choice-metrics.com

Good day.

I am not sure this is a very intelligent question, but I am not sure how to proceed on something. As we don't have priors for our study, I generated the following design to generate priors. The idea was to run two focus groups / small pilot studies (one block for each):

Design
;alts = alt1,alt2,sq
;rows=12
;block=2
;eff=(mnl,d)
;model:
U(alt1) = b2.effects[0|0|0]*dist[8,12,18,5] + b3.effects[0]*size[144,64] +b4[0]*price[-5,5,10,15,25,0] /
U(alt2) = b2*dist + b3*size +b4*price
$

I used block 1 for the first focus group and was planning to use block 2 for the second focus group. Since focus group 1, however, we have decided to increase the number of choice questions from 6 to 8 (so rows = 16) and possible want to change the levels of the price attribute. My dilemma now is how to proceed with focus group 2. Is it critical to use block 2 from our old design above if I want to get priors? Or could I modify the design slightly as mentioned and then still pool the results to estimate priors?

Are the survey results from a design even valid if I only used one of the two blocks?

Any advice would be much appreciated.
Thanks!

A quick answer would be: you are fine pooling data together to obtain the priors, and also only using one block.

The main thing to remember is, that ANY design will eventually give you the true parameter estimates, but some designs find them more quickly (i.e., are more accurate). So you do not need to use the second block if you have new insights. The reason that people have put a large value on blocking in the past is for orthogonality reasons. Since orthogonality is not much of a concern in discrete choice models, it does not matter too much if you do not use the entire design (although you may loose a bit of efficiency).

Theoretically, there could be a slight problem with respect to the scale parameter, as choice tasks with for example a dominated alternative will generally have a larger scale parameter, such that the scale parameter between the choice tasks is not homogeneous and therefore could bias your parameters. This is usually not that much of an issue, and is seldom (actually never) taken into account when generating designs, but it's good to keep it in mind.