Testing a routine-data shortcut below district level
The paper asks whether a coverage estimation method built on routine health data can work at lower administrative levels than it was designed for. Coverage here means the share of the eligible child population that received a service; estimating it requires a numerator (service events recorded) and a denominator (children eligible for that service). In much of sub-Saharan Africa, censuses and vital registration are incomplete or outdated, so the study tests whether routine data can supply both numbers.
The method comes from earlier work by Maina and colleagues. It infers the denominator from a 'benchmark' intervention assumed to reach nearly everyone, so that its recorded service events approximate the eligible population. The study used the first dose of the pentavalent vaccine (Penta1), typically given at 6 weeks in Ghana, as that benchmark, and then estimated coverage for three later vaccinations and utilisation of outpatient diarrhoea care for children under 1 year.
Data came from Ghana's District Health Information Management System (DHIMS2), covering facility records from 2017 to 2020, and estimates were produced for 2020 in the Eastern region, its districts, selected sub-districts and individual health facilities. Household survey data were used to calibrate and evaluate the results, so the routine-data estimates could be compared against survey-derived equivalents.
Regional vaccination estimates sit close to survey values
At the regional level, the routine-data approach produced vaccination coverage that looked credible. Estimated coverage was 90.3% for the second pentavalent dose (Penta2), 92.4% for the third (Penta3) and 83.8% for the first measles-rubella dose (MRV1).
The comparison with survey-derived estimates shows how close these are, and where they are not. For Penta2 the survey estimate was 89.8% (95% CI 82.9–96.7), a deviation of 0.4 percentage points. For MRV1 the survey estimate was 81.3% (72.6–90.0), a deviation of 2.5 points. Penta3 diverged more: the survey estimate was 80.1% (68.7–91.5) against 92.4% from routine data, a deviation of 12.3 points.
That Penta3 gap matters for reading the rest of the paper. The approach is not uniformly accurate even at the top level; it tracked Penta2 and MRV1 closely but overstated Penta3 substantially in this region.
Smaller units bring wider spread and impossible values
As the analysis moved from region to district to sub-district to facility, the estimates became more variable and more often implausible. A coverage estimate above 100% is impossible in principle, since it would mean more doses given than children eligible, and the paper treats such values as a signal that the numerator and denominator no longer describe the same population.
For Penta2, the district-level median across all 33 districts was 93.6% (IQR 87.3–96.2), with 2 districts (6.1%) above 100%. At facility level the median was 96.5% (IQR 92.0–104.5), and the upper quartile already crosses 100%. For Penta3, 9 of 33 districts (27.2%) produced estimates above 100%.
The paper's explanation is cross-boundary health-seeking. The method implicitly assumes that people using a facility live in the administrative unit that facility serves. In smaller units, and especially in facility catchment areas, that assumption is more likely to fail: caregivers may bypass the nearest facility for a higher-tier one, or avoid a facility they perceive as below standard. When that happens, the recorded numerator and the Penta1-derived denominator can refer to different groups of children.
Curative care is underestimated at every level
Diarrhoea service utilisation came out lower than the survey-derived figure at all four levels. At region level the routine-data estimate was 7.8% against a survey estimate of 15.4% (95% CI 8.7–22.1), a deviation of −7.6 points. The district-level median was 7.7% (IQR 5.9–9.6) and the facility-level median 2.3% (IQR 0.0–5.5), with several facilities near zero.
The paper attributes this to a mismatch in what the two numbers count. Vaccinations are preventive and targeted at every eligible child, so even small units record many events. Diarrhoea treatment is curative and used only by children who are both ill and taken for care: the 2017/18 Ghana survey put diarrhoea prevalence among under-1s in the Eastern region at 15.4%, with treatment-seeking at public or private facilities in just 46.5% of episodes. The numerator is therefore far smaller, while the Penta1-derived denominator still represents all children eligible for the service rather than those who used it.
The authors conclude that these utilisation estimates cannot reliably guide allocation of curative resources. They do suggest the Penta1-derived denominator may still be relevant to preventive work, such as resourcing health promotion and education activities in local communities.
Facility estimates may diagnose problems rather than allocate resources
The study is, to the authors' knowledge, the first to apply this approach at the level of individual health facilities. Its central finding is a gradient: the method is relatively robust at the highest levels of aggregation and becomes unstable at the level of individual facilities, which the authors say may preclude using facility estimates for resource allocation.
The spatial analysis supports the cross-boundary explanation without proving it. Global Moran's I values for facility-level coverage were close to zero and not statistically significant — for example −0.073 (p = 0.378) for Penta2 and −0.018 (p = 0.968) for Penta3 — so the estimates did not show strong overall spatial clustering. The paper nonetheless notes low–high and high–low clustering among facility estimates, which it reads as a sign that the factors driving health-seeking may differ between the benchmark and target interventions.
The authors propose a different use for these unstable estimates: a diagnostic role. Facility-level values that depart sharply could flag local areas worth investigating to find out why a facility is over- or under-utilised, rather than being read as direct measures of coverage.
Modelling health-seeking could beat a fixed denominator
The paper's forward-looking conclusion is methodological. Because the fixed Penta1-derived denominator cannot account for why caregivers choose one facility over another, the authors argue that routine health data should be integrated into geostatistical estimation approaches able to model the spatial and aspatial determinants of health-seeking. That, they suggest, could produce more accurate estimates and better inform facility-level resource allocation.
The study also notes practical complications. The benchmark rule recommended by Maina and colleagues is to check against contemporaneous survey data that the prospective benchmark has at least 90% coverage. The authors also raise the possibility that children may receive Penta1 outside the facilities captured, and that the study period overlapped with the early COVID-19 pandemic, which may have caused some differential under-utilisation of certain services.
The scope is an ecological analysis of one Ghanaian region using 2017–2020 facility data to estimate 2020 coverage, calibrated against household surveys. The findings describe how this method behaves across administrative levels in that setting; they are not a validation of the approach elsewhere.
Trust the district picture, question the facility number
The practical split the paper supports is between levels and between service types. For regional and district vaccination coverage, the routine-data approach produced estimates close to survey values for Penta2 and MRV1, and the district medians for Penta2 and Penta3 (93.6% and 96.2%) sit in a plausible range. For diarrhoea treatment, the same approach understated utilisation at every level, so its numbers should not be used to size curative resources.
The most consequential threshold is the 100% line. When a facility or district estimate crosses it, the paper's own interpretation is that the numerator and denominator have come apart, most likely through cross-boundary health-seeking. That makes an implausible value a prompt for local investigation rather than a coverage figure to report.
The paper's priority is therefore not to abandon routine data but to change how it is analysed below district level: use it to reveal local variation and flag anomalies, and pair it with estimation methods that model where people actually seek care before using facility-level numbers to allocate resources.
Sources used in this answer
Child health service coverage estimation using routine health data from Ghana’s Eastern Region
Routine-data coverage estimates were plausible at region and district level but grew more variable and more often exceeded 100% at facility level, while diarrhoea service utilisation was underestimated at all levels.