Have an idea?

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

Should HB analysis be done by segments

We have four segments that we believe will behave very differently.  When we analyze the CBC, should we run HB analysis seperately for each segment?  Or can we put all data through together?
asked Apr 2, 2014 by HongHu Bronze (835 points)
The classic answer to this question was given in the 2001 paper: https://www.sawtoothsoftware.com/downloadPDF.php?file=hbfit.pdf

And, I agree with Keith's comments about the possible use of Covariates.
It is indeed a very interesting workpaper.  

Build on top of this question, what if we have used different values for certain attibutes?  For example, we used three levels for the price attribute.  For different segments, we have used different levels for the low, mid and high price.  So even though the CBC structure is the same, the respondents from different segments actually saw different price levels.  Does this mean that we should treat them as different CBCs?
I think it would depend on how different the price structure was.  If they saw different currencies, but the relative prices were similar, then I'd think it would be safe to toss them together and include a covariate to potentially capture any minor scale/preference differences.  If one group saw a very tight price range and another saw very wide prices, or one group saw low/med/high and another saw high/med/low, then covariates aren't going to help much.
Makes sense.  Thanks Aaron!

2 Answers

+1 vote
Hello, Hong.

Our software now allows the use of covariates in the HB estimation.  I think a good way for you to proceed would be to put all the respondents into the analysis together, but to use the variable identifying segment membership as a covariate.

Using covariates allows (but does not cause) systematic differences between segments to emerge from the analysis.
answered Apr 2, 2014 by Keith Chrzan Platinum Sawtooth Software, Inc. (92,075 points)
Thanks Keith!
+1 vote
Personally, I've found anything quota'd that is subsequently measured in the conjoint, such as 'current main brand' and also potentially sample source to be much more useful than segmentation covariates.

The quota bias partly accounts and controls for any differences due to oversampling particular groups (usually the client brand). Without 'current preference' sheer numbers of cases can 'infect' the smaller part of the sample. Any systematic quotas that are subsequently measured in the conjoint that don't 'naturally' fall out are candidates for this first bias.

Sample source is not always relevant, but it can help to offset systematic bias in sampling, which would be a hidden 'gotcha' if you don't look for it.

As I understand the relationship between HB and covariates, HBs strengths are that it can leverage population data to infer individual scores. Covariates strengths are that they can identify the correct population to infer from.
answered Apr 9, 2014 by Andrew Reynolds Bronze (1,140 points)
Totally agree.  Great point!