john wiley & sons, ltd., the atrium, southern gate, chichester p019 8sq ...

JOHN WILEY & SONS, LTD., THE ATRIUM, SOUTHERN GATE, CHICHESTER P019 8SQ, UK *** PROOF OF YOUR ARTICLE ATTACHED, PLEASE READ CAREFULLY ***

After receipt of your corrections your article will be published initially within the online version of the journal. PLEASE NOTE THAT THE PROMPT RETURN OF YOUR PROOF CORRECTIONS WILL ENSURE THAT THERE ARE NO UNNECESSARY DELAYS IN THE PUBLICATION OF YOUR ARTICLE READ PROOFS CAREFULLY ONCE PUBLISHED ONLINE OR IN PRINT IT IS NOT POSSIBLE TO MAKE ANY FURTHER CORRECTIONS TO YOUR ARTICLE • •

This will be your only chance to correct your proof Please note that the volume and page numbers shown on the proofs are for position only

ANSWER ALL QUERIES ON PROOFS (Queries are attached as the last page of your proof.) •

Please annotate this file electronically and return by email to the production contact as detailed in the covering email. Guidelines on using the electronic annotation tools can be found at the end of the proof. If you are unable to correct your proof using electronic annotation, please list all corrections and send back via email to the address in the covering email, or mark all corrections directly on the proofs and send the scanned copy via email. Please do not send corrections by fax or post. Acrobat Reader & Acrobat Professional

•

You will only be able to annotate the file using Acrobat Reader 8.0 or above and Acrobat Professional. Acrobat Reader can be downloaded free of charge at the following address: http://www.adobe.com/products/acrobat/readstep2.html

CHECK FIGURES AND TABLES CAREFULLY • • • •

Check sizes, numbering, and orientation of figures All images in the PDF are downsampled (reduced to lower resolution and file size) to facilitate Internet delivery. These images will appear at higher resolution and sharpness in the printed article Review figure legends to ensure that they are complete Check all tables. Review layout, titles, and footnotes

COMPLETE COPYRIGHT TRANSFER AGREEMENT (CTA) if you have not already signed one •

Please send a scanned signed copy with your proofs by e-mail. Your article cannot be published unless we have received the signed CTA

AUTHOR SERVICES •

If you have registered this article in Wiley-Blackwell Author Services, the article’s status will be updated shortly after you have returned your proof corrections (you will also receive an e-mail alert if you have opted to receive them). You are entitled to free access to the PDF from Author Services when your article is published online. This free access is considered your PDF offprint, and you will only have access from within Author Services; you will not be sent a PDF. You may also nominate up to 10 colleagues for free access. All accesses from Author Services count towards the usage of your article. For options to order print copies or additional electronic access, please see below.

OFFPRINTS •

Free access to the final PDF offprint of your article will be available via Author Services only. Please therefore sign up for Author Services if you would like to access your article PDF offprint and enjoy the many other benefits the service offers.

Additional reprint and journal issue purchases • • •

Should you wish to purchase additional copies of your article, please click on the link and follow the instructions provided: http://offprint.cosprinters.com/cos/bw/ Corresponding authors are invited to inform their co-authors of the reprint options available. Please note that regardless of the form in which they are acquired, reprints should not be resold, nor further disseminated in electronic or print form, nor deployed in part or in whole in any marketing, promotional or educational contexts without authorization from Wiley. Permissions requests should be directed to mailto: [email protected]

Research Article Received 1 February 2010,

Revised 16 February 2011,

Accepted 16 February 2011

Published online in Wiley Online Library

1

(wileyonlinelibrary.com) DOI: 10.1002/asmb.888

3

Modeling and validating stakeholder preferences with probabilistic inversion R. E. J. Nesloa and R. M. Cookea,b ∗†

9 11 13

F

OO

7

Despite the vast number of models that have been developed for analyzing stakeholders’ preferences, it is difficult to find any true out-of-sample validation for these models. Based on the theory of rational preference, utilities are specific to the individual. Unlike subjective probability, there is no mechanism for changing utilities on the basis of observation, and no operation for getting people’s utilities to converge. The proper goal of stakeholder preference modeling must therefore be the characterization of a population of stakeholders via a distribution over utility functions. Drawing on the theory of discrete choice and random utility theory, we apply probabilistic inversion methods to derive a distribution over utility functions. The utility functions may either attach to the choice alternatives directly, or may be functions of physical attributes. Because the utilities are inferred from discrete choice data, out-of-sample validation is enabled by splitting the data into a test set used to fit the model and a validation set. These techniques are illustrated using discrete choice data for the valuation of health states. Copyright © 2011 John Wiley & Sons, Ltd.

PR

5

21 23 25 27 29 31 33 35

TE

EC

19

Savage [1] formulated axioms for rational preference of an individual, and showed that the preferences of a rational agent can be represented as the expected utility, where the individual’s subjective probability over possible states of the world is unique and the utility function over consequences is affine unique. The representation of a preference relation by a utility function implies that if an individual assigns higher (expected) utility to choice alternative a than to alternative b, then the individual prefers, selects or orders an alternative a above b. The theory of subjective probability has flourished; it is fair to say that the modeling of utility has lagged behind. Without reviewing the activity in this area, suffice to say that, in our opinion, there are two major causes for this. First, modeling techniques, such as multi-attribute utility theory (MAUT), multi-criteria decision making (MCDM), analytical hierarchy process (AHP), etc. focus on capturing ‘the’ utility function over choice alternatives. If an individual’s utility function conformed to additional (rather severe) constraints, such a representation might be possible at an individual level. However, there is no reason to believe that ‘a’ utility function exists for groups of individuals; the search for such is a fool’s errand. Based on the theory of rational decision, the proper goal of utility modeling should be to capture the distribution over utility functions characterizing a group. Second, and perhaps not wholly unrelated, there has been a near total absence of attempts to validate the utility models produced by these various methods. As such, the field of applied utility theory remains parochial. There is a wealth of literature demonstrating that stakeholders often violate the axioms of rational preference. This does not threaten the normative status of the theory any more than the prevalence of invalid inferences imperils logic. However, it does lend urgency to the issue of validation. The prospects for utility theory are brighter within the literature of discrete choice and random utility theory. Thurstone [2] pioneered this field with his celebrated law of comparative judgment. Assuming that utilities are normally distributed over a population of stakeholders, he fits the parameters of this distribution, under various correlation assumptions, using discrete choice data from pairwise comparisons. Later, the logit model was derived by Luce [3--5] under one of the consequences of his choice axioms namely the Independence of Irrelevant Alternatives (IIA). McFadden [6] also derived the logit model

CO RR

17

1. Introduction

UN

15

D

Keywords: out-of-sample validation; discrete choice; random utility; probabilistic inversion; valuation of health states

a Department

of Mathematics, Delft University of Technology, The Netherlands

b Resources for the Future, Washington, DC, U.S.A. ∗ Correspondence to: R. M. Cooke, Department of Mathematics,

Delft University of Technology, The Netherlands.

1

† E-mail:

[email protected]

Copyright © 2011 John Wiley & Sons, Ltd.

Appl. Stochastic Models Bus. Ind. 2011

R. E. J. NESLO AND R. M. COOKE

5 7 9 11 13 15

F

3

under the random utility maximization principle. The logit models assume that the error terms are generalized extreme value (GEV) distributed with mean zero and some constrained covariance matrix. Unhappy with the IIA assumption, McFadden and Train [7] picked up the thread of random utility maximization and extended the standard version of the logit and probit models, to deal with the limitations they pose, whereas goodness-of-fit tests have been developed for many random utility models, true out-of-sample validation is not part of the standard operating procedure. The problem of inferring a distribution over utility functions from discrete choice data is a problem of probabilistic inversion. The theory tells us that each (rational) stakeholder has a utility function; if we knew these utility functions for a group of stakeholders we could predict the distribution of responses in discrete choice situations. We observe the distribution of responses and wish to infer the distribution over utility functions. Even more, we wish to model these utility functions as functions of physical attributes of the choice alternatives, and we wish to infer rather than impose dependence relations between utilities. This program is quite feasible, albeit that the techniques for solving probabilistic inversion (PI) problems are new to this field. A few applications are in press or have been published [8--11]. This paper introduces PI methods within the random utility community. The following section provides an intuitive introduction, followed in Section 3 by an example on stakeholders’ preferences and PI. An recent application on valuing health states is used to illustrate these methods. The application was designed and carried out by Dr V. Flari.‡

OO

1

25 27 29 31 33 35 37

D

TE

23

EC

21

CO RR

19

The goal of stakeholder preference modeling is to derive a distribution over the utility functions on a set of A of choice alternatives characterizing a population S of stakeholders. To motivate the approach, we consider the health state valuation problem. Currently, health states are described and valued using EQ-5D§ EQ-5D is a standardized measure of health states developed by the EuroQol Group in order to provide a simple, generic measure of health for clinical and economic appraisal. Each health state is characterized by five criteria (mobility, self-care, usual activities, pain discomfort, and anxiety depression) which are measurable quantities that increase in a monotonic scale taking values one, two, three. An extended version (i.e. EQ-5D+C; see Table AI in the Appendix) of the system was introduced by Stouthard [12], which we will use. These different health states are possible outcomes of therapeutic procedures, and their valuation is critical in deciding which procedures to support and supply. With six criteria taking three possible values, there are 36 = 729 possible health states. The most direct approach would be to ask a group of stakeholders randomly chosen from the target population, to state their utility values for each health state. To estimate the distribution of utility functions, each stakeholder should apply a standardized utility scale with common zero and unit. The assessment burden for the stakeholders would be forbidding. Although the distribution over utilities of these health states is the immediate goal; we want also to model the utility of health states in terms of the attribute scores. The existing approaches will often ask stakeholders to value the attributes and the attribute scores. This however is problematic for a number of reasons: (1) Whereas we choose health states in choosing a therapeutic procedure, we do not choose attributes as such. (2) The value attached to one attribute (eg ‘mobility’) will depend on the whole set of attributes, as we must know what exactly falls under ‘usual activities’ and ‘self-care’ to avoid double counting. Of course, valuing mobility score 3 versus mobility score 2 assumes that these values are unaffected by the values of other attributes. (3) It is unclear how the resulting distribution over health state utilities would be validated. By adopting a simple model of the utility of a health state in terms of its attribute scores, we can simultaneously lighten the assessment burden and enable validation of the model. The score (utility) of health state i for subject s is modeled as

UN

17

PR

2. Intuitive motivation

39

41 43

u s (ai ) =

6

j =1

s, j ×ci, j ,

6 j =1

s, j = 1, s, j > 0,

(1)

where s, j is the weight for attribute j for subject s and ci, j is the score of health state i on attribute j . If this model is adequate, the distribution of utility functions over the set of stakeholders may be captured as a distribution over attribute weights (1 , . . . , 6 ). If the model is not adequate, a better model must be sought. Instead of asking each stakeholder for his/her weight vector, we will ask them to rank order subsets of the 729 health states, and look for a distribution over weight vectors which recovers the pattern of rankings. Given a single ranking of a set of health states, we could in principle recover

2

‡ Dr

V. A. Flari, Policy and Regulation Programme B, Food and Environment Research Agency, Sand Hutton, York, YO41 1LZ. information can be found in http://www.euroqol.org.

§ Further




Figure 1.

13 15 17 19 21 23 25 27 29 31 33 35 37 39 41

F

OO

PR

11

D

9

TE

7

EC

5

CO RR

3

the set of weight vectors which would yield this ranking under model (1). For a set of rankings, we could take the union of the corresponding sets of weight vectors. The uniform distribution over this set could be taken as our distribution over utility functions. Although this is a feasible approach, it is not the one we adopt. To understand why, we must discuss the discrete choice format. Seven hundred and twenty-nine health states is too many for stakeholders to rank order. A discrete choice format is needed rendering the assessment burden bearable. Best practice suggests that stakeholders can rank at most seven items at a time. The most popular discrete choice format is simple pairwise comparisons: subjects are presented with all pairs of choice alternatives and asked to choose one of the two offered. This is obviously infeasible for 729 alternatives, as there are 265 356 pairs. An economical discrete choice format is acceptable, if it enables validation of the model on which it is based. For the present study, we selected 17 non-dominated health states¶. These 17 health states are broken into five overlapping groups of five, where stakeholders then rank the health states in each group. The overlap structure is shown in Figure 1. Some pairs of states occur in two groups of five, which enable us to screen stakeholders for consistency. A stakeholder who ranks health state i above health state j in group k is called inconsistent if he ranks health state j above health state i in group k +1. With this discrete choice format we gather data which we then use to infer a distribution over the utility values. One of the goals in choosing a discrete choice format is to enable the analyst to identify inconsistent experts. Removing inconsistent experts can produce better results. Depending on the format, it may be unrealistic to expect perfect consistency. In pairwise comparisons, for example, it is usually sufficient that the number of inconsistencies (circular triads) is small enough to reject the hypothesis that a subject is choosing his/her preferences at random. Thus, a solution technique for finding a distribution over utility vectors must be able to deal with inconsistency: even if there is no distribution over utility vectors which exactly reproduces the stakeholders’ responses, it may be possible to minimize the lack of fit and thus arrive at a reasonably well-validated model. Choosing a discrete choice format mixes science and craft, and as this approach to utility quantification is relatively recent, the craft is still evolving. The necessity of dealing with inconsistencies drives our choice of the solution technique. If the problem is feasible, the solution algorithm should converge to a unique solution, if the problem is not feasible, then it should converge to a unique distribution which minimizes lack of fit in some appropriate sense. We proceed as follows. For each group, we define a square preference ranking matrix whose i, j -th entry gives the proportion of stakeholders who ranked health state i in the j -th position, in that group. The task is to find a distribution over the weights that reproduce the preference ranking matrices. Initially, a distribution is chosen over the weights from which a large sample is drawn. For each sample weight vector, we can compute how a stakeholder with that weight vector and model (1) would rank the health states in each five group. The entire sample will lead to five preference ranking matrices which will not agree with those from our stakeholder data. Iterative re-weighting schemes will then assign differential weights to the original sample such that if we resample this distribution using the weights, the resulting preference ranking matrices agree—as much as possible—with those of the stakeholders. If the problem is feasible, then perfect agreement is possible and these techniques converge quickly. The solution is unique and is minimally informative with respect to the initial distribution over the weights. In case the problem is infeasible, these techniques guarantee a solution that is the least infeasible in an appropriate sense. The solution techniques are explained in Section 4. Having obtained a solution using the solution techniques, we validate the solution by recovering the entries of the five preference ranking matrices using only a subset of the entries. In other words, we solve the model using a subset of entries and see if this model successfully predicts the remaining entries. We call this procedure out-of-sample validation. In the following sections, we give detailed information about the discrete choice format, solution techniques, and out-of-sample validation.

UN

1

Overlap structure.



3

attribute scores increase in severity: pain score 3 is worse than pain score 1, etc. Health state i dominates health state j if i s score on each criteria is greater than j s.

¶ The


15 17 19 21 23 25 27 29 31 33

Definition 1 Let X, Y be random vectors taking values in R N and R M , respectively. Further let G : RN → R M be a measurable function. X is called a probabilistic inverse of G at Y if G(X) ∼ Y,where ∼ means ‘has the same distribution as.’ If C is a set of random vectors taking values in R M , then X is an element of the probabilistic inverse of G at C if G(X) ∈ C. There are two main algorithms to carry out PI, namely IPF (Iterative Proportional Fitting) [14--16] and PARFUM (PARameter Fitting for Uncertainty Models) [17--19]. IPF was first described by Kruithof [14] and later rediscovered by Deming and Stephan [20], and many others. Csiszar [21] proved the convergence of IPF in the case of a feasible problem. He shows that if the IPF algorithm converges, then it converges to the unique distribution (called the I -projection) which is minimally informative relative of the starting distribution, within the set of distributions satisfying the marginal constraints. PARFUM was introduced and studied by Cooke [17]. If the problem is feasible, PARFUM converges to a solution which is distinct from the IPF solution. Unlike IPF, PARFUM always converges, and it converges to a solution which minimizes a suitable information functional [18]. The convergence of PARFUM (and its canonical variations) was proved by Matus [22] but has not yet been published. When the problem is feasible IPF is preferred, because of its faster convergence. PARFUM is used when the problem is infeasible, because it insures a solution such that the relative information [23] denoted by I (G(X)|Y) is minimal. The idea now is to infer a distribution over the utility values given the marginal distributions Q over the responses using PI. We denote the distribution over the utility values by P. For the diffuse initial distribution we take u 1 , u 2 , u 3 , u 4 to be independent and uniformly distributed on [0, 1]. This distribution would then yield the following distribution over responses see Table II. It does not comply with I. Note that we could use other distributions as the initial distribution.

UN

35

F

13

OO

11

PR

9

D

7

TE

5

Figure 2 illustrates a mental projection of a stakeholder s faced with two groups of choice alternatives, D1 and D2 . In this simple example, the stakeholder has to choose one most preferred alternative from each set. With the utilities as shown, s prefers alternative a2 from subset D1 and alternative a4 from subset D2 . Section A.1 in the Appendix gives a general formulation of the problem pictured in Figure 2. Another stakeholder might arrange the four alternatives differently on the utility scale, leading to other responses. In practice we cannot observe the utility values that stakeholders assign to the alternatives, but we do observe their preferences in terms of responses to the subsets presented. Suppose that a set of stakeholders induce marginal distributions Q shown in Table I. If we knew the utility function of a stakeholder s ∈ S, then we could obviously predict with certainty how s would respond to a discrete choice problem. Equivalently, if we are given a vector of utility values over the choice alternatives, we can uniquely determine how a stakeholder with that utility would respond in any discrete choice problem. Our problem is to infer a distribution over (standardized) utility functions given a set of responses from stakeholder population S. We solve this problem using a technique called ‘PI.’ PI is similar to ordinary inversion. In ordinary inversion there are quantities x, y, and a function g that maps x to y. The quantity y is observed and then the task is to find an x ∗ such that y = g(x ∗ ). In the probabilistic setup, the quantities x, y are random vectors instead of numbers. There are two formulations to solve a problem of PI namely, the measure theoretic approach and the random variable approach [13] from which we choose the latter.

EC

3

3. Stakeholders’ preference and PI

CO RR

1

Figure 2.

Response to D1 and D2 : u s = (u s (a1 ) , u s (a2 ) , u s (a3 ) , u s (a4 )) , G 1 (u s ) = a2 , G 2 (u s ) = a4 .

Table I. Marginal distribution Q over the responses.

4

D1 D2 Copyright © 2011 John Wiley & Sons, Ltd.

a1

a2

a3

a4

0.5 N/A

0.3 0.25

0.2 0.5

N/A 0.25 Appl. Stochastic Models Bus. Ind. 2011


Table II. Marginal distribution Q over the responses given uniform distribution. D1 D2

a1

a2

a3

a4

1/3 N/A

1/3 1/3

1/3 1/3

N/A 1/3

Table III. Ten input and output samples. u3

u4

G1

G2

0.6047 0.20205 0.11747 0.1898 0.86156 0.14059 0.29232 0.87431 0.36005 0.31374

0.13987 0.34152 0.78388 0.25268 0.89332 0.09522 0.20926 0.65964 0.08888 0.82145

0.8202 0.47065 0.81113 0.04756 0.20379 0.30707 0.73012 0.42908 0.12888 0.00599

0.39849 0.88651 0.36028 0.84957 0.25159 0.28061 0.36256 0.31673 0.03023 0.59636

a3 a3 a3 a2 a2 a3 a3 a1 a1 a2

a3 a4 a3 a4 a2 a3 a3 a2 a3 a2

F

u2

OO

1 2 3 4 5 6 7 8 9 10

u1

PR

Sample

Table IV. Marginal distribution Q over the responses given P0 .

7 9 11 13 15 17 19

21

TE

which are presented in Table IV. Evidently, these probabilities do not comply with the target probabilities in I. An I -projection is an operation that adjusts a joint distribution such that it meets a given marginal constraint. The IPF procedure successively I -projects onto each marginal, and repeats until convergence is reached. If the problem is feasible IPF converges to a P∗ which satisfies all constraints and is minimally informative with respect to the starting distribution P0. PARFUM on the other hand averages the I -projections of each marginal to obtain the next iterate. The I -projection of a distribution P0 onto the margins of G 1 is defined as follows, where Q k,i is the (k, i) entry of Table I.

with

IG 1 (P0 ) = IG 1 =a3 (IG 1 =a2 (IG 1 =a1 (P0 ))) ⎧ ⎪ ⎨ Pl0 ∗ ⎩

Pl0 ,

Q k,i P (G k = ai ) 0

, G k (l) = ai ,

(2)

(3)

G k (l) = ai .

The I -projection of P0 onto the margins of G 2 is computed in the same way as G 1 , but P0 replaced by IG 1 (Pl0 ). Table V shows how an I -projection of P0 onto the margins of G 1 is computed. Note that only the third column is the weight that sum to 1, as (2) requires cycling through all values of G i . Copyright © 2011 John Wiley & Sons, Ltd.


5

25

N/A 0.2

P0 (G 1 = a1), P0 (G 1 = a2 ), P0 (G 1 = a3 ), P0 (G 2 = a2 ), P0 (G 2 = a3 ), P0 (G 2 = a4 )

IG k =ai (Pl0 ) = ⎪ 23

0.5 0.5

EC

5

0.3 0.3

a4

CO RR

3

0.2 N/A

a3

With either IPF or PARFUM it is possible to find a distribution P over the utility values u 1 , u 2 , u 3 , u 4 such that this distribution complies with the distribution over the responses Q. As mentioned, IPF and PARFUM are sample-based algorithms for obtaining a probabilistic inverse. These algorithms re-weight samples from a starting distribution to get a distribution over the samples satisfying the constraints I. We first draw a number of samples from the starting distributions and then compute the responses to D1 , D2 for each sampled utility vector. These constitute the values of the functions G 1 , G 2 . For demonstration purposes we use 10 samples, but a far greater number of samples are needed in real applications. Table III shows 10 samples for u 1 , u 2 , u 3 , u 4 from a uniform distribution together with the computed outputs G 1 , G 2 . The samples (u 1 (l), u 2 (l), u 3 (l), u 4 (l), G 1 (l), G 2 (l)), on the l-th sample (l-th virtual stakeholder), are drawn from the starting joint distribution P0 . Each sample has probability 0.1 under P0 . The successive joint distributions obtained after m-th iterate of IPF or PARFUM will be denoted by Pm . Next, we compute the marginal distributions for outputs of G 1 and G 2

UN

1

a2

D

D1 D2

a1


Table V. I -projection of the marginal of G 1 ; I1,2 ◦ I1,1

denotes IG 1 =a2 ◦ IG 1 =a1 (P0 ), etc. IG 1 =a1 (P0 )

I1,2 ◦ I1,1

I1,3 ◦ I1,2 ◦ I1,1

0.1

0.1

0.1

0.1

0.2 ∗0.1=0.04 0.5 0.2 ∗0.1=0.04 0.5 0.2 ∗0.1=0.04 0.5

0.1

0.1

0.1

0.3 ∗0.1=0.1 0.3 0.3 ∗0.1=0.1 0.3

0.1

0.1 0.1

0.2 ∗0.1=0.04 0.5 0.2 ∗0.1=0.04 0.5

0.25

0.25

0.1 0.5 ∗0.1=0.25 0.2 0.5 ∗0.1=0.25 0.2

0.25

G 1 = a3

0.4437669 0.4998888 0.4999959 0.4999998 0.5

0.2896825 0.3001261 0.3000048 0.3000002 0.3

0.2665505 0.1999851 0.1999993 0.2 0.2

Table VII. Marginal distributions G 1 , G 2 .

1 3

G 1 = a2

0.3416667 0.4744914 0.5000717 0.5000018 0.5

0.2958333 0.2919279 0.2999113 0.2999979 0.3

G 2 = a3

G 2 = a4

0.25 0.25 0.25 0.25 0.25

0.5 0.5 0.5 0.5 0.5

0.25 0.25 0.25 0.25 0.25

G 1 = a3

G 2 = a2

G 2 = a3

G 2 = a4

0.3625 0.2335807 0.200017 0.2000003 0.2

0.35 0.2847194 0.2500172 0.2500003 0.25

0.455 0.4803188 0.4999107 0.499998 0.5

0.195 0.2349619 0.2500721 0.2500017 0.25

CO RR

1 5 25 50 75

G 1 = a1

EC

Step m

G 2 = a2

D

G 1 = a2

TE

1 5 10 15 17

G 1 = a1

0.1

PR

Table VI. Marginal distributions G 1 , G 2 . Step m

0.25

0.3 ∗0.1=0.1 0.3

0.1

F

0.1

0.1

OO

0.1

The I -projections onto G 1 , G 2 are respectively equal to (0.04, 0.04, 0.04, 0.1, 0.1, 0.04, 0.04, 0.25, 0.25, 0.1) and (0.0488, 0.0714, 0.0488, 0.1786, 0.0556, 0.0488, 0.0488, 0.1389, 0.3049, 0.0556). The next joint distribution P1 is equal to last I -projection. P1 = (0.0488, 0.0714, 0.0488, 0.1786, 0.0556, 0.0488, 0.0488, 0.1389, 0.3049, 0.0556).

7 9

The marginal distributions for G 1 , G 2 given several IPF iteration steps are presented in Table VI. After 17 iterations IPF has converged. Let PIPF be the solution of the IPF procedure. PIPF is then equal to

UN

5

PIPF = (0.03625, 0.055, 0.03625, 0.195, 0.0525, 0.03625, 0.03625, 0.145, 0.355, 0.0525).

13

6

15

(5)

For PARFUM, the next iterate is computed by first I -projecting onto the margins of G 1 , G 2 and then averaging the result. The next iterate P1 is equal to P1 = (0.07, 0.0825, 0.07, 0.1125, 0.0917, 0.07, 0.07, 0.1667, 0.175, 0.0917).

11

(4)

(6)

The marginal distributions for G 1 , G 2 given several PARFUM iteration steps are presented in Table VII. Finally after 75 iterations, PARFUM converges. The PARFUM solution is then equal to PPARFUM = (0.0363, 0.0547, 0.0363, 0.1953, 0.0524, 0.0363, 0.0363, 0.1453, 0.3547, 0.0524).

(7)

As expected, IPF converges faster to a solution than PARFUM. Note that the sample weights obtained from IPF (5) are slightly different from those obtained using PARFUM (7). The utilities of the choice alternatives under IPF and PARFUM Copyright © 2011 John Wiley & Sons, Ltd.



Table VIII. Means and standard deviations of utility values using IPF.

Mean Standard Deviation

a1

a2

a3

a4

0.40629 0.25053

0.32981 0.27967

0.25088 0.23545

0.36642 0.32350

Table IX. Means and standard deviations of utility values using PARFUM. a2

a3

a4

0.4063 0.2506

0.3298 0.2796

0.2510 0.2356

0.3665 0.3234

Table X. Correlation coefficients of the utility values using IPF.

1 3

a3

1.00 0.78 0.24 0.23

0.78 1.00 0.69 0.16

0.24 0.69 1.00 0.60

9

D

In many applications we do not simply want utility values for alternatives, we want to model the utility values as functions of underlying physical variables. The paradigm case is MAUT, where utility is expressed as a weighted combination of physical attributes, such as price, weight, reliability, maintainability, etc. The reservations regarding standard utility modeling expressed in the introduction can be dispelled to some extent within the PI approach. First, we may assume that the utility of stakeholder s for alternative ai , u s (ai ) can be expressed as some function

u s (ai ) =

17 19 21 23 25 27 29

M j =1

s, j ×ci, j .

(9)

MAUT assumes that are normalized weights. The solution algorithms using IPF and PARFUM proceed exactly as before: we begin with a diffuse starting distribution over from which a large number of samples are drawn. Using (9) we compute the joint distribution of utility values associated with the starting distribution. IPF and/or PARFUM are applied to re-weight the starting distribution to comply with the discrete choice data. We hasten to add that a wide variety of functional forms would be tractable. For example, we might add quadratic and interaction terms of arbitrary order to (9) without compromising solvability. Constraints on the parameters , (non-negativity, normalization) can also be imposed. Sampling the resulting joint distribution of we obtain the joint distribution of utilities for all alternatives, characteristic for S . Note that no assumption regarding the dependence of utility values across S is imposed; rather, dependencies emerge from the fitting algorithms themselves. Of course the simple linear form (9) has distinct advantages; because of the linearity of expectation, the expected utility of an alternative can be simply computed by plugging in the expected values of . To exploit the modeling freedom afforded by PI, and to conform with sound science, it is essential to evaluate and validate model forms. As mentioned in the introduction, the near absence of validation has rendered the field of utility theory more confessional than scientific. To effectuate validation, we use part of the discrete choice data to derive the distributions over the parameters , or indeed over the utilities u s (ai ). For example we used the ranks that received more than 30% of the vote to infer a distributions over the parameters using IPF. Then we use these to predict the empirical distribution of

UN

15

(8)

where ci is a vector of ‘criteria scores’ which depend on i but not on s, and s is a vector of parameters which depend on s but not on i. The population of stakeholders S would be described by a distribution over s . The most familiar form is the standard M AU T expression:

CO RR

13

EC

u s (ai ) = (ci , s ),

11

0.23 0.16 0.60 1.00

TE

7

a4

are also close (see Tables VIII and IX). The resulting joint distribution over the choice alternatives induces correlations, as shown in Table X for IPF . Thus, a stakeholder who values a2 highly is also likely to value a1 highly, but little can be said about his/her valuation of a4 . More information on iterative methods for PI may be found in [19]. 3.1. Model validation with IPF and PARFUM

5

OO

a2

PR

a1 a2 a3 a4

a1

F

Mean Standard deviation

a1

7

The

correlation matrix X is not semi-positive definite due to rounding and the small number of samples used.




13 15 17

19 21 23 25 27

F

11

OO

9

As mentioned in Section 2, this valuation study uses the extended version EQ-5D+C (see Table AI in the Appendix) introduced by Stouthard [12]. Although the EuroQol descriptive system is non-disease specific, health states can be associated with diseases and/or disease stages via the use of questionnaires filled by patients, proxies and/or physicians. Owing to human variability, as well as variability in symptoms and stages of a disease, a particular disease could be associated with a number of health states [24]. The number of health states used is 17. The health states were the scenarios that stakeholders participating in the current study judged. The number of participating stakeholders was 19. A great majority (17/19) of stakeholders, represented a panel of ‘non-health care professionals’ defined as people with academic background but with no health-care professional experience. This did not exclude any health-care personal experience; however, they were asked to declare this in advance. The scenarios were presented in five groups of five. Figure A1 in the Appendix illustrates the questions asked to rank the scenarios per group. Scenarios overlapped among the groups; the last two in each group were repeated as the first two in the consecutive group (see Figure 1). This design ensured that we could test stakeholders for consistency in their results and to test if we get similar weights for the criteria in each five group. The 17 scenarios are non-dominated.

PR

7

4. Application: valuing health states

4.1. The model

We used two models in the analysis. The first is the so-called unmodeled scores, where u s (ai ) = u i as in the IPF and PARFUM example of Section 3, taking values between zero and one. The second is the linear model formulated by Equation (9). An overview of the health criteria is illustrated by Figure A2 in the Appendix. The criteria took values 1, 2, 3 and we assumed that a higher values are always worse. As a result, the health scores will take values between −1 and −3. The unmodeled scores can be seen as a composition of both observed and unobserved criteria and will be used as a benchmark. If we can fit the unmodeled scores, but fail to fit linear model then we know the lack of fit is due to the linear model. The number of health states used is 17 so A = {a1, . . . , a17 }. The five groups of five lead to the following the discrete choice problem:

D

5

TE

3

responses in the remaining part of the discrete choice data. This is true out-of-sample validation, and should be recognized as an essential part of any application. The application discussed in the following section illustrates these ideas. There are many issues to be addressed with regard to utility model validation; the approach presented below is not the last word; it is intended to introduce this subject, not close it.

EC

1

D = {D1 , . . . , D5 },

CO RR

D1 = {a1 , a2 , a3 , a4 , a5 },

D2 = {a4 , a5 , a6 , a7 , a8 },

(10)

D3 = {a7 , a8 , a9 , a10 , a11 }, D4 = {a10 , a11 , a12 , a13 , a14 },

D5 = {a13 , a14 , a15 , a16 , a17 }.

31 33

Stakeholders were asked to rank the health states within each group. The task was to find a distribution over the weights (1 , . . . , 6 ) which reproduces the distributions over the rankings.

UN

29

P(1 , . . . , 6 | u(aki ) is j -th ranked in u(Dk )) =

#{s ∈ S | s ranks aki in j -th position in Dk } . #S

(11)

If an stakeholder ranks health state ai above a j in group k then we consider an stakeholder to be inconsistent if he ranks a j above ai in group k +1. 4.2. Results and validation

35 37

8

39

4.2.1. Model adequacy. Of the 19 stakeholders who participated in the study, 13 consistently ranked all of the health states. Using Kendall’s coefficient of concordance W [25, 26] we also looked if the stakeholders rankings are random in each group. If W is one, then each stakeholder has assigned the same order to the choice alternatives. If W is zero, then there is no overall agreement among the stakeholders, and their responses may be regarded as essentially random. Intermediate values of W indicate a greater or lesser degree of agreement among the various responses. Copyright © 2011 John Wiley & Sons, Ltd.



1

Let R(ai , s) be the rank given to health state ai by stakeholder s. Then, the sum of ranks given to ai is R(ai ) = R(ai , e)

(12)

e

3

The mean value of these sum ranks is equal to R¯ = 12 m(n +1)

5

(13)

with m the number of stakeholders and n the number of health states. The sum of squared deviations is defined as nS =

n

¯ 2. (R(ai )− R)

(14)

i=1

W=

19 21 23

25 27 29 31

PR

D

17

TE

15

(15)

EC

13

4.2.2. Criteria weights. We fitted the model to each health state group separately to see how it affects the weights for the criteria. Figures 3–5 illustrate the low, mean, and high values for the weights. The low and high values are computed by respectively subtracting and adding one standard deviation to the means of the weights. Criterion Pain Discomfort seems to be the most important factor followed by Cognitive Functioning, whose values seem to be stable for all groups. However, if we look at the weights from fitting the rank data in each group we notice that criterion Pain Discomfort no longer is the most important factor, but Self-care. We could also read the most important criteria from the cumulative distribution over the weights, see Figure 6. The cumulative distribution over the weights is computed from the

CO RR

11

12S . m 2 (n 3 −n)

The null hypothesis that stakeholders choose ranks at random can be tested in terms of the values for S given n and m. Friedman [26] derived a table which contains the critical values ∗ of S at 5% significance level, for n between 3 and 7 and m between 3 and 20. For each group we computed the values of S and W , shown in Table XI. Friedman’s table does not contain the critical values for S given m = 19; we therefore used the values for m = 20. The null hypothesis would be rejected at the 5% level for m = 20 if S>468.5. With this criterion the null hypothesis is rejected for all groups. We fitted both the rank data of all stakeholders and the rank data of the consistent stakeholders to the MCDM model (9). Each point is an alternative-rank in each of the sets Dk . Each alternative could possibly be ranked in any of the five positions; yielding 5×5×5 points; however, the actual number of points is smaller as the stakeholders confined some alternatives to a smaller number of ranks, and zero probabilities are not plotted. The observed probabilities of rankings are on the horizontal axis, and the probabilities recovered by the linear model of the utilities are on the vertical axis. We used linear regression as goodness-of-fit measure for our method. The slope tells us how accurate the predictions are on average, and the coefficient of determination tells us how much of the variation of the observed frequencies is explained by the predicted frequencies. Both fits have a accuracy of around 95% and a high coefficient of determination (R 2 = 0.896), (R 2 = 0.964) which suggests that the preferences of the stakeholders are consistent with the linear model (9). We continue the analysis using the rankings of the consistent stakeholders.

Table XI. Values of S and W for each group.

UN

9

F

The coefficient of concordance is then equal to

OO

7

S W

Group 1

Group 2

Group 3

Group 4

Group 5

1166 0.323

1854 0.514

1744 0.483

1282 0.355

1597 0.443

9

Figure 3.

Low-value weights obtained from fitting experts’ rankings.




Mean-value weights obtained from fitting experts’ rankings.

Figure 5.

High-value weights obtained from fitting experts’ rankings.

Figure 6.

Cumulative distribution of the weights.

CO RR

EC

Color Online, B&W in Print

TE

D

PR

OO

F

Figure 4.

Table XII. Correlation coefficients of the weights. Mobility

UN

Mobility 1 Self-care −0.2812 Usual act. 0.1018 Pain discomfort 0.1044 Anxiety depression −0.4669 Cognitive functioning −0.1695

1 3 5

10

7

Self-care

−0.2812 1 −0.5247 −0.3103 0.1577 −0.5574

Usual activities 0.1018 −0.5247 1 −0.0104 −0.0146 −0.0362

Pain discomfort 0.1044 −0.3103 −0.0104 1 −0.4643 −0.1107

Anxiety depression −0.4669 0.1577 −0.0146 −0.4643 1 −0.1880

Cognitive functioning −0.1695 −0.5574 −0.0362 −0.1107 −0.1880 1

distribution over the weights obtained from the distribution over the rankings. The most important criterion is represented by the right-most distribution. We say that the stakeholder weights have interactions if, for example, knowledge that a stakeholder assigns high weight to the Self-care criterion gives significant information regarding weights for other criteria. Detailed analysis of interactions is not undertaken, but the correlation matrix presented in Table XII suggests that the mobility–anxiety depression; selfcare–cognitive functioning; self-care–usual activities; pain discomfort–anxiety depression interactions are rather strong. The requirement that the weights sum to one imposes an overall negative correlation. Copyright © 2011 John Wiley & Sons, Ltd.


Prediction of rank percentages using rank percentages greater than 30%.

Figure 8.

Average prediction of rank percentages less than 30%.

5

7 9 11

13 15 17


4.2.4. Unmodeled scores. Fitting the unmodeled scores yields a near perfect fit (R 2 = 0.999). This means that we can assign a distribution over the utilities for the 17 health states, such that the probabilities of observing each health state at each rank in each Dk is predicted nearly perfectly. This merely says that our population of stakeholders can be modeled as rational in the sense of Savage. The unmodeled scores presented by Figure A4 are assumed to take values between −1 and 0. The variance of the unmodeled scores seems to be higher than the variance of the health scores from the linear model (9). Note that the ranks of the health states have slightly changed with respect to the scores from the linear model. 4.2.5. Validation. Finally, we validated the results by fitting the model to the ranks that got more than 30% of the votes from the stakeholders, and then predicted the remaining ranks using the fitted model. At a first glance, from Figure 7 the prediction does not look good at all; but if we zoom in on the ranks that got less than 30% and take the average of these predicted ranks we get a fit that is not so bad, see Figure 8. However, the unmodeled scores seems to give better prediction both overall and on the unpopular ranks, see Figures 9 and 10. The out-of-sample validation is not overwhelming; the distribution over weights derived from the ‘popular rankings’ (i.e. those rankings assigned by more than 30% of the stakeholders, did not accurately predict the percentages for the ‘unpopular Copyright © 2011 John Wiley & Sons, Ltd.


11

19

CO RR

3

4.2.3. Health state scores. Fitting the MCDM model not only gives us statistics about the disability weights, but also statistics about the health scores. In the introduction we mentioned that utility is affine unique, which we have used to transform the health scores to the zero one interval. Initially the health scores given the MCDM model vary between −3 and −1, which have no tangible meaning. However, if we standardize these scores we can think of them as the relative impact on health, see Figure A3.

UN

1

EC

TE

D


PR

Figure 7.

OO

F


Prediction of rank percentages using rank percentages greater than 30%.

3

5

7 9

11

rankings’ (i.e. rankings assigned by less than 30% of the stakeholders). Nonetheless, the average predictions do align with the observed predictions. Without undertaking an in-depth analysis, it seems that the unpopular ranks concerned primarily the lowest ranked health states, and one possible explanation for these results is that the preference judgments for the ‘less important’ health states were less discerning.

5. Conclusion

PI methods can be used to infer a distribution over utility functions based on discrete choice data. This type of application is rather new and more experience in real applications is needed. In particular, questions regarding the optimal format of the discrete choice data, the best approach to out-of-sample validation is still largely open. Equally important is how best to model utilities in terms of physical attributes. It is hoped that out-of-sample validation will eventually yield utility models with a solid scientific foundation.

UN

1

Average prediction of rank percentages less than 30%.

CO RR

Figure 10.

EC

TE

D

Color Online, B&W in Print Color Online, B&W in Print

PR

Figure 9.

OO

F


Appendix A: Figures and Tables A.1. Definitions

12

13 15

Discrete choice or random utility models describe and analyze the preferences of a group of stakeholders S for set of choice alternatives A = {a1, . . . , a N }. The preferences of each stakeholder are denoted in the customary way as ai a j . If necessary to distinguish s ∈ (S), we will write s . Copyright © 2011 John Wiley & Sons, Ltd.



Table AI. The extended, six-dimensional version of the original EuroQol descriptive system, i.e. EQ-5D+C. Value

Mobility

Self-care

1

No problems in walking about

No problems with self-care

2

Some problems in walking about Confined to bed

Some problems washing or dressing self Unable to wash or dress self

No problems with performing usual activities (e.g. work, study, housework) Some problems with performing usual activities Unable to perform usual activities

Pain Discomfort

Anxiety Depression

Cognitive functioning

No pain or discomfort

Not anxious or depressed

No problems in cognitive functioning

Moderate pain or discomfort

Moderately anxious or depressed Extremely anxious or depressed

Moderate problems in cognitive functioning

Extreme pain or discomfort

Severe problems in cognitive functioning

3 5 7 9 11 13

15

Definition A.1 D is called a discrete choice problem on A if (1) A = a1 , . . . , a N is a finite non-empty set of N choice alternatives. (2) D = {D1 , . . . , D K |Di ⊆ A, Di = ∅, i = 1, . . . , K }.

A familiar type of a discrete choice problem is paired comparisons. Choice alternatives are presented in N2 pairs from which stakeholders pick their preferred alternative. The response of a stakeholder to a discrete choice problem may take many forms. For example, (s)he may choose a unique preferred alternative from each set Dk ∈ D (strict choice) or a set of non-dominated alternatives in Dk ∈ D (non-dominated choice), or (s)he may order the elements of Dk ∈ D, such that the response is a permutation ∈ Dk ! (strict preference order), Copyright © 2011 John Wiley & Sons, Ltd.


13

17

Savage’s [1] theory of rational decision ensures that the preferences of a single rational stakeholder can be expressed in terms of expected utility. Each s ∈ S may be assigned a utility function over choice alternatives that is unique up to a positive affine transformation, i.e. unique up to choice of zero and unit. If u : A→ R is a utility function for a given stakeholder, then cu +d, c > 0, d ∈ R, is also a utility function for this stakeholder. We will assume that our set of stakeholders have utility functions which can be assigned the same unit. This means that there are two consequences, say g and b, not necessarily belonging to the choice set A, such that all stakeholders agree that g is strictly preferred to b. Each stakeholder would then choose respectively g and b as the unit and zero of his utility scale. We call such a set of stakeholders orientable. For an orientable set of rational stakeholders, the utilities over A may be represented as standardized A-vectors taking values in [0, 1], that is, as elements of [0, 1]A .

UN

1

Questions group 1.

CO RR

Figure A1.

EC

TE


D

PR

OO

F

3

Usual Activities

PR

OO

F


Criteria values per health state.

Figure A3.

Standardized health scores obtained from fitting the model to the discrete choice data.

UN

CO RR


EC

TE

D

Figure A2.

1

or (s)he may produce an ordered partition of elements Dk where the alternatives in each element of the partition are equivalent (non-dominated preference order). These are captured in

3

Definition A.2

5 7 9

14

11

• A strict choice response r = (r1 , . . . ,r K ) to discrete choice problem D is a set of mappings rk : S → Dk , k = 1, . . . , K . D • A non-dominated choice response r = (r1 , . . . ,r K ) to discrete choice problem D is a set of mappings rk : S → 2+k , with Dk 2+ = 2 Dk \∅, k = 1, . . . , K . • A strict preference order response r = (r1 , . . . ,r K ) to discrete choice problem D is a set of mappings rk : S → Dk ! where Dk ! is the set of permutations of Dk , k = 1, . . . , K . • A non-dominated preference order response r = (r1 , . . . ,r K ) to discrete choice problem D is a set of mappings rk : S → k where k is the set of ordered partitions of Dk , k = 1, . . . , K . Copyright © 2011 John Wiley & Sons, Ltd.


9 11 13

15 17 19

D

References 1. 2. 3. 4. 5.

21 23


Q1

TE

7

EC

5

6. 7. 8.

25 27

9.

29

10.

31

11. 12.

33 13.

35 14. 15.

39

16.

Savage L. The Foundations of Statistics. Wiley: New York, 1954. Thurstone LL. A law of comparative judgement. Psychological Review 1927; 34:278--286. Luce RD. Individual Choice Behavior: A Theoretical Analysis. Wiley: New York, 1959. Terry ME, Bradley R. Rank analysis of incomplete block designs: I. the method of paired comparisons 1952. Marschak J. Binary choice constraints on random utility indicators. Cowles Foundation Discussion Papers 74, Cowles Foundation, Yale University, 1959. McFadden D. Conditional logit analysis of qualitative choice behavior. Frontiers in Econometrics, 1974; 198--272. McFadden D, Train K. Mixed MNL models for discrete response. Journal of Applied Econometrics 15(5):447--470. Neslo REJ, Micheli F, Kappel CV, Selkoe KA, Halpern BS, Cooke RM. Modeling stakeholder preferences with probabilistic inversion application to prioritizing marine ecosystem vulnerabilities. Real-time and Deliberative Decision Making, NATO Science for Peace and Security Series C: Environmental Security. Springer: Netherlands, 2008; 265--284. Cooke RM. Obtaining distributions from groups for decisions under uncertainty. In Making Essential Choices with Scant Information: Frond-end Decisions Making in Major Projects, Samset K, Williams TM, Sunneva KJ (eds). Palgrave Macmillan: 2009; 257--274. Teck S, Halpern B, Kappel C, Micheli F, Selkoe K, Crain C, Martone R, Shearer C, Arvai J, Fischhoff B, Murray G, Neslo R, Cooke R. Using expert judgment to estimate marine ecosystem vulnerability in the california current. Ecological Applications 2000. Kurowicka D, Bucura C, Havelaar A, Cooke RM. Probabilistic inversion in priority setting of emerging zoonoses. Risk Analysis 2010. Stouthard MEA, Louise E-BM, Gouke BJ. Disability weights for diseases. a modified protocol and results for a western European region. European Journal of Public Health 2000; 10(1):24--30. Kraan BCP, Bedford TJ. Probabilistic inversion of expert judgements in the quantification of model uncertainty. Management Science 2005; 51(6):995--1006. Kruithof J. Telefoonverkeersrekening. De Ingenieur 1937; 52(8):15--25. Stephan FF. An iterative method of adjusting sample frequency tables when expected marginal totals are known. Annals of Mathematical Statistics 1942; 13:166--178. Fienberg SE. An iterative procedure for estimation in contingency tables. Annals of Mathematical Statistics 1970; 41(3):907--917.



Q2

Q3

15

37

CO RR

3

The set of responses to D for all s ∈ S will be denoted by rD . Many other response forms are conceivable, but the above are the most straightforward. Note the difference between strict and non-dominated choice. In the standard versions of the random utility theory, when a stakeholder chooses element ai from a set {a1 , . . . , an }, this is interpreted to mean that ai is at least as good as the other elements. On this choice data alone we cannot distinguish strict preference from equivalence in preference, while this might not be severe in modeling the preferences of one individual, for populations of stakeholders, such ambiguity can cause problems. Thus, if 50% of stakeholders preferred a red bus to a blue bus, and 50% preferred the blue to red bus, this might either mean that everyone had strict preferences evenly divided over the population, or alternatively it might mean that everyone in the population was indifferent to the busses’s color, and was choosing one color at random. Failure to distinguish these cases can cause problems in modeling the preferences of the population. These issues are important and have dominated a good deal of the discrete choice literature. Nonetheless, they are not the point of the present study. The tools that we develop for deriving a distribution over utility functions, given a distribution of responses to a discrete choice problem, apply equally well for strict and non-strict preference. However, allowing for equivalence in preference considerably complicates the notation, as can be inferred from Definition A.2. We therefore restrict attention in this study to strict preference.

UN

1

Unmodeled health scores.

PR

Figure A4.

OO

F



3 5 7 9 11 13

UN

CO RR

EC

TE

D

PR

OO

15

17. Cooke RM. Parameter fitting for uncertain models: modeling uncertainty in small models. Reliability Engineering and Systems Safety 1994; 44(1):89--102. 18. DU C, Kurowicka D, Cooke RM. Techniques for generic probabilistic inversion. Computational Statistics and Data Analysis 2006; 50(5):1164--1187. 19. Kurowicka D, Cooke RM. Uncertainty Analysis with High Dimensional Dependence Modelling. Wiley: New York, 2006. 20. Deming EW, Stephan FF. On a least squares adjustment of a sampled frequency table when the expected marginal totals are known. Annals of Mathematical Statistics 1940; 11(4):427--444. 21. Csiszar I. divergence geometry of probability distributions and minimization problems. Annals of Probability 1975; 3(1):146--158. 22. Matus F. On iterated averages of i-projections. bielefeld: Statistiek und informatik, University. Not published, 2007. 23. Kullback S. Information Theory and Statistics. Wiley Inc.: New York, 1959. 24. Ellis JJ, Eagle KA, Kline-Rogers EM, Erickson SR. Validation of the EQ-5D in patients with a history of acute coronary syndrome. Current Medical Research and Opinion 2005; 21(8):1209--1216. 25. Kendall MG, Smith BB. The problem of m rankings. The Annals of Mathematical Statistics 1939; 10(3):275--287. 26. Friedman M. A comparison of alternative tests of significance for the problem of m rankings. Annals of Mathematical Statistics 1940; 11(1):86--92.

F

1

16 Copyright © 2011 John Wiley & Sons, Ltd.


Author Query Form Journal Name : Query No

Q1

Q2 Q3

Proof Page/lineno

ASMB

Article Number: 888

Details required

There is a cross-reference to color in the text for appendix figures. As the figures will be printed in black and white, please modify the text accordingly. Please provide DOI number if volume and page range are not yet available for Refs. [4, 10, 11]. Please provide publisher location for Ref. [9].

Authors response

`

USING eANNOTATION TOOLS FOR ELECTRONIC PROOF CORRECTION Required Software to eAnnotate PDFs: Adobe Acrobat Professional or Acrobat Reader (version 8.0 or above). The Latest version of Acrobat Reader is free: http://www.adobe.com/products/acrobat/readstep2.html Once you have Acrobat Reader 8, or higher, open on your PC you should see the Commenting Toolbar:

****(If the above toolbar does not appear automatically go to Tools>Comment & Markup>Show Comment & Markup Toolbar)****

1. Replacement Text Tool — For replacing text. Strikes a line through text and opens up a replacement text box.

How to use it: 1. Highlight a word or sentence 2. Select “Replace Selected Text” from the Text Edits fly down button 3. Type replacement text in blue box

2. Cross-out Text Tool — For deleting text. Strikes a red line through selected text.

How to use it: 1. Highlight a word or sentence 2. Select “Cross Out Text for Deletion” from the Text Edits fly down button

3. Highlight Tool — For highlighting a selection to be changed to bold or italic. Highlights text in yellow and opens up a text box.

How to use it: 1. Highlight desired text 2. Select “Add Note To Selected Text” from the Text Edits fly down button 3. Type a note detailing required change in the yellow box

4. Note Tool — For making notes at specific points in the text Marks a point on the paper where a note or question needs to be addressed.

How to use it: 1. Select the Sticky Note icon from the commenting toolbar 2. Click where the yellow speech bubble symbol needs to appear and a yellow text box will appear 3. Type comment into the yellow text box

USING eANNOTATION TOOLS FOR ELECTRONIC PROOF CORRECTION

5. Drawing Markup Tools — For circling parts of figures or spaces that require changes These tools allow you to draw circles, lines and comment on these marks.

How to use it: 1. Click on one of shape icons in the Commenting Toolbar 2. Draw the selected shape with the cursor 3. Once finished, move the cursor over the shape until an arrowhead appears and double click 4. Type the details of the required change in the red box

6. Attach File Tool — For inserting large amounts of text or replacement figures as a files. Inserts symbol and speech bubble where a file has been inserted. How to use it: 1. Right click on the Commenting Toolbar 2. Select “Attach a File as a Comment” 3. Click on paperclip icon that appears in the Commenting Toolbar 4. Click where you want to insert the attachment 5. Select the saved file from your PC or network 6. Select type of icon to appear (paperclip, graph, attachment or tag) and close

7. Approved Tool (Stamp) — For approving a proof if no corrections are required. How to use it: 1. Click on the Stamp Tool in the toolbar 2. Select the Approved rubber stamp from the „standard business‟ selection 3. Click on the text where you want to rubber stamp to appear (usually first page)

Help For further information on how to annotate proofs click on the Help button to activate a list of instructions:

COPYRIGHT TRANSFER AGREEMENT Date:

Contributor name:

Contributor address: Manuscript number (if known): Re: Manuscript entitled (the “Contribution”) for publication in published by

(the “Journal”) (“Wiley-Blackwell”).

Dear Contributor(s): Thank you for submitting your Contribution for publication. In order to expedite the editing and publishing process and enable Wiley-Blackwell to disseminate your Contribution to the fullest extent, we need to have this Copyright Transfer Agreement signed and returned as directed in the Journal’s instructions for authors as soon as possible. If the Contribution is not accepted for publication, or if the Contribution is subsequently rejected, this Agreement shall be null and void. Publication cannot proceed without a signed copy of this Agreement.

A. COPYRIGHT 1. The Contributor assigns to Wiley-Blackwell, during the full term of copyright and any extensions or renewals, all copyright in and to the Contribution, and all rights therein, including but not limited to the right to publish, republish, transmit, sell, distribute and otherwise use the Contribution in whole or in part in electronic and print editions of the Journal and in derivative works throughout the world, in all languages and in all media of expression now known or later developed, and to license or permit others to do so. 2. Reproduction, posting, transmission or other distribution or use of the final Contribution in whole or in part in any medium by the Contributor as permitted by this Agreement requires a citation to the Journal and an appropriate credit to Wiley-Blackwell as Publisher, and/or the Society if applicable, suitable in form and content as follows: (Title of Article, Author, Journal Title and Volume/Issue, Copyright © [year], copyright owner as specified in the Journal). Links to the final article on Wiley-Blackwell’s website are encouraged where appropriate.

B . R E TA I N E D R I G H T S Notwithstanding the above, the Contributor or, if applicable, the Contributor’s Employer, retains all proprietary rights other than copyright, such as patent rights, in any process, procedure or article of manufacture described in the Contribution.

C. PERMITTED USES BY CONTRIBUTOR 1. Submitted Version. Wiley-Blackwell licenses back the following rights to the Contributor in the version of the Contribution as originally submitted for publication: a. After publication of the final article, the right to self-archive on the Contributor’s personal intranet page or in the Contributor’s institution’s/ employer’s institutional intranet repository or archive. The Contributor may not update the submission version or replace it with the published Contribution. The version posted must contain a legend as follows: This is the pre-peer reviewed version of the following article: FULL CITE, which has been published in final form at [Link to final article]. b. The right to transmit, print and share copies with colleagues. 2. Accepted Version. Reuse of the accepted and peer-reviewed (but not final) version of the Contribution shall be by separate agreement with WileyBlackwell. Wiley-Blackwell has agreements with certain funding agencies governing reuse of this version. The details of those relationships, and other offerings allowing open web use are set forth at the following website: http://www.wiley.com/go/funderstatement. NIH grantees should check the box at the bottom of this document.

3. Final Published Version. Wiley-Blackwell hereby licenses back to the Contributor the following rights with respect to the final published version of the Contribution: a. Copies for colleagues. The personal right of the Contributor only to send or transmit individual copies of the final published version to colleagues upon their specific request provided no fee is charged, and further-provided that there is no systematic distribution of the Contribution, e.g. posting on a listserve, website or automated delivery. For those Contributors who wish to send high-quality e-prints, purchase reprints, or who wish to distribute copies more broadly than allowed hereunder (e.g. to groups of colleagues or mailing lists), please contact the publishing office. b. Re-use in other publications. The right to re-use the final Contribution or parts thereof for any publication authored or edited by the Contributor (excluding journal articles) where such re-used material constitutes less than half of the total material in such publication. In such case, any modifications should be accurately noted. c. Teaching duties. The right to include the Contribution in teaching or training duties at the Contributor’s institution/place of employment including in course packs, e-reserves, presentation at professional conferences, in-house training, or distance learning. The Contribution may not be used in seminars outside of normal teaching obligations (e.g. commercial seminars). Electronic posting of the final published version in connection with teaching/training at the Contributor’s institution/place of employment is permitted subject to the implementation of reasonable access control mechanisms, such as user name and password. Posting the final published version on the open Internet is not permitted. d. Oral presentations. The right to make oral presentations based on the Contribution. 4. Article Abstracts, Figures, Tables, Data Sets, Artwork and Selected Text (up to 250 words). a. Contributors may re-use unmodified abstracts for any non-commercial purpose. For on-line uses of the abstracts, Wiley-Blackwell encourages but does not require linking back to the final published versions. b. Contributors may re-use figures, tables, data sets, artwork, and selected text up to 250 words from their Contributions, provided the following conditions are met: (i) Full and accurate credit must be given to the Contribution. (ii) Modifications to the figures, tables and data must be noted. Otherwise, no changes may be made. (iii) The reuse may not be made for direct commercial purposes, or for financial consideration to the Contributor. (iv) Nothing herein shall permit dual publication in violation of journal ethical practices. CTA-PS

D. CONTRIBUTIONS OWNED BY EMPLOYER 1. If the Contribution was written by the Contributor in the course of the Contributor’s employment (as a “work-made-for-hire” in the course of employment), the Contribution is owned by the company/employer which must sign this Agreement (in addition to the Contributor's signature) in the space provided below. In such case, the company/employer hereby assigns to Wiley-Blackwell, during the full term of copyright, all copyright in and to the Contribution for the full term of copyright throughout the world as specified in paragraph A above. 2. In addition to the rights specified as retained in paragraph B above and the rights granted back to the Contributor pursuant to paragraph C above, WileyBlackwell hereby grants back, without charge, to such company/employer, its subsidiaries and divisions, the right to make copies of and distribute the final published Contribution internally in print format or electronically on the Company’s internal network. Copies so used may not be resold or distributed externally. However the company/employer may include information and text from the Contribution as part of an information package included with software or other products offered for sale or license or included in patent applications. Posting of the final published Contribution by the institution on a public access website may only be done with Wiley-Blackwell’s written permission, and payment of any applicable fee(s). Also, upon payment of Wiley-Blackwell’s reprint fee, the institution may distribute print copies of the published Contribution externally.

E. GOVERNMENT CONTRACTS In the case of a Contribution prepared under U.S. Government contract or grant, the U.S. Government may reproduce, without charge, all or portions of the Contribution and may authorize others to do so, for official U.S. Govern-

ment purposes only, if the U.S. Government contract or grant so requires. (U.S. Government, U.K. Government, and other government employees: see notes at end.)

F. C O P Y R I G H T N O T I C E The Contributor and the company/employer agree that any and all copies of the final published version of the Contribution or any part thereof distributed or posted by them in print or electronic format as permitted herein will include the notice of copyright as stipulated in the Journal and a full citation to the Journal as published by Wiley-Blackwell.

G . C O N T R I B U T O R ’ S R E P R E S E N TAT I O N S The Contributor represents that the Contribution is the Contributor’s original work, all individuals identified as Contributors actually contributed to the Contribution, and all individuals who contributed are included. If the Contribution was prepared jointly, the Contributor agrees to inform the co-Contributors of the terms of this Agreement and to obtain their signature to this Agreement or their written permission to sign on their behalf. The Contribution is submitted only to this Journal and has not been published before. (If excerpts from copyrighted works owned by third parties are included, the Contributor will obtain written permission from the copyright owners for all uses as set forth in WileyBlackwell’s permissions form or in the Journal’s Instructions for Contributors, and show credit to the sources in the Contribution.) The Contributor also warrants that the Contribution contains no libelous or unlawful statements, does not infringe upon the rights (including without limitation the copyright, patent or trademark rights) or the privacy of others, or contain material or instructions that might cause harm or injury.

CHECK ONE BOX: Contributor-owned work AT TA C H A D D I T I O N A L S I G N AT U R E PA G E S A S N E C E S S A RY

Contributor’s signature

Date

Type or print name and title

Co-contributor’s signature

Date

Type or print name and title

Company/Institution-owned work (made-for-hire in the course of employment)

Company or Institution (Employer-for-Hire)

Date

Authorized signature of Employer

Date

U.S. Government work

Note to U.S. Government Employees A contribution prepared by a U.S. federal government employee as part of the employee’s official duties, or which is an official U.S. Government publication, is called a “U.S. Government work,” and is in the public domain in the United States. In such case, the employee may cross out Paragraph A.1 but must sign (in the Contributor’s signature line) and return this Agreement. If the Contribution was not prepared as part of the employee's duties or is not an official U.S. Government publication, it is not a U.S. Government work.

U.K. Government work (Crown Copyright)

Note to U.K. Government Employees The rights in a Contribution prepared by an employee of a U.K. government department, agency or other Crown body as part of his/her official duties, or which is an official government publication, belong to the Crown. U.K. government authors should submit a signed declaration form together with this Agreement. The form can be obtained via http://www.opsi.gov.uk/advice/crown-copyright/copyright-guidance/ publication-of-articles-written-by-ministers-and-civil-servants.htm

Other Government work

Note to Non-U.S., Non-U.K. Government Employees If your status as a government employee legally prevents you from signing this Agreement, please contact the editorial office.

NIH Grantees

Note to NIH Grantees Pursuant to NIH mandate, Wiley-Blackwell will post the accepted version of Contributions authored by NIH grant-holders to PubMed Central upon acceptance. This accepted version will be made publicly available 12 months after publication. For further information, see www.wiley.com/go/nihmandate. CTA-PS

WILEY AUTHOR DISCOUNT CARD As a highly valued contributor to Wiley’s publications, we would like to show our appreciation to you by offering a unique 25% discount off the published price of any of our books*. To take advantage of this offer, all you need to do is apply for the Wiley Author Discount Card by completing the attached form and returning it to us at the following address: The Database Group John Wiley & Sons Ltd The Atrium Southern Gate Chichester West Sussex PO19 8SQ UK In the meantime, whenever you order books direct from us, simply quote promotional code S001W to take advantage of the 25% discount. The newest and quickest way to order your books from us is via our new European website at:

http://www.wileyeurope.com Key benefits to using the site and ordering online include: • Real-time SECURE on-line ordering • The most up-to-date search functionality to make browsing the catalogue easier • Dedicated Author resource centre • E-mail a friend • Easy to use navigation • Regular special offers • Sign up for subject orientated e-mail alerts So take advantage of this great offer, return your completed form today to receive your discount card. Yours sincerely,

Verity Leaver E-marketing and Database Manager

*TERMS AND CONDITIONS This offer is exclusive to Wiley Authors, Editors, Contributors and Editorial Board Members in acquiring books (excluding encyclopaedias and major reference works) for their personal use. There must be no resale through any channel. The offer is subject to stock availability and cannot be applied retrospectively. This entitlement cannot be used in conjunction with any other special offer. Wiley reserves the right to amend the terms of the offer at any time.

REGISTRATION FORM

FOR 25% BOOK DISCOUNT CARD To enjoy your special discount, tell us your areas of interest and you will receive relevant catalogues or leaflets from which to select your books. Please indicate your specific subject areas below. Accounting • Public • Corporate

[] [] []

Architecture

[]

Business/Management

[]

Chemistry • Analytical • Industrial/Safety • Organic • Inorganic • Polymer • Spectroscopy

[ [ [ [ [ [ [

] ] ] ] ] ] ]

Computer Science • Database/Data Warehouse • Internet Business • Networking • Programming/Software Development • Object Technology

[ [ [ [ [ [ [

] ] ] ] ] ] ]

Encyclopedia/Reference • Business/Finance • Life Sciences • Medical Sciences • Physical Sciences • Technology

[ [ [ [ [ [

] ] ] ] ] ]

Engineering • Civil • Communications Technology • Electronic • Environmental • Industrial • Mechanical

[ [ [ [ [ [ [

] ] ] ] ] ] ]

Earth & Environmental Science

[]

Hospitality

[]

Finance/Investing • Economics • Institutional • Personal Finance

[ [ [ [

] ] ] ]

Genetics • Bioinformatics/Computational Biology • Proteomics • Genomics • Gene Mapping • Clinical Genetics

[ [ [ [ [ [ [

Life Science

[]

Landscape Architecture

[]

Mathematics/Statistics

[]

Manufacturing

[]

Material Science

[]

Psychology • Clinical • Forensic • Social & Personality • Health & Sport • Cognitive • Organizational • Developmental and Special Ed • Child Welfare • Self-Help

[ [ [ [ [ [ [ [ [ [

Physics/Physical Science

[]

] ] ] ] ] ] ]

Medical Science • Cardiovascular • Diabetes • Endocrinology • Imaging • Obstetrics/Gynaecology • Oncology • Pharmacology • Psychiatry

[ [ [ [ [ [ [ [ [

] ] ] ] ] ] ] ] ]

Non-Profit

[]

] ] ] ] ] ] ] ] ] ]

[ ] I confirm that I am a Wiley Author/Editor/Contributor/Editorial Board Member of the following publications:

SIGNATURE: ………………………………………………………………………………………………………………………………………………… PLEASE COMPLETE THE FOLLOWING DETAILS IN BLOCK CAPITALS: TITLE AND NAME: (e.g. Mr, Mrs, Dr) ………………………………………………………………………………………………………… JOB TITLE: …………………………………………………………………………………………………………………………………………………… DEPARTMENT: …………………………………………………………………………………………………………………………………………….. COMPANY/INSTITUTION: …………………………………………………………………………………………………………………………… ADDRESS: ……………………………………………………………………………………………………………………………………………………. ……………………………………………………………………………………………………………………………………………………………………… ……………………………………………………………………………………………………………………………………………………………………… ……………………………………………………………………………………………………………………………………………………………………… TOWN/CITY: ………………………………………………………………………………………………………………………………………………… COUNTY/STATE: …………………………………………………………………………………………………………………………………………. COUNTRY: …………………………………………………………………………………………………………………………………………………… POSTCODE/ZIP CODE: ………………………………………………………………………………………………………………………………… DAYTIME TEL: ……………………………………………………………………………………………………………………………………………… FAX: ……………………………………………………………………………………………………………………………………………………………… E-MAIL: ………………………………………………………………………………………………………………………………………………………… YOUR PERSONAL DATA We, John Wiley & Sons Ltd, will use the information you have provided to fulfil your request. In addition, we would like to: 1.

Use your information to keep you informed by post, e-mail or telephone of titles and offers of interest to you and available from us or other Wiley Group companies worldwide, and may supply your details to members of the Wiley Group for this purpose.

[ ] Please tick the box if you do not wish to receive this information 2.

Share your information with other carefully selected companies so that they may contact you by post, fax or e-mail with details of titles and offers that may be of interest to you.

[ ] Please tick the box if you do not wish to receive this information. If, at any time, you wish to stop receiving information, please contact the Database Group ([email protected]) at John Wiley & Sons Ltd, The Atrium, Southern Gate, Chichester, West Sussex PO19 8SQ, UK. E-MAIL ALERTING SERVICE We offer an information service on our product ranges via e-mail. If you do not wish to receive information and offers from John Wiley companies worldwide via e-mail, please tick the box [ ]. This offer is exclusive to Wiley Authors, Editors, Contributors and Editorial Board Members in acquiring books (excluding encyclopaedias and major reference works) for their personal use. There should be no resale through any channel. The offer is subject to stock availability and may not be applied retrospectively. This entitlement cannot be used in conjunction with any other special offer. Wiley reserves the right to vary the terms of the offer at any time.

Ref: S001W