Confounding by indication of the safety of de-escalation in community-acquired pneumonia: a simulation study embedded in a prospective cohort.

Background: Observational studies have demonstrated that de-escalation of antimicrobial therapy is independently associated with lower mortality. This most probably results from confounding by indication. Reaching clinical stability is associated with the decision to de-escalate and with survival. However, studies rarely adjust for this confounder. We quantified the potential confounding effect of clinical stability on the estimated impact of de-escalation on mortality in patients with community-acquired pneumonia. Methods: Data were used from the Community-Acquired Pneumonia immunization Trial in Adults (CAPiTA). The primary outcome was 30-day mortality. We performed Cox proportional-hazards regression with de-escalation as time-dependent variable and adjusted for baseline characteristics using propensity scores. The potential impact of unmeasured confounding was quantified through simulating a variable representing clinical stability on day three, using data on prevalence and associations with mortality from the literature. Results: Of 1,536 included patients, 257 (16.7%) were de-escalated, 123 (8.0%) were escalated and in 1156 (75.3%) the antibiotic spectrum remained unchanged. The adjusted hazard ratio of de-escalation for 30-day mortality (compared to patients with unchanged coverage), without adjustment for clinical stability, was 0.36 (95%CI: 0.18-0.73). If 90% to 100% of de-escalated patients were clinically stable on day three, the fully adjusted hazard ratio would be 0.53 (95%CI: 0.26-1.08) to 0.90 (95%CI: 0.42-1.91), respectively. The simulated confounder was substantially stronger than any of the baseline confounders in our dataset. Conclusions: With plausible, literature-based assumptions, clinical stability is a very strong confounder for the effects of de-escalation. Quantification of effects of de-escalation on patient outcomes without proper adjustment for clinical stability results in strong negative bias. As a result, the safety of de-escalation remains to be determined.