Difference in level balance when mfederov algorithm is used

[VaishnaviP]

Dear Michiel,
I am extremely thankful for your regular help. I have a small doubt regarding dominant alternatives in the choice scenarios.

With the above corrected code, I have prepared different number of choice scenarios such as 24, 36 and 48 in total with 6 choice sets in each block. Although I didn't get strictly dominant alternatives (i.e., with all attributes of one alternative on a higher side compared to the other), but a considerable number of choice sets have dominant alternatives. For example, for alternative1 of a choice set, only 1 or 2 attributes out of 7 have lesser improvement levels compared to alternative 2 which is ultimately becoming dominant. In case of 48 choice scenarios, almost 40%-50% of them are dominant. Surprisingly, this has better d-efficiency compared to other number of choice sets 24 and 36. Of all, in 36 choice sets case, minimum of 33% dominant alternatives are present and in case of 24, it is 50%.

1. Should we remove such choice sets with dominant alternatives or what will be the remedy for this? Or is this normal to happen so?
2. On this basis, can I select 36 for my study?
3. Should the code be further changed?

Thanks,Vaishnavi

[Michiel Bliemer]

This happens because you are using (near) zero priors, so Ngene does not know to what extent each attribute contributes to utility. In choice models, only the total utility of an alternative matters, i.e. the same across all attributes, so there is no way to pick up the kind of dominance you are referring to with zero priors. Once you have informative priors this situation will likely improve. Why not do a pilot study with the design and not worry about these weakly dominant alternatives, estimate parameters, and generate a new design with informative priors?

I usually do not remove any choice tasks from the design because that may have unpredictable consequences (i.e., certain attribute levels may no longer appear or it could create high correlations or an unidentified model). If you do, then it is best to save the design and let Ngene evaluate the efficiency of the reduced design using ;eval.

Michiel