How psychological distance of a study sample in discrete choice experiments affects preference measurement: a colorectal cancer screening case study

Abstract

Purpose

The purpose of this study was to investigate to what extent the outcomes of a discrete choice experiment (DCE) differ based on respondents’ psychological distance to the decision at hand.

Methods

A DCE questionnaire regarding individuals’ preferences for genetic screening for colorectal cancer (CRC) within the Dutch national CRC screening program was created. The DCE contained nine D-efficient designed choice tasks and was distributed among two populations that differ in their psychological distance to the decision at hand: 1) a representative sample of the Dutch general population aged 55–65 years, and 2) a sample of Dutch individuals who attended an information appointment regarding colonoscopies following the detection of blood in their stool sample in the CRC screening program. The DCE consisted of four attributes related to the decision whether to participate in genetic screening for CRC: 1) risk of being genetically predisposed, 2) risk of developing CRC, 3) frequency of follow-up colonoscopies, and 4) survival. Direct attribute ranking, dominant decision-making behavior, and relative importance scores (based on panel MIXL) were compared between the two populations. Attribute level estimates were compared with the Swait and Louviere test.

Results

The proportion of respondents who both ranked survival as the most important attribute, and showed dominant decision-making behavior for this attribute, was significantly higher in the screened population compared to the general population. The relative importance scores of the attributes significantly differed between populations. Finally, the Swait and Louviere test also revealed significant differences in attribute level estimates in both the populations.

Conclusion

The study outcomes differed between populations depending on their psychological distance to the decision. This study shows the importance of adequate sample selection; therefore, it is advocated to increase attention to study sample selection and reporting in DCE studies.

Keywords:

• discrete choice experiment
• preferences
• stated preferences
• sample
• psychological distance
• genetic screening

Supplementary material

The role of the scale parameter

According to the method proposed by Swait and Louviere,¹ the role of the scale parameter when comparing two data sets can be determined by confirming that β₁ = β₂ and µ₁ = µ₂, where β represents the attribute level estimates and µ represents the accompanying standard error. In order to do so, a four-step specific procedure that contains two hypotheses needs to be followed.

1. An MNL model is fitted within both the population data sets separately. For both those models, the log likelihood is collected (L₁ and L₂).¹

2. The attribute level codes in one of the data sets are multiplied by a trial version of the expected scale factor. The two data sets are then combined and the log likelihood for this pooled data set is determined (L_µ).¹ This routine is repeated for different trial versions of the expected scale factor. This second step will result in a list of log likelihoods of which 1 represents the model with the best fit.

3. The following hypothesis is tested: l_A =−2^*[L_µ−(L₁ + L₂)] < χ² with (K+1) degrees of freedom.¹ In this hypothesis, the log likelihood of the optimal model in step 2 (while accounting for a specific scale parameter) is compared with the log likelihoods of the two separate models from step 1. This value is compared with the χ² value of the number of parameters in the model plus 1.¹ If this hypothesis is rejected, the differences in attribute level estimates between both the models are statistically significant. If this hypothesis is accepted, the attribute level estimates of both the models do not differ significantly and testing for the scale factor can be continued.

4. Both the data sets from step 1 are pooled and an MNL model is fitted. The log likelihood of this model (L_p) is then compared with the log likelihood of the optimal model found in step 2 (L_µ), using the following hypothesis: λ_B = −2[L_p − L_µ] < χ² with (K+1) degrees of freedom.¹ If this hypothesis is rejected, the scale parameter is statistically different from 1 and differences between the two models of both the data sets are explained by scale. If this hypothesis is accepted, the models of both the data sets are equal and the scale parameter does not differ from 1.

Reference

• SwaitJLouviereJThe role of the scale parameter in the estimation and comparison of multinomial logit modelsJ Mark Res1993303305314
   Web of Science ®Google Scholar

Acknowledgments

Authors would like to thank Josepine Fernow of the Centre for Research Ethics & Bioethics from Uppsala University for her assistance with writing this manuscript. This paper was presented at the International Choice Modelling Conference and at the International Academy of Choice Modelling Conference as an oral presentation with interim findings. The abstract was published on the International Choice Modelling Conference website (http://www.icmconference.org.uk/ index.php/icmc/ICMC2017/paper/view/1071) and printed within an abstract book (attendee access only) for both the conferences.

Disclosure

The authors report no conflicts of interest in this work.