Starting search from existing design using mfederov

[ejdwebb]

Hi,

I managed to run some code for a while successfully, but then had my computer crash. I've tried re-starting from the last design I reached using the ;start command. The code runs without error, and begins by successfully evaluating the previous design, but as soon as NGene starts looking for new designs the D-error becomes much larger. Essentially it looks like the starting design is being ignored and NGene is starting from scratch.

I'm searching for an efficient design using the mfederov algorithm.

Am I missing something, or is this some sort of bug?

Thanks in advance,
Ed.

[Michiel Bliemer]

Would you be able to post the syntax here on the forum (or send it to contact@choice-metrics.com if you do not want to show it publicly)? Then we will investigate what is going wrong. Maybe there is an issue in the syntax, but if there is a bug we would like to find it and fix it as soon as possible.

Michiel

[ejdwebb]

Hi Michel,

Sure, the code is below. If it's not already a known bug I'd bet on it being my mistake since I'm relatively new to NGene. I've emailed the previous design file to the email address you gave.

Thanks!
Ed.

Design;

alts = alt1, alt2;
rows = 40;
block = 4;
eff = (rp,d);

alg = mfederov;
start=temp_design_12-03-19.ngd;

bdraws = halton(250);
rdraws = halton(250);

reject:
alt1.enrollment>=alt2.enrollment and alt1.completion>=alt2.completion and alt1.fruitnveg>=alt2.fruitnveg and alt1.stafftime<=alt2.stafftime and alt1.setupcost<=alt2.setupcost and alt1.runningcost<=alt2.runningcost,
alt2.enrollment>=alt1.enrollment and alt2.completion>=alt1.completion and alt2.fruitnveg>=alt1.fruitnveg and alt2.stafftime<=alt1.stafftime and alt2.setupcost<=alt1.setupcost and alt2.runningcost<=alt1.runningcost;

model:
U(alt1) = b_enrollment[n,(u,0,1),(u,0.1,0.5)]*enrollment[6,8,10] + b_completion[n,(u,0,1),(u,0.1,0.5)]*completion[50,70,80] + b_enrollment_x_completion[0.1]*enrollment*completion + b_fruitnveg[n,(u,0,1),(u,0.1,0.5)]*fruitnveg[0.5,1,2] + b_stafftime[n,(u,-1,0),(u,0.1,0.5)]*stafftime[4,8,12] + b_setupcost[n,(u,-1,0),(u,0.1,0.5)]*setupcost[15,20,30] + b_runningcost[n,(u,-1,0),(u,0.1,0.5)]*runningcost[15,20,30]/

U(alt2) = b_enrollment*enrollment + b_completion*completion + b_enrollment_x_completion*enrollment*completion + b_fruitnveg*fruitnveg + b_stafftime*stafftime + b_setupcost*setupcost + b_runningcost*runningcost
$

[Michiel Bliemer]

Thank you.

I identified an issue in the start design (the first column should contain the respondent ID, i.e. setting all values to 1 in this case) and also your priors are not appropriate (multiplying cost attributes with a value between 15 and 30 with a parameter between 0 and 1 (and possibly beyond due to Bayesian draws) results in very large partworths. These parameters are too large and will be problematic. I suggest starting with small values, possibly even using zeros everywhere. Finally, I would advice optimising for an MNL model (and evaluating possibly for an RPPANEL model). The RP model is usually not considered a very good model since it does not account for the fact that each respondent faces multiple choice tasks.

I got the syntax to run fine by using zero priors and changing the input file as described above. I do indeed see that the initial design has a lower D-error than subsequent designs generating by Ngene and I am checking with our software developer what is happening.

Michiel

[ejdwebb]

Hi Michiel,

Thanks very much for the comments on the design, I'll revisit it in the light of your suggestions.

Best,
Ed.