choice-metrics.com

Dear Professor Bliemer and Ngene users:
I am dealing with an unlabelled design and I have everything more or less clear but some customisation that I have included. This is the syntax for the pilot (the priors were setting following in some way the illuminating paper of Professor Bliemer -On determining priors for the generation of efficient stated choice experimental designs-.):
Design
;alts = A*, B*, SQ
;rows = 24
;eff = (mnl,d,mean)
;bdraws = sobol(2000)
;bseed = 12345
;block = 4,minsum,noimprov(60 secs)
;store = all
;model:
U(A) = b1.dummy[(u,-0.45,-0.2)|(u,-0.35,-0.10)|u,-0.125,-0.05]*X1[3,2,1,0]
+b2.dummy[(u,0.025,0.075)]*X2[1,0]
+b3[(u,0.10,0.17)]*X3[0,1,2,3]
+b4[(u,-0.2,-0.10)]*X4[0,1,2,3]
+b5[(u,0.1,0.3)]*X5[0,1,2,3] /
U(B) = b1*X1
+b2*X2
+b3*X3
+b4*X4
+b5*X5/
U(sq) = asc[0]$

I have some concerns regarding the following:
1. ¿Would you find this customisation -"minsum,noimprov(60 secs)"- suitable for the design?
;bdraws = sobol(2000)
;bseed = 12345
;block = 4,minsum,noimprov(60 secs)
2. When using uniform distributions for getting the experimental design for the pilot stage or in situations of high uncertainty because you do not have any clue about the priors, would it be always advisable to use "mean" for the command -“;eff = (mnl,d,mean)”-?
3. Would you introduce non-linear effects directly in the experimental design for the pilot study? Or better using linear effects at the beginning and then after the pilot study is done defining non-linear priors if suitable for the final design?
Thank you so much Professor Bliemer for your generosity
Kind regards

1. Yes you may be able to improve the blocking of the design with this command, although it will not affect the design efficiency. I think 60 seconds is not needed, probably 10 seconds is more than enough.

2. Yes I think "mean" is almost always appropriate for uniform distributions. The alternative is "median", but the median is only needed if there is a worry that draws for the priors may result in very large (positive or negative) values. This typically only happens with normal distributions (which are unbounded).

3. You can define them in the pilot study or later if you believe they are relevant. Leaving them out simply means that you are not optimising for them, but interaction effects are typically less important than main effects anyway. The reason for putting them in is sometimes to make sure that you can estimate them (i.e. to ensure that the interactions in the design are not perfectly correlated with other attributes, but this is rarely happens so this is typically not a real concern).

Some suggestions:
* You specified [u,-0.125,-0.05], which denotes a random parameter, but I think you mean [(-u,-0.125,-0.05)] that is a Bayesian prior.
* You could consider using ;bdraws = gauss(3) which does the around 2000 draws as well but in a smarter way.

Michiel

Dear Professor Bliemer and NGENE users,
Thank you so much for your answers… To be honest it is not always so easy to get illuminating advising in other forums as it is here, and I appreciate that. I would like to share with the NGENE community a clarification and a couple of new general questions:
Regarding the 3rd question, I meant linear vs non-linear effects (not interaction effects), for example:
U(A) = b1.dummy[(u,-0.45,-0.2)|(u,-0.35,-0.10)|(u,-0.125,-0.05)]*X1[3,2,1,0]
vs
U(A) = b1[(u,-0.3,-0.1)]*X1[0,1,2,3]
I am quite confident about the existence of a potential break point in level 3 or even 2 regarding the rest (the d-error is much higher with the non-linear specification vs linear one which makes sense) but I have certain degree of uncertainty and maybe it could be better to begin with linear and then when having the pilot check for the existence of non linear effects for getting the priors for the final experimental design...

The new questions are more general but I would like to share them since they could be relevant to practitioners.
A) If I am not mistaken when the SQ alternative is included but there is no associated utility function the utility is considered by default as 0 (none option). In this case dealing with farmers, I have a status quo alternative (not diversification in crops) where all the attributes would take 0 for most of the farmers (due to technical limitations I cannot implement a pivot design. Nonetheless, the status quo of each farmer would be collected but it is supposed to be very homogeneous) so I have specified U(sq) = asc[0]$. When I do that the efficiency of the design worsens regarding no specification.

Design
;alts = A*, B*, SQ
.
.
.
U(sq) = asc[0]$
vs
Design
;alts = A*, B*, SQ
.
.
. $
B) The technician in Latin America is advising me (taking into account the farmers literacy) that it would be advisable displaying choice cards with the status quo and only one alternative instead of two. Has anyone experience with this type of design dealing with farmers in developing countries? Is the loss of statistical power compensated by a potential efficiency gain in choice tasks? Do you have any recommendations (regarding for example the number of choice cards or any advisable change related to the design) when using choice tasks with only one alternative?

Thank you so much in advance to the NGENE community
Best regards

If you believe that nonlinear effects could be important, then I would design the experiment assuming nonlinear effects, which also allows estimating linear effects.

When including a no-choice or a status quo alternative, there needs to be an alternative-specific constant, either in alternatives A and B, or in the SQ alternative. Leaving it out is not appropriate. Therefore, including asc is correct. Do not worry about the efficiency going down, this is merely because adding asc means that you have to estimate an additional parameter, and therefore you will need to capture more information in your design. In your case, you could also specify the following where you leave out the utility formulation of the SQ:

Design
;alts = A*, B*, SQ
.
U(A) = asc[0] + ...
U(B) = asc[0] + ...
.
. $

Note that typically the status quo alternative is fully specified with attributes, but I understand what you are saying in that the levels vary acros the population and therefore you simply set it to zero.

I am not familiar with farmers in developing countries, but showing only a single alternative and a status quo alternative would be fine I suppose. Showing 2 alternatives and a status quo would mean showing 15 attribute levels, which could become complex. In order to decrease complexity, you can also choose to use graphics/pictures as much as possible to decrease the cognitive burden. And of course in a face-to-face interview the interviewer can explain everything in more detail. When you use only 1 alternative instead of 2 you will loose quite a bit of efficiency, which will mean that you may need to ask more questions. But I believe that it is important that the farmers understand the questions, so I would think that creating a design that is well understood and not too difficult is more important than efficiency (unless your sample size is very restricted, then you need to balance efficiency and complexity).

Michiel