MaxDiff Analysis

Top  Previous  Next

When you click MaxDiff Analysis, a score ranging from 0 to 100 is computed for each item for each respondent.  The scores sum to 100 across MaxDiff items.  The scores are ratio scaled, meaning that an item with a score of 10 is twice as preferred (or important) as an item with a score of 5.  The most typical visualization is to list the items from highest to lowest scores for the sample as a whole, or for key subgroups of respondents.


MaxDiff is like a "beauty contest" among the items.  The utility scores are developed based on the relative comparisons among the items in the study.  However, there is no information available regarding whether the items are all very much liked or very much disliked by an individual.  The scores are only on a relative scale.  For this reason, it is helpful to include a wide variety of items in your study ranging from less desirable (less important) to more desirable (more important) items.   Adding a level that has direct monetary meaning or a level that represents the status quo (neutral) can also help with MaxDiff's relativity issue, since the item scores can be compared to the scores for those specific reference points.


When you click MaxDiff Analysis, Discover's MaxDiff uses a type of individual-level score estimation called empirical Bayes.  Empirical Bayes is extremely fast and produces high quality results when you follow the recommendations in the software for number of questions asked of each respondent (so that each item appears around three or more times per respondent).  For more information about Discover-MaxDiff analysis, see Discover-MaxDiff: How and Why It Differs fromLighthouse Studio’s MaxDiff Software.


NOTE: For advanced users who wish to estimate MaxDiff scores using hierarchical Bayes (HB), under Collect go to Data | Advanced Settings | Export MaxDiff Design & Choices.  This will allow you to export a CSV file in the proper format for reading into Sawtooth Software's standalone CBC/HB software for HB estimation (which requires a separate software subscription).


Segmentation Analysis


It's common to compare groups of people based on their MaxDiff scores (after you click MaxDiff Analysis, choose Segment by on the Scores tab).  The groups could be based on questions you included in your Discover survey, or you can also import a CSV file that contains new data to use (under Collect go to Data | Variables | Import Data).


Note for advanced users: Some researchers like to use MaxDiff results to find latent segments (or clusters) representing groups of people who have similar needs or opinions.  You can export the individual-level item scores from Discover-MaxDiff to a CSV file, which you can submit to your favorite clustering package.  We should note, however, that many statisticians prefer to develop latent segments for MaxDiff data using Latent Class choice analysis.  You can export the raw MaxDiff data to a CSV file appropriate for opening within Sawtooth Software's standalone Latent Class package (under Collect go to Data | Advanced Settings | Export MaxDiff Design & Choices).  Sawtooth Software's Latent Class package requires a separate software subscription.

Page link: