Accounting for length of hospital stay in regression models in clinical epidemiology

摘要

In hospital epidemiology, logistic regression is a popularmodel to study risk factors of hospital-acquiredinfections. One key issue in this analysis is how toincorporate the time dependency of acquiring an infectionduring the hospital stay. In the applied literature,researchers often simply adjust for the entire lengthof hospital stay, which also includes the time afterinfection. A further issue is that discharge and deathare competing events for hospital-acquired infections.After discussing the limitations of logistic regressionadjusted for length of stay in this setting, we comparethis approach with appropriate analyses incorporatingcompeting risks and with an illness–death model withhospital-acquired infection as an intermediate event.The cumulative incidence function, cause-specific hazardratios, and subdistribution hazard ratios are consideredas reference measures. Real-life and simulateddata are used to demonstrate biases and limitationsassociated with logistic regression adjusted for lengthof stay. We conclude that logistic regression adjustedfor length of stay should not be used when investigatinghospital-acquired infections and that appropriate methods involving the use of multistate models shouldbe used to capture the time dependency in time-to-eventsettings, especially in the presence of competing events.