Number of rows? Bayesian vs. zero priors?

[bbullington]

Hi there! I'm new to Ngene, so I wanted to make sure that I'm doing this right. I'm designing a DCE with 6 attributes (2 with 3 levels, 4 with 2 levels). Each participant will be presented with 11 choice sets. I don't have any pilot data, so I was trying out zero priors and Bayesian priors for the first participants and then adjusting the priors based on those initial results. Here is the code I've worked with so far:

0 priors:
Design
;alts = alt1, alt2
;rows = 66
;block = 6
;eff = (mnl,d)
;model:
U(alt1)=b1[0]*A[-1,0,1]+ b2[0]*B[-1,0,1] + b3[0]*C[-1,1] + b4[0]*D[-1,1] +
b5[0]*E[-1,1] + b6[0]*F[-1,1] /
U(alt2)=b1*A + b2*B + b3*C + b4*D + b5*E + b6*F
$

Bayesian priors:
Design
;alts = alt1, alt2
;rows = 66
;block = 6
;eff = (mnl,d)
;model:
U(alt1)=b1[(u,-1,1)]*A[-1,0,1]+ b2[(u,-1,1)]*B[-1,0,1] + b3[(u,-1,1)]*C[-1,1] + b4[(u,-1,1)]*D[-1,1] +
b5[(u,-1,1)]*E[-1,1] + b6[(u,-1,1)]*F[-1,1] + b7[(u,-1,1)]*G[-1,1] /
U(alt2)=b1*A + b2*B + b3*C + b4*D + b5*E + b6*F + b7*G
$

Here are my questions:
1. How should the number of rows be determined? I know that it has to be divisible by 2 and 3 (the number of levels), so I just multiplied by 11 (number of choice sets per participants), but that feels random. Is there a benefit to having more rows and more blocks? The total sample size of our study is 136, so with 66 rows, ~23 people will be answering each block.

2. Do you recommend using 0 priors or Bayesian before we have actual priors from pilot data?

3. Does everything else seem right?

Thank you so much for your help! I am very appreciative.

[Michiel Bliemer]

1. A block of 11 choice tasks means that each block cannot be attribute level balanced, so you would preferably choose a block size that is divisible by 2 and 3, i.e. either choose 6 or 12. This would mean a design size of 72. Given that you are only estimating 6 parameters I think that you do not need a design of such a large size, you could consider using ;rows = 36 and ;block = 3, but there is nothing wrong with using a design size of 72 if you are happy creating 6 blocks (i.e. 6 versions of the survey).

2. Both are fine, I could not say which one I prefer. You are using attribute level coding that do not seem to reflect the actual levels to be used in the model (I am not sure what -1,1 would mean). In an efficient design you should not use design coding or orthogonal coding but instead use the actual attribute levels (e.g. cost in dollars).

3. Yes it further looks fine!

Michiel

[bbullington]

Michiel,

Thank you so much for your response! It's very helpful. In this study, the levels of the attributes are all categorical ie church vs. clinic vs. community center. Does that mean I should not be using the efficient design? Now I'm confused! I'll do some more reading, but would appreciate any insights.

[Michiel Bliemer]

Yes you can use categorical variables but you will have to use an appropriate coding scheme. Your 2-level attributes seem effects coded (using -1,1) but your 3-level attributes are not.

For dummy coding for your 2-level and 3-level attribute, please use something like:
b3.dummy[0] * C[1,0]
b1.dummy[0|0] * A[1,2,0]
where 0 can be replaced with a Bayesian prior. This means that your first prior is associated with level 1 and the second prior (for the 3-level attribute) is associated with level 2 while level 0 is the reference level.

For effects coding, please use something like:
b3.effects[0] * C[1,0]
b1.effects[0|0] * A[1,2,0]

For the 3-level attributes, you will be estimating 2 parameters.

Michiel