Have an idea?

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

Utility Estimation in Discover-CBC

i refer to the paper 'Discover-CBC: How and why it differs from SSI Web´s CBC Software'.
In that paper you write about the Utility Estimation in Discover-CBC.
Discover uses a 'maximum likelihood estimation via individual level logit with Bayesian smoothing toward population parameters via Empirical Bayes'.
Is it right that Discover first computes the maximum likelihood for a Multinomial Logit Model (for each individual) or is a individual level logit something else? And how is Empirical Bayes linked to that first step?
Do you have a more detailed description with equations for that whole computations?
For my master´s thesis i have to describe the way the utilities are estimated very detailed and mathematically.
So i would be so thankful if you could give me more detailed information.
asked Apr 4, 2017 by Elena

1 Answer

0 votes
The first step is to compute the pooled (aggregate) logit solution.  For that step, we use the standard and well-known aggregate MNL with the Newton-Raphson algorithm to find the maximum likelihood solution.  This is well-documented in other texts and articles in the industry.

The next step is as described in the document you reference from us.  We reformat the data to look like a "chip allocation" CBC, based on the lambda set in our software.  (The lambda determines the amount of information contributed by the pooled solution vs. the individual-level choices.)

Next, purely individual-level MNL is done again using the Newton-Raphson algorithm to find maximum likelihood (only using the reformatted chip-allocation tasks within each individual).  However, in each step the of the algorithm, the utilities are constrained (tied) for any levels within attributes that have prior preference order.   Prior order can be established by the researcher by setting an attribute to have best-to-worst or worst-to-best order.  Or, for non-ordered attributes (like Brand or Color) a prior ratings question could have been included in the questionnaire that asked respondents to rate each of the levels of, say, Brand on a 5-point scale.
answered Apr 4, 2017 by Bryan Orme Platinum Sawtooth Software, Inc. (132,290 points)