Characterising patterns in routinely reported longitudinal HIV data in South Africa using a Bayesian multiplicative interaction model

Introduction We consider an analytical problem of characterising patterns and identifying discrepancies between database systems for longitudinal aggregated healthcare data involving multiple facilities. Methods We used routinely collected data on the registered number of people living with HIV who initiated antiretroviral treatment (ART) in 69 South African facilities in 2019; reported in the Three Interlinked Electronic register (Tier.net) and the District Health Information System. A Bayesian multiplicative interaction model quantified the average time effect as realised through the heterogeneous facility-specific slopes and quantified discrepancies between the two database sources. Results The estimated average trends showed a slight dip in June and a large dip in December. The estimated slopes identified clusters of facilities based on their ranges of fluctuations over time. The differences in average monthly ART initiations between the two database sources had a median of 1.6 (IQR 0.8–3.3), while 3 outlying facilities differed by at least 10 ART initiations between the 2 sources. Conclusion Multiplicative interaction models are a powerful tool for quantifying average trends over time and for evaluating discrepancies between reporting systems for multiple facilities with heterogeneous time slopes. The Bayesian framework enables efficient estimation for a very large number of parameters.