Shapley Values (SV)

Shapley Values (SV) is relatively well know in the academic world, but only a limited number of consulting agencies adopted it, due to the limitations of most commercial algorithms (e.g., speed of analysis, maximum number of attributes that can be analysed).

  • SV analysis is a multivariate statistical approach where an overall metric (e.g., likelihood to recommend) is modeled against various product features.
  • SV allow assessing the contribution of each attribute by accounting for the potential overlap between attributes.
  • SV overcomes the limitations of correlation and regression analysis:
    • all attributes are considered in the model at the same time;
    • SV can handle correlated attributes and is well suited to estimate their contribution towards an overall metric.
  • Therefore, SV is particularly indicated for the assessment of drivers, allowing the estimation of the impact of each attribute:
    • SV algorithm is robust, with results being consistent even when some attributes are removed from the analysis or others are added;
    • the approach is ideal for comparisons in time (between waves) or space (across markets).

The functions derived.importance and derived.importance.manyvar available in R-swDrivers allow Shapley Values analysis and can handle most projects, even when many drivers (30+) need to be assessed.

In SV analysis, there are two main deliverables:

  • Impact scores: presented as percentages of contribution towards the overall metric (e.g., 30% impact). They are rescaled to add up to 100% to facilitate interpretation and actionability.
  • A goodness-of-fit indicator (similar to the R-square of linear regression) to indicate to what extent the attributes included in the analysis explain the overall metric. A high score suggests that no significant driver has been missed when developing the questionnaire.