Page 1 of 1

D-efficient design (7 Attributes, 2 are correlated)

PostPosted: Mon Nov 19, 2018 3:38 am
by SophieTsai
Hi everyone,

Hoping you could shed some light on the design problem I have below.

In my current research project, we would like to design a discrete-choice experiment for 7 attributes, with one caveat, two of the attributes are strongly correlated.

We decided to recreate the experimental design that was used before and published, in which, seven attributes were used and two attributes (say A and B) were highly correlated.

In the paper, the study team used a main effects D-efficient experimental design with 30 hypothetical paired profiles. According to one of the authors, 50% of the respondents answered questions composed of only A attribute plus the 5 attributes, the other half of the respondents answered questions with only B attribute plus the 5 attributes. And the seven attributes were analyzed together in the results.

Can anyone please let me know what kind of design is this and how to go about designing it? Thank you so much!

Many Thanks,
Sophie

Re: D-efficient design (7 Attributes, 2 are correlated)

PostPosted: Mon Nov 19, 2018 8:35 am
by Michiel Bliemer
You describe a design blocked in two parts in which six out of seven attributes are shown, where in the first block attribute B is left out and in the second block attribute A is left out. It sounds like a very specific case of a partial profile design in which each choice task contains a subset of attributes instead of the full profile.

If attributes A and B are indeed highly correlated then it may be difficult to estimate coefficients in a joint model with 7 attributes.

Creating such a design can be done by generating two blocks of choice tasks and calculating the joint Fisher information on the combined data set to estimate a model with all attributes. The ;fisher command In Ngene may be useful. Alternatively, you could create an external candidate set (in Excel) in which you create overlapping attribute levels, i.e. choice tasks in which A is the same across both profiles while B is different, and vice versa. You can then use the modified Federov algorithm, e.g. ;alg = mfederov(candidates = spreadsheetfile). In order to get blocks of the same size, you will likely need to impose some (soft or hard) attribute level balance constraints.

Michiel