Many studies have attempted to determine the disease severity and patterns of COVID-19. However, at the beginning of the pandemic, the complex patients’ trajectories were only descriptively reported, and many analyses were worryingly prone to time-dependent-, selection-, and competing risk biases. Multi-state models avoid these biases by jointly analysing multiple clinical outcomes while taking into account their time dependency, including current cases, and modelling competing events. This paper uses a publicly available data set from the first wave in Israel as an example to demonstrate the benefits of analysing hospital data via multi-state methodology.
We compared the outcome of the data analysis using multi-state models with the outcome obtained when various forms of bias are ignored. Furthermore, we used Cox regression to model the transitions among the states in a multi-state model. This allowed for the comparison of the covariates’ influence on transition rates between the two states. Lastly, we calculated expected lengths of stay and state probabilities based on the multi-state model and visualised it using stacked probability plots.
Compared to standard methods, multi-state models avoid many biases in the analysis of real-time disease developments. The utility of multi-state models is further highlighted through the use of stacked probability plots, which visualise the results. In addition, by stratification of disease patterns by subgroups and visualisation of the distribution of possible outcomes, these models bring the data into an interpretable form.
To accurately guide the provision of medical resources, this paper recommends the real-time collection of hospital data and its analysis using multi-state models, as this method eliminates many potential biases. By applying multi-state models to real-time data, the gained knowledge allows rapid detection of altered disease courses when new variants arise, which is essential when informing medical and political decision-makers as well as the general population.