OBJECTIVES: Patient characteristics (age, sex, smoking status, baseline FEV-1 percent predicted) can have important consequences for the prognosis of Chronic Obstructive Pulmonary Disease (COPD). A Markov model for COPD was developed that allowed different patient prognoses and consideration was given to how such patient heterogeneity be analysed and presented. METHODS: A four state Markov model of COPD progression (mild, moderate, severe COPD and a dead state) was structured using the American Thoracic Society's FEV-1 thresholds for the definition of disease. Time to progression through the states was modelled as a function of age, sex, baseline FEV-1 and smoking status. Frequency of disease exacerbations was modelled as a function of the disease state. Utility values for the health states were taken from the literature and costs were estimated from the literature and expert opinion. Treatment effects were estimated from emerging clinical trial data. Lifetime cost and QALY outcomes were predicted from the model for 2068 subjects for whom information on prognostic factors was available. RESULTS: Evaluating the model at the mean of the prognostic factors for the population of interest gave costs of $41,000 and $66,000 for the control and treatment groups. The corresponding QALY estimates were 2.6 and 3.5 leading to an estimated ICER of $27,000 per QALY gained. However, averaging across the 2068 individual estimates yielded $43,000 and $72,000 for costs and 3.7 and 4.6 for QALYs in the control and treatment groups respectively, generating an ICER estimate of $32,000 per QALY. CONCLUSIONS: These results clearly demonstrate how, in the presence of heterogeneity, evaluating models at the average values of important prognostic factors can lead to serious bias compared to averaging over individual-based predictions. This bias is due to the non-linearities inherent in most Markov models and is exacerbated once uncertainty in parameter estimation is included in a fully probabilistic framework.