Evaluating signals generated in a large-scale sequence symmetry analyses: macrolides and heart failure; and NSAIDs and pneumonia

Abstract

Objective Using US claims data and the most up-to-date pharmacoepidemiological study design tools we aimed to investigate two safety signals for (1) macrolides and heart failure; and (2) non-steroidal anti-inflammatory drugs (NSAIDs) and pneumonia generated from a large-scale screening analysis using a self-controlled sequence symmetry design in Danish data. Methods We used IBM Marketscan data to conduct two new-user, active-comparator cohort studies. In the macrolides example, the exposure was clarithromycin or azithromycin and the comparator was amoxicillin/clavulanate, in patients with sinusitis. In the NSAIDs example, the exposure was oral NSAIDs and the comparator was topical diclofenac, in patients with osteoarthritis. We used Cox proportional hazards regression to estimate hazard ratios (HR) and 95% confidence intervals (CI) to adjust for approximately 50 investigator-specified confounders in a propensity score (PS) matched analysis. In a secondary analysis, we used high-dimensional PS (hd-PS) to adjust for 200 additional proxy confounders. Results We had 1,012,364 propensity score matched patients exposed to clarithromycin or azithromycin versus amoxicillin/clavulanate. With 162 outcomes among clarithromycin or azithromycin exposed patients and 134 among amoxicillin/clavulanate, the HR for overall heart failure was 1.14 (95% CI 0.90-1.43). In the NSAIDs example, we included 94,490 patients after propensity score matching. With 794 pneumonia outcomes among oral NSAID patients and 700 among topical diclofenac, we found HR 0.98 (95% CI 0.89-1.09). Some upward bias was suspected as larger HRs were observed in the days immediately following exposure for both the macrolides and NSAIDs examples. We found similar results in the hd-PS matched analyses for both examples. Conclusion Our findings for NSAIDs and pneumonia suggest the original signal may have been due to protopathic or detection bias. Our analyses for macrolides and heart failure with short-term follow-up also suggest bias, although we encourage further research.

High‐throughput screening for prescribing cascades among real world statin initiators

“…Prior work has suggested corrections for multiple testing may not be advantageous in this setting 21,47,48 . Nevertheless, we acknowledge that our findings are hypothesis‐generating, and need additional confirmation in a traditional cohort study is necessary for confirmation 49 …”

“…21,47,48 Nevertheless, we acknowledge that our findings are hypothesis-generating, and need additional confirmation in a traditional cohort study is necessary for confirmation. 49…”