Regression discontinuity analysis for pharmacovigilance: statin example reflected trial findings showing little evidence of harm

Abstract

Objectives: The study aims to explore the use of regression discontinuity analysis (RDA) to examine effects of prescription of statins on total cholesterol and adverse outcomes (type 2 diabetes, rhabdomyolysis and myopathy, myalgia and myositis, liver disease, CVD, and mortality).

Study design and setting: We conducted a prospective cohort study using the Clinical Practice Research Datalink including patients with QRISK scores of 10 to 30 in 2010 to 2013 who were last followed-up in October 2016. Comparing patients with QRISK≥20 and QRISK<20, we explored RDA assumptions, provided proof of concept analyses (total cholesterol as outcome), and investigated the effect of statins prescription on adverse outcomes.

Result: RDA confirmed statin prescription reduced total cholesterol (Mean difference (MD) -1.33 mmol/L, 95%Confidence Interval (CI) -1.93 to -0.73). RDA provided little evidence for adverse effects on diabetes, myalgia and myositis, liver disease, CVD, or mortality. The RDA analysis findings are similar to RCT results. Findings from non-RDA analysis agree with published observational studies.

Conclusion: RDA can be used with large routine clinical datasets to provide evidence on effects of medications which are prescribed according to a threshold. Testable RDA assumptions were satisfied, but confidence intervals were wide, partly due to the low compliance with the prescribing threshold.

Keywords: Cardiovascular disease; Epidemiology; Health services research; QRISK score; Regression discontinuity analysis; Statins.

Fig. 1

Assumption 1: Discontinuity in the probability of being prescribed statins at the QRISK score threshold.

Fig. 2

Assumption 2: Individual values of the assignment variable are not manipulated – histogram of QRISK scores.

Fig. 3

Assumption 3: Exposure groups are exchangeable at the cut-off-distributions of key measured confounders by QRISK score in the adhering population.

Fig. 4

RDA of all outcomes. RD = Risk difference, CI = confidence interval, IV = Instumental variable; n = 11,758 corresponds to bandwidth 10–30, n = 5,568 corresponds to bandwidth 15-25.