The time horizons of formal decision analyses

Clinical decision analyses use time horizons that vary from hours to the patient's entire life. Analyses of decisions with a lifetime horizon commonly use Markov models, which simulate the patient's lifespan by dividing it into equal periods (cycles). At each cycle, the model exposes a hypothetical cohort to the competing hazards of normal aging and of the disease in question (disease-specific hazards), and the results are presented as years of life expectancy. This paper highlights two limitations of lifetime Markov models that have been ignored in recent publications. First, since there are no readily available data on changes in disease-specific hazards over time, these hazards are often derived from short-term follow-up studies, and assumed to be constant over the patient's entire life. Second, results may be better presented in terms of health states (i.e. proportions of patients expected to recover completely, recover with a disability or die) rather than life expectancy. Although well-known, these two limitations require re-emphasis. They may be avoided by restricting the time horizon of decision analyses and presenting results as health states as well as life expectancies. When a lifetime horizon is necessary, the performance of Markov models may be improved by the using of time-variant disease-specific hazards derived from long-term follow-up studies, or from theoretical models that simulate more closely the disease progression over time, rather than assuming constant disease-specific hazards.