Sequential Regression: A Neodescriptive Approach to Multicollinearity

Classical regression analysis uses partial coefficients to measure the influences of some variables (regressors) on another variable (regressand). However, a descriptive point of view shows that these coefficients are very bad measures of influence. Their interpretation as an average change of the regressand is only valid if the regressors are weakly correlated, and they are useless when the degree of multicollinearity is high. Despite these obvious flaws there is a lack of alternative ideas to measure influences. On that score this paper proposes two new coefficients of influence: (1) A supplementary coefficient measures the additional influence of a regressor when certain variables are already taken into account. (2) A particular coefficient, which is a mean of certain supplementary coefficients, allocates the influence of a regressor within the collective influence of all regressors. Both new coefficients can directly be interpreted as average changes of the regressand.