Efficient design using Effects coding

[reynold707]

Hello there,

I am trying to set up a syntax using the following code;

Design
;alts = alt1, alt2
;rows = 12
;eff = (mnl,d)
;model:
U(alt1) = B1.effects[0.003|0.004] * A[50,100,150] + B2.effects[0.003|0.004]*B[50,100,150] + B3[-0.025] * C[5,15,25]/
U(alt2) = B1 * A + B2 * B + B3 * C $

I just added the effects coding but i'm experiencing an inconvenience. Without the effect coding the number of respondents needed lies around 10-20. But with the effect coding this number rises to 600.000-800.000.

I dont understand why the number of respondents rises this hard.

Also any comment on the design is welcome ofcourse. This syntax will be the basis for a new study for Vehicle to Grid addaptation (drafting electricity from electric cars back to the network). unfortunately priors for A and B and the values of A and B are the same. I dont know if this is the reason why i need so many respondents?

Greetings Chris

[Michiel Bliemer]

Hi Chris,

Effects coding uses values of 0 and 1 for A and B, while your priors have been set to very small values (0.004), which clearly contribute little to utility when multiplied with one. If an attribute is almost irrelevant, you cannot estimate a coefficient for it.

Likely your priors make sense for attribute levels of 50 and 100, but you probably forgot to put in priors that make sense for effects coding. Please use priors from a pilot study in which you estimated the coefficients using effects coding instead of linear coding. Or simply convert the values so that they have similar utility impacts (0.003 * 50 = 0.15, 0.004 * 100 = 0.4, and 150 contributes -0.15-0.4 = -0.55 to utility). Note that the last level in Ngene refers to the base level. I am not sure the priors you provide make much sense (they seem strange), so maybe look a bit more into how effects coding works.

Michiel