Multistate modeling in long-term follow-up studies, with application to a multiple sclerosis cohort.

Objectives: In long-term follow-up studies, individuals often experience multiple types of events. Standard survival models focus on just one type, limiting the scope of the analysis. In contrast, multistate models (MSMs) investigate multiple event types simultaneously. MSMs, though straightforward to implement, are not common in clinical studies. This article demystifies multistate modeling and demonstrates helpful techniques for long-term follow-up studies.

Study Design And Setting: A case study (1996-2017) of 9124 individuals with multiple sclerosis (MS) illustrates several multistate modeling techniques: choosing the time scale, using time-dependent strata and time-varying coefficients in intensity-based analyses, and analysis of risk using pseudovalue regression with landmarking. We apply these techniques with an illness-death model (states are out-of-hospital, hospitalized, and dead) to investigate the association between exposure to disease-modifying drugs (DMDs) and hospitalizations, while accounting for the competing risk of death.

Results: Using intensity-based analyses, we found exposure to DMDs was associated with reduced hazard of hospitalization soon after the first demyelinating/MS-related event (index date). The hazard ratio (HR) of hospitalization was 0.79 (CI: 0.71, 0.88) 2 years after the index date. This HR increased by 2.83% (CI: 1.65%, 4.02%) annually. We did not find evidence of association between exposure and the hazard of discharge. Pseudovalue regression reveals the association between exposure and time out-of-hospital differed by comorbidity levels: individuals with high comorbidity burden experienced greater benefits.

Conclusion: We demonstrated the benefits of MSMs, showed that the approach is straightforward to implement, and described some potential issues in model interpretation.

Plain Language Summary: Even though many outcomes are relevant when studying diseases, most studies on drug efficacy investigate a drug's effect on only a single outcome. Multistate modeling techniques take a broader view and allow the drug's effects on multiple types of events to be quantified which can elucidate a greater understanding of the impact of the drug. This paper applies multistate modeling to health-care data of British Columbia residents living with multiple sclerosis. In doing so, the association between drugs that are prescribed to treat multiple sclerosis and three outcomes are explored: the frequency of hospitalizations, the lengths of hospital visits, and the chances of dying (inside and out of the hospital). The impact of these drugs for individuals living with comorbidities is studied as well.