VoxEU Column Health Economics

The value of variation in clinical practice under uncertainty

Clinical practice guidelines recommend treating all patients with similar attributes the same way. This column argues that, under conditions of uncertainty or ambiguity, this may be bad advice. Treating similar patients differently provides two benefits. The first is diversification – assigning similar patients to different treatments limits the consequences of choosing an inappropriate treatment. The second benefit is that randomly assigning treatments helps clinicians learn which ones are most effective.

Commentators on delivery of healthcare regularly decry ‘unwarranted’ or ‘inappropriate’ variation in clinical practice and recommend uniform treatment of observationally similar patients. This recommendation, often expressed in clinical practice guidelines, is well-motivated if knowledge of treatment response is strong enough to determine optimal treatments. However, much patient care is delivered with substantial uncertainty. Considering patient care under uncertainty, this column uses basic ideas of decision theory to explain why uniform treatment of similar patients may be counterproductive.

Foundations for recommendations regarding variation in clinical practice

There have been two prominent recommendations regarding variation in clinical practice. One calls for systematically different treatment of patients who differ in observed attributes. The other calls for uniform treatment of observationally similar patients.

The recommendation for differentiation of patients by observed attributes stems from a longstanding recognition that evidence on patient medical history, symptoms, and test results may help to diagnose illness and predict treatment response. Moreover, health professionals increasingly emphasise the relevance to treatment choice of genetic variability and patient preferences. Writers focusing on genetic variability often use the term ‘personalised’ medicine, while those focusing on patient preference use the term ‘patient-centred’ medicine (Hamburg and Collins 2010, Bardes 2012).

The recommendation for uniform treatment of similar patients stems from a conviction that variation in treatment of such patients indicates a deficiency in clinical practice. Wennberg (2011: 687) defines ‘unwarranted’ variation in care as variation that: “isn’t explained by illness or patient preference.” The UK National Health Service (2011) gives its Atlas of Variation in Healthcare the subtitle Reducing unwarranted variation to increase value and improve quality. A report on clinical practice guidelines by the US Institute of Medicine (2011: 26) states: “Trustworthy CPGs have the potential to reduce inappropriate practice variation.”

While medical discussions of the two recommendations largely take them to be self-evident, economists who study patient care as an optimisation problem have provided an analytical foundation in settings where knowledge of treatment response suffices to determine optimal treatments (Phelps and Mushlin 1988, Meltzer 2001, Basu and Meltzer 2007, Manski 2013). These articles study optimal decision-making by a health planner who chooses treatments for a population of patients. They assume that treatment is individualistic and that a planner with an additive welfare function wants to optimise treatment.

The optimisation problem has a simple solution. The planner should divide patients into groups having the same observed attributes. He should assign all patients in a group to the treatment that yields the highest within-group mean welfare. Thus, it is optimal to differentially treat patients with different observed attributes if different treatments maximise their within-group mean welfare. Patients with similar attributes should be treated uniformly. These findings motivate the two recommendations.

Assessing the recommendations under uncertainty

Much healthcare is delivered with substantial uncertainty about treatment response. Considering treatment of cancer, Mullins et al. (2010: 59) observe that “there is considerable uncertainty surrounding the clinical benefits and harms associated with oncology treatments.” The Institute of Medicine (2011: 33) calls attention to the assertion in 1992 by the Evidence-Based Medicine Working Group that “clinicians must accept uncertainty and the notion that clinical decisions are often made with scant knowledge of their true impact.”

Are the prevalent recommendations regarding variation in clinical practice well-motivated when treatment decisions are made under uncertainty? To address the question, it is useful to consider how a health planner might choose treatments. Elementary decision theory shows that there is no uniquely optimal way to make decisions under uncertainty, but there are various reasonable ways. Bayesian decision theory suggests that the planner place a subjective probability distribution on the values of treatment response deemed possible and maximise subjective expected welfare. Research on decision-making under ambiguity studies ways of making reasonable choices in the absence of a subjective distribution on unknown quantities, developing such ideas as maximin and minimax-regret decision-making.

Assuming that the welfare function is additive and treatment is individualistic, in Manski (2009) I show that a planner who uses any of the above decision criteria would divide patients into groups having the same observed attributes and then apply the criterion to the patients in the group. The resulting treatment decision may vary with patient attributes. This finding justifies the recommendation to differentiate treatment of observationally different patients.

The remaining question is whether a planner using these decision criteria would adhere to the recommendation calling for uniform treatment of similar patients. There exist special cases in which he might do so, but the general answer is negative. A planner cannot systematically differentiate the treatment of observationally similar patients, but he can do so randomly. That is, the planner can specify treatment assignment probabilities in the manner of a randomised clinical trial and use these probabilities to choose treatments for individual patients. In Manski (2009), I show that there are two reasons to do so: diversification and learning. These provide complementary static and dynamic rationales to randomly vary treatment of observationally similar patients.

Diversification and learning

Financial diversification is a familiar recommendation for portfolio allocation. A portfolio is diversified if an investor allocates positive fractions of wealth to different investments. Analogously, treatment is diversified if a health planner randomly assigns similar patients to different treatments. In both cases, diversification enables a decision-maker facing uncertainty to limit the consequences of inadvertently making inferior choices.

Economists commonly use the expected utility criterion to motivate financial diversification. Consider an investor who places a subjective probability distribution on investment returns, whose utility function places decreasing marginal value on each additional dollar earned, and who maximises expected utility. A classical result is that such a risk-averse investor may choose a diversified portfolio. Analogously, a risk-averse health planner may diversify his treatment decisions.

In Manski (2009), I consider treatment choice under ambiguity, focusing on the minimax-regret criterion. The regret of a treatment allocation is the loss in welfare resulting from choosing this allocation rather than the best allocation. The best allocation yields zero regret, so a planner would like to minimise regret. However, a planner facing ambiguity does not know the best allocation. The minimax-regret criterion selects an allocation that minimises the maximum regret that could potentially materialise. I show that a planner who does not know which of two treatments is best and who uses the minimax-regret criterion always diversifies, assigning a positive fraction of patients to each treatment. The specific fraction assigned to each treatment depends on the available knowledge of treatment response.

Whereas diversification provides a static rationale for randomisation of treatment, learning provides a dynamic rationale. Consider a health planner who makes treatment decisions in a sequence of periods, facing new groups of patients each period. Such a planner may observe the outcomes of early treatment decisions and use this evidence to inform treatment choice later on. Diversification is advantageous for learning treatment response because it generates randomised experiments. As time passes and evidence accumulates, the planner can revise the fraction of patients assigned to each treatment in accordance with the available knowledge. I have called this idea ‘adaptive diversification’.

Thus, when a health planner makes decisions under uncertainty, it may be counterproductive to treat similar patients uniformly. In the short run, uniform treatment prevents society from limiting errors by diversifying treatment. In the long run, it prevents learning about treatment response because clinical practice produces no evidence about response to unused treatments (see Manski 2013 for a formal analysis of this problem). Treatment variation is valuable because it enables diversification and learning.

Implications for clinical practice guidelines

In decentralised healthcare systems, government health agencies and medical professional societies publish guidelines that aim to influence patient care. Clinical practice guidelines typically recommend uniform treatment of similar patients. Commentators exhort clinicians to adhere to such guidelines. Thus, clinical practice guidelines seek to have a decentralised healthcare system behave as if treatments were chosen by a planner who knows about treatment response. This perspective is evident in the definition of clinical practice guidelines given by the Institute of Medicine (2011: 4), which states: “Clinical practice guidelines are statements that include recommendations intended to optimise patient care that are informed by a systematic review of evidence and an assessment of the benefits and harms of alternative care options.” The report does not address how a clinician might reasonably behave if uncertainty about treatment response makes optimisation of care infeasible.

When treatment response is uncertain, I think it unwise for clinical practice guidelines to recommend uniform treatment of similar patients. Instead, they should help clinicians recognise that treatment choice may reasonably depend on how one interprets the available evidence and on the decision criterion that one uses. Rather than discourage variation in clinical practice, guidelines should encourage variation that appropriately diversifies treatment and that yields new evidence on treatment response.


Bardes, C (2012), “Defining ‘Patient-Centered Medicine’”, New England Journal of Medicine, 366: 782–783. 

Basu, A and D Meltzer (2007), “Value of information on preference heterogeneity and individualized care”, Medical Decision Making, 27: 112–127. 

Hamburg, M and F Collins (2010), “The Path to Personalized Medicine”, New England Journal of Medicine, 363: 301–304. 

Institute of Medicine (2011), Clinical Practice Guidelines We Can Trust, Washington, DC: National Academies Press.

Manski C (2009), “Diversified Treatment under Ambiguity”, International Economic Review, 50: 1013–1041.

Manski, C (2013), “‘Diagnostic Testing and Treatment under Ambiguity: Using Decision Analysis to Inform Clinical Practice”, Proceedings of the National Academy of Sciences, 110: 2064–2069.

Meltzer, D (2001), “Addressing Uncertainty in Medical Cost-Effectiveness: Implications of Expected Utility Maximization for Methods to Perform Sensitivity Analysis and the Use of Cost-Effectiveness Analysis to Set Priorities for Medical Research”, Journal of Health Economics, 20: 109–129.

Mullins, D, R Montgomery, and S Tunis (2010), “Uncertainty in Assessing Value of Oncology Treatments”, The Oncologist, 15(supplement 1): 58–64. 

National Health Service (2011), The NHS Atlas of Variation in Healthcare

Phelps, C and A Mushlin (1988), “Focusing technology assessment using medical decision theory”, Medical Decision Making, 8: 279–289. 

Wennberg, J (2011), “Time to Tackle Unwarranted Variations in Practice”, BMJ, 342: 687–690.

1,470 Reads