Have an idea?

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

Interaction effect in HB estimates.. Is it good?

Hello Everyone,

I have a question on the interaction effect calculation while estimating utilities in CBC-HB.

1. Is it a good practice even if it improves the RLH of the estimates?
2. In the estimates, all the interaction effect utilities are calculated against every respondent.
For willingness to pay calculation, how is it calculated then?
Without the interaction effects, WTP = (Sum of relevant individual partworth utilities)/price utility coefficient.

When interaction effect is calculated, how is it computed then?
The sum of the part worth utilities = main effects utilities + interaction effects where the main attributes in consideration are there?

Kindly advise.

asked Apr 14, 2018 by sacharya (490 points)

1 Answer

0 votes
Best answer

Improving RLH by itself isn't good enough.   If you really want to know whether to include an interaction I suggest you use the Model Explorer.  This is free software you can download here:  http://www.sawtoothsoftware.com/support/downloads/tools-scripts

For estimating WTP in the presence of interactions, I recommend using the share equalization approach.  Run a simulation with two concepts.  Have these be identical except as described below and have them be realistic concepts that might contain your attribute and level of interest.  Have one of the  two (otherwise identical) concepts NOT contain the attribute level you want to value and have the concept appear at some price X.  Have the other concept contain the level you are trying to value and keep increasing the price (X) until the two concepts have equal shares.  This is the share-equalizing price estimate of WTP.  

You can do this in more involved ways if you like (e.g. put these two concepts in a full competitive set of concepts to capture the effects of competition) but the logic is the same - concept A without level of interest and at a lower price versus concept B that HAS the level of interest but at a higher price.  The difference in price between the higher price with the level of interest and the lower price without it is an estimate of WTP.
answered Apr 14, 2018 by Keith Chrzan Platinum Sawtooth Software, Inc. (93,025 points)
selected Apr 14, 2018 by sacharya
Hi Keith,
thanks for your fine reply.
I have a small question though. To run the Model explorer, I need to have the I have CBCHBCon.exe and this seems to be not there.
I have Lighthouse Studio 9.5.3.

For the second suggestion on share equalization, I am working on it at the moment. Will update you once I am done with this.
Hmmmm, if there's a file missing please contact support@sawtoothsoftware.com.
I will raise this with the support. thanks again
Hi Keith,
I tried your recommendation. Very smart..It works.
Just a small question to be sure I am doing the right thing -
If I have to don't want an attribute level as you mentioned above, will it be "N/A" or "null" (null is the attribute level defined within the design)?
Yes. I assume you mean you have attribute present and not present (null), right?
Just to clarify, not present = null.

Sorry, I just realized how bad my English was in the previous comment. Sorry for the confusion though.
So, I will choose "N/A" then instead of "null" as both the options are shown in the drop down menu against each attribute.
Hi Keith, Good morning.
I am bit confused on the results I am getting when I use N/A or Null in the simulation settings.
If I use N/A, the share of preference is high compared to Null. I guess it has to do with the negative utility for null. If I use N/A, it means the utility scores are completely ignored in the calculation. Is this correct inference?

So, to find the optimal product combinations which matches the share of preference, should I use N/A or Null ?

Kindly advise.
Null.  You want the utility for having the worse level of the attribute rather than the N/A setting which turns the attribute off (and which as you have seen is NOT the same as having the "absent" level of the attribute).
Thanks again Keith for your help as always. I applied "null" always now.

The reason why I applied "N/A" to some of the combinations is due to prohibitions initially applied in the design as the service attributes were never shown to the respondents.

For "Null", I applied it only when the combination contained those service attributes and was not shown due to randomness.

But I apply Null now to sections when I am simulating. The only warning I get is the "Prohibition" as I explained above.

Thanks again.