A second set of analyses was conducted using the best-minus-worst scores (Marley and Louviere, 2005 ; Flynn, 2010a (link)). Within the OMEP design, each attribute level appeared on four occasions within the 16 scenarios. It could therefore have been picked as best up to four times and as worst up to four times for each person. These best-minus-worst scores are calculated for each attribute by determining the number of times that a person picked an attribute level (for example, ‘I am able to be completely independent’) as best and subtracting from that the number of times that they picked it as worst. Scores can therefore range from −4 (never picked as best and always picked as worst) to +4 (never picked as worst and always as best). Scores for respondents provide an immediate indication of which attribute levels they value. Figure 1 gives an example of the scores for one individual.
It is important to adjust for heterogeneity in both preference and variance scale at the level of individual respondents (Louviere et al., 2000 ; Swait and Adamowicz, 2001 ; Louviere et al., 2002 ; Fiebig et al., 2010 ). Variance scale is concerned with how consistent individuals are in making their choices: some individuals are more consistent and others are less so. If this is not adjusted for, people may be thought to have different preferences where, in fact, their preferences are similar but they are just less consistent in making them. Although this makes the analysis considerably more complex, it is vital in estimating a set of population values (Flynn et al., 2010 (link)), because not accounting for this sort of heterogeneity leads to bias in the mean estimates obtained from limited dependent variable (such as probit and logit) models (Yatchew and Griliches, 1985 ). A series of cluster analyses based on functions of the best-minus-worst scores was conducted, with further details provided in Appendix 2 (Supporting Information). These did not provide the final capability scores but were essential in ensuring that the main scale-adjusted latent class analyses (SALCs) did not give spurious solutions, such as finishing at a local maximum of the likelihood function.
It is important to adjust for heterogeneity in both preference and variance scale at the level of individual respondents (Louviere et al., 2000 ; Swait and Adamowicz, 2001 ; Louviere et al., 2002 ; Fiebig et al., 2010 ). Variance scale is concerned with how consistent individuals are in making their choices: some individuals are more consistent and others are less so. If this is not adjusted for, people may be thought to have different preferences where, in fact, their preferences are similar but they are just less consistent in making them. Although this makes the analysis considerably more complex, it is vital in estimating a set of population values (Flynn et al., 2010 (link)), because not accounting for this sort of heterogeneity leads to bias in the mean estimates obtained from limited dependent variable (such as probit and logit) models (Yatchew and Griliches, 1985 ). A series of cluster analyses based on functions of the best-minus-worst scores was conducted, with further details provided in Appendix 2 (Supporting Information). These did not provide the final capability scores but were essential in ensuring that the main scale-adjusted latent class analyses (SALCs) did not give spurious solutions, such as finishing at a local maximum of the likelihood function.
Free full text: Click here
