choice-metrics.com

Dear Ngene users,

We have conducted a pilot survey and estimated an MNL model. The results seem fine but five of the 11 estimated parameters are highly insignificant. I know that data from pilot studies most likely lead to insignificant estimates but the problem is whether to use them as priors or not. I learn from the forum that we must be very cautious when using these estimates and suggestions have been put. Based on these suggestions, I specified the design as follows. I have some doubt whether I am correct or not, can you help me? Thanks indeed!


Design
; alts = alt1*, alt2*, SQ*
; rows = 36
; eff = (mnl,d, mean)
; block = 3
; Model :
U(alt1) = nut.effects[0.734|0.588]*nutrition[1,2,3]
+fsafe.effects[0.359|0.383]*foodsafety[1,2,3]
+recom.effect[(n,0.12,0)|(n,0.3,0)|(n,0.15,0)]*recommend[1,2,3,4]
+locate.effects[(n,0.1,0)|(n,0.17,0)]*location[1,2,3]
+cost[-0.008]*price[50,60,70,85,105,120]/
U(alt2) = nut*nutrition+fsafe*foodsafety+recom*recommend+locate*location+cost*price/
U(SQ) = b[0]$

And note that the insignificant variables are recom and locate, and it is only for these attributes I used Bayesian distribution, the rest are fixed as they are significant.

I am sorry for some silly questions but running the above syntax yields some overlapping levels in four to five scenarios (especially for nut and fsafe attributes). I guess this might be related to the priors but those fixed priors are obtained from the pilot survey.