149 14. Estimation of Non-market Forest Benefits Using Choice Modelling Jungho Suh This module is concerned with some fundamental features and conditions of choice modelling applications to non-market valuation. Choice modelling is an advance on the contingent valuation method (CVM), and is creating strong interest among researchers, and has much potential for non-market valuation in multiple use forestry. To undertake a choice modelling application for estimating non-market values, potential practitioners need to understand the theoretical issues and practicalities involved in applying the technique. These include the statistical foundation of choice modelling, strict rules for the experimental design and ways of utilising the estimates. This module first outlines the characteristic features of choice modelling in terms of the method of evaluating resource use alternatives, compared with contingent ranking, contingent rating and CVM. Some basic assumptions and considerations needed in designing choice sets are then examined. Roles and rules of focus groups are next introduced. Some choice modelling applications made in forestry research are then briefly reviewed. Finally, ways of extrapolating welfare measures from a choice modelling application are reviewed. More complex statistical issues are explained in three appendices. 1. THE CHOICE SET FORMAT OF goods rather than non-market goods. CHOICE MODELLING Conjoint analysis is founded on the theory of consumer preference in an attempt to Choice modelling originated from conjoint describe how consumers choose between analysis, and is also a variation on similar products, for example, beers, contingent valuation. In comparison to coffees and soft drinks. Respondents are CVM, conjoint analysis describes options by asked to rank or rate or choose from a set decomposing them into a number of of multiple product profiles. Setting prices attributes, and presents respondents with a for the products was not necessarily the choice between j available options (j = 1, primary concern of conjoint analysis in 2,⋅⋅⋅, J). This situation can be made quite marketing studies. In this sense, the fact realistic by mirroring actual market choice that conjoint analysis was eventually that may depend upon a number of developed to value non-market public attributes. goods can be dubbed a ‘paradigm change’ in the field of economic valuation. If the number of available options is too large, the full options are divided into The rationale of conjoint analysis several sets. Then, respondents are asked applications for estimating environmental to rank, rate or choose their preferred non-market values is that it is possible to combination from each set. Moreover, the estimate the amount that people are willing number of sets can be increased to as to pay to achieve a greater amount of one many as each respondent can answer or more environmental attribute, given that within a limited time. For this reason, it is the dollar cost is treated as one of the said that one of the major advantages of characteristics for a non-market good. In conjoint analysis compared to CVM is that fact, the price factor does not represent an many options provide a large number of inherent attribute of a commodity under observations so that fewer respondents are consideration. Rather, the price presents required to yield results within acceptable dollar costs that are traded off for proposed confidence limits. changes in attribute levels. This is why Mitchell and Carson (1989) classified Conjoint techniques have been widely used conjoint analysis as a ‘hypothetical and in marketing studies dealing with market indirect’ approach. 150 Socio-economic Research Methods in Forestry In contingent ranking, respondents rank cardinal measurement of utility (Morrison et three or more options from most to least al. 1996). Since individual rating scales in preferred. In the contingent rating contingent rating applications reveal only application, respondents are asked to rate relative value between the respondents, it is each option separately on a given rating necessary to assume that rating scales scale instead of ranking the options. For being used are consistent across example, consider the case of a protected individuals (Rolfe and Bennett 1996). forest area as illustrated in Table 1, where Similarly, contingent ranking suffers j inconsistent ordinal measurement of utility zk represents the kth attribute of the jth j across individuals. option and z p is the price factor. The protected area is here defined as a 2. ASSUMPTIONS AND combination of attributes, each of which CONSIDERATIONS IN THE may take various levels. If a respondent EXPERIMENTAL DESIGN prefers the jth option { z j ,z j ,⋅⋅⋅ ,z j } to the 1 2 k Any type of market or non-market good can other options, a higher ranking or rating is be described by a range of characteristics. assigned to the jth option. Compared to In applications of choice modelling, a contingent ranking, contingent rating number of hypothetical profiles are created contains cardinal information. In choice- by combining distinct levels of attributes, based conjoint analysis, respondents are which must represent a wide range of asked only to choose their highest characteristics of the object being valued. preference from among several options – The number of attributes and their levels for example the set of choices presented in determines the total number of distinct Table 1. Carson et al. (1994) called this profiles. A full factorial design includes all method ‘choice modelling’ to distinguish it combinations of the attribute levels, where from contingent ranking or rating. In some every level of a given attribute is combined literature, the term ‘environmental choice with all levels of every other attribute. In experiments’ is used rather than choice general, if there are m factors and n levels modelling, especially by the Canadian of each, nm unique combinations can be group of practitioners. made. If factor space S has k factors, S = (z , z ,···,z ), and each factor z has L It is notable that a dichotomous CVM 1 2 k k k possible levels, then S has L × L ×··· × L question is the same as a binary choice 1 2 k possible combinations. modelling one except for the pricing format (Bennett and Carter 1993; Roe et al. 1996; Question formats based on complete Stevens et al. 2000). Consider a choice factorial design quickly become impractical modelling question with only one alternative due to the cost of administering the survey, (two options), as in Table 2, where not to speak of the respondents’ confusion respondents are asked whether to accept and fatigue, as the number of either factors the new option, comparing to the current or levels of the individual factors increases. status option. Note that one of z represents k Indeed, in many cases, a choice modelling the WTP amount (z ). It can be seen that p researcher is simply unable to conduct a the question is virtually identical to that of a survey using a large number of profiles. dichotomous CVM where respondents are Hence, the researcher is forced to adopt a asked whether they would be willing to pay fractional factorial design, where only some z for the same change in z . p k of the combinations of factor levels are included. In choice modelling practice, a Choice modelling has an advantage over selected fractional factorial design is again contingent rating in the sense that the broken into a number of separate choice former is free of metric bias with which the sets. Rolfe and Bennett (1996) noted that latter is plagued. Metric bias occurs when a the number of choice sets should not be too respondent values an amenity according to onerous for a single respondent. They a different metric or scale than the one suggested that choice sets be divided into intended by the researcher (Mitchell and manageable blocks, with each block Carson 1989). This bias also relates to the allocated to a sub-sample of respondents. problems of interpersonal comparison of Estimation of Non-market Forest Benefits Using Choice Modelling 151 Table 1. A choice format with several scenarios with various levels of attributes Option ( j ) Attribute (zk) z1 z2 … zk zp 1 1 1 … 1 1 z1 z2 zk z p 2 2 2 … 2 2 z1 z2 zk z p : : : : : : J J J … J J z1 z2 zk z p Table 2. A binary choice modelling question format Option Attribute (zk) z1 z2 … zk zp 0 0 0 0 Current situation z1 z2 … zk z p 1 1 1 1 New option z1 z2 … zk z p Louviere (1988) warned that one must be the minimum number of options that should cautious of fractional designs because a appear in each choice set is three. strictly additive utility function, known as the orthogonality assumption, underlies choice Second, extreme care is called for modelling. The orthogonality assumption regarding the levels and range of the means that choice modelling estimates only payment variable. Lareau and Rae (1989, the main effect of each attribute on the pp. 729–730) in an empirical study of the overall utility, assuming that all interaction contingent ranking technique warned that “if effects between attributes are zero. Thus, prices are too low, respondents order choice modelling questions should be options by focusing mainly on the designed to comply with the orthogonality environmental attributes, while if prices are assumption. Further explanation on the too high, respondents order options orthogonal experimental design is provided according to the price attribute.” This in Appendix A. warning is applicable to choice modelling studies. Respondents would choose an For designing a choice modelling option by focusing mainly on the questionnaire, a few other considerations environmental attributes if prices are too are required. First, the number of choices low and by focusing on the financial available in a choice set should be attribute if prices are too high.