Unlabeled design Syntax & Statistical Efficiency Criterion

[lich93]

Dear moderators,

I am creating a DCE design for pilot study now, and for a nationwide survey later on. The design is as follows:

Design
;alts = FVRmax, FVRsom, FVRstq
;rows = 15
;block = 3
;eff = (mnl,d)
;model:
U(FVRmax) = b1*ATR[2,1,0] + b2*ADM[2,1,0] + b3*AFR[2,1,0] + b4*OFI[2,1,0] + b5*PLFP[2,1,0] + b6*SFM[2,1,0] + b7*TAX[2,1,0,-1,-2] /
U(FVRsom) = b1*ATR + b2*ADM + b3*AFR + b4*OFI + b5*PLFP + b6*SFM + b7*TAX /
U(FVRstq) = ACSFVRstq
$

As you see that I have 7 attributes and 3 alternatives including status quo (FVRstq), maximum improvement (FVRmax) and some improvement (FVRsom). For all attributes, I have 3 levels except attribute # 7 – Tax/Fee in which I plan to have 5 levels including -10%, -5%, 0%, +5%, 10% where 0% is a status quo level for that attribute. For the Tax attribute, I have changed the coding to 2 for +10%, 1 for +5%, -1 for -5%, and -2 for -10% as I actually do not know how to include the reference value for the tax attribute (e.g., average annual municipal tax $4000) in the efficient design. If you can help with the syntax for this, it will be a great help.

After reading a couple of posts in the forum, I realize that the orthogonal design is not really appreciated nowadays, and it is even better to start with an efficient design by setting up zero priors. If this is correct, I would like to proceed with an efficient CE design for the pilot study. After getting the results from the pilot study I can construct a Bayesian efficient design with priors for a nationwide survey and data collection.

My questions are as follows:

1. Do you think that my design is okay statistically and follows MNL model efficiency criteria?
2. Should I proceed with 5 levels or 3 levels for the tax attribute although I would like to proceed with 5 levels?
3. Is there any way I can reduce the number of choice sets (e.g., 6 to 8) by following efficient design criterion?
4. With blocking do we actually compromise the level of efficiency? Or, theoretically and practically, blocking doesn’t have any severe consequences in the results.
5. How can I add the status quo option in the choice scenarios as a third alternative for the respondents? Do I have to add status quo alternative manually? Please see attached two choice scenarios generated by the said syntax. I want to add the status quo option there.
6. When I use the aforesaid design, the design kept running and the MNL D-error goes down to 0.129245, and A -error to 0.138017 but the system keeps running. Is this a normal behaviour for an efficient design? Is there any benchmark level for D-error or A-error?

Sorry for my long email. I am new to choice metrics and I am working on my Ph.D. thesis using Ngene. Your help in making an efficient design for my Ph.D. research will be greatly appreciated.

I look forward to getting a quick reply.

Sincerely,
Liton

[Michiel Bliemer]

1. Unlike for an orthogonal design, for an efficient design you need to specify the utility functions and levels just as you would estimate the model. Therefore, your attribute levels need to either represent the actual values shown to respondents, or you need to use dummy coding. If there are some categorical variables, then you have no choice but to use dummy coding (or effects coding).

2. Three or five are both fine.

3. You are estimating 7 parameters with 3 alternatives, which means that you need a theoretical minimum of 4 choice tasks but it is always a good idea to include more variation in your data by somewhat increasing the number of choice tasks. If dummy coding is used, the number of choice tasks will need to be significantly larger.

4. Blocking does not affect efficiency, it only decides which choice tasks are shown to each respondent.

5. You will be able to find many examples of status quo alternatives here on the forum. Instead of using a constant, you include exactly the utility function for the SQ alternative but you fix its attribute levels. You can direct set the tax of $4000 as a level for the tax attribute, and use values around $4000 as the levels for the other alternatives.

6. There are billions of possible designs and it takes a very long time for Ngene to go through the entire space of possible designs. Ngene simply keeps running until you as the user stop the searching process. At some point the efficiency does not improve (much) anymore and you can manually stop the search process. No there is no benchmark for D-error or A-error, the lower the better but different studies have different optimal values and we do not know beforehand what is a good value.

Michiel