choice-metrics.com

Hello,

I am currently using the demo version of NGene to run an explicit partial profile design, aiming to estimate the D-Efficiency which will influence my decision to purchase a license. However, I encountered some issues that I hope you can help me with.

Initially, I ran the code for 3000 randomly sampled rows from an externally created dataset. The process took 3-4 hours to reach a D-Error value that remained unchanged. This first run only considered single-level interactions (i.e., the attributes themselves) in the utility function. My rationale was that the dataset, designed to vary only 2 attributes between alternatives while keeping the rest constant, would inherently handle two-way interactions.

To verify this, I adjusted the syntax to explicitly include two-way interactions in the utility function. However, after 3 hours, the iteration history table still shows "Undefined" for the MNL D-Error, with no observable improvement.

Could you please elaborate on the potential causes of this issue? Your assistance would be greatly appreciated.

Thank you very much in advance.

Beyza

There could be several reasons for an undefined D-error, including an unidentified model, not enough rows in the design, too many restrictions, or multicollinearity in your candidate set. To resolve an unidentified model you need to specify utility functions whereby all parameters can be identified. But the first thing I would try is to increase the number of rows because adding interaction effects increases the number of parameters and hence more rows may be needed.

Michiel

Hello,

Thank you very much! Lastly, when we purchase the license, we are going to get the D-Efficiency value, too, right (not only the MNL D-Error value)?

Beyza

D-efficiency is only meaningful for orthogonal designs and where all attributes are coded using orthogonal-polynomial coding with zero priors. In all other cases, the minimum D-error value is not known and therefore it is not possible to meaningfully compute D-efficiency. Your design is not orthogonal so you should only look at the D-error.

Michiel