State selection in Markov models for panel data with application to psoriatic arthritis

Area of Science:

• Biostatistics
• Epidemiology
• Health Outcomes Research

Background:

• Continuous-time Markov multistate models are vital for analyzing disease progression and treatment effects.
• Defining disease states and categorizations within these models presents significant uncertainty.
• Different model categorizations can lead to varied conclusions regarding covariate effects and event times.

Purpose of the Study:

• To address the challenge of comparing multistate models with different categorizations and information scales.
• To adapt and apply modified Akaike's and cross-validation criteria for comparing the predictive performance of these models.
• To enable robust model comparison when standard likelihood-based methods are not applicable.

Main Methods:

• Adapted a modified Akaike's criterion and a cross-validatory criterion to compare models based on shared information.
• Utilized Hidden Markov models for implementing the comparison criteria.
• Applied the methodology to panel data, specifically analyzing the Health Assessment Questionnaire score in psoriatic arthritis.

Main Results:

• Developed a method to compare multistate models with differing structures and information content.
• Demonstrated the applicability of the adapted criteria for models fitted to panel data.
• The procedure is implementable within the R package 'msm'.

Conclusions:

• The proposed criteria effectively compare the predictive abilities of multistate models, even with different state definitions.
• This approach provides a reliable method for selecting appropriate models in disease progression studies.
• Facilitates more accurate understanding of disease dynamics and treatment impacts in complex health conditions.