ABSTRACT
ImportanceCommunication and adoption of modern study design and analytical techniques is of high importance for the improvement of clinical research from observational data. ObjectiveTo compare (1) a modern method for causal inference including a target trial emulation framework and doubly robust estimation to (2) approaches common in the clinical literature such as Cox proportional hazards models. To do this, we estimate the effect of corticosteroids on mortality for moderate-to-severe coronavirus disease 2019 (COVID-19) patients. We use the World Health Organizations (WHO) meta-analysis of corticosteroid randomized controlled trials (RCTs) as a benchmark. DesignRetrospective cohort study using longitudinal electronic health record data for 28 days from time of hospitalization. SettingsMulti-center New York City hospital system. ParticipantsAdult patients hospitalized between March 1-May 15, 2020 with COVID-19 and not on corticosteroids for chronic use. InterventionCorticosteroid exposure defined as >0.5mg/kg methylprednisolone equivalent in a 24-hour period. For target trial emulation, interventions are (1) corticosteroids for six days if and when patient meets criteria for severe hypoxia and (2) no corticosteroids. For approaches common in clinical literature, treatment definitions used for variables in Cox regression models vary by study design (no time frame, one-, and five-days from time of severe hypoxia). Main outcome28-day mortality from time of hospitalization. Results3,298 patients (median age 65 (IQR 53-77), 60% male). 423 receive corticosteroids at any point during hospitalization, 699 die within 28 days of hospitalization. Target trial emulation estimates corticosteroids to reduce 28-day mortality from 32.2% (95% CI 30.9-33.5) to 25.7% (24.5-26.9). This estimate is qualitatively identical to the WHOs RCT meta-analysis odds ratio of 0.66 (0.53-0.82)). Hazard ratios using methods comparable to current corticosteroid research range in size and direction from 0.50 (0.41-0.62) to 1.08 (0.80-1.47). Conclusion and RelevanceClinical research based on observational data can unveil true causal relationships; however, the correctness of these effect estimates requires designing the study and analyzing the data based on principles which are different from the current standard in clinical research. Key PointsO_ST_ABSQuestionC_ST_ABSHow do modern methods for causal inference compare to approaches common in the clinical literature when estimating the effect of corticosteroids on mortality for moderate-to-severe coronavirus disease 2019 (COVID-19) patients? FindingsIn an analysis using retrospective data for 3,298 hospitalized COVID-19 patients, target trial emulation using a doubly robust estimation procedure successfully recovers a randomized controlled trial (RCT) meta-analysis benchmark. In contrast, analytic approaches common in the clinical research literature generally cannot recover the benchmark. MeaningClinical research based on observational data can unveil true causal relations. However, the correctness of these effect estimates requires designing and analyzing the data based on principles which are different from the current standard in clinical research. Widespread communication and adoption of these analytical techniques are of high importance for the improvement of clinical research.
