This section examines the effect of preference uncertainty, strength of preference and presentation effects on valuation and forecasting. A central concern for capturing preference uncertainty in CVM as well as CE studies apart from getting further information on respondents’ preference level is the effect uncertainty can have on willingness to pay (WTP) estimate. Several studies have been found across these two valuation methods where the effects of preference uncertainty as well as the different methods of preference elicitation and calibration on the WTP estimate have been outlined.
Comparing preference uncertainty adjusted WTP across the literature with a dichotomous choice (DC) WTP estimate, Akter et al. (2008) expected the uncertainty adjusted WTP to be lower than the DC estimate as capturing preference uncertainty is considered to remove hypothetical bias which is thought to result in a higher WTP estimate. However, examining the results from different studies it was found that this expectation was not valid across all different elicitation and calibration methods. With the numeric certainty scale (NCS) elicitation method, almost all the studies that the authors reviewed found that using the different recoding methods, lower welfare estimates were obtained except for one study which obtained a higher estimate. With the polychotomous choice (PC), variable results were obtained across different studies.
Comparing PC, NCS and the composite certainty scale (CCS – which combines verbal expressions with numerical and graphical interpretations), (Akter and Bennett, 2010) found that a greater number of ‘yes’ responses were obtained with the PC format compared to the DC method while CCS obtained a greater number of certain responses. Comparing the certainty calibrated mean WTP with that obtained from the DC WTP estimate it was found that all calibrated mean WTP were lower (except in one case) than the DC mean WTP estimate. With the PC method, the mean WTP was found to be 117% higher than the DC WTP estimate when ‘maybe yes’, ‘probably yes’ and ‘definitely yes’ were recoded as ‘yes’ while in other cases where certainty cut-off points were included, the mean WTP reduced
from 3-83% compared to the DC WTP estimate. The lowest change in WTP estimate compared to the DC method was found in case of the CCS method (where a reduction of 3-32% was obtained) due to the relatively small calibration scale used compared to the other methods. Thus, this study showed that different methods of preference elicitation along with different preference calibration methods can cause significant variation in the WTP estimate.
In another study where respondents were asked a double bounded (DB) WTP question on different starting bid amounts followed by an open ended (OE) question which was subsequently followed by a five point scale (extremely unlikely, fairly unlikely, not sure, fairly likely and extremely likely) on respondents’ likelihood of paying on the OE WTP question, Akter et al. (2009) found that the mean WTP from DB question without uncertainty was 23 Euro/flight while from the OE WTP without uncertainty was 43 Euro. Moreover, a substantial variation in the WTP amounts was obtained based on the different recoding methods used. The authors also found that the OE WTP values were significantly clustered around the five point scale categories which implied that not accounting for respondent uncertainty resulted in increased estimation errors.
Alberini et al. (1997) and Alberini et al. (2003) emphasised the effect of recoding method on the WTP estimate using the MBDC response format. The two studies revealed that based on the analytical and recoding method used, the WTP estimate varied substantially. Wang (1997) found that treating ‘don’t know’ as ‘no’ or deleting it lowered the mean WTP compared to the threshold models. The effect of different recoding methods as well as different estimation methods was also observed in Broberg and Brannlund (2008) who obtained a substantially high WTP amount as all levels of uncertainty were recoded as ‘yes’ responses. Comparing different bid orders Alberini et al. (2003) found that the descending order of the bid panel yielded higher mean welfare estimates while including the uncertain responses in the model also substantially increased the welfare estimates.
Examining the effect of uncertainty level on WTP, Boman (2009) found that individuals who are uncertain have a greater propensity to accept higher bids than his/her maximum WTP. This finding has been previously stated by Hanemann et
al. (1996) who found that the preference uncertainty parameter in the SP study is quite high indicating that this parameter captures much of the variation in the individual WTP and hence the mean WTP amount when accounting for preference uncertainty is lower than that obtained from the conventional model which does not incorporate uncertainty information. Li and Mattsson (1995) found that the overall mean WTP estimate was reduced by about six times compared to the model without preference uncertainty implying that not accounting for preference uncertainty can cause serious upward bias of the WTP estimate.
Comparing the MBDC format with different recoding approaches along with the payment card and dichotomous choice, Welsh and Poe (1998) found that very substantial difference in the mean WTP estimate was obtained based on the different specifications of the switching interval. When the lower end of the switching interval was chosen as the highest amount at which the respondent chose
‘definitely yes’, the MBDC method gave a lowest mean WTP of $16.70 while when the lower end of the switching interval was taken as the highest amount at which the respondent chose ‘probably yes’, a mean WTP of $39.56 was obtained and with that for ‘not sure’, a highest WTP of $92.96 was obtained. The mean WTP obtained from the payment card method was $36.64 while that obtained from the DC was $98.40. Comparing the results obtained from the different elicitation methods as well as the specification of the switching interval for the MBDC, it can be seen that the mean WTP estimate is highly sensitive to both the issues.
Moreover, results from the DC yield a higher mean WTP compared to the MBDC method with the ‘not sure’ model which reiterates the previous findings that not accounting for preference uncertainty can result in an upwardly biased WTP estimate.
By asking respondents whether they are ‘definitely sure’ or ‘probably sure’ of paying a particular price which was followed by a numeric certainty scale where 0 corresponds to ‘not sure’ and 10 corresponds to ‘very sure’, Blomquist et al. (2009) sought to find a point on the numeric scale which provides the same WTP estimate as that obtained from the ‘definitely sure’ response for three different disease management programs. It was observed that the same mean WTP estimate as obtained from the ‘definitely sure’ response was obtained for numeric scale 9.9, 10
and 9.0 for each of the programs. While the authors aimed to find the numeric certainty level that relates to the ‘definitely sure’ response, the numeric level given by the respondent can be highly subjective as well as context specific and hence, this finding cannot be extrapolated to other studies.
Brouwer (2009) found that using a post-decisional NCS question where respondents were asked to indicate on a 100 points scale (with 10 percent intervals) whether they are ‘not at all certain’ or ‘completely certain’, respondents who are 100% certain of their WTP amount are willing to pay more compared to respondents who are not 100% certain. Thus the author concluded that not accounting for preference uncertainty could result in an over or underestimation of the welfare measure based on the presence (or absence) of preference uncertainty.
Moreover, as the statistical efficiency of the welfare estimate was found to be lower for respondents who were uncertain, the author concluded that accounting for self-reported uncertainty in econometric models leads to smaller and less precise welfare estimates.
Loureiro and Loomis (2008) asked a follow-up certainty question where respondents were asked to indicate their level of certainty on a scale from 1 (not sure) – 10 (totally sure) on a referendum WTP question. Analysing the data by classifying it into two major segments of respondents – the ‘uncertain’ and the
‘certain’ class using the finite mixture model, the authors found that the WTP estimate conditional on the segment was affected such that the ‘certain’ segment had a fairly higher estimate than that obtained from the ‘uncertain’ segment with the mean WTP of individuals belonging to the ‘certain’ group being Euro 82.14 while for those in the ‘uncertain’ group being Euro 54.08.
Examining the results obtained from the above studies it can be seen that while not accounting for preference uncertainty (such as in the DC method) can result in upwardly biased WTP estimate, ‘certain’ respondents are willing to pay more than
‘uncertain’ respondents while the modelling method employed can have a significant effect on the WTP estimate obtained as can be seen in Welsh and Poe (1998).
Shaikh et al. (2007) has shown that based on the analytical model used, the preference uncertainty adjusted WTP can be higher or lower than the standard model. For the Weighted Likelihood Function Method (where initial DC question was followed by a NCS method on a 0-100 interval and responses were recoded based on the certainty score), asymmetric uncertainty model (where using a scale from 1-10 respondents were asked a certainty follow-up question and all ‘yes’
responses were recoded as ‘no’ if the respondent was not completely certain) and fuzzy model (where the post-decisional confidence of a response was used to determine the membership values of the WTP and willingness not to pay (WNTP) fuzzy sets while the intersection of the estimated fuzzy sets determined the membership values of the ‘comfort’ level of the associated welfare estimates) approaches, incorporating respondent uncertainty was found to lower the WTP estimate while in the case of the Random Valuation Model (an individual’s value of an amenity was considered as a random variable with unspecified probability distribution and thus error was incorporated into the model) and Symmetric Uncertainty Model (based on the certainty score given by the respondent, the responses were recoded - thus ‘yes’ with 60% certainty was recoded as 0.6), the opposite result was obtained.
Thus, studies in CVM have revealed significant effects on the WTP estimate and hence on attribute valuation based on the elicitation and calibration methods used.
While comparisons of alternative preference elicitation methods is relatively more examined with the contingent valuation literature, fewer studies can be found within CE which have compared the effect of alternative preference elicitation forms on valuation and forecasting. A major area of interest in CE however has been the comparison of different modelling techniques on model fit as well as attributes valuation.
Within choice experiments, in a comparison of the effect of different treatments of the uncertain responses by recoding and eliminating them, Koseinius (2009) found that compared to the base model, recoding the uncertain responses to business-as-usual or the best available alternative decreased the mean WTP while eliminating it increased the WTP estimate. Compared to the benchmark model where respondents’ uncertainty is ignored, Lundhede et al. (2009) found that models with
uncertain choice eliminated slightly increased WTP while models with asymmetrically recoded (uncertain choices were recoded as status-quo) method decreased the total WTP for a change from the current situation. In case of symmetric recoding (uncertain choice was recoded as the best alternative different from the one chosen) as well as for models with certainty level (incorporated through a scale parameter), the effect on WTP was insignificant. Applying the five point Likert scale where respondents were asked to make a binary choice as well as indicate their strength of preference, Swallow et al. (2001) compared results obtained from the binary and ordered logit models to find that the ordered logit model outperformed the binary logit model in estimating the WTP.
In another application of the five point Likert scale in CE, Whelan and Tapley (2006) adapted different analytical techniques to model the data. The authors combined the ‘definitely’ and ‘probably’ levels while eliminating the ‘no preference’ response to form a binary choice, converted the five point scale to a three alternative choice by including the ‘no preference’ responses but combining the ‘definitely’ and ‘probably’ levels as well as analyse the five point scale through ordered logit. The authors noted that though comment on overall model fit was difficult to make as the dependent variable varied across the different specifications, the scale of the model coefficients decreased as the strength of preference was incorporated into the model, implying that the statistical precision of the estimation increased as preference uncertainty levels were incorporated though relative attribute valuations across the different techniques were found to be quite constant.
While comparison of alternative preference elicitation format in choice experiments on valuation and forecasting is extremely rare, the above studies have revealed that preference uncertainty can have substantial effect on valuations and methods of elicitation as well as calibration can play a significant role in the WTP estimate.
While forecasting is a significant factor in marketing and transport studies which can be affected by the method of preference elicitation, a major concern in the environmental literature has been the effect of preference uncertainty on valuation.
Comparison of different elicitation formats in the contingent valuation literature reveals that the method of preference elicitation does affect valuation and the WTP
estimate. In terms of forecasting, the method of recoding and the treatment of uncertain responses can play a significant role in the forecast.
In a study examining alternative catchment management strategies where an attribute was defined as ‘species lost’ or ‘species present’, Kragt and Bennett (2009) found that significant differences in the implicit prices between the two versions were obtained with higher valuation obtained when the attribute was framed as ‘lost’ though with lesser precision of the WTP estimate, compared to when it is described as ‘present’.
In order to examine how framing affects the valuation of mortality risk reduction, Rheinberger (2009) conducted a split-sample survey where one half of the respondents were asked their WTP for reduction of up to 16 fatalities per year with reference to the 7.5 million residents of Switzerland while the other half was offered an identical choice where avoided fatalities was reframed in reference to the 500 road fatalities that occur in the country. Comparing the value of statistical life (VSL) across the two framing methods, it was found that the VSL was CHF 8.44 million where risk reduction was framed in reference to road fatalities while it was CHF 7.12 million when it was framed in reference to the residential population.
However, while the values obtained were statistically significant, large standard error implied that the framing effect on VSL values could be random. Analysing the effect of different framing effects on the different sample based on the interaction with socio demographic variables, the author found that the different sub-samples adapted different techniques to form perceptions of risk reduction which further affected their choice though the effect on valuation of the traffic safety programs from the different framing methods was not statistically significant.
In another study conducted by Howard and Salkeld (2009) to compare the framing effects within health context, potential benefits and harm of screening tests were presented in both positive (number of cancer/large polyps found) and negative (number of cancer missed) terms. Three attributes – cancers found, large polyps found and test prediction where thus framed as: a) number of cancers found and missed, b) number of large polyps found and missed and c) people correctly reassured that they do not have cancer and unnecessary colonoscopies. Comparing
the parameter estimates of cancer found with cancer missed when the specificity attribute of test accuracy was expressed as ‘reassured’ or ‘unnecessary tests’, the results found that with the framing ‘reassured’, the framing did not significantly affect the valuation of the attribute while in the case where it was specified as
‘unnecessary test’, the sensitivity attribute was found to significantly affect valuations. The same results were obtained when framing effects of cancer test were analysed on the number of polyps found or missed. Moreover, the WTP for test attributes was found to be significantly affected by framing. While the WTP for one extra cancer found was $11.45, for one fewer cancer missed was $9.81.
Thus, the study revealed that attribute framing significantly affected respondents’
WTP.
Though some of the above studies have focused on framing effects as an extension of the gain-loss asymmetry literature found on risks (Kahneman and Tversky, 2000), it is nonetheless significant to observe empirical evidence of the theory especially as applied within CE. The other studies have also revealed that different framing methods can affect attribute perception as well as valuation. While the above studies have focussed on the effect of framing on valuation, an important aspect that has been rarely examined is the effect of framing on respondent’s level of preference uncertainty. The examination of this aspect thus encompasses one of the novel aspects of this research.
As uncertainty can be defined in different ways which can have an implication on the method of analysis employed, the following section will outline the different forms of uncertainty representation along with the relevance of this classification in the research.
3.4.1 Uncertainty Representations
Besides the work conducted in utility theory in context to risk and uncertainty as outlined in the previous sections, a more general classification of uncertainty can also be found in the literature. When employing analytical methods different from probabilistic measures, this classification becomes especially pertinent as will be observed later in the section.
An interesting classification of uncertainty is given in Parsons (2001) who uses Smithson’s and Smets’ taxonomy (which classifies sources of ‘ignorance’) to characterise types and sources of uncertainty. Based on Smithson’s taxonomy, three causes of uncertainty specified are vagueness, probability and ambiguity.
While uncertainty due to probabilistic causes can be more generally viewed as risk, uncertainty from vagueness arises from fuzziness and non-specificity (Parsons, 2001). Based on Smets’ taxonomy, the state of ‘ignorance’ is classified into uncertainty and imprecision. In this case, while uncertainty encompasses whether a state is subject to randomness or is believable, it is imprecision that results from fuzziness (Parsons, 2001). This understanding of uncertainty is especially pertinent to the thesis. While most choice and behavioural decision theory literature focuses on uncertainty as associated with probabilistic risk, the interpretation of uncertainty in this research arise from vagueness and ambiguity, along with the interaction between the two.
Vagueness can be defined as something that is not well-defined while ambiguity is caused to exist due to lack of information that distinguishes between two alternatives (Parsons, 2001). These definitions hold significant insight in understanding the concept of uncertainty as pertaining to this research.
In case of this research which aims to apply the location and linguistic methods of attribute representation, the concepts of vagueness and ambiguity are especially interesting. Vagueness, the quality of being poorly defined, is a rather subjective term as it is based on how well the respondent relates the attribute to the method of its representation. While instinctively, the term might arouse negative
In case of this research which aims to apply the location and linguistic methods of attribute representation, the concepts of vagueness and ambiguity are especially interesting. Vagueness, the quality of being poorly defined, is a rather subjective term as it is based on how well the respondent relates the attribute to the method of its representation. While instinctively, the term might arouse negative