This presentation is part of: Poster Session - CEA: Methods and Applications; Health Services Research
PROBABILISTIC ANALYSIS AND COMPUTATIONALLY EXPENSIVE MODELS: NECESSARY AND REQUIRED?
Susan Griffin, MSc, BSc, Karl Claxton, PhD, Neil Hawkins, MSc, and Mark Sculpher, PhD. University of York, Centre for Health Economics, York, United Kingdom
Purpose: To demonstrate the necessity of characterising uncertainty in second-order non-linear models with computationally expensive model structures and patient-level simulation. Methods: The recent methodological guidance on the assessment of health technologies issued by the National Institute of Clinical Excellence (NICE) requires probabilistic sensitivity analysis (PSA) as the appropriate way to quantify decision uncertainty. This requirement may be computationally expensive in more complex models structures, particularly those that employ patient-level simulation. However, where models are second-order non-linear, accounting for second-order uncertainty is required to estimate expected costs and effects as well as decision uncertainty. A review of published NICE technology appraisals identified all cases in which the economic evaluation was undertaken using patient-level simulation. Each case was examined to see if the model structure implied second-order non-linearity, whether PSA was performed, and, where this was not the case, whether alternative modelling approaches were available such as less computationally expensive model structures or the use of emulators. Results: All of the model structures examined implied second-order non-linearity. The majority of cases did not attempt to account for second-order uncertainty by performing PSA. The reasons cited included the computational expense of using patient-level simulation, and the complexity of the model structure exceeding current availability of data. However, the possibility of implementing less computationally expensive models was demonstrated in two areas: i) two appraisals examined pharmacotherapy for epilepsy, one employed patient-level simulation without PSA due to computational expense, the other used an equivalent semi-Markov model structure where PSA was feasible; ii) one appraisal of treatments for osteoporosis used emulators (Gaussian processes) to address both first- and second-order uncertainty in a complex model structure employing patient-level simulation. In a number of other cases, the use of computational expensive methods was due to modelling treatment switching, rather than evaluating the full range of possible clinical strategies, which may have been more appropriate. Conclusions: Where model structures exhibit second-order non-linearity, incorporating second-order uncertainty is required for unbiased estimates of expected cost and effect, as well as to characterise decision uncertainty. Given that PSA is necessary and required, the computational expense of alternative model structures should be considered. Where complex and expensive model structures are required further work is needed to identify suitable computational methods and emulators.
