Have an idea?

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

Average Importance for an alternative-specific design CBC


I would like to know how I can interpret the average importance scores from the HB analysis for an alternative specific design (CBC).

For example, customers can choose between two alternatives with the following attributes when purchasing a product:

1. Delivery Offer
primary attribute (delivery type): Delivery
common attribute: Price
conditional attributes: Delivery time; Shipping costs

2. Pick-Up Offer
primary attribute (delivery type): self pick up
common attribute: Price
conditional attributes: Distance to the store; Service at the store

After the HB analysis my average importance scores are the following:
delivery type: 10
Price: 21
delivery time: 14
shipping costs: 16
Distance to the store: 30
Service at the store: 9

How can I interpret these values now? For example, the distance to the store has a average importance of 30%. However, not every offer has this attribute (only the pick-up offers).

The average importance score describes how much difference an attribute could make in the total utility of the choice alternative in a "normal CBC". But how is the importance of an attribute interpreted if it is not present in every alternative?

Thanks a lot!
asked Jan 3, 2020 by anonymous

1 Answer

+1 vote
Best answer
I would compute and interpret the importance scores within the context of each branch of the alternative-specific design.

For example, if my design involved bus, car, and train.  Each alternative has its specific attributes that only show up with it.  Also, there may be a few generic attributes that apply to all three alternatives.

I would take the raw HB utilities into your favorite analysis program...it could even be a spreadsheet program like Excel.  Then, compute the difference in utility for each attribute for each respondent (except for the principal attribute which contains the levels bus, car, and train).  Next, summarize the importance scores for Bus, by making the range of utilities for each attribute that apply to bus (including any generic attributes) sum to 100%.  Do the same for car and train.  

This makes the importance scores sum to 100% for the bus decision, the car decision, and the train decision.  You interpret the importance scores only within each of the principal alternatives, such as "assuming respondents are considering choosing the bus, what is the importance breakdown within the attributes that apply to bus that drives their decision"?
answered Jan 3, 2020 by Bryan Orme Platinum Sawtooth Software, Inc. (198,715 points)
Thanks for your answer Bryan!

by "raw HB utilities" do you mean the individual utilities (raw) or (zc diffs)?
After your answer, I would go on like this:
I would compute the difference between the attribute-level with the highest and the attribute-level with lowest utility per attribute (except for the principal attribute) for each respondent.

After that,  I would average these ranges of all respondents for each attribute. These "average ranges of utilities" for each attribute are then my basis to summarize the importance scores, by making the "average range of utilities" for each attribute that apply to bus (e.g.) sum to 100%.

Do I have that right?
By raw, I meant "raw".  But, you'll get the same answer using zero-centered diffs.

For each respondent, you would apportion the differences for attributes within each alt-spec category (such as bus) to sum to 100%.  That way each respondent gets equal weight into the population mean.
I'm very sorry but I'm not sure I got that right.

You mean I should compute individual attribute utility ranges and then I have to summarize the importance scores for bus (train and car) for each respondent and compute the average after that?
Yes, for each respondent and for each level of the principal attribute (e.g., bus, train, car), you should compute the importance scores for the applicable attributes such that they sum to 100% for each respondent.  E.g., bus attributes sum to 100%; train attributes sum to 100%; car attributes sum to 100% for each respondent.  Then, average those importance scores across respondents.
And...remember that you cannot compare the importance scores for an attribute that applies to bus to the importance score for an attribute that applies to train.  The importance scores should only be compared within the same alt-spec category.  E.g., only compare bus attributes to bus attributes, train attributes to train attributes, etc.
Thank you very much!