Efficient design with rp and rppanel

[dov0101]

Hello,

we want to make a CE study with two labelled varying options and the reference (status quo) fixed alternative.

We have the results of a pilot survey that allowed us to obtain the priors of the parameters from the MNL model estimation.
The GEO and COL parameters were insignificant in the pilot,

We are designing the CE using Ngene and our current syntax is as follows:

Code: Select all

Design   
;alts = BSPout*, BSPsem*, ref
;rows = 24
;block = 2, minsum
;eff = (mnl, d)
;rep = 500
;rdraws = halton (200)
;alg = mfederov
;model:   
U(BSPout) = bBPSout[n, 0.8906360,  0.1151940]
        + b1[n,-0.0065247,  0.0011050]                                      *PRICE[70, 120, 170]
        + b2.effects[n, 0.0569746,  0.2068940 | n, 0.2671585,  0.2083617]   *GEO[1,2,3]
        + b3.effects[n, 0.0951933,  0.1356148]                              *COL[1,2]
/   
U(BSPsem) =  bBPSsem[n, 0.6322091,  0.1163940]
        + b1                  *PRICE               
        + b2.effects          *GEO     
        + b3.effects          *COL 
/
U(ref)  = b1                *PRICEref[45]
$


It seems to work perfectly with both D-error and sample size minimisation when optimized assuming an MNL model (with fixed parameters).

However, we get an "Undefined" result when I try to optimize using the rp option, and no error message.

I would appreciate your suggestions on this matter.

Thanks a lot.

[Michiel Bliemer]

Optimising for rppanel models is tricky and typically requires running algorithms for a very long time, usually the best approach is to optimise for the MNL model and evaluate for the panel mixed logit model, which you can do with your existing syntax.

In order to evaluate the D-error for the rppanel model, Ngene simulates random choices (via the ;rep command). In some cases, this leads to negative definite covariance matrix for which the D-error is negative and Ngene returned Undefined. This may randomly occur and changing the number of simulations may avoid the issue. For example, I used 200 simulated respondents (design replications) with ;rep = 200 that this provided a finite D-error. I also suggest using the more efficient gaussian draws and perhaps add ;con since the constants in your model seem important.

Code: Select all

Design   
;alts = BSPout*, BSPsem*, ref
;rows = 24
;block = 2, minsum
;eff = (mnl, d)
;con
;rep = 200
;rdraws = gauss(3)
;alg = mfederov
;model:   
U(BSPout) = bBPSout[n, 0.8906360,  0.1151940]
        + b1[n,-0.0065247,  0.0011050]                                      *PRICE[70, 120, 170]
        + b2.effects[n, 0.0569746,  0.2068940 | n, 0.2671585,  0.2083617]   *GEO[1,2,3]
        + b3.effects[n, 0.0951933,  0.1356148]                              *COL[1,2]
/   
U(BSPsem) =  bBPSsem[n, 0.6322091,  0.1163940]
        + b1                  *PRICE               
        + b2.effects          *GEO     
        + b3.effects          *COL
/
U(ref)  = b1                *PRICEref[45]
$

Michiel

[dov0101]

Hello Michiel,

thanks for the hints.

perhaps add ;con since the constants in your model seem important

Correct, the constants are the main parameters we need to estimate and also we want to evaluate whether an information treatment would significantly change their value, compared to the control group.

We plan to survey 400 people, providing extra information to about 200 randomly chosen respondents.

We hope that the effect can be strong enough to be detectable with this design.

Warm regards.
Maurizio