Efficient design with require property and interaction effec

[Michiel Bliemer]

Estimating dummy and effects coded coefficients always require larger sample sizes for statistical efficiency than estimating linear effects, so often in pilot studies you may not find priors that are statistically significant. That means that your priors will be unreliable, and you should not pay too much attention to the S-estimates, which will also be unreliable because they heavily depend on the priors. If you are collecting data from a sufficiently larger sample size in the main data collection phase it is not so much an issue if you have good priors, you will mainly need good priors if your sample size is very small, since then you need all the efficiency you can get. Many people use simply priors set equal to zero to optimise the design, which works fine (but at some loss of efficiency).

Further, if there are no clear rankings coming out in the attribute levels, then you will likely not need to worry about dominance either.

If your data collection sample size is large, then I would just accept that the priors are highly unreliable, and use Bayesian priors close to zero with a wide variation around them to optimise the design.

Michiel

[Nonka]

Thank you very much again!
Then I will not go for more test responses and use close to zero priors.

Best wishes,
Nonka