Announcing a Significant New Breakthrough: The ICE
Module for Estimating Individual-Level Utilities from CBC
Choice-Based Conjoint is receiving a lot of attention lately.
Researchers and clients alike value the realistic context of
making choices from sets of available products. However, using
CBC has been bittersweet. Because choices are an inefficient
way to collect information about preferences, we have been
compelled to use aggregate analysis.
Researchers and academics have argued that respondents are
unique. The world does not consist of clean market segments.
Aggregate models which assume no respondent differences
cannot be optimal. Last year we made an important step toward
recognizing respondent heterogeneity with the release of our
Latent Class Module. But no software has been available
for estimating individual-level utilities from Choice data until
The ICE (Individual Choice Estimation) Module is an
add-on to the CBC system. With ICE, you'll get:
Our goal at Sawtooth Software is to provide sophisticated
research tools for those who are not necessarily specialists in
statistics. The ICE Module upholds that tradition. As with all
our products, the ICE Module comes with a 60-day money-back
- Rapid estimation of individual-level utilities
- More accurate simulation models
- A powerful market simulator (as used in ACA/CVA),
- an easy-to-use user interface with pull-down menus
- automatic sensitivity analysis and multi-scenario batch processing
- respondent-level weighting
segmentation scenarios without having to re-run logit multiple times
Ten Reasons Why You Need ICE
- Increased Model Accuracy. If your respondents have unique preferences, you'll
get more accurate market simulations with ICE. We've analyzed dozens of real and synthetic
data sets and confirmed that ICE generally improves predictions for holdout concepts relative to
aggregate logit and latent class.
One data set was collected by Huber and Zwerina (1996). They interviewed 50 MBA students
and asked 30 choice questions. The best Mean Absolute Error (MAE) values for share
predictions for aggregate logit, Lclass and ICE were 12.9%, 6.3%, and 2.7%. If we only use
aggregate logit, our aggregate share predictions are off by an average of 12.9 share points. With
ICE our predictions only miss by 2.7 share points. We confess that this data set is atypical, and
represents the most dramatic margin of victory we've seen for ICE. Even so, results across many
studies show ICE generally gives more accurate predictions.
Another data set was contributed by Pathfinder Strategies, in Australia (1997). Sample size was
880, with 21 tasks per respondent. The best MAE results were 8.2% for aggregate logit, 8.1%
for lclass and 7.9% for ICE. Diagnostics from lclass suggested this market was relatively
homogeneous. Even so, ICE predictions were slightly more accurate than aggregate modeling.
- Greater Flexibility in Simulations. Having individual-level utilities means not
having to run logit separately for different segments before computing segment-based share
predictions. ICE uses the same simulator as used in Sawtooth Software's ACA and CVA
conjoint systems. This is an easy-to-use tool you can confidently give your clients. You can
merge segmentation variables and toggle quickly between segments, process multiple scenarios,
and even conduct automatic sensitivity analysis. ICE also gives capabilities for including
customized calibration concepts (as offered in ACA) in your CBC interview. Calibration
concepts scale the data for use in Purchase Likelihood simulations, which aren't available in
- No Need to Be Limited to Small Designs. The ICE module leverages information
from all respondents to estimate results for the individual. Designs don't need to be kept
impractically small. A CBC study with six attributes and 30 total levels can be comfortably
implemented if you ask around 20 to 30 tasks per individual. We have seen ICE estimation
match or exceed aggregate modeling accuracy with as few as 8 tasks per respondent, though we
would never suggest using so few tasks with ICE.
- Estimation is Relatively Fast. Recently, leading academics and researchers have
been investigating the use of Hierarchical Bayes for estimating individual-level utilities from
Choice. Their run times have been measured in terms of dozens of hours for typical data sets.
The typical ICE run takes about 30 minutes on a Pentium 166 MHz processor.
- Reduces IIA Problems. A long-standing complaint against logit has been its IIA
assumption, often referred to as the red bus/blue bus problem. Very similar products tend to
capture too much net share in competitive logit simulations. With individual-level modeling,
this effect is less problematic. If you use the First Choice model, it is entirely resolved.
- Cleaner Data. Conjoint analysis can suffer from inconsistent respondents and
from utilities which seem out of order (reversals) for variables such as price. ICE reports a
goodness of fit for each individual, and the ICE simulator can discard those cases which don't
meet your specified threshold. Utility constraints can also be imposed for attributes such as price
which have a definite order of preference. Both of these safeguards can improve the quality of
- Individual-level Utilities Are Valuable for Additional Analyses. ICE utilities
may be used in standard cross-tabs and for additional multivariate modeling using software of
- Straightforward Significance Tests. With individual-level models, statistical tests
based on sampling variation are straightforward. The market simulator reports standard errors
for shares of choice. Additionally, individual-level utilities, attribute importances and shares of
choice are available in ASCII files for performing t-tests.
- If You Own Lclass, ICE Is the Logical Extension. Lclass users should be pleased
to know that their Lclass Module still offers unique value to Choice analysis. Even though ICE
can estimate utilities "from scratch," it usually benefits from a latent class starting point. We
also believe that latent class is a robust approach for investigating market structure and assigning
segment membership using choice data.
- ICE Is Easy to Use. The ICE system has a simple menu system which leads you
each step of the way. ICE uses your existing data file (studyname.CHO) and attribute definitions