generating priors

[medard kakuru]

Greetings Michiel

I have two questions though they seem to be general.

in generating orthogonal designs to use in the pilot survey, do I need to include the "opt-out" option or it only appears during efficient design?

secondly, while analyzing data for the pilot study to obtain priors, do I include the other covariates like consumer characteristics or I use the product characteristics only? in addition, if the parameter value turns out to be insignificant, do I still use it as a prior?

thank you.

regards,

medard

[Andrew Collins]

Hi Medard

In answer to your questions:

1. You would not include the opt-out alternative for an orthogonal design, as there are no considerations of relative utility (important for efficient designs), and no attributes in the alternative for which levels need to be assigned.

2. You may wish to include covariates, then handle them using the approach outlined in Section 8.4 of the manual "Including covariates in generating efficient designs".

3. If marginally significant, then you may wish to use the parameter as a prior. I would be very careful about using an insignificant parameter, as it could have a very detrimental impact on the design. You may want to include a prior of small magnitude but correct sign, if for example you are checking for dominance (see section 8.8).

Andrew

[Michiel Bliemer]

Adding to Andrew's response, it is quite likely that from a pilot study not all parameters will be significant. However, you can use the standard error as a measure of uncertainty for this parameter.

So suppose parameter b is estimated as 0.05 and has a standard error of 0.04. This parameter is therefore not statistically signficant (t-ratio = 0.05/0.04 < 1.96). However, you can use this information as a Bayesian prior in Ngene:
b[(n,0.05,0.04)]

Michiel

[medard kakuru]

hi,

can I say the I will use a binary logit to analyze data from pilot survey since I don't have the third alternative of opt-out?
secondly, can I use the RPL on the pilot survey data with two alternatives?




medard

[Andrew Collins]

If there is still an opt-out alternative in the choice tasks that you present (even if not specified when generating an orthogonal design), then you will not be estimating a binary logit model. If the choice tasks have no opt-out, then it will be binary logit.

RPL can be used with two alternatives, yes. You may have difficulties estimating significant parameters, if there are not many choice responses in the pilot.

[medard kakuru]

Hi Andrew,
thank you, together with your entire team



regards medard

[medard kakuru]

hi,

the last time I asked about use of priors are that are not significant, you told me to use a Bayesian efficient design for a prior that is marginally insignificant. now suppose I have a prior that is very insignificant, do I still use it in that way?

secondly, I ran a binary logit with no constant and got all parameters significant except one. however, when I introduce the constant, most of these parameters become insignificant but the parameter for the constant is significant. which of the two parameters can I use?

[Michiel Bliemer]

If the prior is very insignificant, you will have to use common sense. Do you know the sign of the attribute? Then make it for example (u,-1,0) for a negative parameter, depending on the size of the attribute. If you do not know the sign, perhaps use (n,0,1), with a standard deviation depending on the size of the attribute.

If you have two unlabelled alternatives, you use no constant. If you have labelled alternatives, you can use a constant. These are typically the rules that you go by in estimation.

[medard kakuru]

hullo Michiel,
I got your feedback and it answers my question. thank you.

regards,

medard

[medard kakuru]

Hi,

below is the syntax I used to generate a Bayesian efficient design.

Design
;alts = alt1*, alt2*,alt3*
;rows = 24
;block = 6
;eff = (mnl,d,mean)
;model:
U(alt1) = colour.dummy[(n,-1.16,0.35)|(n,-1.19,0.37)|(n,-1.62,0.36)] * colour[0,1,2,3] + b3[(u,-1,0)] * taste[0,1]+ b4[(n,-0.58,0.23)] * texture[0,1] + b5[(n,-0.54,0.21)] * smell[0,1] + b6[(n,0,1)] * size[0,1] + b7[(n,-0.001,0.0002)] * price[400,650,1000, 1650] /
U(alt2) = colour * colour + b3 * taste + b4 * texture + b5 * smell + b6 * size + b7 * price
$

I got the following warning messages from the output.
"Warning: Two alternatives were specified for alternative dominance checking, but do not have the same priors, and so will not be checked. 'alt1', 'alt3'
Warning: Two alternatives were specified for alternative dominance checking, but do not have the same priors, and so will not be checked. 'alt2', 'alt3'

where could I have gone wrong?


regards,

medard