Dominance Analysis for the Relative Importance of Predictors

Description

Ranks predictors by comparing their added contribution to R-squared across all subset models, giving complete, conditional and general dominance that sums to the full-model R-squared.

Key-driver questions ask which attributes matter most, yet standardised regression weights are not importance measures: with correlated predictors, suppressor effects can make an important driver look trivial or even negative. Budescu defined importance by contribution to fit, compared directly between pairs of predictors across all subset models. Predictor X completely dominates Y if it adds more to R-squared in every subset model containing neither. Because complete dominance can leave pairs unranked, Azen and Budescu added conditional dominance (mean added contribution within each model size) and general dominance (the average of those means across sizes). General dominance values partition the full-model R-squared, which makes them easy to report as shares of explained variance, and they coincide with the Lindeman, Merenda and Gold (Shapley-value) decomposition. In exploratory work, criticality analysis complements this by bootstrapping how often a predictor appears in the best subset models.

Attributes

Loss type
Siloed
Effort
Days

Steps

  1. Fit all 2^p - 1 subset regressions of the outcome on the p candidate drivers and record each R-squared (practical up to about 15-20 predictors).
  2. For each predictor and each subset not containing it, compute its added contribution to R-squared.
  3. Compare every pair across all subsets that contain neither to establish complete dominance where it holds.
  4. Average contributions within each model size for conditional dominance, then across sizes for general dominance; check general dominance sums to the full R-squared.
  5. Bootstrap the analysis (for example 1,000 resamples) to show how often each dominance relation and rank reproduces.
  6. Report the dominance matrix and the R-squared shares, stating that importance is relative to this predictor set.

Sources

  • The SAGE Handbook of Quantitative Methods in Psychology (Millsap & Maydeu-Olivares, 2009)

Also known as

  • Budescu dominance analysis
  • Complete, conditional and general dominance
  • Criticality analysis
  • Dominance analysis

Categories

  • Analytics & Behaviour
  • Statistics & Modelling