in child survival across nine African adopters, on the order of
Why I built this
User fees are one of health financing's oldest arguments. The 1987 Bamako Initiative institutionalized them across Africa; the abolition wave this analysis studies unwound them. I spent four years at UNICEF producing the maternal and child mortality statistics these models run on, and grew up partly in two of the countries in the treatment table. The debate surfaced constantly in that work; a clean causal answer never did. This is me going back for it.
Read next
- Who benefited? — the mechanism: wealth-quintile survey data show poorest-quintile service use rising faster after abolition than in the never-treated controls, for every adopter country-service pair
- Policy brief — the three-minute version for a decision-maker
- Explore the data — every indicator, every country, every year used in this analysis
- Methods & data quality — treatment coding decisions, estimator details, the data error we caught in the World Bank series, and everything that could be wrong with this analysis