Target trial emulation with multi-state model analysis to assess treatment effectiveness using clinical COVID-19 data

Abstract

Background: Real-world observational data are an important source of evidence on the treatment effectiveness for patients hospitalized with coronavirus disease 2019 (COVID-19). However, observational studies evaluating treatment effectiveness based on longitudinal data are often prone to methodological biases such as immortal time bias, confounding bias, and competing risks.

Methods: For exemplary target trial emulation, we used a cohort of patients hospitalized with COVID-19 (n = 501) in a single centre. We described the methodology for evaluating the effectiveness of a single-dose treatment, emulated a trial using real-world data, and drafted a hypothetical study protocol describing the main components. To avoid immortal time and time-fixed confounding biases, we applied the clone-censor-weight technique. We set a 5-day grace period as a period of time when treatment could be initiated. We used the inverse probability of censoring weights to account for the selection bias introduced by artificial censoring. To estimate the treatment effects, we took the multi-state model approach. We considered a multi-state model with five states. The primary endpoint was defined as clinical severity status, assessed by a 5-point ordinal scale on day 30. Differences between the treatment group and standard of care treatment group were calculated using a proportional odds model and shown as odds ratios. Additionally, the weighted cause-specific hazards and transition probabilities for each treatment arm were presented.

Results: Our study demonstrates that trial emulation with a multi-state model analysis is a suitable approach to address observational data limitations, evaluate treatment effects on clinically heterogeneous in-hospital death and discharge alive endpoints, and consider the intermediate state of admission to ICU. The multi-state model analysis allows us to summarize results using stacked probability plots that make it easier to interpret results.

Conclusions: Extending the emulated target trial approach to multi-state model analysis complements treatment effectiveness analysis by gaining information on competing events. Combining two methodologies offers an option to address immortal time bias, confounding bias, and competing risk events. This methodological approach can provide additional insight for decision-making, particularly when data from randomized controlled trials (RCTs) are unavailable.

Keywords: Bias; COVID-19; Multi-state models; Observational data; Target trial emulation.

Fig. 1

Multi-state model for COVID-19 progression. Notes: Five possible states with seven transitions and the number of patients for each transition were defined. The ICU status was modelled as an intermediate state represented in a multi-state model

Fig. 2

Weighted estimated cause-specific cumulative hazards from the normal hospital ward using the Nelson-Aalen estimator. a-d Illustrates transitions from normal ward. a From normal ward to ICU. b From normal ward to in-hospital death. c From normal ward to discharge home. d From normal ward to discharge to another HCF

Fig. 3

Weighted results for transition rates starting from hospital admission. Notes: a Non-X-treated arm. b X-treated arm. State 1: Normal ward. State 2: Admission to ICU. State 3: In-hospital death; State 4: Discharge home. State 5: Discharge to another HCF