different coeffients for one attribut

[mdann]

hello everyone,

to create an efficient design I create a pretest with an orthogonal design within the attribut alt= b1+ b1*A[1,2,3,4]+ b2 * B[....].... where the levels stand for ordinal aspekts. 1 means ripping, 2 half ripping, 3 litter, 4 rubber mat (I dont`t know if eine use the right translation, but the content is important).

To create a mixlogit model I coded this attribut to dummy variables. As results I got for each level one a b-coefficient and the standard deviation. e.g.:

means rippimg: b=-0,8, std=0,2
half ripping: b=-0,3, std=0,15
litter: b=0,4, std= 0,1
rubber mat: b=0,2, std=0,3

Q1: Means that for creating an efficient design, that each level has to be transformed to a new dummy coded attribut like: ??

...b1[n,-0,8,0,2] * A[0,1] + b2[n,-0.3, 0.15] * B[0,1]..

Q2: Or is there a possibility to put all coefficients together so that I have already one attribut as before???

Q3: I saw in some syntaxes a vertical bar. What does this mean?

Thank you in advance!!

[Michiel Bliemer]

The syntax b1.dummy[x|y|z] * A[1,2,3,4] means that A = 4 is the reference level (with a zero utility attached), A = 1 has a utility (coefficient) of x attached, A = 2 has a utility (coefficient) of y attached, and A = 3 has a utility (coefficient) of z attached.

In your case, your mixed logit model would be something like:

b1.dummy[n,0.8,0.2|n,-0.3,0.15|n,0.4,0.1] * A[1,2,3,4]

Note that the number of coefficients (3) is always one less the number of attribute levels (4), because the last level is the reference level. Therefore, I do not understand your estimation results, because one of the attribute levels should have a zero coefficient as the reference level. So which attribute level did you use as reference?

Michiel

[mdann]

Hey Michiel,

thank you a lot!

I got 4 coefficients by changing the basic in a second process.

Micha

[Michiel Bliemer]

It does not make sense to change the reference level and obtain the final coefficient, as changing the reference level will impact upon all parameter estimates. You have to select a single reference level and stick with it, and only estimate 3 coefficients, you cannot estimate 4 coefficients. You can use effects coding or orthonormal coding instead of dummy coding if you wish.