Simulactio

CURRENT OPINION

Pharmacoeconomics 2005; 23 (4): 323-332 1170-7690/05/0004-0323/$34.95/0  2005 Adis Data Information BV. All rights reserved.

Pharmacoeconomic Analyses Using Discrete Event Simulation
J. Jaime Caro
Caro Research Institute, Boston, Massachusetts, USA

Abstract

To date, decision trees and Markov models have been the most common methods used in pharmacoeconomicevaluations. Both of these techniques lack the flexibility required to appropriately represent clinical reality. In this paper an alternative, more natural, way to model clinical reality – discrete event simulation – is presented and its application is illustrated with a real world example. A discrete event simulation represents the course of disease very naturally, with few restrictions. Neither mutuallyexclusive branches nor states are required, nor is a fixed cycle. All relevant aspects can be incorporated explicitly and efficiently. Flexibility in handling perspectives and carrying out sensitivity analyses, including structural variations, is incorporated and the entire model can be presented very transparently. The main limitations are imposed by lack of data to fit realistic models. Discreteevent simulation, though rarely employed in pharmacoeconomics today, should be strongly considered when carrying out economic evaluations, particularly those aimed at informing policy makers and at estimating the budget impact of a pharmaceutical intervention.

It is by now well accepted that pharmacoeconomic analysis[1,2] and the setting of health policy[3] require a model of the disease andits management.[4,5] In exceptional circumstances, a single study – a randomised clinical trial, for example – may provide all of the necessary information on costs and outcomes. Even then, a model is indispensable to address the economic impact of the intervention in actual practice[6] because, at a minimum, the experimental data must be applied to a realworld setting in order to reflect clinicalreality.[7] These economic models have typically been structured and analysed using decision trees.[8] This technique has been very successfully applied despite recognition a decade ago of the severe limitations of this approach when applied to medical problems.[9] In particular, decision trees impose a rigid structure

based on mutually exclusive ‘outcomes’, they do not explicitly considertime, and they are very inefficient because every analysis requires computing all possible pathways (the ‘branches’), sometimes multiple times. Thus, an alternative approach – the Markov model – has been increasingly employed.[10] Instead of basing the model on mutually exclusive outcomes as the decision tree does, a Markov model represents the course of a disease in terms of mutually exclusive‘health states’ and the transitions among them.[11] While this technique considers time more explicitly and can be analysed very efficiently, it retains some structural rigidity that can make appropriate representation of clinical reality difficult. The requirement that all the aspects of the disease be denoted by a ‘state’ in a Markov model forces the

324

Caro

analyst to render features thatare naturally continuous as discrete, such as weight gained (yes or no rather than the actual amount of weight gain), severity in specific categories, and so on. To begin to approximate their continuous nature, an unwieldy number of states have to be specified. For example, nearly 20 states are required just to reflect weight changes of ±40 pounds in increments of 5 pounds. This problem iscompounded if the implications of the state change over time, as each instance then generates a new state. For example, the first month after a hospitalisation for psychiatric illness would be one state, the second month another, and so forth. A similar proliferation is imposed if the subsequent course of the disease depends on previous history, such as when the risk of adverse events depends on prior...