Objective: To apply a causal analysis approach to estimate the intent-to-treat and per-protocol effects in a safety outcome study of the 30-day risk of liver injury associated with antibiotic use (macrolides, penicillin-based antibiotics, and fluoroquinolones).
Research design and methods: For each antibiotic class, we constructed a pooled cohort of treated episodes matched with untreated episodes using an electronic medical record database from a university hospital. High-dimensional propensity scores were calculated using baseline patient characteristics, concomitant medications and medical history as surrogate confounders. Intent-to-treat hazard ratios (HRs) were estimated using inverse probability of treatment weighted discrete hazard models that ignored subsequent treatment changes. Per-protocol HRs were calculated using inverse probability of treatment and censoring weighted models after terminating each episode's observation at the first treatment change.
Results: For macrolides, the intent-to-treat HR (95% confidence interval) was 1.22 (0.75-1.98) and the per-protocol HR was 1.22 (0.67-2.22). For penicillin-based antibiotics, the intent-to-treat HR was 4.01 (3.16-5.08) and the per-protocol HR was 7.25 (5.58-9.41). For fluoroquinolones, the intent-to-treat HR was 1.60 (1.27-2.03) and the per-protocol HR was 1.69 (1.23-2.30).
Conclusion: Researchers should clearly define the target estimands, and carefully estimate and interpret both effects.
Keywords: Causal inference; cohort study; electronic medical records data; observational study; pharmacoepidemiology; propensity score.