ABSTRACT
During most of 2020, the COVID-19 pandemic gave rise to considerable and growing numbers of hospitalizations across most of the U.S. Typical COVID-19 hospitalization data, including length of stay, intensive care unit (ICU) use, mechanical ventilation (Vent), and in-hospital mortality provide clearly interpretable health care endpoints that can be compared across population strata. They capture the resources consumed for the care of COVID-19 patients, and analysis of these endpoints can be used for resource planning at the local level. Yet, hospitalization data embody novel features that require careful statistical treatment to be useful in this context. Specifically, statistical models must meet three goals: (i) They should mesh with and inform mathematical epidemiologic or agent-based models of the COVID-19 experience in the population. (ii) They need to handle administrative censoring of hospitalization experience when data are extracted and downloaded for a given patient before that patients hospitalization experience has terminated. And, (iii) models need to handle risks for competing events, the occurrence of one blocking the possibility of the other(s). For example, live discharge from the hospital "competes with" (i.e., blocks) in-hospital mortality. We have adapted approaches from the survival analysis literature to address these challenges in order to better understand and quantify the population experience in hospital with respect to length of stay, ICU, Vent use and so on. Using hospitalization data from a large U.S. metropolitan region, in this report, we show how standard techniques from survival analysis can be brought to bear to address these challenges and yield interpretable results. In the breakout/discussion, we will discuss formulation, estimation and inference, and interpretation of competing risks models.