Decision Rule Robust Design

[Giovanna]

Hi all!

I am trying to generate a Db-efficient design robust for estimating RUM and RRM models (following Van Cranenburgh & Collins, 2019).
I have already performed a pilot study to gather Bayesian priors.
My experiment is labelled, I have three alternatives (L, P, B) which vary on 2 attributes:
Att1, Cert [0,1]
Att2, Price [1.39,1.89,2.39,2.89]
The issue is that every time I specify non-zero priors in the RRM model I am not able to come up with a design. Indeed, I get this error if I use the swap algorithm:
"No valid design has been found after 1000 evaluations. There may be a problem with the specification of the design. A common problem is that the choice probabilities are too extreme (close to 1 and 0), perhaps because some or all of the prior values are too large. Also, it is generally a good idea to start with a simple design (MNL, non-Bayesian), then add complexity. If you press stop, a design will be reported, which may assist in diagnosing the problem."

I am using the following syntax

Code: Select all


Design
;alts(RUM)=l_alt,p_alt,b_alt
;alts(RRM)=l_alt,p_alt,b_alt
;rows=16
;block=4
;eff=0.5*RUM(mnl,d,mean,rum)+0.5*RRM(mnl,d,mean,rrm)
;bdraws=gauss(5)
;alg=swap

;model(RUM):
U(l_alt)     = a1 [(n,1.10,0.11)] + b1.dummy[(n,0.40,0.09)]*cert[1,0] + b2[(u,-1.02,0.10)]*price[1.39,1.89,2.39,2.89] /
U(p_alt)    =  b1.dummy*cert   + b2  *price /
U(b_alt)    = a3 [(n,0.53,0.12)] + b1.dummy *cert  + b2  *price         

;model(RRM):
U(l_alt)     = a1 [(n,0.78,0.09)]  + b1.dummy [(n,0.27,0.06)]*cert[1,0] + b2[(u,-0.64,0.06)]*price[1.39,1.89,2.39,2.89] /
U(p_alt)    = b1.dummy  *cert  + b2    *price /
U(b_alt)    = a3 [(n,0.36,0.08)]  + b1.dummy   *cert + b2   *price     
$



Conversely, If I specify zero priors for the RRM model parameters, I am perfectly able to generate a design.

Moreover, I have also tried to replicate the design used in the study: Van Cranenburgh, S., & Collins, A. T. (2019). New software tools for creating stated choice experimental designs efficient for regret minimisation and utility maximisation decision rules. Journal of Choice Modelling, 31(April), 104–123. To be specific, I have used the third ngene example. Unexpectedly, I get the same issue (also the same error) and I don’t understand why.

Currently, I am using the updated version of Ngene (1.2.1).
What am I doing wrong or missing? Any kind of support or guidance will be really appreciated!
Many thanks in advance

Giovanna

[Michiel Bliemer]

I will ask Andrew Collins to answer this question as he is co-author of the article, but I suspect that the alternative-specific constants a1 and a3 are the issue for the RRM model. If I remove these constants, the syntax runs fine.

Michiel

[Giovanna]

Dear Professor Bliemer,
Thank you very much for your answer and for suggesting that the ASCs may be the issue.
Am I right in assuming that since my experiment is labelled, I can’t remove these constants?

I hope Professor Collins will find the time to reply to my post.

Very best regards
Giovanna

[Michiel Bliemer]

In a labelled experiment you would need those constants yes, you should not remove them. What I do not know is whether the RRM model (at least how it is implemented in Ngene) actually supports labelled experiments,

Andrew Collins has responded that he will look at the issue and will comment shortly.

Michiel