Hi all,
I'd like to analyze data of a binary discrete choice model. Respondents, sampled in four districts of three different towns, were administered three conjoint experiments, each with five attributes and up to four levels. I would like to make the most of individual-level characteristics (gender, age, etc.) while accounting for possible variations across towns and district. I'd like to be able to make statements such as "On average, respondents in Subgroup1 are X% more likely than demographically similar respondents in Subgroup2 to choose an alternative containing attribute A of level 1, all else equal, and the difference in likelihood is statistically significant."
I've been navigating the net and found different approaches and was hoping to get some guidance on the right one to use
Stata conjoint command: allows to compute the Average Marginal Component Effects (AMCE) and Marginal means, as well as Conditional Marginal Means (as CAMCE doesn't seem to be the right approach to study subgroup preferences (Leeper, T.J., Hobolt, S.B. and Tilley, J., 2020. Measuring subgroup preferences in conjoint experiments. Political Analysis, 28(2), pp.207-221)
The paper suggests to use a number of pairwise omnibus tests to investigate statistical difference between subgroups at different attribute level by performing a nested model comparison of two models: one with the attribute, the second with additional interactions between the subgroup covariate and the feature. However, the conjoint command doesn't seem to allow for interactions, control variables or multilevel stratification
Multilevel mixed-effects logistic regression: using the Stata command melogit, which would allow me to account for the stratified sampling (as opposed to mixlogit).
though I'm not sure where to add intragroup correlation at the respondent level (I have three observations per respondent) as the following code doesn't seem to work
I'm also wary of using a different approach given AMCE/MM seem to be the standard in conjoint analysis.
Finally, in addition to the binary variable indicating whether the alternative was chosen, the data also contains information on levels of satisfaction for that alternative should it be offered (five-point Likert-scale), as well as attributes most and least preferred for each alternative. I guess to better compare the results, it would make sense to use the multilevel mixed-effects logistic regression for the binary outcome, and command meologit when using the likert scale variable as a dependent variable:
Thanks for your help
I'd like to analyze data of a binary discrete choice model. Respondents, sampled in four districts of three different towns, were administered three conjoint experiments, each with five attributes and up to four levels. I would like to make the most of individual-level characteristics (gender, age, etc.) while accounting for possible variations across towns and district. I'd like to be able to make statements such as "On average, respondents in Subgroup1 are X% more likely than demographically similar respondents in Subgroup2 to choose an alternative containing attribute A of level 1, all else equal, and the difference in likelihood is statistically significant."
I've been navigating the net and found different approaches and was hoping to get some guidance on the right one to use
Stata conjoint command: allows to compute the Average Marginal Component Effects (AMCE) and Marginal means, as well as Conditional Marginal Means (as CAMCE doesn't seem to be the right approach to study subgroup preferences (Leeper, T.J., Hobolt, S.B. and Tilley, J., 2020. Measuring subgroup preferences in conjoint experiments. Political Analysis, 28(2), pp.207-221)
Code:
conjoint binary attribute1 attribute2 attribute3 attribute4 attribute5, estimate(amce) id(id) subgroup(subgroup)
Multilevel mixed-effects logistic regression: using the Stata command melogit, which would allow me to account for the stratified sampling (as opposed to mixlogit).
Code:
melogit binary i.attribute1 i.attribute2 i.attribute3 i.attribute4 i.attribute5 cov1 cov2 i.district, nofvlabel || town: i.district, covariance(unstructured)
Code:
melogit binary i.attribute1 i.attribute2 i.attribute3 i.attribute4 i.attribute5 cov1 cov2 i.district, nofvlabel || town: i.district|| id:, vce(cluster id)
Finally, in addition to the binary variable indicating whether the alternative was chosen, the data also contains information on levels of satisfaction for that alternative should it be offered (five-point Likert-scale), as well as attributes most and least preferred for each alternative. I guess to better compare the results, it would make sense to use the multilevel mixed-effects logistic regression for the binary outcome, and command meologit when using the likert scale variable as a dependent variable:
Code:
meologit likert i.attribute1 i.attribute2 i.attribute3 i.attribute4 i.attribute5 cov1 cov2 i.district, nofvlabel || town: i.district|| id:, vce(cluster id)
