Insignificant bayesian priors

[eliD]

Dear all,

after collecting pilot data (based on an Efficient design) I am now in the process to design the final Bayesian design, and I have concerns on how to deal with insignificant coefficient estimates.
More in detail, from the MNL estimates I obtained all significant coefficients, except for one. Therefore, based on the suggetsions found in this forum I run the Bayesian design setting all priors using the means and std errors (divided by two) for all significant parameters, while I set zero mean and the std error (divided by two) for the insignificant beta. By doing so I obtained a final design with S-estimate=39, which I suppose is good enough.

However, I was wandering whether it is correct to keep this non-significant attribute in the final design, because the pilot data seem to indicate that this product characteristic does contribute to consumer utility. These results are based on just a few respondents, but I tested two different experimental conditions and this specific attribute is non-significant in both of them.

Can I decide to eliminate this attribute from the final bayesian, keeping all other priors equal or in this case should repeat the pilot test without this attribute?

Thank you in advance for your help.

eli

[Michiel Bliemer]

You can either set the prior for this attribute to zero, or you can keep the statistically insignificant parameter estimate as a best guess while it has a large standard error (which is fine).

Parameters are either statistically not significant because (i) the attribute is not relevant, or (ii) the sample size is too small. I would not remove attributes after a pilot study since sample sizes of pilot studies may not be large enough to pick up all relevant attributes.

Michiel

[eliD]

Dear Michiel,
Thank you so much for your help.
The final Bayesian design seems perfect!

Thanks again
Elisa