choice-metrics.com

Dear Michiel,
I generate the DCE and the priors are from pilot study as follows

Code: Select all

design
;alts=alt1*,alt2*,alt3
;rows=6
;eff =(mnl,d,mean)
;con
;bdraws = gauss(3)
;model:
U(alt1)=b0 [(n,2.334779,0.555722)]+b1.dummy[(n,-1.0134, 0.353174)|(n,-0.88372,0.43772)]*visit[1,2,0]+b2.dummy[(n,-0.82938, 0.4533)]*hap[1,0]+b3.dummy[(n, 0.147476, 0.495515)|(n,-0.8025,0.543363)]*plan[1,2,0]+ b4.dummy[(n,0.693785,0.26066)]*excer [1,0]+ b5.dummy[(n,0.139101,0.501286)]*behav[1,0]+ b6.dummy[(n, -0.37844, 0.584715)]*menu[1,0]/
U(alt2)= b0 [(n,2.334779, 0.555722)]+b1* visit +b2*hap +b3* plan +b4* excer+b5*behav+b6*menu
$


D-error =0.855071 (18 min). How to improve d-error for this design? I have a limitation of the sample is around 140.
Sincerely yours,
Sukunta Muadthong

I get a much larger D-error than 0.85 when I run that syntax. This is due to the fact that your number of rows is too small, you need more variation to efficiently estimate the 9 parameters in your model. I propose that you use 12 rows and block the design in 2 versions such that each respondent still only faces 6 choice tasks. You can also increase the number of choice tasks you give to a respondent to increase efficiency. If you only have 140 respondents, you may actually consider giving 12 choice tasks to a single respondent.

Further note that your Bayesian priors have a high degree of unreliability as indicated by their standard deviations, therefore sample size estimates will also have a high degree of uncertainty.

Code: Select all

design
;alts=alt1*,alt2*,alt3
;rows=12
;block=2
;eff =(mnl,d,mean)
;con
;bdraws = gauss(3)
;model:
U(alt1) = b0[(n,2.334779,0.555722)]
        + b1.dummy[(n,-1.0134, 0.353174)|(n,-0.88372,0.43772)]   * visit[1,2,0]
        + b2.dummy[(n,-0.82938, 0.4533)]                         * hap[1,0]
        + b3.dummy[(n, 0.147476, 0.495515)|(n,-0.8025,0.543363)] * plan[1,2,0]
        + b4.dummy[(n,0.693785,0.26066)]                         * excer [1,0]
        + b5.dummy[(n,0.139101,0.501286)]                        * behav[1,0]             
        + b6.dummy[(n, -0.37844, 0.584715)]                      * menu[1,0]
        /
U(alt2) = b0
        + b1 * visit
        + b2 * hap
        + b3 * plan
        + b4 * excer
        + b5 * behav
        + b6 * menu
$

This syntax generates a design with a Bayesian D-error of 0.82 (fixed D-error of 0.66). You can estimate all parameters, but some may not be statistically significant. If your prior value are correct, the first dummy coded parameter of b3 and b5 may not be statistically significant with 140 respondents, but all other ones would be.

Michiel

Dear Prof. Michiel,
Thank you so much for your reply. I added the row from 12 to 18 with block=2 (The respondents can handle the 9 choice-set). Ngene was running for 13 hours. It generates a D-error of 0.552143. Sp estimates: b3(d0)=96.39298, b5(d0)=58.58133. Is it ok for the main survey?

Sincerely yours,
Sukunta

Assuming that the priors are close to the true values, the design should be fine. There is not much else you can do to improve the design with the information you have.

Michiel

Dear Prof. Michiel,
Let me inform you that I used the same the priors, but adding the number of the row and take more time for running Ngene.
Thank you so much.
Sukunta Muadthong