Regarding Probabilistic Analysis and Computationally Expensive Models- Necessary and Required?
Jul 1, 2007, 00:00
10.1111/j.1524-4733.2007.00176.x
https://www.valueinhealthjournal.com/article/S1098-3015(10)60617-1/fulltext
Title :
Regarding Probabilistic Analysis and Computationally Expensive Models- Necessary and Required?
Citation :
https://www.valueinhealthjournal.com/action/showCitFormats?pii=S1098-3015(10)60617-1&doi=10.1111/j.1524-4733.2007.00176.x
First page :
Section Title :
Open access? :
No
Section Order :
11
To the Editor—In a recent article [1], the authors acknowledge that simulation (so-called “patient-level” simulations) can appropriately handle the realities required to correctly model most human disease processes but fret that “computational expense” will limit the provision of probabilistic sensitivity analyses (PSA) currently in vogue. While we recognize that full characterization of uncertainty is necessary to suitably inform decisions about the adoption of new health technologies, we do not share the authors’ concern. It is the case that a model that is built at the required depth [2] to reasonably inform decision-makers and deal with uncertainty (including structural), variability, and heterogeneity will involve a large number of computations and these can become a hindrance to conducting PSA. The solution does not lie, however, with accepting unnecessary compromises as the authors’ suggest. Although cohort Markov models may involve fewer calculations, they require gross oversimplifications making them rarely suitable for informing real decisions. Embarking on patches and partial, theoretically weak, “semi-Markov” constructs and emulators adds complexity without solving the problem. Instead, we need to build computationally efficient simulations and use well-established techniques [3–5] to reduce nuisance variance and thus the number of replications required.
The practical problems of increased computation pale besides those caused by the significant untenable assumptions often required in a cohort model. For example, age is usually an important determinant of the course of illness and, in Markov processes, the “mean” age assigned to the subcohort in each state is increased typically by one time unit each cycle, flagrantly ignoring that since age affects the transition probabilities, the “mean” age will not increase in such a linear fashion. There are many more reasons than those noted by the authors for preferring individual simulation: it allows for proper correlation among characteristics—the well-known “clustering” of risk factors, for example [6,7]; it permits correct implementation of competing risks without recourse to such peculiar solutions as applying risk equations in random order [8]; it does not force the occurrence of only a single transition within an often lengthy cycle—people can develop an illness, be hospitalized, suffer a complication and die, all within a day, if appropriate; varying event sequences can be appropriately represented and medical decisions which are not probabilistic can be fully incorporated; physiologic parameters can change over time in clinically relevant individual manner; limitations in resources can be reflected directly as well as uncertainty in their use and costs, to name just a few; and all of this can be done without a massive proliferation of “states,” tunnel constructs, and outside-of-model programming.
Indeed, after building more than 100 Markov models, many with semi-Markov constructs, and now more than 30 simulations, we find the latter tool to be much easier to use, and the models are faster to construct and much more transparent when completed. This comes at very little, if any, increase in computation time. For example, our semi-Markov diabetes model took 2.6 min (in Fortran) to compute a cohort over a lifetime while our much more complex diabetes simulation takes 2.1 min to do the same for 1000 individual patients. Therefore, PSA can be done as easily in the complex simulation. One can create a model of type 2 diabetes that is “instantaneous,” but the result will reflect the disease and its management so poorly that it will not be useful to decision-makers.
If we as a profession aspire to inform real decision-makers and actually influence decisions, then we must see to it that our models are as realistic as required and that the information we provide is as detailed and rich as the questions that are posed to us. Persuading decision-makers and analysts to opt for inferior techniques does both them and our field a disservice. It is time for us to join the rest of the scientific world and embrace simulation as the standard modeling tool. —J. Jaime Caro, MDCM, FRCPC, FACP, Caro Research Institute, and Division of General Internal Medicine and Department of Epidemiology, Biostatistics and Occupational Medicine, McGill University, Montreal, PQ, Canada; Denis Getsios, BSc, Montreal, PQ, Canada; Jörgen Möller, BEng, Eslov, Sweden.
References
1 Griffin S, Klaxton K, Hawkins N, Sculpher M. Probabilistic analysis and computationally expensive models: necessary and required? Value Health 2006; 9:244–52.
2 Eddy DM. Accuracy vs transparency in pharmacoeconomic modeling: finding the right balance. Pharmacoeconomics 2006;24:837–44.
3 Koopman J. Modeling infection transmission. Annu Rev Public Health 2004;25:303–26.
4 Tafazzoli A, Roberts SD, Ness RM, Dittus RS. A comparison of screening methods for colorectal cancer using simulation modeling. In: Kuhl ME, Steiger NM, Armstrong FB, Jones JA, eds. Proceedings of the 2005 Winter Simulation Conference. Piscataway, NJ: IEEE, 2005; pp. 2236–45.
5 Barnes DJ, Hopkins TR. The impact of programming paradigms on the efficiency of an individual-based simulation model. Simul Model Pract Theory 2003;11:557–69.
6 Quaglini S, Stefanelli M, Boiocchi L, et al. Cardiovascular risk calculators: understanding differences and realising economic implications. Int J Med Inform 2005;74:191–9.
7 Aizawa Y, Kamimura N, Watanabe H, et al. Cardiovascular risk factors are really linked in the metabolicsyndrome: the phenomenon suggests clustering rather than coincidence. Int J Cardiol 2006;109:213–18.
8 Clarke PM, Gray AM, Briggs A, et al. A model to estimate the lifetime health outcomes of patients with type 2 diabetes: the United Kingdom Prospective Diabetes Study (UKPDS) Outcomes Model (UKPDS no. 68). Diabetologia 2004;47:1747–59.
Categories :
- Methodological & Statistical Research
- Missing Data
- Modeling and simulation
- PRO & Related Methods