Re: [biogeme] D-efficient experimental design: defining prior values
Michael Nassar
Thank you for the explanation.
Hi Elisa
If you have a sufficiently large sample, you can use any design you want. Many famous authors use completely random designs. If you don’t want to use a random design (you may have reasons for this, such as fear of issues with reviewers who don’t understand design theory, fear of dominate alternatives that may arise), you might consider using an orthogonal design if you have absolutely no idea about what the parameters might be. As I teach my students, orthogonal designs have worked of the past 40 years, and just because we know more now, it doesn’t mean they have stopped working all of a sudden. An orthogonal design is really a local optimal design with zero parameters (under certain assumptions).
If you want to use an efficient design, then you can use Bayesian priors. For example, I know price should be negative but I don’t know the exact value it will take, you might use a uniform distribution with negative values. This is often better than assuming negative priors, but it is a matter of choice really.
With respect to the paper Michel mentions, I have to laugh. We use that paper in our courses on how not do designs. To give you just one issue with the paper, look at the travel times and costs, and priors they use. The travel time ranges between 30 and 72 minutes (with an average of 51 minutes) and cost from $1 to $22 (with an average of $11.5). They then use priors of -0.33 and -1 respectively. Just using the averages, the marginal contribution for travel time is -0.33*51 = -16.83 and for cost -1*11.5 = -11.5. These are ridiculous values to assume for a logit model. So whilst I agree with Michel – they conclude random designs work better than efficient designs, the reason they make this conclusion is because the priors they assume are so silly that it can be no other way. They rigged the result to get the conclusion.
Look at the end of the day, what you do with the design is up to you. Efficient designs properly generated should perform better than other designs (random and orthogonal) but what most people miss is what does better mean – it simply means sample size, which as per my original comment is probably a more important consideration.
John
From: bio...@googlegroups.com <bio...@googlegroups.com> On Behalf Of Michael Nassar
Sent: Saturday, January 27, 2024 12:05 PM
To: neerajsax...@gmail.com
Cc: michel.b...@epfl.ch; elisa....@unibg.it; Biogeme <bio...@googlegroups.com>
Subject: Re: [biogeme] D-efficient experimental design: defining prior values
Michael Nassar 3...@gmail.com
On Tue, Jan 23, 2024, 4:46 PM Neeraj Saxena <neerajsax...@gmail.com> wrote:
Hi Elisa
I would like to add onto the advice provided by Prof. Bierlaire.
I encountered the same situation almost a decade ago during my PhD, and the tips I followed that time were (Hope the following aren't obsolete now):
1. Use values from previous literature if it aligns well with your study
2. If Step#1 is not applicable, assume prior values closer to zero, but with directionality (+ or -) based on expected outcome. E.g. In my domain, the coefficient for Travel Time is generally negative. If this is what I expect in a new study, I'll start with -0.1 as its beta value
3. Generate a D-efficient design and launch a PILOT SURVEY
4. Run any Discrete Choice Model (e.g. Logistic Regression, Mixed Logit, etc.) on the Pilot data to get beta values
5. Use beta values from Step#4 to update the D-efficient design in Step#3.
Hope this helps!
Good Luck with your study :)
Best
Neeraj
On Tue, 23 Jan 2024 at 02:20, 'Michel Bierlaire' via Biogeme <bio...@googlegroups.com> wrote:
This is a chicken and egg problem, isn’t it?
Have a look at this paper, that provides a critical analysis of those designs: https://link.springer.com/article/10.1007/s11238-017-9647-3
One of their conclusions is that a random design (which is the easiest to generate) performs as well as any design…
> On 22 Jan 2024, at 14:01, Elisa ALESSIO <elisa....@unibg.it> wrote:
>
> Dear all,
>
> My name is Elisa, and I am a Ph.D. student. I am currently working on the development of efficient stated choice experimental designs. In the process, I encountered a challenge related to determining the prior values to be incorporated into the utility function. Since I haven't conducted the discrete choice experiment yet, I can't estimate these values analytically. Therefore, I am wondering how to calibrate them in a way that allows me to obtain the most efficient choice set.
>
> I would appreciate any advice that can help me optimally determine these values.
>
> Best regards,
> Elisa
>
> --
> You received this message because you are subscribed to the Google Groups "Biogeme" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to biogeme+u...@googlegroups.com.
> To view this discussion on the web visit https://groups.google.com/d/msgid/biogeme/69945668-8286-45ec-a639-0937eae0763dn%40googlegroups.com.
Michel Bierlaire
Transport and Mobility Laboratory
School of Architecture, Civil and Environmental Engineering
EPFL - Ecole Polytechnique Fédérale de Lausanne
http://transp-or.epfl.ch
http://people.epfl.ch/michel.bierlaire
--
You received this message because you are subscribed to the Google Groups "Biogeme" group.
To unsubscribe from this group and stop receiving emails from it, send an email to biogeme+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/biogeme/D96A9D04-AECA-4ABF-9E02-F9163C31075A%40epfl.ch.
--
Dr Neeraj Saxena, PhD (Transport Engineering)
T: 0481207876
Google Scholar: https://scholar.google.com.au/citations?user=2a8hpZkAAAAJ&hl=en
--
You received this message because you are subscribed to the Google Groups "Biogeme" group.
To unsubscribe from this group and stop receiving emails from it, send an email to biogeme+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/biogeme/CANnaUtQG6s2mNfmk3%3D9BmbaCKBOiRM_cQiry845-tN1mb0ZS%3DA%40mail.gmail.com.--
You received this message because you are subscribed to the Google Groups "Biogeme" group.
To unsubscribe from this group and stop receiving emails from it, send an email to biogeme+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/biogeme/CAHXRE9qPtCSCX%3DRS76d_nzka9sDTk4gwRPVp5bkkt%3DNHDi_cnw%40mail.gmail.com.
Michael Nassar
Michael Nassar 3...@gmail.com