Is the unexplained component necessarily discrimination?

No. It can reflect many modeled and unmodeled differences.

Does changing the reference group change the total gap?

No, but it can change how the components are allocated.

Is every Kitagawa decomposition a regression?

No. The original rate decomposition uses composition shares and stratum-specific rates.

AI Glossary / Statistics and probability / Advanced

Kitagawa - Oaxaca - Blinder decomposition

Definition

Kitagawa - Oaxaca - Blinder decomposition is a family of methods for expressing a difference between groups as components associated with their composition and with within-group rates or fitted relationships. Kitagawa's original approach concerns differences in aggregate rates across strata; the Oaxaca - Blinder formulation commonly uses linear regressions to decompose a difference in outcome means. The methods are related, but their setups should not be collapsed into one universal formula. A decomposition is primarily an accounting description of a modeled gap; causal explanations require additional assumptions and evidence.

What it means

An overall difference can arise because groups contain different mixes of observations, because outcomes differ within comparable categories, or both. For example, two service regions may differ in the mix of request types and in resolution times within each type. Decomposition makes those contributions explicit under a selected reference structure. In regression versions, the terms are often called explained and unexplained. These labels mean explained by the included variables and coefficients under that model, not fully understood versus mysterious, and not justified versus unjustified.

How it works

For a standard linear mean decomposition, include an intercept in the predictor vector and write the fitted group means as x̄AᵀβA and x̄BᵀβB. Using group B's coefficients as a reference, their difference equals (x̄A − x̄B)ᵀβB + x̄Aᵀ(βA − βB). The first term uses composition differences at B's coefficients; the second uses coefficient differences at A's composition. Other reference coefficients give other allocations while preserving the total gap. A threefold decomposition separates a composition term, a coefficient term and their interaction. For stratified rates, analogous calculations use category shares and category-specific rates rather than requiring a regression model.

Three practical examples

Service comparisons: separate an average-duration gap into differences associated with request mix and fitted within-mix relationships.
Workforce research: describe a mean outcome gap using explicitly selected predictors and reference coefficients, without equating the remainder automatically with discrimination.
Demographic analysis: separate differences in crude rates into population-composition and stratum-specific rate components. Substantive interpretation depends on the variable choices and the population being compared.

Think of it this way

Think of comparing two baskets whose average values differ. Part of the difference may come from what each basket contains, while another part comes from the values attached to those contents. A reference rule specifies which basket's values are used to price the composition difference. Changing that rule can change the allocation between components even though the observed total difference remains the same.

Worked example

Suppose group A has mean predictor 6 and fitted equation y = 10 + 3x, while group B has mean predictor 4 and fitted equation y = 8 + 2x. Their fitted means are 28 and 16, giving a gap of 12. Using B's coefficients, the composition component is (6 − 4) × 2 = 4. The coefficient component is the intercept difference 2 plus 6 × (3 − 2) = 6, totaling 8. Thus 4 + 8 = 12. Using A's coefficients instead gives composition 2 × 3 = 6 and coefficient component 2 + 4 × 1 = 6, again totaling 12. A threefold B-based allocation gives composition 4, coefficient difference at B's composition 6 and interaction 2. All add to the same gap. None of these arithmetic decompositions proves that changing x would causally change the outcome by the associated component.

Illustrative case study

Fictional case study: Cypress Services studies average resolution times across two regions. Its decomposition attributes part of the gap to the mix of requests and part to different fitted relationships. Reviewing the data reveals an omitted complexity measure and different recording practices. The team adds sensitivity analyses and avoids labeling the coefficient component as pure operational inefficiency. It reports the reference group and explains how alternative references change component sizes. The analysis helps organize questions for investigation while leaving causal and measurement issues explicit.

Common misconceptions and mistakes

Common mistakes include calling the unexplained component a direct measure of discrimination or inefficiency, treating the explained component as inherently legitimate, and hiding the reference-coefficient choice. Omitted predictors, measurement error, functional form and selection into the sample can affect components. Poor overlap can make counterfactual combinations rely on extrapolation. Detailed contributions for categorical predictors can depend on coding and normalization unless handled carefully. Components can be negative or exceed the total gap in magnitude because they can offset one another; percentages are especially unstable when the overall gap is near zero.

Business use

Businesses can use decomposition to structure comparisons and identify where more detailed investigation may be useful. Define the groups, outcome scale and included variables before interpreting contributions. Report uncertainty, reference choices and sensitivity to model specification. Distinguish a raw mean gap from a fitted gap when the modeling setup does not reproduce group means exactly. If the analysis will guide an intervention, develop the causal question separately and examine whether the chosen variables are causes, consequences or proxies. Accounting identities alone do not determine effective or appropriate actions.

Check your understanding

Why do both 4 + 8 and 6 + 6 represent the same example?

Answer: They use different reference coefficients to allocate the same total difference of 12. The reference choice changes the accounting components, not the underlying fitted group means.

Related terms

Standardization
Regression coefficients
Omitted-variable bias
Causal inference
counterfactual mean
regression coefficient
composition effect
interaction

Learn next

compare twofold and threefold decompositions on a small dataset and inspect sensitivity to the reference coefficients. Then study overlap, categorical-variable normalization and uncertainty estimation before assigning substantive meaning to individual components.

Categorical variable

Sources

Sources: B. Jann, The Blinder–Oaxaca Decomposition for Linear Regression Models, https://doi.org/10.1177/1536867X0800800401. U.S. Census Bureau, Standardization and Decomposition of Rates: A User's Manual, https://www2.census.gov/library/publications/1993/demographics/p23-186.pdf. R oaxaca documentation, https://search.r-project.org/CRAN/refmans/oaxaca/html/oaxaca.html. Worked example and fictional case are original.

Previous: Kinematics Next: KL annealing