Appendix 3. 1 Proof of the theorem in section 3,2,2
4. DATA AND ESTIMATION PROCEDURES
4.1 REVEALED AND STATED PREFERENCE DATA
4.1.2 Stated preference data,
Since Stated preference data is the result of a controlled experiment it can overcome most of the difficulties related to RP data. In particular:
1) The ranges of the alternative attributes can be extended to values not obseiwed in the market.
2) Since the experimental design is controlled by the researcher multicollinearity can be avoided. This will improve the precision of the parameter estimates.
3) There is no measurement error in the data as the attributes are specified by the researcher (there may however be differences in perception of the attribute values).
4) The choice-set is pre-specified by the researcher.
5) The experiment can include attributes and alternatives that do not exist in the market at the present.
SP data also have some benefits that are not directly related to the problems with RP data outlined above. An advantage of the SP methodology is that it is both feasible and common to present several choice tasks to the respondents in the SP survey. As a result each sampled individual provides more information about his or her preferences compared to RP data, which typically consist of one observation per respondent (with the exception of costly travel diaries that follow respondents over a period of time). Thus collecting SP data is in general more efficient than collecting RP data.*
A further difference between RP and SP methods is that in the SP methodology there aie several ways of eliciting the respondents’ preferences. The method that most closely resembles the choice process observed in the market is to instruct the respondent to choose her preferred alternative. This approach is referred
to as the stated choice (SC) method in the stated preference literature. Alternative approaches are the rank and rate methods in which the respondents rank the available alternatives or rate the alternatives following a given semantic scale (a typical question would be “On a scale from 1 to 10 how do you rank alternative A?”). The rank and rate methods collect more information about the individuals’ preferences compared to the SC method. There are, however, also some drawbacks to the rank and rate approaches (see Willumsen and Ortuzar, 2001 for a discussion). In the SP application in chapter 7 the individuals were asked to choose their preferred option and the following discussion will therefore concentrate on this approach.
As previously mentioned the main concern when it comes to the use of SP data in modelling choice behaviour is that the choices observed are hypothetical. As a consequence, SP data does not in general depict the market equilibrium and cannot easily reflect changes in personal constraints (e.g. work location, income and information availability). Some critics have gone as far as claiming that SP data have no value, since “hypothetical questions result in hypothetical answers”. This is clearly an exaggeration given that a growing number of studies focusing on the external validation of SP models (see chapter 5 for a review) suggest that a well-designed SP design can elicit preferences similar to those observed in the market. On the other hand the fact that SP data are a result of a hypothetical choice situation should not be ignored, and the SP questionnaire should be carefully designed in order to reduce the likelihood of bias in the responses. Much effort has been devoted to identify the sources of bias in the choice variable that may be present as a result of the
' It should be pointed out, however, that some recent studies have collected RP data using new data collection techniques such as GPS, which makes it possible to collect several observations per respondent given access to GPS technology.
hypothetical choice situation (see, for example, Fowkes and Preston, 1991). The sources of bias include:
1) Policy bias. Respondents may be inclined to answer strategically in order to achieve tlieir desired policy response. The goal of the researcher is to make the experiment sufficiently complex to make it difficult for the respondents to bias their answers in order to influence the results of the study (and hence the policy recommendations derived fi-om the study) in a straightforward manner.
2) Justification bias. Respondents may choose a particular alternative in order to justify their current behaviour. Justification bias is difficult to identify, especially since it resembles choice inertia, which is congruent with actual choice behaviour.
3) Self selectivity bias. It is possible that the characteristics of survey respondents differ from those of the overall sample. This is especially likely to be a problem if the response rate of the survey is low. It should be noted that this type of bias may be present in surveys of all types, not just SP surveys. In the case of an SP experiment designed to forecast the demand for a new seiwice there is clearly more incentive for likely users of the service to respond.
4) Non-commitment bias. The respondents to the survey are not committed to behave in the way that they have responded. This is related to the policy and justification bias discussed above.
In addition the SP experiment might suffer fi*om factors such as learning (learning effects) and boredom (fatigue effects) (McFadden, 1986). In the presence of such effects preferences are unstable over the sequences of choices performed by the individual, which may lead to biased parameter estimates. It has been shown that the likelihood of learning and fatigue effects increases with the complexity of the experiment (Sælensminde, 2001). Consequently there is a trade-off between reducing the likelihood of response bias by making the experiment sufficiently complex, and reducing the potential for learning and fatigue effects by making the design relatively easy to complete. It should be noted that the potential bias due to learning and fatigue effects can be reduced by presenting the choice scenarios to respondents in a randomised order.
In spite of these difficulties the use of stated preference data should not be readily dismissed given its strength in forecasting changes in behaviour by incorporating a wider range of attribute levels as well as having the flexibility to introduce new alternatives and attributes. Since it is evident that both RP and SP data have their advantages there has been a growing interest in combining the different types of data to provide more robust parameters for the choice model. We will describe two methods for combining the two data types in section 4.4.
Table 4,1. Comparison of RP and SP data
RP data SP data
Preference Choice behaviour in actual market. Cognitively congruent with actual behaviour
Preference statement for hypothetical scenarios. May be cognitively incongruent with actual behaviour
Alternative
s non-existing alternatives are not Actual alternatives. Response to observable
Generated alternatives. Can elicit preference for new (non-existing) alternatives
Attributes May include measurement eirors Correlated attributes
Ranges of attributes are limited
No measurement errors
Multicollinearity can be avoided by design
Ranges of attributes can be extended
Choice set Ambiguous in many cases Pre-specified Number of
responses responses from an individualDifficult to obtain multiple Repetitive questioning is easily implemented Response
format Preference information available is “choice” Various response formats (e.g., choose one, ranking, rating) aie possible
Source: Morikawa, 1994