Labelled choice experiment, too many constraints?

[Michiel Bliemer]

Great! Happy to hear that.

Michiel

[LaïlaS]

Dear Professors,

As stated in my previous message, thanks to all your advice my code is running.

However I posted that message before noticing that first, it runs endlessly (multiple days) and second, the iteration history window only displays one iteration with an "MNL D-Error Undefined" (even after a current number of evaluation of 17545313 and a current number of invalid designs of 202500)... Which means that the model parameters are not identifiable (as I read in previous posts).

Does this mean that there is a problem with my model? Is it because all the levels of each alternative are determined by the label (i.e. the price, the CO2 emissions etc. are all strictly dependent of the time of heating system)? Would it help if I had more types of heating systems (I have 13 for the moment)?

Should I post something in the other part of the forum that concerns the design of stated choice experiment rather than this one about Ngene itself?

Thank you again so much for your help,

Laïla

[Michiel Bliemer]

Ngene will always run endlessly until the user stops the search process. Given that there exist typically billions of possible designs, Ngene simply keeps evaluating designs until the user stops the process.

Having an Undefined D-error is indeed problematic and means that the model is not identifiable with the data set collected through the experiment, meaning that certain parameters cannot be estimated. This can be caused by two things:
(1) Problematic utility functions
(2) Problematic constraints

In your case, I do not see any issue with your utility functions, so that means that you have likely imposed constraints in your candidate set that create perfect correlations between certain attribute levels. So please have a careful look at the constraints you impose. You can usually see in Ngene which are the problematic ones by inspecting the AVC matrix of the design with Undefined (infinite) D-error. In the rows/columns of parameters that cannot be estimated you will see extremely large values, indicating that the associate attribute levels do not vary in the design or are perfectly correlated with another attribute.

Looking at your constraints in your earlier post, I think you have created a lot of perfectly correlated attributes. For example, the first constraint:
if(optionA.TEC=1,optionA.INV=11000)

Since you are trying to estimate parameters for both TEC and INV you cannot perfectly correlate them, you should use something like:

if(optionA.TEC=1,optionA.INV=[9000,11000,13000])

or something like that, such that your attributes have some room to play. Otherwise, you will not be able to estimate both the TEC parameter and the INV parameter. You have created such perfect correlations with your other attributes as well it seems, such that everything is perfectly correlated with TEC.

Michiel