ApolloChoiceModelling forum

Dear Stephane,

in my DCE I have three linear coded attributes. When I dummy code the attribute levels for A2 and A3 it seems that there might be a box-cox-transformation necessary as the connections between the levels don´t seem to be linear but diminishing, see below.

b0 -0.4386
b_A1 -0.5713
b_A2_level1 1.2657
b_A2_level2 0.8188
b_A2_level3 0.3729
b_A2_level1 0.0000
b_A3_level1 0.5924
b_A3_level2 0.3442
b_A3_level3 0.2207
b_A3_level4 0.0000

I did neither find any information on the box cox transformation in the apollo manual nor in this forum. Can you advise on the logik how to implement it in the code for a MNL or MMNL model?
Furthermore, I have read in the literature that box cox can only be used for positive values, but my linear coefficients (see below) are negative. Is it therefore not possible to use box-cox-transformation in this case?

b0 0.25984
b_A1 -0.63877
b_A2 -0.16785
b_A3 -0.07220

Thank you for your help.
Best,
J.

Hi

first, in relation to the sign needing to be positive, this refers to the attribute, not the parameter. So as long as your attribute has only positive levels, this is not an issue.

What you would need to do is write something like this:

b_A1 * ( A1 ^ lambda_A1 - 1 ) / lambda_A1

but with only 4 levels, you might as well just dummy code it and cover the non-linearity that way

Stephane