choice-metrics.com

Dear all,

After conducting a panel on an unlabeled design and running a pilot (36 resp.* 8 scenarios each) , I have used the coefficients resulted from a MNL estimation of the pilot (which were significant) as prior means for a Bayesian design as follows:
Design
;alts = alt1, alt2, alt3, alt4
;rows=72
;block=9
;orth=seq
;eff=(mnl,d,fixed)
;bdraws = halton(500)
;model:
U(alt1) = C1+ b2.dummy[(u,3.366,12.67)|(u,7.776,15.026)]*carModel[0,1,2]+b3[(u,-2.907, -2.065)]*costhour[5,6,7]+b4[(u,-9.963, -6.677)]*costMile.43]+b5[(u ,-0.127,-0.059)]*WalkTime[5,10,15]+b6[(u,-0.574,-0.392)]*available[1,4,6]/
U(alt2) = c2+ " ------------" / (same atts, and leves)
U(alt3) = c4+ " ------------" $ (ame atts and levels)

Not surprisingly, when I run this design, no valid design is find, and and it warns me for too large prior means..
Any advice? cab I scale all the prior means by say divide all priors by 10...
Thanks for ant advice.
Best
Anat

Looking at the priors that you put in, they cannot be correct, they are way too large. Did you obtain them through estimation from pilot data? Then please check your results, as they cannot be correct, this cannot be an outcome of a model estimation. Also, values for uniformly distributed priors cannot be an outcome from estimation, only values for normally distributed Bayesian priors can.

Thank you very much Michiel!