Bayesian structural time series, an alternative to interrupted time series in the right circumstances

Abstract

ObjectivesCompare two approaches to analyzing time series data—interrupted time series with segmented regression (ITS-SR) and Bayesian structural time series using the CausalImpact R package (BSTS-CI)—highlighting advantages, disadvantages, and implementation considerations. Study Design and SettingWe analyzed electronic health records using each approach to estimate the antibiotic prescribing reduction associated with an educational program delivered to Australian primary care physicians between 2012 and 2017. Two outcomes were considered: antibiotics for upper respiratory tract infections (URTIs) and antibiotics of specified formulations. ResultsFor URTI indication prescribing, average monthly prescriptions changes were estimated at −4,550; (95% confidence interval, −5,486 to −3,614) and −4,270; (95% credible interval, −5,934 to −2,626) for ITS-SR and BSTS-CI, respectively. Similarly for specified formulation prescribing, monthly average changes were estimated at −7,923; (95% confidence interval, −15,887 to 40) for ITS-SR and −20,269; (95% credible interval, −25,011 to −15,635) for BSTS-CI. ConclusionDiffering results between ITS-SR and BSTS-CI appear driven by divergent explanatory and outcome series trends. The BSTS-CI may be a suitable alternative to ITS-SR only if the explanatory series represent the secular trend of the outcome series before the intervention and are equally affected by exogenous or confounding factors. When appropriately applied, BSTS-CI provides an alternative to ITS with more readily interpretable Bayesian effect estimates.