Designing a discrete choice experiment (choice set size and parameter estimation)

[Maria99]

Hello everybody,

I have a question about designing a discrete choice experiment. My design contains:

• 2 unlabeled alternatives (Alt 1 and Alt 2)
• 5 attributes (x1, x2, x3, x4, x5) with three levels (0 = underachieved, 1 = sufficient, 2 = overachieved)
• 1 cost-attribute (x6) with four levels (0 = 0, 1 = 1%, 2 = 2%, 3 = 3%)
• Interaction effects between each of the 5 attributes (x1 – x5) and the cost attribute (x6)
• Restriction to avoid comparing Alt 1 with better attributes and lower costs with an Alt 2 with worse attributes and higher costs (and vice versa)

Following the book “Marketing Research Methods in SAS” by Warren Kuhfeld, I created the following code:

%mktruns(3 3 3 3 3 4, interact=x1*x6 x2*x6 x3*x6 x4*x6 x5*x6); 

%mktex(3 3 3 3 3 4, seed = 200, n=144, interact=x1*x6 x2*x6 x3*x6 x4*x6 x5*x6);


%macro res;
	g1 = (x[1,1:5])[+]; * Attributes in alt 1;
	g2 = (x[2,1:5])[+]; * Attributes in alt 2;
	bad = bad + (g1 > g2 & x[1,6] < x[2,6]); * Better attributes in 1 and lower price in 1;
	bad = bad + (g2 > g1 & x[2,6] < x[1,6]); * Better attributes in 2 and lower price in 2;
	%mend;


%mktlab(data=design, int=f1-f2)
proc print; run;
%choiceff(data=final, model=class(X1-X6), nsets=72, maxiter=20, seed=200, flags=f1-f2, options=relative, restrictions=res, resvars=X1-X6, beta=zero);
proc print; by set; id set; run;
%mktblock(data=best, nalts=2, nblocks=8, seed=200, maxiter=20);

The final results are as follows:

Finally, I still have the following questions:

1. Is it correct to use the %mktex macro with n = 144 and the %choiceff macro with nsets = 72?  Unfortunately, I don't understand exactly how the recommended size of the kandidate set relates to the size of the choice sets.
2. Is my design able to estimate all parameters?

Thank you so much!!

[WarrenKuhfeld]

Is it correct to use the %mktex macro with n = 144 and the %choiceff macro with nsets = 72?  Unfortunately, I don't understand exactly how the recommended size of the kandidate set relates to the size of the choice sets.

Sure. The candidates are simply profiles that might be included in the final design. 144 seems like a great number. Other numbers would work too.

Is my design able to estimate all parameters?

Yes. Your D-efficiency is greater than zero, so yes. There is other output not shown that shows the parameters.

[Maria99]

Thank you so much for this quick response, Warren! I run my code again. I assume this is the output you meant in your answer:  

I think I still miss something, and three new questions came to my mind:

1. Why do I only see 13 parameters in the list, although I want to estimate 19 parameters (5 attributes x 3 levels + 1 attribute x 4 levels = 19 parameters)? Can I still estimate all 19 parameters?
2. I tried to optimize by design by increasing my nsets within my %choiceff from nsets = 72 to nsets = 144. As a result, the D-Efficiency is higher, and my variances from my parameters are lower. Would you recommend me to make this optimization, or would you say my design with nsets = 72 is actually pretty good so its not necessary? I am not sure because the Relative D-Efficiency is still very similar to each other.

1. Finally, you wrote in your answer “The candidates are simply profiles that might be included in the final design.” So do I understand correctly that the following would also work: %mktex macro with n = 144 and the %choiceff macro with nsets = 144?

Thank you so much for your help!

[WarrenKuhfeld · Accepted solution]

I just noticed that you posted a follow up question. If you have an n-level factor, there are n-1 parameters. Have you looked at https://support.sas.com/techsup/technote/mr2010c.pdf ? I know there is a lot there, and it is not the easiest topic, but in my biased opinion, it is a pretty good source of info on all this. I don't think you want 144 choice sets. Your call not mine, but with issues getting subjects, I would think you would want a smaller design.

[Maria99]

Thank you again for answering my follow-up questions!! Your book really helped me a lot, and I will take a closer look at the chapter again!