choice-metrics.com

Dear all,

I designed a pilot choice experiment and used very conservatives priors based on expert opinion- for example 0.0001 for a level of attribute X.

I collected data and obtained priors from the pilot study. Should I feel comfortable using priors from the pilot in the main study?- for example, the coefficient for the same level of attribute X is estimated to be 1.2.

Thank you.
Nasrin

It depends on the standard errors or t-ratios of the parameter estimates. If the t-ratios indicate that all parameters are statistically significant, then the parameters are likely okay and you can use (n,estimate,standarderror) as Bayesian prior for each coefficient. If the t-ratios indicate that the parameters are quite unreliable, you may consider shrinking them a little towards zero, i.e. by dividing all parameters by a certain factor (i.e. 1.5).

Michiel

Hi Michiel,

Thank you for clarification. I have two more questions and would be pleased if you could answer

1- If a parameter has an unexpected sign, do you recommend using a uniform distribution as Baysian priors for that specific attribute?

2- When shrinking all parameters by a factor, do you also divide SD by that factor?

Thank you!