Background
Vaccine effectiveness (VE) evidence is often delayed in low-resource settings because PCR capacity and serology coverage are limited. We aimed to develop a sentinel-clinic framework that produces timely, policy-grade VE estimates using minimal laboratory inputs while preserving validity across districts and seasons.
Methods
We combined routinely collected influenza-like illness (ILI) encounter data from sentinel primary-care networks and emergency departments with rapid antigen testing (RAT) results at presentation and a targeted serology subsample for calibration. The analytic backbone was a test-negative design with doubly robust estimation such as outcome regression and inverse-probability weighting, and transportable modeling to generalize VE from sentinel catchments to district populations using census-linked covariates. Syndromic severity scores were derived from standardized symptom clusters and vital-sign thresholds; a RAT timing proxy was used to model diagnostic sensitivity decay. Monte Carlo randomization tests with permutation within clinic-week strata and evaluated robustness to residual confounding. Model comparison used log-likelihood ratio (LLR) tests; discrimination of vaccinated vs unvaccinated infection risk used AUC from cross-validated risk scores.
Results
Across 18 sentinel clinics, 24,612 ILI visits were captured over two seasons; 9,842 received RAT and 1,210 contributed paired serology. Median age was 34 years (IQR 18–57). Infection positivity increased with delayed testing (β for days-since-onset = 0.21, SE 0.03 on log-odds scale), consistent with clinic presentation lags. After calibration, overall VE against RAT-confirmed infection was 41 (AUC 0.73), with stronger protection when serology indicated recent boosting (median severity score difference −0.38, IQR −0.62 to −0.11). Transported district-level VE varied by crowding and care-access proxies, showing a measurable policy contrast such as districts in the highest crowding quintile had 9-point lower VE than the lowest quintile under the same product mix. LLR favored models incorporating severity and timing proxies over antigen-only models (LLR 18.6). Monte Carlo p-values remained <0.01 for the VE signal under clinic-week permutations, without inflation of effect sizes. Sensitivity analyses showed peak attenuation of VE estimates by 6–8 points when serology calibration was removed, indicating controlled bias rather than overclaiming.
Conclusions
A sentinel, minimal-lab VE system can yield timely, transportable estimates that remain statistically stable under permutation stress-tests and provide interpretable, district-actionable contrasts for immunization policy in resource-constrained settings.