Pilot with insignificant coefficients and Bayesian priors

[claudiab]

Dear Ngene team,
I am in the process of designing a Bayesian efficient design. I just conducted my pilot study and estimated a MNL model. I obtained insignificant parameters (effects coded) and I wonder whether I should use these values as Bayesian priors in the Bayesian design.
Specifically, my main concern is that I do not know the sign of these attribute parameters. As such, following suggestions I found on this forum, instead of using the mean values of the insignificant coefficients as priors, I generated an efficient design using Bayesian priors with mean=0 with the standard errors obtained from the MNL estimates, e.g. (n,0,0.13), (n,0,0.06). In this way, I obtained a Bayesian design with S estimate = 25 which, I suppose, is good. Here is the code I used:

design
;alts = Alt1, Alt2, Alt3
;rows = 16
;eff = (mnl,d, mean)
;block = 2
;bdraws = gauss(3)
;model:
U(Alt1) = b1.effects[(n,0.54,0.09)|(n,-0.19,0.1)]*loc[2,1,0]
+ b2.effects[(n,0.28,0.12)|(n,0,0.13)|(n, 0,0.13)]*hops[3,2,1,0]
+ b3.effects[(n,0,0.06)]*org[1,0]
+ b4[(n,-0.16,0.02)]*price[7.99,9.99,10.99,12.99,16.99]/
U(Alt2) = b1*loc
+ b2*hops
+ b3*org
+ b4*price/
U(Alt3) = b0[(n,-2.78,0.28)]
$

I hope that this approach is correct and any further suggestions would be very appreciated.
I apologize in advance if this issue has already been discussed and thanks a lot for the support.
Claudia

[Michiel Bliemer]

If the parameters are not statistically significant, you indeed want to be careful putting in the estimates as means, since if the priors are large, this could have an adverse effect on the efficiency of the design. So it is always good to be conservative, and putting in 0 is the most conservative choice.

You do not need to worry much about the sign since you are not checking for dominant alternatives (there is no * behind your alternatives in your ;alts command), so having -0.001 and +0.001 is essentially the same as 0.000. So if you feel that your means are not too large, then you could just use them directly as means in your Bayesian priors, or you may wish to be a bit more conservative and divide them by two.

[claudiab]

Michiel,

thank you very much for your kind reply. Very helpful.

Claudia