Statistics Glossary → Proportional reduction in error measures (lambda, Goodman-Kruskal τ, uncertainty coefficient)
Proportional reduction in error measures (lambda, Goodman-Kruskal τ, uncertainty coefficient)concept
Measures of association that show by what percentage knowing one nominal variable reduces the error in predicting a second nominal variable.
The proportional reduction in error (PRE) idea is PRE = (E₁ − E₂) / E₁. E₁ is the prediction error without knowing the independent variable, and E₂ is the error when it is known. Goodman-Kruskal λ uses the rule 'always predict the most frequent category'. Goodman-Kruskal τ uses the rule 'predict categories at random in proportion to their observed frequencies'. The uncertainty coefficient (Theil's U) measures error with entropy. All three range from 0 to 1, where 0 means no improvement and 1 means perfect prediction.
These measures are asymmetric: the answer to 'how well does knowing A predict B?' can differ from the reverse, so the dependent variable must be stated. Symmetric versions also exist. λ can be 0 even when there is a real association, if the same category remains the most frequent at every level of the independent variable. In that case τ or the uncertainty coefficient is more sensitive. Unlike chi-square-based measures such as Cramér's V, PRE measures have a direct 'improvement in prediction' interpretation.
An everyday example
Of 100 voters, 50 vote for party A, 30 for B and 20 for C. Without knowing the region, predicting 'A' for everyone gives 50 errors. If 40 of the 50 northern voters choose A and 25 of the 50 southern voters choose B, predicting A in the north and B in the south cuts the errors to 10 + 25 = 35. λ = (50 − 35) / 50 = 0.30: knowing the region reduces prediction error by 30%.
In MerQur
In the chi-square test of independence, the 'λ, GK τ, uncertainty coefficient' checkbox gives the nominal PRE measures; in the crosstab, 'Additional outputs: Association measures' provides lambda and the uncertainty coefficient for nominal variables.
Analyses where this concept appears
Chi-Square IndependenceCrosstab
Related concepts
Ordinal association measures (gamma, tau-b, tau-c, Somers' d)Cramér's V and Cohen's w