How estimate the priors in the pilot study syntax?

[sukunta]

Dear Prof.Michiel Bliemer,
Happy New Year 2020 ^___^.
My pilot model has two attributes with two dummy levels, one attribute with three dummy levels and two attributes with four continuous levels. I estimate the prior value of cost attribute -0.0001 and the screening format 0.0001 (The qualitative data proposed 17 customers like type 1 and the less like type 0).
My syntax is

Code: Select all

design
;alts=alt1,alt2,alt3
;rows=12
;eff=(mnl,d)
;model:
U(alt1)=b1.dummy[0.0001]*format[1,0]+b2.dummy[0]*con_ser[1,0]+b3[0]*time[10,20,30,40]+b4.dummy[0|0]*area[1,2,0]+b5[-.0001]*cost[0,100,200,300]/
U(alt2)=b1*format+b2*con_ser+b3*time+b4*area+b5*cost/
U(alt3)=b0[0]
$


My questions are
1. Are the prior values appropriate?
2. If I finish the pilot study, the prior values are the coeffients of the model, or not? and if the coefficient of the model not significant, could I use the prior values from the pilot study?
3. When I stop the running program? How long?
Sincerely yours,
Sukunta Muadthong

[Michiel Bliemer]

1. If you are not checking for dominant alternatives, then you can simply set all priors equal to zero. If you would like to check for dominant alternatives, then you should use:

;alts = alt1*, alt2*, alt3

together with small positive or negative values as you do for cost and format. Does time not have a sign? And is there an ordering for area? And how about the other attributes? If you know the sign of the priors or ordering of the dummy coded variables then you can add these small positive/negative values together with the command above.

2. When you estimate an MNL model using data from the pilot study, you obtain parameter estimates beta and standard errors se. Then you can use normally distributed Bayesian priors (n,beta,se), e.g. b5[(n,-0.01,0.005)] if the cost parameter is -0.01 with a standard error of 0.005. You can also do this for parameters that are not statistically significant. For Bayesian priors, you need to add something like ;bdraws = gauss(3) and set ;eff = (mnl,d,mean).

3. You stop running when the (Bayesian) D-error does not decrease much anymore. For this pilot study syntax this will only take one minute or so. When running syntax with Bayesian priors, it may take a few hours.

Michiel

[sukunta]

Dear Prof.Michiael,
Thank you for your prompt reply.
Sincerely yours,
Sukunta