Hello,
I am writing my thesis on crowdsourcing platforms, how to manage and design.
We have run a conjoint analysis to test the crowd's preferences on 4 design attributes. Also, we have run a survey questioning the crowd about their willingness to participate, for that we used the 1-7 likert scale.
Now, the conjoint data was very insightful to know which attribute is more important, but I am interested in knowing if those preferences also meant a higher participation in crowdsourcing contests.
I want to run a regression analysis using Stata or SPSS, wanted to know if there is any way to translate the conjoint data into a likert scale that I can include in my models.
The conjoint study was done on QuestionPro.
IMO Not unless you have the cards ranked in preference order for each respondent. Then you might have a chance. Good luck, D. Booth
Hello Milad,
The Conjoint Analysis template in Question Pro uses a rating scale (7="Definitely will buy", 6="Probably will buy", etc.) in response to a set of options that are made up of combinations of your four design attributes. With four attributes, each at two levels (yes/no, high/low, etc) then you have 16 profiles in a full-factorial, which would be reduced to eight profiles in a main-effects-only fractional-factorial design. So each of your respondents gave you 8 ratings for your different profiles. Something like this from each person:
- a b c d Rating
- -1 -1 -1 -1 4
- 1 -1 -1 1 5
- -1 1 -1 1 3
- 1 1 -1 -1 7
- -1 -1 1 1 4
- 1 -1 1 -1 6
- -1 1 1 -1 2
- 1 1 1 1 1
Is that correct?
You already have a regression model here. Conjoint Analysis, of the type you have used, is simply multiple regression applied to experimental design data. (The main difference is the jargon - it kept a lot of overpaid consultants happily occupied throughout the 1980's and 1990's)
Rating is your Dependent Variable, and your four attributes are the Independent Variables.
- If you run the regression at the level of the individual person, then you have a total 8 degrees of freedom total, and five df for regression, leaving only 3 df for error. In this case, you'd ignore the significance figures for the regressors (attributes) and use Adjusted R2 for goodness of fit. From my experience, I throw out any respondents who have individual-level Adjusted R2 less than 0.70.
- Alternatively, you can run your regression with all respondents together - this assumes that each respondent is a random variation around a group-level utility function. Unless your respondents are very similar to each other with respect to crowdsourcing preferences, expect rather low regression coefficients.
- Alternatively, you can run a cluster analysis on the rating scales for all of your respondents and find groups of respondents (market segments) who have similar patterns of preferences for crowdsourcing, and then run group-level regression models for each cluster.