choice-metrics.com

Hi All

I am generating a partial profile design using a constructed candidate set and the modified Fedorov algorithm. But the design search results seem quite weird – it seems like it should be doing better quicker.
Here’s what I get just running the syntax below in terms of design search in the output box:

Evaluation Time MNL D-Error
1 2:12:32 PM, 1/16/2025 0.086856
4382 3:14:14 PM, 1/16/2025 0.086523
8763 3:51:40 PM, 1/16/2025 0.086307

Curiously it took 4381 to get the first design that improved the D-Error and then a further 4381 before you get the next design that improved the D-Error. This isn’t just a fluke I’ve run similar syntax with a slightly different candidate set and got this same evaluation sequencing i.e. 2nd set to improve the D-Error was exactly double the number of evaluations as the first set.

I then just did a series where I re-ran the syntax but just stopped at the first evaluation. So presumably this is just another set of random choices from the candidate set.
This is what I got for the D-Error for 5 evaluations like this:

0.086052
0.087535
0.086037
0.086264
0.086264

So for 4 of the 5 extra (presumably just random choice) evaluations I got a better D-Error than the best design generated after 8763 evaluations of the Modified Fedorov algorithm.

This seems really odd OR my idea of what might be going on is very off-track (which is probably quite likely).

Here’s the syntax:
Design
;alts = Option1*, Option2*, Option3*
;rows = 200
;block = 20
;eff = (mnl,d)
;alg = mfederov(candidates = Candidate_Student_4.csv)
;model:
U(Option1) = b1.dummy[-0.1 | -0.2 | -0.3 | -0.4] * Price[1,2,3,4,5] +
b2.dummy[0.025 | 0.05 ] * GPVisit[1,2,3] +
b3.dummy[0.01 | 0.04 | 0.07] * NonGPVisit[1,2,3,4] +
b4.dummy[0.01 ] * Ambulance[1,2] +
b5.dummy[0.01 | 0.02 | 0.03 ] * TravelIns[1,2,3,4] +
b6.dummy[0.01 | 0.02 ] * DentalEmerg[1,2,3] +
b7.dummy[0.01 | 0.07 | 0.13 | 0.2] * DentalPrevent[1,2,3,4,5] +
b8.dummy[0.01 ] * Optical[1,2] +
b9.dummy[0.01 | 0.02 | 0.03 ] * Physio[1,2,3,4] +
b10.dummy[0.01 | 0.02 | 0.02 | 0.03 ] * VideoDoctor[1,2,3,4,5] +
b11.dummy[0.01 | 0.02 | 0.07 ] * Vaccine[1,2,3,4] +
b12.dummy[0.01 ] * PharmacyDemand[1,2] +
b13.dummy[0.01 | 0.02 ] * H&W[1,2,3] +
b14.dummy[0.01 | 0.02 ] * PartnerDiscount[1,2,3] +
b15.dummy[0.01 ] * Accomodation[1,2] +
b16.dummy[0.01 | 0.02 ] * Service[1,2,3] +
b17.dummy[0.01 ] * Repatriation[1,2] /
U(Option2) = b1 * Price +
b2 * GPVisit +
b3 * NonGPVisit +
b4 * Ambulance +
b5 * TravelIns +
b6 * DentalEmerg +
b7 * DentalPrevent +
b8 * Optical +
b9 * Physio +
b10 * VideoDoctor +
b11 * Vaccine +
b12 * PharmacyDemand +
b13 * H&W +
b14 * PartnerDiscount +
b15 * Accomodation +
b16 * Service +
b17 * Repatriation /
U(Option3) = b1 * Price +
b2 * GPVisit +
b3 * NonGPVisit +
b4 * Ambulance +
b5 * TravelIns +
b6 * DentalEmerg +
b7 * DentalPrevent +
b8 * Optical +
b9 * Physio +
b10 * VideoDoctor +
b11 * Vaccine +
b12 * PharmacyDemand +
b13 * H&W +
b14 * PartnerDiscount +
b15 * Accomodation +
b16 * Service +
b17 * Repatriation
$

Cheers
Roger

No there is no problem, this is the result of how the modified Fedorov algorithm works. It first determines a random design, and then it replaces the FIRST choice task with ALL choice tasks in your candidate set. You likely have a candidate set of around 5000 rows since Ngene only reports a better design after it goes through the entire candidate set size. Then it replaces the SECOND row with ALL choice tasks in the candidate set. So after twice doing this, your design will still not be very efficient, which is why you can also easily find another random design with a similar D-error. Since you have a large number of rows (200) and likely also a large candidate set size, you will need to run the script very long, I would say at least 200 * 5000 = 1 mln iterations, so that each row can be replaced at least once.

I generally never use candidate set sizes larger than 2000 rows because it makes the algorithm too slow. And if you really need 200 rows then most algorithms will be slow.

In summary, there is nothing wrong with the algorithm, just let it run for a long time

Michiel

Hi Michiel

Many thanks, I think I understand.

You'd think it would be a whole lot more efficient if it took 100 random start points and started the algorithm at the one with the lowest D-Error.

Anyway, you have given me the info to find something that will work in a reasonable time.

Cheers
Roger

The swapping algorithm does what you suggest, but the modified Fedorov algorithm (which comes from the literature) does not do that as it systematically goes through the candidate set and for typical design sizes and candidate set sizes this works well.In your case, with a huge design size and huge candidate set size, the modified Fedorov algorithm becomes really inefficient as you indicate.

In your situation, it would probably be much faster to split your candidate set in, say, 4 pieces, and then run 4 instances of Ngene (you can simply start Ngene 4 times and run them in parallel using a separate cpu), each searching the best 50 rows in each of the four portions of the candidate set. And then combining the 4*50 = 200 rows. Just a thought.

Michiel

Hi Michiel

Thanks for your advice - all good.

Cheers
Roger