Have an idea?

Visit Sawtooth Software Feedback to share your ideas on how we can improve our products.

Is there a way to estimate only relevant interaction terms if using alt-spec variables?

Hi there,

We are currently designing quite a complicated shelf display conjoint (with 40+ SKUs). We've managed to get our client to simplify what needs to be tested quite a bit but we are still faced with one challenge: we have a number of promotional price attributes which are alternative-specific to certain sets of SKUs. Each attribute has one promotional attribute associated with it but these attributes differ between SKUs. We were hoping to be able to estimate interaction effects for these but because they're alternative-specific we end up with an 'ill-conditioned' message and huge standard errors in the test report because the software is trying to estimate interaction effects which don't exist (and have no data associated with them).
Is there a way to estimate only the relevant interaction terms when using alt-spec variables to avoid throwing the main effects estimates? Some kind of partial coding of these which could be done prior to estimation?

Many thanks.
asked Jan 12, 2017 by anonymous

1 Answer

0 votes
You could potentially reformat your data to get around this. With CBC/HB, you could code the entire interaction by hand. For analysis purposes, I might be tempted to add a new attribute that indicated the SKU grouping, then create an alternative-specific attribute for each set of SKU's. Then the software could estimate the interaction between the subset of SKU's and the promotional attribute associated with that set since each promo attribute would be fully interacted with the SKU's that pertained to it. You'd need to estimate the SKU Grouping utilities to tie the whole thing together, though, so this might complicate analysis a bit.
answered Jan 18, 2017 by Aaron Hill Silver Sawtooth Software, Inc. (9,745 points)
Hi Aaron,

Thank you very much for your reply.  That's a great idea - we'll try this. The other appreach we are thinking about is to perhaps recode the alt-spec attributes so that they're split into a larger number - and then we could make them specific to smaller sets of SKUs. But with this approach we might be more limited in terms of the levels we go down to as we effectively only have the interaction effects (no larger main effects). Your method might mean we could get more at the individual SKU level potentially if we have the  larger main effects in there as well.

Thanks again,