Introduction: the DCE method is used to measure the preferences of a set of (mainly) nominal features that make the research plan in the form of scenarios that are a combination of these features. The number of scenarios that the subject sees and assess is the result of all possible combinations of all possible features - and various DCE analysis algorithms, for example, a parial orthogonal plan or conjoint reduction allows to make only necessary scenarios to be presented to be kept to a minimum questions set. In other words, we get such a combination of features to evaluate in the form of ready-made scenarios that are the minimum set necessary to calculate the preference (utility) indicators for a full set.
Doubt: I have seen several works so far in which the research DCE plan was very complex and contained a very large set of features; even the aforementioned methods of limitation gave too large a set of minimal questions necessary to present to the subject to even have cognitive abilities to assess them. Therefore, the researchers decided to divide the (obtained orthogonal) minimum set into several subsets and present each one separately to a different group of subjects.
My question: since the question-reduction methods for DCE themselves assume that obtained set of questions is the minimum number of questions that the subject must see in order to get the correct results for all possible combinations of features (preferences/utility), is it not an error/incorret method to divide the minimum set into subgroups? That is - if the subgroups see only part of the minimum plan, are we allowed to correctly conclude about the full combination of features?
Also: I understand that it is technically possible to do it and obtain indicators (the statistical package will accept it), but is the assumption of consistency of the DCE result not violated here?
Thank You for all sugestions
As far as I understand, experimental design is key for implementing a DCE. Michiel Bliemer from Sydney is dedicating most of his research to that topic. Particularly, pushing the frontiers of Bayesian Efficient Design addresses the issue (or the doubt) you mentioned above. Attribute combinations and the construction of valid choice sets are subject to piloting and estimating priors.
Bliemer, M. C., Rose, J. M., & Hess, S. (2008). Approximation of Bayesian efficiency in experimental choice designs. Journal of Choice Modelling, 1(1), 98-126.
Bliemer, M. C., & Rose, J. M. (2010). Construction of experimental designs for mixed logit models allowing for correlation across choice observations. Transportation Research Part B: Methodological, 44(6), 720-734.
Bliemer, M. C., & Collins, A. T. (2016). On determining priors for the generation of efficient stated choice experimental designs. Journal of choice modelling, 21, 10-14.
I hope this goes into the direction you were asking for.
Cheers,
Chris
Thank You for that answers Cordula Hinkes
and Christoph Schulze . It took me some time to read these articles, understand them and read other articles they refer to. Although they brought a lot of knowledge, I don't think that's not an answer to my question.I will try to present my problem/doubt again. Suppose I have a DCE research plan 3x3x2x2 = 36 scenarios, let's call them from A1 to A36. Different tools for subplans (partial orthogonal and so on) can give me a smaller set, say only A2, A5, A8 ... A31. If they are well chosen, they retain the efficancy (i mean efficient plan) type of research plan. And we were at this point able to limit the full factorial plan to (for example) half of the scenarios.
And my doubt arises with idea of dividing mentioned partial subset it into more subsets. For example - researches concludes that even half od scenarios - from 36 to 18 is still too much for respondent, and decides to show half of the respondents half of the subset and to other half of the subjects the second half of the subset. If half of the respondents only see solutions A2, A5 and A8 (for example), then relative to the partial and efficient plan that was "unbiased", we get a plan in which factors are overestimated due to the incompleteness of the partial plan.
If the partial plan can be divided between the sub-sets, it means that it was not initially minimal. If the plan is minimal, then its division should not be included because it leads to a large errors in study.
It seemed to me that it would be enough just for a sufficiently large sample to randomly make new sets, but if it is not about randomly choising sub-subset every time, how to justify the above problem? How is this technically solved?
The indicated publications did not explain to me this relation between the minimum subset and more subsets.
I also dont understand maybe. I think Cordula answers the question. You are looking for blocks I believe. Assign respondents randomly to blocks. Use a large sample. Then pooling the data is fine under reasonable assumptions.
Dear Jens Rommel and Cordula Hinkes
- thank You for additional comments, now i understand. I confused the blocking procedure with the "manual" division of a set between subgroups of subjects. I now understand that one can keep a D-efficient scheme by creating blocks with software prepared for this. I got interested in NGene, but it probably won't be as comfortable for me as the R I use.Already i use R packages: AlgDesign and choiceDes and exploring dcm.design.cand function. It looks very promising. Do you have any tips or information on how to choose the number of blocks x comparisons x sets?
Can only second Cordula. No strict rules. In Transport Economics, there are studies in which respondents get 50+ choice sets, but the norm in Environmental Economics or Marketing is much lower. Most studies I know of do 6-8 choice sets as well (as stated by Cordula). But think of trade-offs. Maybe you can drop attributes or levels also if you would have to use to many blocks on a small sample.
Agree to what Cordula and Jens mentioned before. If you are into R, it might be worth it to look into the Apollo package. It was released quite recently by DCE experts and has many functions. Not sure if it does experimental design, but definitely purposeful for estimation of all kinds of models.
- Badges
- Science topic
- Similar topics
- Filing