Abstract
OBJECTIVES: To introduce multilevel repeated measures (RM) models and compare them with multilevel difference-in-differences (DID) models in assessing the linear relationship between the length of the policy intervention period and healthcare outcomes (dose-response effect) for data from a stepped-wedge design with a hierarchical structure. DESIGN: The implementation of national essential medicine policy (NEMP) in China was a stepped-wedge-like design of five time points with a hierarchical structure. Using one key healthcare outcome from the national NEMP surveillance data as an example, we illustrate how a series of multilevel DID models and one multilevel RM model can be fitted to answer some research questions on policy effects. SETTING: Routinely and annually collected national data on China from 2008 to 2012. PARTICIPANTS: 34 506 primary healthcare facilities in 2675 counties of 31 provinces. OUTCOME MEASURES: Agreement and differences in estimates of dose-response effect and variation in such effect between the two methods on the logarithm-transformed total number of outpatient visits per facility per year (LG-OPV). RESULTS: The estimated dose-response effect was approximately 0.015 according to four multilevel DID models and precisely 0.012 from one multilevel RM model. Both types of model estimated an increase in LG-OPV by 2.55 times from 2009 to 2012, but 2-4.3 times larger SEs of those estimates were found by the multilevel DID models. Similar estimates of mean effects of covariates and random effects of the average LG-OPV among all levels in the example dataset were obtained by both types of model. Significant variances in the dose-response among provinces, counties and facilities were estimated, and the 'lowest' or 'highest' units by their dose-response effects were pinpointed only by the multilevel RM model. CONCLUSIONS: For examining dose-response effect based on data from multiple time points with hierarchical structure and the stepped wedge-like designs, multilevel RM models are more efficient, convenient and informative than the multilevel DID models.