MaxDiff Magnitude

In MaxDiff analysis, respondents are asked to evaluate a dozen or more items (such as attributes, statements, or brands) in sets of typically four or five at time. Within each set, they simply choose the best and worst items. Generally, an HB (hierarchical Bayes) application of multinomial logic analysis (MNL) is used to estimate individual-level scores for MaxDiff experiments. Other similar methods that develop individual-level scores by leveraging information across a sample of respondents may be employed (e.g. Mixed Logit or Latent Class with C Factors offered by Latent Gold software). However, for certain situations researchers may need to compute a reasonable set of individual-level scores on-the-fly, at the time of the interview, without the ability or benefit of combining the respondent.s data with a sample of other respondents.

One measure to consider when comparing scores from different methods is the standard deviation across the sample (a reflection of heterogeneity). The standard deviation is naturally larger for higher-scoring items and smaller for lower-scoring items. One way to examine the variation of the score relative to its absolute magnitude is to compute a Coefficient of Variation, which is done by taking the standard deviation of the item divided by its population mean. The results are shown below, again comparing Counts vs. HB for the 15 items.