Effects of categorizing continuous variables in decision-analytic models.

Area of Science:

• Health economics
• Biostatistics
• Epidemiological modeling

Background:

• Markov models are essential for simulating disease progression and intervention impacts.
• Continuous risk factors require categorization in discrete-state Markov models, potentially introducing bias.
• Assessing the trade-offs between bias and model complexity is crucial for accurate health economic evaluations.

Purpose of the Study:

• To evaluate the bias introduced by categorizing continuous risk factors in Markov models.
• To assess the relationship between the number of risk categories and model bias.
• To determine the impact of disease incidence, mortality, and relative risk on bias.

Main Methods:

• Developed a generic Markov cohort model to simulate disease progression.
• Defined bias as the percentage change in life expectancy gain from an intervention.
• Compared outcomes using 2 to 15 risk factor categories versus a continuous variable.

Main Results:

• Categorization overestimated life expectancy gains (positive bias) in the base case.
• Bias decreased as the number of risk categories increased, not exceeding 4% absolute bias.
• Bias magnitude varied with disease incidence, mortality, and relative risk, with increased bias at higher relative risks.

Conclusions:

• Categorizing continuous risk factors in Markov models results in less than 4% absolute bias when using at least two categories.
• This approach is acceptable under assumptions of normally distributed risk factors and moderate relative risks.
• The findings support the use of categorized risk factors in Markov models, balancing accuracy and practical data limitations.