Decision Analysis, Common Errors Made in Conducting

Comparators

Every decision analysis compares at least two options. If the decision is a clinical one (e.g., how should localized prostate cancer be treated?) all feasible and practical options should be considered. These might include doing nothing (or active surveillance), surgery, radiation, brachytherapy, or cryotherapy, and more. If the decision is a policy decision (say, whether a national human papillomavirus vaccination program should be funded), the same criteria apply: Feasible and practical options might include no vaccination, universal vaccination, vaccination targeted at high-risk groups, vaccination targeted at specific age groups, and more. Feasible and practical are clearly subject to interpretation, but the key ideas are that all options that stand a realistic chance of being implemented (feasibility) should be examined, given the resources available to address the problem (practicality).

The decision analysis neophyte often is reluctant to include many options because of concerns that the model will become unmanageably complex. As a result, many models consider only the two or three most intuitively attractive options. Options such as “do nothing” or “supportive care only,” or alternate frequencies or intensities of an intervention may be avoided. This is acceptable if the goal is to gain experience in modeling, but it is not acceptable if the goal is to choose the best therapeutic or policy option.

More advanced analysts may also inappropriately constrain the potential options considered. This may be because of a desire to adhere closely to the best quality evidence published in high-impact journals. Or it may be a strategic decision to put a new drug or device in the best possible light by choosing a plausible but weak comparator or by avoiding comparisons across types of interventions (e.g., comparing drugs only to drugs but not to surgery). Regardless of the reason, inappropriately constraining the set of comparators is a common and serious error in modeling.

Model Structure

Decision models represent potential outcomes of alternate strategies using models, which may be simple decision trees, discrete-time state-transition (i.e., Markov) models, discrete-event simulation models, or dynamic infectious disease models. Models may be simple or complex, but should correspond to an underlying theory or biological model of disease.