Decision SupportDecision-network polynomials and the sensitivity of decision-support models
Introduction
Decision trees, influence diagrams, Bayesian networks and event trees support the solution of decision analysis problems in several applications. Their use is nowadays facilitated by a number of software programs (see Bielza et al., 2011, Jensen and Nielsen, 2007).1 Computer implementation enables analysts to develop sophisticated codes that incorporate a variety of aspects of the problems under investigation (Dillon, Paté-Cornell, & Guikema, 2003). However, model complexity exposes analysts and decision-makers to the risk of a partial understanding of the model input–output response. Then, deriving insights about the structure of the model becomes essential to make robust conclusions and inferences.
Researchers have developed methods for exploring the informational content of decision-support models. For Bayesian networks several of the most recent findings rest on the fundamental result that the input-mapping in Bayesian networks is a multilinear polynomial (Castillo et al., 1996, Castillo et al., 1997). For instance, multilinearity is crucial for arithmetic and decision circuits (Bhattacharjya and Shachter, 2012, Darwiche, 2003). However, there are unique challenges in sensitivity analysis for influence diagrams due to the non-linearities created by maximization operations for making decisions (Bhattacharjya & Shachter, 2010, p. 1). Indeed, this issue is transversal to all decision-support models that include a maximization (or minimization) operator. The maximization operation induces piecewise-definiteness and impairs differentiation. Thus, properties of Bayesian networks cannot be transferred directly to models such as influence diagrams, decision trees and decision circuits. The difficulties associated with differentiation then raise broader issues concerning the methodology for deriving managerial insights from decision-support models.
In this work, we conduct a systematic investigation into the mathematical properties of decision-support models. We show that the decision-theoretical principles underlying their construction (Savage, 1954) lead to a piecewise defined input–output mapping. Each piece corresponds to an available strategy and is multilinear in probabilities and utilities. We call the mapping a decision-network polynomial.
Non-differentiability occurs at those values of the model inputs (probabilities/utilities) for which the decision-maker is indifferent among alternative strategies. Using the terminology of Howard (1968) (see also Bhattacharjya & Shachter, 2010),2 this result suggests that differentiation finds its natural application in an open-loop analysis, i.e., when the preferred strategy (or any other strategy) is under scrutiny. In a closed-loop analysis differentiation might not be possible. We then employ sensitivity measures based on an orthogonal decomposition via finite difference operators (Borgonovo, 2010), that do not require differentiability. These finite change sensitivity indices allow us to obtain a clear understanding of the model response, by apportioning the change in expected utility to the individual effects and interactions among the model inputs.
The next step is the analysis of the model structure when consequences are monetary and certainty equivalents are the output of interest. Findings show that the input–output mapping remains piecewise-defined but each piece is composite multilinear. We link high order derivatives of certainty equivalents polynomials analytically to the derivatives of the corresponding decision-network polynomials. Finite change sensitivity indices acquire a direct interpretation as the monetary gain (loss) associated with the model input variations.
The well known case-study of Felli and Hazen (2004) helps us illustrating how the approach can be used for: (i) understanding direction of change, (ii) quantifying the relevance of interactions; and (iii) identifying the model inputs on which to focus managerial attention during implementation (Eschenbach, 1992, pp. 40-41).
The remainder of the paper is organized as follows. Section 2 offers a literature review. Section 3 presents decision-network polynomials. Section 4 investigates the link between indifference and differentiation. Section 5 discusses the presence of interactions and specializes finite change sensitivity indices to the case of decision-network polynomials. Section 6 discusses results for certainty equivalents. Section 7 illustrates the derivation of decision-making insights through a case study in the medical sector. Section 8 offers conclusions. All proofs are in Appendix A.
Section snippets
Review and taxonomy
This section offers a concise overview of relevant literature on decision support models and their sensitivity analysis. For a broad overview of these models we refer to Bielza et al. (2011).
Bayesian networks are among the most widely used decision-support models for the factorization of probability distributions. Their applications range from reliability analysis to genetics (Darwiche, 2010, Jensen and Nielsen, 2007). Technically, a Bayesian network is a directed acyclic graph , where N
Decision-network polynomials
The purpose of this section is to obtain the formal representation of the input–output mapping in decision-support models starting from the theoretical principles of expected utility theory. In fact, decision support models are developed consistently with the normative perspective of decision-making, namely, the expected utility model of von Neumann and Morgenstern (1944) and the subjective expected utility model of Savage (1954) (Smith & von Winterfeldt, 2004, p. 561).
In this respect, we
Piecewise-definiteness, differentiability and indifference among strategies
Piecewise-definiteness represents the main departure of decision-network polynomials from Bayesian-network polynomials. The differentiability of a piecewise-defined function is a subtle matter. A general result is presented in Proposition 1 of Borgonovo and Peccati (2010) states that not only the regularity of the pieces, but also by the order of their contact along the frontier of each admissible domain determines differentiability. For decision-network polynomials, the following holds. Proposition 2 Let
Interactions and high-order derivatives
In the remainder of this work, we consider one of the most typical sensitivity analysis exercises. We have a decision support model , where , and we select two points of interest in and . One usually calls and base case and sensitivity case, respectively. We write , and . Then, we evaluate the model at and computing the change .
For decision-support models, three results are possible:
- (a)
if ;
- (b)
Certainty equivalents
In several applications, consequences are monetary payoffs and a utility function u over monetary consequences might be assessed. With a slight abuse of notation, let be the monetary payoff of the corresponding consequence and set the consequent utility value. In order to express the decision-maker’s risk attitude, without loss of generality, we can assume monotone and convex (or concave). An important feature of this representation is the possibility of utilizing
Managerial insights through a case study in the medical sector
In this section, we discuss the derivation of managerial insights through a well-known case study. The case-study is presented in Felli and Hazen (2004), and refers to the giant cell arthritis problem (GCA) of Buchbinder and Detsky (1992), which describes a realistic-size application (Felli & Hazen, 2004, p. 101). We refer to Buchbinder and Detsky, 1992, Felli and Hazen, 2004 for a thorough overview. The treatment of the GCA problem foresees a prednisone therapy, which may lead to collateral
Conclusions
We have derived the representation of the input–output mapping in decision-support models from the principles of expected utility theory. We have called this representation a decision-network polynomial. We have seen that it is a piecewise defined function of probabilities and utilities. Each piece is multilinear and infinitely many times differentiable at points interior to an admissible domain. Differentiability is lost at indifference points. At these points, directional derivatives exist,
References (48)
- et al.
A review of representation issues and modeling challenges with influence diagrams
Omega
(2011) Sensitivity analysis with finite changes: An application to modified eoq models
European Journal of Operational Research
(2010)- et al.
A distance measure for bounding probabilistic belief change
International Journal of Approximate Reasoning
(2005) - et al.
Decision making with hybrid influence diagrams using mixtures of truncated exponentials
European Journal of Operational Research
(2008) - et al.
Sequential valuation networks for asymmetric decision problems
European Journal of Operational Research
(2006) - et al.
Scenario-based portfolio selection of investment projects with incomplete probability and utility information
European Journal of Operational Research
(2012) Mathematical foundations of event trees
Reliability Engineering & System Safety
(1998)- et al.
A differential semantics for jointree algorithms
Artificial Intelligence
(2004) - et al.
Sensitivity analysis in decision circuits
- et al.
Three new sensitivity analysis methods for influence diagrams