choice-metrics.com

Hi,
I’ve done a pilot study of an unlabeled design on food choice and health labelling, and I’m trying to use the coefficients to get an efficient design with fixed priors. I have a couple of questions I need help with:
Firstly, I’ve realized I made a mistake when designing the pilot… I added an alternative-specific constant to my unlabeled design. I think this has meant that Ngene checked for repeated alternatives and row repetition, but NOT for dominance. Therefore, there were some dominated alternatives in my design (4 out of 24 rows, 2 per block, so I hope not disastrous). I’d checked, but there were always some dominated alternatives, so I thought that was just the best that could be done with my syntax!
The coefficients are of the expected sign and mostly significant…

Would you still recommend using them for the pilot?

Many thanks for the help!

I add below the original syntax, and the coefficients


Coefficients are shown as effects coded, alt3 is an opt-out
; alts = alt1*, alt2*, alt3
; rows = 24
; block = 2
; eff = (mnl, d)
; alg = swap(stop=total(10 mins))


; model:

U(alt1) = b1 + b2[-0.0001]*price[60,80,100,120]
+ b3.effects[0.0001]*low_salt[1,0]
+ b4.effects[0.0001]*safety_certified[1,0]
+ b5.effects[0.0001]*no_antibiotics[1,0]
+ b6.effects[0.0001]*originSA[1,0]

/
U(alt2) = b2*price
+ b3*low_salt
+ b4*safety_certified
+ b5*no_antibiotics
+ b6*originSA

$


VARIABLES Priors
Price -0.00170 (0.00264)
SaltContentD 0.307*** (0.0649)
SafetycertificationD 0.377*** (0.118)
RoutineantibioticuseD 0.0141 (0.0534)
OriginD 0.182*** (0.0566)

Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

Given that you have an opt-out alternative and two unlabelled alternatives, you need to add constant b1 to alt2 as well and estimate its coefficient. Your parameter estimates do not include an estimate for b1, which is essential when you have an opt-out alternative. You can also remove b1 from alt1 and alt2 and use U(alt3) = b1[...].

I am not too worried about the choice tasks with dominant alternatives, your priors should be good enough I think.

Michiel

Thanks a lot for the advice, and the quick response!

The constant b1 is 2.07.. (b1=1 for alts1 and alt2)
I had understood from the manual that it's not necessary to add in the constant prior value, as it would be ignored in calculating efficiency measures.

1. Is that the case, or should I also add it in as a prior?

I also had a couple of questions about the price range. The prior for the price variable seems really low, elasticity= -0.018 as opposed to to -0.6 from a demand model in the literature. I think this is because we used a price range that was too small.

2. Would it make sense to use a larger price range for the rest of the survey?
3. In that case, would you recommend using the prior for the price variable from the pilot, or from the literature?
4. We were also thinking of using 6 levels for the price variable as opposed to 4, although not sure if that could create problems...

Thanks a lot in advance
S

I do not understand what you mean with "constant b1 is 2.07.. (b1=1 for alts1 and alt2)". You either have a constant for alt3, OR you have a constant for alt1 and all2 (the same constant). You cannot put constants in all alternatives.

1. You should specify the model including constants when you have a labelled (no choice) alternative. The fact that efficiency is not optimised to estimating the constant does not mean that the constant has no influence on the efficiency of the other parameters. Constants influence choice probabilities, and choice probabilities influence efficiency. If you add ;con to the syntax than Ngene will also try to minimise the standard error of the constant, if you leave it out then Ngene will only minimise the standard errors of the attribute coefficients, but either way, the constants influence the standard errors of all coefficients and therefore need to be specified if there are labelled alternatives.

2. Maybe, but this depends on the study itself. If you feel that respondents are not trading off on price then you can consider increasing the range.

3. You could consider using the following Bayesian prior for price, based on a range of twice the standard error from the parameter estimate, capped at 0: b2[(u,-0.007,0)]. If you take it from the literature, make sure that you only transfer WTP values since each study has scale differences and therefore you cannot directly use absolute parameter estimates. Also be mindful of currency differences etc.

4. Either is fine. A larger number of attribute levels can be useful if you would like to estimate nonlinearities in the model, e.g. log(price), or if you think that differences between two consecutive levels is too large and may make price a dominant attribute. If you increase the range, increasing the number of levels may be a good idea.

Michiel

Yes, sorry, there was a typo in my post. I meant the same constant for Alt1 and Alt1, like you say.
Thanks a lot for your help, very grateful!
Regards
S