Additional ranking, rating, or sorting tasks are completed outside the MaxDiff exercise and then added (augmented) as new choice tasks to each respondent's MaxDiff data for utility estimation. For example, respondents might be asked to rank-order the 6 items chosen "best" across the previous 6 MaxDiff sets. The rank-order judgments may be exploded into paired comparisons (or other related coding approaches) for those 6 items and added to the choice data set. Augmented approaches can lead to even more accurate measurement of the top few items for each respondent, assuming the augment focuses on obtaining more information about best items. Other augmentations are possible and have been proposed, including augments that focus on obtaining more precision for the worst few items. Typical analysis is HB-MNL; aggregate logit or latent class MNL may also be used.

Originally called "Adaptive MaxDiff" by the author (Orme) in 2006, but to avoid confusion it's probably easier to think of it as Tournament MaxDiff, because it proceeds similar to a round-robin tournament in sports competitions. For each respondent, items that are selected "worst" are dropped from that same respondent's later MaxDiff sets. Later sets compare winners vs. winners, until an overall winning item is identified. The sets can decrement in complexity, from 6 items at a time, to 5 items, etc. until a final ranking is done among the remaining 2, 3, or 4 winning items. The utilities are typically estimated via HB-MNL, though aggregate logit and latent class MNL are also possible. The benefits include increased respondent engagement in the exercise and improved accuracy at the individual level for the best items per respondent.

Most MaxDiff approaches lead to relative (ipsative) scores such that we don't know whether any of the items is good or bad in any absolute sense. This often isn't a big issue, especially if a wide variety of items have been included in the experiment. With Anchored MaxDiff, we can establish whether each of the items is above or below some threshold anchor representing an important/not important, good/bad, or buy/not buy threshold. For anchoring, we ask additional questions wherein respondents indicate whether selected items are important/not important, good/bad, etc. Respondents do not need to evaluate all items with respect to the anchor. The additional anchoring questions are added to the choice data set as additional comparisons vs. the "anchor item" (threshold). Utility estimation may be done with HB-MNL, aggregate logit, or latent class MNL. The utility of the anchor threshold is typically set to zero, such that items with positive utilities indicate that they are important, good, or a "buy." Items with negative utilities indicate items that are not important, not good, or not a "buy." Anchored MaxDiff may be combined with any of the other six flavors of MaxDiff described in this article.

Used when the number of items is about 40 or more. The approach involves showing each item to each respondent about 1 or fewer times. For example, with 60 items in the study, 15 sets per respondent and 4 items per set, each item can show 1x per respondent. Or, with 120 items in the study, 15 sets per respondent and 4 items per set, it then takes 2 respondents to show each item once. Because the data are so sparse (or even missing) at the individual level, we estimate scores typically via pooled analysis such as aggregate logit or a latent class MNL. However, some authors have shown that HB-MNL can do a reasonable job even if each item is shown just 1x per respondent. (Though more accurate estimates certainly could be obtained by showing each item 2x or 3x per respondent.)

Used when the number of items is about 40 or more. The approach involves randomly (or via a blocked design) drawing a subset of the items (such as 30 out of 60 items) for each respondent such that each item drawn per respondent can be shown 2x or 3x to each respondent. There are two potential advantages for Express MaxDiff over Sparse MaxDiff: 1) respondents don't need to orient themselves to so many items within the same questionnaire leading to some potential cognitive efficiencies, 2) if running HB-MNL, a fit statistic (RLH) can be computed for each respondent leading to the possible identification of inconsistent respondents. Despite these potential benefits, we've found under a variety of tests that Express MaxDiff performs a bit worse than Sparse MaxDiff. The typical approach for analysis is aggregate logit to summarize scores for the sample.

Used when the number of items is about 50 or more, extending potentially into the 100s of items. Like Express MaxDiff, we draw a subset of the items (such as 30 out of 60 items) for each respondent such that each item drawn per respondent can be shown 2x or 3x to that respondent. Unlike Express MaxDiff, this is an across-respondents adaptive approach that draws items for respondents (using Thompson Sampling) that have tended to be preferred by earlier respondents. Thus, better items for the population are oversampled dramatically compared to worse items. When the goal is to identify the top few items for the population, Bandit MaxDiff can be 2x to 4x more efficient than Sparse or Express MaxDiff.