We have conducted a conjoint survey of people's preferences when shopping for salmon. We have defined our salmon products in terms of "price" (20p steps from £3.60 to £5), "eco-certification" (none, green statement and certified) and "Brand" (high, medium and low) and have produced a csv file of our results with "eco-certification" and "Brand" coded as binary variables.
We now wish to carry out a conjoint analysis on this data, to derive a model in the form: probability (choice) = a* 'price' + b* 'green statement' + c* 'certified' + d* 'high' + e* 'medium' + error
'none' and 'low' are not included in the model as they are taken to be our base variables. We then need to divide all our coefficients by 'a' to derive a 'willingness to pay' estimate, for which we will need to calculate the associated standard error. Finally, we have a load of associated demographic data for each respondent: gender, age etc. We want to test whether males and females respond differently for example. To do this, we know that we need to either generate 2 separate models, or, the method we intend to pursue, include gender as an interaction term with each of the attributes in the model above.
We are working in 'R', but having read through the conjoint package information it sounds like each respondent has to have been shown every possible choice, whereas each of our respondents chose between 3 different products for 4 repeats (with different sets of 3 products). I have also read through the m.logit pdf, but cannot work out from this what function I need to use, or whether our data is currently in the right format.
Please could you advise us a) how to run the initial conjoint analysis and derive the stated model, b) how to calculate the SE associated with the derived WTP estimate, and c) how to then run the conjoint analysis again with the added interactions from the demographic variable? If it would be loads easier in STATA or another program, we are open to switching, but as we are all familiar with R we would prefer to use that.
Best Answer
The mlogit package can analyse this data, but you need to make sure that the data is transformed to the appropriate shape (either long or wide, depending on how your data look). See the mlogit.data function on page 24 of the documentation. Once you've specified the shape of the data, you can use the mlogit function. This function allows for respondents to have seen different alternatives in each task.
As for your other question, there are many way to test whether males and females had different preferences but an interaction term is likely to be the most straightforward.