OBJECTIVE: Markov models are often used as part of cost-effectiveness analysis to extrapolate from short-term experimental evidence. Most Markov models used in economic evaluation typically assume fixed transition probabilities with respect to time. This is because implementation of time-dependant probabilities for all states is difficult in widely used modeling software. This is a limiting assumption when transitions are clearly time dependent. In such circumstances, semi-Markov processes, with time-dependent probabilities, may be necessary to provide reliable estimates of cost-effectiveness. METHODS: The implementation of a semi-Markov process will be illustrated using a recent analysis of anti-epileptic drug sequences for the National Institute for Clinical Excellence. In this case study, the probability of treatment failure was dependant on the time spent on the current treatment. Although such models can be implemented in Excel or other modeling applications, the large number of states required makes this cumbersome. As an alternative, the model was realized using 'R', a statistical programming language, using a three-dimensional transition matrix, where the third dimension represents time spent in the current state. The ability of R to manipulate n-dimensional numeric arrays allowed the complex model to be easily implemented. RESULTS: The use of a semi-Markov process to model cost-effectiveness in epilepsy allowed the reported natural history of the condition to be accurately reflected. This was achieved efficiently and transparently using the R statistical programming language. Furthermore, the alternative (and commonly used) assumption of fixed transition probabilities with respect to time generated important differences in cost-effectiveness results compared to the semi-Markov process. CONCLUSIONS: Semi-Markov process models may be useful in modeling a wide range of treatment processes. By adding further dimensions to the transition matrix, the transition probabilities could be made dependant on other aspects of patients' history providing a useful alternative to discrete event simulation, where increased speed of execution will aid probabilistic modeling.