Purpose: To assess the tradeoffs between model bias and complexity when categorizing continuous risk factors in decision-analytic models.

Methods: We developed a generic decision-analytic Markov cohort model in which subjects are distributed into disease-free states defined by a categorized risk factor. From each disease-free state we assign a unique probability of developing disease based on the average risk factor value and assume a logistic function of disease risk. Persons may die either from disease or from other causes. The outcome of interest is life expectancy with versus without a hypothetical intervention, which changes the average risk factor value by a percentage. We ran the model while varying the number of risk factor categories from 2 to 15, using 15 as the gold standard against which to compare results. We evaluated model results for changes in four key model parameters: overall disease risk; risk factor effect; disease-specific mortality; and intervention effect. Bias was defined as: (LEGAin_NMax-LEGain_N)/LEGain_N, where LEGain is the life expectancy gain from the hypothetical intervention, N is the number of categories used, and NMax is 15.

Results: The bias that was associated with using fewer than 15 categories never exceeded an absolute value of 0.85%, and always approached zero as the number of categories increased. For most parameter variations the bias was negative, indicating that using fewer categories overestimates expected intervention benefit. A positive bias – suggesting an underestimate of benefit – resulted only when the overall annual probability of disease was very high (10-15%) or the risk factor effect was very small, given that other parameters remain constant. Bias was most sensitive to changes in the probability of disease. Bias decreased as disease risk increased, except for extremely large disease risks in which case the bias was largest but of the opposite sign. The absolute bias increased as either risk factor effect or intervention effect increased, but decreased as disease-specific mortality increased. The strongest bias resulted when disease risk and risk factor effect were both high. With all parameter variations, the absolute bias never exceeded 0.45% when at least 3 categories were used.

Conclusions: Categorizing continuously-valued risk factors in decision-analytic models has a negligible effect on model outcomes, remaining at less than 1% absolute bias even when only 2 categories are used.