Measurement Error Correction and Multidimensional Aggregation in Epidemiological Surveillance Systems: Application to the Ebola Virus Disease Epidemic in the Democratic Republic of the Congo

Epidemiological surveillance systems produce decision-making indicators from data that may be affected by under-reporting, misclassification and notification delays. These errors can propagate through the extraction, transformation, loading and multidimensional aggregation stages. This article proposes a mathematical framework that integrates measurement error correction into an ETL–data warehouse–OLAP decision architecture. The central problem is to determine whether the correction of a misclassification error can be performed before or after aggregation without changing the resulting indicator. We show that the Rogan–Gladen estimator commutes with the roll-up operator only when the sensitivity and specificity parameters are homogeneous across the aggregated strata. A numerical instantiation based on a synthetic dataset calibrated on published parameters of the Ebola virus disease epidemic in North Kivu and Ituri (DRC, 2018–2020) illustrates this non-commutation and its consequences for the estimation of incidence and the reproduction number. These results lead to a design rule: when surveillance quality is heterogeneous, correction must be performed at the finest level at which the error parameters can be regarded as homogeneous, before any aggregation, because this is the only strategy that remains unbiased whatever the spatial distribution of cases. A bias–variance analysis (Theorem 4) and a Monte Carlo uncertainty analysis (5,000 replicates, seven scenarios) further show that aggregate-level correction can have a slightly smaller variance but becomes biased when cases concentrate in poorly covered zones, and that the corrected indicators are highly sensitive to the assumed specificity when test positivity is close to the false-positive floor (a misspecification of 0.005 in specificity changes cumulative incidence by more than 100%). No individual-level surveillance data were used: the synthetic data are fully specified by a generation algorithm and a parameter table provided for reproducibility. The proposed framework thus provides a formal basis for designing surveillance decision-support systems that explicitly account for data quality.