Double Machine Learning for Inequality Decompositions: Reference Outcomes, Orthogonal Scores and Limited Overlap

arXiv:2511.13433 · econ.EM · Submitted 2025-11-17 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: We're looking at the paper: "Double Machine Learning for Inequality Decompositions".

Warren: Here's the gist of the paper. The paper reformulates the Kitagawa-Oaxaca-Blinder decomposition using potential outcomes and a doubly robust estimator based on double machine learning to avoid common support issues and model misspecification.

Ruth: Decomposition Framework The Kitagawa-Oaxaca-Blinder decomposition splits the difference in means between two groups into an explained part, due to observable factors, and an unexplained part.

Warren: This framework is extended using potential outcomes to highlight the critical role of the reference outcome.

Ruth: The observed difference in means can be decomposed into three parts: ∆rX that can be explained by a different composition in terms of characteristics between both groups, ∆r,1S that is the unexplained advantage of group 1, and −∆r,0S that is the unexplained disadvantage of group 0.

Title and authors: Ruth: So, let's go over who wrote this and what the title actually says. It's "Double Machine Learning for Inequality Decompositions: Reference Outcomes, Orthogonal Scores and Limited Overlap."

Warren: The authors are Emmanuel Flachaire and Bertille Picard. They bring expertise from both econometrics and machine learning, which is exactly what you need when you’re trying to build these kinds of robust decomposition methods.

Iris: I’m interested in the "Orthogonal Scores" part of the title. That suggests they are using a specific mathematical setup to make sure their estimation method is orthogonal to some errors in the models.

Ruth: Right, and "Limited Overlap." That hints at the practical reality—that you can't always have perfect overlap between groups when you try to apply these methods across many characteristics.

Warren: Essentially, this paper proposes a new way to decompose inequalities that handles those complexities better than the old methods. It’s about making sure your estimates for unexplained parts are solid, not just based on simple regression.

Iris: So, what does this mean for the identification part? If you can estimate the unexplained part without needing common support assumptions, that opens up a lot of possibilities for real-world data analysis.

Ruth: It does. The old methods often required strong assumptions about how well your propensity score models fit, and this paper suggests you can relax those requirements by using machine learning to estimate those nuisance parameters.

The paper's summary: Warren: So, in terms of what the paper actually does, they start with the standard Kitagawa-Oaxaca-Blinder decomposition but reformulate it entirely. They use potential outcomes to emphasize how critical that reference outcome choice is for understanding the unexplained part.

Iris: They break down the observed difference into three parts: what you can explain from observable characteristics, plus a part that’s an unexplained advantage for one group, and another part that’s an unexplained disadvantage for the other.

Ruth: And they focus on estimating that last, unexplained piece using their new framework. They use a double machine learning procedure based on Neyman orthogonality to achieve a doubly robust estimator.

Warren: That means if your initial model for the outcome or the propensity score is slightly wrong, this estimator still gives you decent results because it’s designed to be insensitive to those small mistakes.

Iris: They specifically look at an equilibrium reference outcome, defined as a weighted sum of the two potential outcomes where weights depend on how much each group shares characteristics. That seems like a smart way to define a baseline when you have many variables.

Ruth: It’s designed to avoid trimming or extrapolation, which are often necessary in other methods when you try to use many characteristics in the decomposition. They show that this method is quite flexible across different types of data, categorical or continuous.

The paper's improvements: Iris: So what are the main improvements they’re pushing? It seems like the biggest improvement is getting rid of those common support assumptions when estimating the unexplained part.

Warren: Right. They show that you can estimate that difference without having to assume a certain overlap in the data distributions, which is a big deal for real-world datasets where you don't always know the exact boundaries of group membership.

Ruth: And they tackle model misspecification by using machine learning techniques to estimate those complex non-linear functions instead of relying on simple parametric models for every step.

Iris: They use double machine learning, which uses cross-fitting and Neyman orthogonality to ensure the resulting estimator is robust and achieves root n convergence rates, meaning it gets accurate as your data gets bigger.

Warren: It’s designed to be consistent if you correctly specify either the outcome model or the propensity score model, which is what makes it doubly robust. This sounds like a lot of work for an estimation strategy.

Ruth: And they show that using this equilibrium reference outcome, specifically r equals two, leads to more stable and precise results compared to just picking a reference outcome of zero or one. That’s a solid empirical finding they highlight.

Conclusion: Ruth: So to wrap up the paper "Double Machine Learning for Inequality Decompositions: Reference Outcomes, Orthogonal Scores and Limited Overlap," the main point is that using this equilibrium reference outcome allows for a doubly robust estimator with Neyman orthogonality.

Warren: This combination lets researchers estimate the unexplained component without having to rely on common support or simple low-dimensional parametric models. It’s an alternative path to analyzing distributional differences that is much more flexible.

Iris: From my econometric view, it means we have a stronger tool for identifying the causal effects of observable characteristics when those characteristics are complex and numerous. The sensitivity they point out to irrelevant variables is something we have to keep in mind though.

Ruth: That’s right, the authors themselves flag that the decomposition based on this reference outcome can still be sensitive to including irrelevant explanatory variables in your analysis. So researchers need to be careful about what they’re putting into those models.

Warren: It sounds like a powerful tool for people trying to get reliable estimates of inequality components in large datasets, even when the underlying data structure is complicated. That's where this work sits right now.

Emmanuel Flachaire, Bertille Picard

Aix-Marseille University · Univ Rennes, Ensai, CNRS, CREST-UMR 9194

econ.EM

Submitted: 2025-11-17

Updated: 2026-11-07

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 91/100

The gist: The paper reformulates the Kitagawa-Oaxaca-Blinder decomposition using potential outcomes and a doubly robust estimator based on double machine learning to avoid common support issues and model…

Key concepts

Kitagawa-Oaxaca-Blinder decomposition
This framework splits the difference in means between two groups into an explained part due to observable characteristics and an unexplained part. It is extended using potential outcomes to show how the reference outcome dictates this decomposition.
Equilibrium Reference Outcome (r=2)
This reference outcome is a weighted average of potential outcomes, defined by the proportion of individuals sharing characteristics between groups. Using this specific outcome stabilizes results and removes the need for common support assumptions when identifying unexplained differences.
Double Machine Learning
This technique uses Neyman orthogonality and cross-fitting to estimate parameters simultaneously. It provides a doubly robust estimator, meaning it remains accurate even if the models used to estimate nuisance parameters (like outcome or propensity scores) are slightly incorrect.
Neyman Orthogonality Condition
This mathematical condition ensures that the estimator derived from machine learning is robust. When applied to the equilibrium reference outcome, this property guarantees that the resulting estimator has good statistical properties and achieves root-n convergence rates.

Terminology

Summary

The paper reformulates the Kitagawa-Oaxaca-Blinder decomposition using potential outcomes and a doubly robust estimator based on double machine learning to avoid common support issues and model misspecification. p. 0

Where the paper says it

Decomposing Inequalities using Machine Learning and Overcoming Common Support Issues Emmanuel Flachaire∗ and Bertille Picard† November 2025 Abstract The Kitagawa-Oaxaca-Blinder decomposition splits the difference in means between two groups into an explained part, due to observable factors, and an unexplained part. In this paper, we reformulate this framework using potential outcomes, highlighting the critical role of the reference outcome.

Decomposition Framework

The Kitagawa-Oaxaca-Blinder decomposition splits the difference in means between two groups into an explained part, due to observable factors, and an unexplained part. This framework is extended using potential outcomes to highlight the critical role of the reference outcome. p. 0

Where the paper says it

Decomposing Inequalities using Machine Learning and Overcoming Common Support Issues Emmanuel Flachaire∗ and Bertille Picard† November 2025 Abstract The Kitagawa-Oaxaca-Blinder decomposition splits the difference in means between two groups into an explained part, due to observable factors, and an unexplained part. In this paper, we reformulate this framework using potential outcomes, highlighting the critical role of the reference outcome.

The observed difference in means can be decomposed into three parts: ∆rX that can be explained by a different composition in terms of characteristics between both groups, ∆r,1S that is the unexplained advantage of group 1, and −∆r,0S that is the unexplained disadvantage of group 0. p. 3
Where the paper says it

In the simplest case, it can be the outcome received in the other group. Let ∆obs be the expected observed difference in outcomes: ∆obs : = E[Y obs | D = 1] − E[Y obs | D = 0] (2) = E[Y (1) | D = 1] − E[Y (0) | D = 0] By adding and subtracting E Y (r) | D = 0 and E Y (r) | D = 1 , we obtain ∆obs = E[Y (r) | D = 1] − E[Y (r) | D = 0] | {z } ∆r X + E[Y (1) | D = 1] − E[Y (r) | D = 1] | {z } ∆ r,1 S − E[Y (0) | D = 0] − E[Y (r) | D = 0] | {z } ∆ r,0 S (3) Thus, a difference in means between two groups can be decomposed in three parts which we denote by: • ∆r X that can be explained by a different composition in terms of characteristics between both groups.

Reference Outcome Selection

The choice of reference outcome significantly impacts the decomposition results. Classical choices include using the potential outcome of one group as a reference, which resembles an ATT or ATU, or using an equilibrium reference outcome defined by a propensity score-weighted combination of the two potential outcomes. p. 1

Where the paper says it

This is achieved by changing the reference outcome. Instead of choosing the outcome of one of the two groups as the reference outcome, we propose using a propensity score-weighted combination of the two potential outcomes as a reference, following Neumark (1988) in the standard linear framework.

The equilibrium reference outcome is defined as Y(2) = wX Y(1) + (1 - wX) Y(0), where weights are defined by the proportion of individuals who share the same set of characteristics in each group, i. p. 5
Where the paper says it

In the following, we extend the reference outcome in (6) to the case of a set of characteristics X, which may consist of many variables, either categorical or continuous. For an individual with a given set of characteristics X, the equilibrium reference outcome is defined as the weighted average: Y (2) = wX Y (1) + (1 − wX) Y (0) (7) where the weights are defined by the proportion of individuals who share the same set of characteristics in each group, that is, wX = P(D = 1 | X) (8) In this more general setting, Y (2) is nonlinear and the estimation of the equilibrium outcome requires non-parametric or machine learning methods (see the following subsection and section 4).

e., wX = P(D = 1 X).

Identification Strategies

The identification strategy focuses on estimating the unexplained part ∆rS, which is defined as a difference between an observable mean and a counterfactual mean. Several strategies are considered, including outcome regression, Inverse Probability Weighting (IPW), and Augmented Inverse Probability Weighting (AIPW). p. 9

Where the paper says it

• Inverse Probability Weighting (IPW): The reference outcome is obtained by reweighting through the conditional assignment model to one of the two groups P D = d | X). • Augmented Inverse Probability Weighting (AIPW): combining both strategies, to get an estimator which is consistent if at least one of the models is well-specified.

For the equilibrium reference outcome r = 2, the identification strategy shows that Assumption 2 [common support] is not required to identify the unexplained part of the observed difference in outcomes. p. 2
Where the paper says it

Section 3 is devoted to the identification strategy. The appropriate estimators are derived and it is shown that those based on the equilibrium reference outcome do not require the common support condition.

Machine Learning Estimation

The paper utilizes machine learning techniques, specifically a double machine learning procedure based on Neyman orthogonality and cross-fitting, to estimate the unexplained part of the decomposition. This method is shown to be robust to small mistakes in the estimation of nuisance parameters and achieves root-n convergence rates for debiased estimators. p. 2

Where the paper says it

Chernozhukov et al. (2018) provided a so-called double machine learning procedure to recover debiased estimators that achieve root-n convergence rates for the Average Treatment Effect (ATE) and Average Treatment Effect on the Treated (ATT).

The score function for the equilibrium reference outcome satisfies the Neyman orthogonality condition, ensuring that the estimator is robust. p. 0
Where the paper says it

To address limitations like common support and model misspecification, we extend Neumark’s (1988) weighted reference approach with a doubly robust estimator. Using Neyman orthogonality and double machine learning, our method avoids trimming and extrapolation.

Empirical Findings

Empirical applications show that with reference outcomes r = 0 or r = 1, results are sensitive to the choice of estimator and estimation method, but with the equilibrium reference outcome (r = 2), the results are more stable and more precise. The double machine learning method provides estimators robust to misspecification of regression models with valid inference. p. 0

Where the paper says it

To address limitations like common support and model misspecification, we extend Neumark’s (1988) weighted reference approach with a doubly robust estimator. Using Neyman orthogonality and double machine learning, our method avoids trimming and extrapolation.

Application examples suggest that the part of the difference in means which is unexplained by the set of characteristics is quite large when using this method compared to parametric approaches. p. 2
Where the paper says it

This reference outcome is sensitive to gender composition within sectors and is consistent with the standard concept of equal redistribution conditional on a set of characteristics. We extend this reference outcome to the case of a large set of characteristics, either categorical or continuous, and derive several estimators of the unexplained part of the mean difference.

Practical Considerations

Practical considerations include calibration, trimming, and the use of pooled regression for the equilibrium reference outcome. Calibration issues in machine learning models require attention because estimated probabilities may not match empirical probabilities. p. 1

Where the paper says it

This implies divisions by the conditional probability of belonging to the reference group, which is feasible only when these probabilities are bounded away from zero (hence the common support assumption). Intuitively, therefore, it is necessary for individuals to have comparable counterparts in the reference group, to avoid unstable weights driven by very small estimated probabilities.

Trimming is often recommended when propensity scores are close to zero or one for IPW estimators, but this is not required for the equilibrium reference outcome r = 2. The sensitivity of the decomposition based on Neumark’s reference outcome to including irrelevant explanatory variables is highlighted. p. 0
Where the paper says it

This improves flexibility and robustness, as illustrated by two empirical applications. Nevertheless, we also highlight that the decomposition based on the Neumark reference outcome is particularly sensitive to the inclusion of irrelevant explanatory variables.

Conclusion

The main contribution is analyzing the unexplained component using this equilibrium reference outcome, demonstrating that it allows for a doubly robust estimator satisfying Neyman orthogonality, enabling estimation via double machine learning. This provides an alternative to extrapolation or trimming strategies, relaxing two key assumptions: the common support condition and the need for correctly specified, low-dimensional parametric models. p. 1

Where the paper says it

Instead of choosing the outcome of one of the two groups as the reference outcome, we propose using a propensity score-weighted combination of the two potential outcomes as a reference, following Neumark (1988) in the standard linear framework. To understand why changing the reference outcome helps relaxing these standard assumptions simultaneously, we underline the idea that the common support hypothesis is needed whenever both reweighting techniques are used and the outcome of one of the two groups is taken as the reference outcome.

The paper concludes that applied researchers must carefully reflect on the relevance of the chosen variables, even when using these data-driven methods to analyze inequalities.

References

The references listed include Angrist and Pischke (2009), Arrow (1972), Becker (1957), Blinder (1973), Busso et al. p. 1

Where the paper says it

1 Introduction The Kitagawa-Oaxaca-Blinder decomposition (Kitagawa 1955; Oaxaca 1973; Blinder 1973) is widely used in empirical studies to study differences in distributional statistics between two groups or two periods (DiNardo et al. , 1996; Firpo et al.

p. 1
Where the paper says it

1 Introduction The Kitagawa-Oaxaca-Blinder decomposition (Kitagawa 1955; Oaxaca 1973; Blinder 1973) is widely used in empirical studies to study differences in distributional statistics between two groups or two periods (DiNardo et al. , 1996; Firpo et al.

(2014), Chernozhukov et al. p. 2
Where the paper says it

A more appropriate approach has been developed in recent literature on treatment effects. Chernozhukov et al.

p. 2
Where the paper says it

A more appropriate approach has been developed in recent literature on treatment effects. Chernozhukov et al.

p. 2
Where the paper says it

A more appropriate approach has been developed in recent literature on treatment effects. Chernozhukov et al.

p. 2
Where the paper says it

A more appropriate approach has been developed in recent literature on treatment effects. Chernozhukov et al.

(2018), DiNardo et al. p. 1
Where the paper says it

1 Introduction The Kitagawa-Oaxaca-Blinder decomposition (Kitagawa 1955; Oaxaca 1973; Blinder 1973) is widely used in empirical studies to study differences in distributional statistics between two groups or two periods (DiNardo et al. , 1996; Firpo et al.

(1996), Firpo et al. p. 1
Where the paper says it

1 Introduction The Kitagawa-Oaxaca-Blinder decomposition (Kitagawa 1955; Oaxaca 1973; Blinder 1973) is widely used in empirical studies to study differences in distributional statistics between two groups or two periods (DiNardo et al. , 1996; Firpo et al.

p. 1
Where the paper says it

1 Introduction The Kitagawa-Oaxaca-Blinder decomposition (Kitagawa 1955; Oaxaca 1973; Blinder 1973) is widely used in empirical studies to study differences in distributional statistics between two groups or two periods (DiNardo et al. , 1996; Firpo et al.

p. 1
Where the paper says it

1 Introduction The Kitagawa-Oaxaca-Blinder decomposition (Kitagawa 1955; Oaxaca 1973; Blinder 1973) is widely used in empirical studies to study differences in distributional statistics between two groups or two periods (DiNardo et al. , 1996; Firpo et al.

p. 1
Where the paper says it

1 Introduction The Kitagawa-Oaxaca-Blinder decomposition (Kitagawa 1955; Oaxaca 1973; Blinder 1973) is widely used in empirical studies to study differences in distributional statistics between two groups or two periods (DiNardo et al. , 1996; Firpo et al.

(2009, 2018), Fortin et al. p. 1
Where the paper says it

, 1996; Firpo et al. , 2009, 2018).

p. 1
Where the paper says it

, 1996; Firpo et al. , 2009, 2018).

p. 1
Where the paper says it

, 1996; Firpo et al. , 2009, 2018).

p. 1
Where the paper says it

, 1996; Firpo et al. , 2009, 2018).

(2011), Fortin (2008), Freedman (1981), Hastie et al. p. 1
Where the paper says it

These standard assumptions can be relaxed by using reweighting techniques (Fortin et al. , 2011) and machine learning methods (Strittmatter & Wunsch, 2021).

p. 1
Where the paper says it

These standard assumptions can be relaxed by using reweighting techniques (Fortin et al. , 2011) and machine learning methods (Strittmatter & Wunsch, 2021).

p. 1
Where the paper says it

These standard assumptions can be relaxed by using reweighting techniques (Fortin et al. , 2011) and machine learning methods (Strittmatter & Wunsch, 2021).

p. 1
Where the paper says it

These standard assumptions can be relaxed by using reweighting techniques (Fortin et al. , 2011) and machine learning methods (Strittmatter & Wunsch, 2021).

(2005), Imbens and Rubin (2015), Rosenbaum and Rubin (1983), Rubins 1974, Kitagawa 1955, Oaxaca 1973, Oaxaca and Sierminska (2025), Neumark (1988). p. 9
Where the paper says it

[ignorability] For r ∈ {0, 1}, Y (r) ⊥⊥ D | X (15) where ⊥⊥ denotes statistical independence (Rosenbaum & Rubin, 1983). It is also known as the conditional independence assumption or unconfoundedness (Imbens, 2004).

The references listed include Angrist, J. D., & Pischke, J. p. 4

Where the paper says it

In this case, ∆0,0 S = 0. This leads to the decomposition: ∆obs = E[Y (0) | D = 1] − E[Y (0) | D = 0] | {z } ∆0 X + E[Y (1) | D = 1] − E[Y (0) | D = 1] | {z } ∆ 0,1 S (ATT) (4) Note that this is exactly the decomposition derived in Angrist and Pischke (2009), specifically in equation (3.

p. 4
Where the paper says it

In this case, ∆0,0 S = 0. This leads to the decomposition: ∆obs = E[Y (0) | D = 1] − E[Y (0) | D = 0] | {z } ∆0 X + E[Y (1) | D = 1] − E[Y (0) | D = 1] | {z } ∆ 0,1 S (ATT) (4) Note that this is exactly the decomposition derived in Angrist and Pischke (2009), specifically in equation (3.

p. 4
Where the paper says it

In this case, ∆0,0 S = 0. This leads to the decomposition: ∆obs = E[Y (0) | D = 1] − E[Y (0) | D = 0] | {z } ∆0 X + E[Y (1) | D = 1] − E[Y (0) | D = 1] | {z } ∆ 0,1 S (ATT) (4) Note that this is exactly the decomposition derived in Angrist and Pischke (2009), specifically in equation (3.

-S. (2009), Arrow, K. p. 5
Where the paper says it

2 The equilibrium outcome Building on the taste-based discrimination framework of Becker (1957) and Arrow (1972), Neumark (1988) shows that the Kitagawa-Oaxaca-Blinder decomposition can be derived from a theoretical model of employers’ discriminatory tastes. Setting Y (r) = Y (1) reflects the idea of “pure” discrimination against women, while setting Y (r) = Y (0) reflects the idea of “pure” nepotism toward men, that is, a willingness to hire men even at relatively higher wages.

p. 5
Where the paper says it

2 The equilibrium outcome Building on the taste-based discrimination framework of Becker (1957) and Arrow (1972), Neumark (1988) shows that the Kitagawa-Oaxaca-Blinder decomposition can be derived from a theoretical model of employers’ discriminatory tastes. Setting Y (r) = Y (1) reflects the idea of “pure” discrimination against women, while setting Y (r) = Y (0) reflects the idea of “pure” nepotism toward men, that is, a willingness to hire men even at relatively higher wages.

p. 5
Where the paper says it

2 The equilibrium outcome Building on the taste-based discrimination framework of Becker (1957) and Arrow (1972), Neumark (1988) shows that the Kitagawa-Oaxaca-Blinder decomposition can be derived from a theoretical model of employers’ discriminatory tastes. Setting Y (r) = Y (1) reflects the idea of “pure” discrimination against women, while setting Y (r) = Y (0) reflects the idea of “pure” nepotism toward men, that is, a willingness to hire men even at relatively higher wages.

J. (1972), Becker, G. p. 5
Where the paper says it

2 The equilibrium outcome Building on the taste-based discrimination framework of Becker (1957) and Arrow (1972), Neumark (1988) shows that the Kitagawa-Oaxaca-Blinder decomposition can be derived from a theoretical model of employers’ discriminatory tastes. Setting Y (r) = Y (1) reflects the idea of “pure” discrimination against women, while setting Y (r) = Y (0) reflects the idea of “pure” nepotism toward men, that is, a willingness to hire men even at relatively higher wages.

p. 5
Where the paper says it

2 The equilibrium outcome Building on the taste-based discrimination framework of Becker (1957) and Arrow (1972), Neumark (1988) shows that the Kitagawa-Oaxaca-Blinder decomposition can be derived from a theoretical model of employers’ discriminatory tastes. Setting Y (r) = Y (1) reflects the idea of “pure” discrimination against women, while setting Y (r) = Y (0) reflects the idea of “pure” nepotism toward men, that is, a willingness to hire men even at relatively higher wages.

p. 5
Where the paper says it

2 The equilibrium outcome Building on the taste-based discrimination framework of Becker (1957) and Arrow (1972), Neumark (1988) shows that the Kitagawa-Oaxaca-Blinder decomposition can be derived from a theoretical model of employers’ discriminatory tastes. Setting Y (r) = Y (1) reflects the idea of “pure” discrimination against women, while setting Y (r) = Y (0) reflects the idea of “pure” nepotism toward men, that is, a willingness to hire men even at relatively higher wages.

S. (1957), Blinder, A. p. 5
Where the paper says it

2 The equilibrium outcome Building on the taste-based discrimination framework of Becker (1957) and Arrow (1972), Neumark (1988) shows that the Kitagawa-Oaxaca-Blinder decomposition can be derived from a theoretical model of employers’ discriminatory tastes. Setting Y (r) = Y (1) reflects the idea of “pure” discrimination against women, while setting Y (r) = Y (0) reflects the idea of “pure” nepotism toward men, that is, a willingness to hire men even at relatively higher wages.

p. 5
Where the paper says it

2 The equilibrium outcome Building on the taste-based discrimination framework of Becker (1957) and Arrow (1972), Neumark (1988) shows that the Kitagawa-Oaxaca-Blinder decomposition can be derived from a theoretical model of employers’ discriminatory tastes. Setting Y (r) = Y (1) reflects the idea of “pure” discrimination against women, while setting Y (r) = Y (0) reflects the idea of “pure” nepotism toward men, that is, a willingness to hire men even at relatively higher wages.

p. 5
Where the paper says it

2 The equilibrium outcome Building on the taste-based discrimination framework of Becker (1957) and Arrow (1972), Neumark (1988) shows that the Kitagawa-Oaxaca-Blinder decomposition can be derived from a theoretical model of employers’ discriminatory tastes. Setting Y (r) = Y (1) reflects the idea of “pure” discrimination against women, while setting Y (r) = Y (0) reflects the idea of “pure” nepotism toward men, that is, a willingness to hire men even at relatively higher wages.

S. (1973), Busso, M. p. 1
Where the paper says it

1 Introduction The Kitagawa-Oaxaca-Blinder decomposition (Kitagawa 1955; Oaxaca 1973; Blinder 1973) is widely used in empirical studies to study differences in distributional statistics between two groups or two periods (DiNardo et al. , 1996; Firpo et al.

, DiNardo, J. p. 1
Where the paper says it

1 Introduction The Kitagawa-Oaxaca-Blinder decomposition (Kitagawa 1955; Oaxaca 1973; Blinder 1973) is widely used in empirical studies to study differences in distributional statistics between two groups or two periods (DiNardo et al. , 1996; Firpo et al.

, & McCrary, J. (2014), Chernozhukov et al. (2018), Chernozhukov et al. p. 2
Where the paper says it

Chernozhukov et al. (2018) provided a so-called double machine learning procedure to recover debiased estimators that achieve root-n convergence rates for the Average Treatment Effect (ATE) and Average Treatment Effect on the Treated (ATT).

p. 2
Where the paper says it

Chernozhukov et al. (2018) provided a so-called double machine learning procedure to recover debiased estimators that achieve root-n convergence rates for the Average Treatment Effect (ATE) and Average Treatment Effect on the Treated (ATT).

p. 2
Where the paper says it

Chernozhukov et al. (2018) provided a so-called double machine learning procedure to recover debiased estimators that achieve root-n convergence rates for the Average Treatment Effect (ATE) and Average Treatment Effect on the Treated (ATT).

(2015), DiNardo et al. p. 1
Where the paper says it

1 Introduction The Kitagawa-Oaxaca-Blinder decomposition (Kitagawa 1955; Oaxaca 1973; Blinder 1973) is widely used in empirical studies to study differences in distributional statistics between two groups or two periods (DiNardo et al. , 1996; Firpo et al.

p. 1
Where the paper says it

1 Introduction The Kitagawa-Oaxaca-Blinder decomposition (Kitagawa 1955; Oaxaca 1973; Blinder 1973) is widely used in empirical studies to study differences in distributional statistics between two groups or two periods (DiNardo et al. , 1996; Firpo et al.

p. 1
Where the paper says it

1 Introduction The Kitagawa-Oaxaca-Blinder decomposition (Kitagawa 1955; Oaxaca 1973; Blinder 1973) is widely used in empirical studies to study differences in distributional statistics between two groups or two periods (DiNardo et al. , 1996; Firpo et al.

(1996), Firpo et al. (2009, 2018), Fortin et al. (2011), Fortin (2008), Freedman (1981), Hastie et al. (2005), Imbens and Rubin (2015), Imbens, G. p. 9
Where the paper says it

This condition states that the potential outcome is independent of group membership D, conditional on covariates X. It implies mean independence: E[Y (r)|X] = E[Y (r)|X, D = 0] = E[Y (r)|X, D = 1] (16) Therefore, within the subpopulation of individuals with the same covariates, the difference in the distributions of the observed outcomes between individuals from the two groups fairly represents the difference in means in this subpopulation (Imbens & Rubin, 2015).

p. 9
Where the paper says it

This condition states that the potential outcome is independent of group membership D, conditional on covariates X. It implies mean independence: E[Y (r)|X] = E[Y (r)|X, D = 0] = E[Y (r)|X, D = 1] (16) Therefore, within the subpopulation of individuals with the same covariates, the difference in the distributions of the observed outcomes between individuals from the two groups fairly represents the difference in means in this subpopulation (Imbens & Rubin, 2015).

p. 9
Where the paper says it

This condition states that the potential outcome is independent of group membership D, conditional on covariates X. It implies mean independence: E[Y (r)|X] = E[Y (r)|X, D = 0] = E[Y (r)|X, D = 1] (16) Therefore, within the subpopulation of individuals with the same covariates, the difference in the distributions of the observed outcomes between individuals from the two groups fairly represents the difference in means in this subpopulation (Imbens & Rubin, 2015).

W. (2004), Rosenbaum and Rubin (1983), Rubin, D. p. 9
Where the paper says it

[ignorability] For r ∈ {0, 1}, Y (r) ⊥⊥ D | X (15) where ⊥⊥ denotes statistical independence (Rosenbaum & Rubin, 1983). It is also known as the conditional independence assumption or unconfoundedness (Imbens, 2004).

p. 9
Where the paper says it

[ignorability] For r ∈ {0, 1}, Y (r) ⊥⊥ D | X (15) where ⊥⊥ denotes statistical independence (Rosenbaum & Rubin, 1983). It is also known as the conditional independence assumption or unconfoundedness (Imbens, 2004).

p. 9
Where the paper says it

[ignorability] For r ∈ {0, 1}, Y (r) ⊥⊥ D | X (15) where ⊥⊥ denotes statistical independence (Rosenbaum & Rubin, 1983). It is also known as the conditional independence assumption or unconfoundedness (Imbens, 2004).

(1974), Strittmatter and Wunsch (2021), Kitagawa, E. p. 2
Where the paper says it

, 2011), double machine learning can thus be used to estimate it with good statistical properties (Strittmatter & Wunsch, 2021). Our main contribution can be summarized as follows: drawing on Rubin’s conceptual framework (Rubin, 1974) and reformulating the decomposition problem in terms of potential outcomes, we rewrite the unexplained part when the reference outcome is directly inspired by what is proposed by Neumark (1988).

p. 2
Where the paper says it

, 2011), double machine learning can thus be used to estimate it with good statistical properties (Strittmatter & Wunsch, 2021). Our main contribution can be summarized as follows: drawing on Rubin’s conceptual framework (Rubin, 1974) and reformulating the decomposition problem in terms of potential outcomes, we rewrite the unexplained part when the reference outcome is directly inspired by what is proposed by Neumark (1988).

M. (1955), Oaxaca 1973, Oaxaca and Sierminska (2025), Neumark (1988). p. 1
Where the paper says it

1 Introduction The Kitagawa-Oaxaca-Blinder decomposition (Kitagawa 1955; Oaxaca 1973; Blinder 1973) is widely used in empirical studies to study differences in distributional statistics between two groups or two periods (DiNardo et al. , 1996; Firpo et al.

p. 1
Where the paper says it

1 Introduction The Kitagawa-Oaxaca-Blinder decomposition (Kitagawa 1955; Oaxaca 1973; Blinder 1973) is widely used in empirical studies to study differences in distributional statistics between two groups or two periods (DiNardo et al. , 1996; Firpo et al.

The references listed include Angrist, J. D., & Pischke, J.-S. (2009), Arrow, K. J. (1972), Becker, G. S. (1957), Blinder, A. S. (1973), Busso et al. (2014), Chernozhukov et al. (2018), Chernozhukov et al. (2015), DiNardo et al. (1996), Firpo et al. (2009, 2018), Fortin et al. (2011), Fortin (2008), Freedman (1981), Hastie et al. (2005), Imbens and Rubin (2015), Imbens, G. W. (2004), Rosenbaum and Rubin (1983), Rubin, D. (1974), Strittmatter and Wunsch (2021), Kitagawa, E. M. (1955), Oaxaca 1973, Oaxaca and Sierminska (2025), Neumark (1988).

The references listed include Angrist, J. D., & Pischke, J.-S. (2009), Arrow, K. J. (1972), Becker, G. S. (1957), Blinder, A. S. (1973), Busso et al. (2014), Chernozhukov et al. (2018), Chernozhukov et al. (2015), DiNardo et al. (1996), Firpo et al. (2009, 2018), Fortin et al. (2011), Fortin (2008), Freedman (1981), Hastie et al. (2005), Imbens and Rubin (2015), Imbens, G. W. (2004), Rosenbaum and Rubin (1983), Rubin, D.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Develop a double machine learning method for estimating treatment effects that achieves root-n convergence rates by utilizing Neyman orthogonality and cross-fitting to estimate both nuisance functions simultaneously. This allows AI systems to estimate parameters of interest with good statistical properties even when underlying outcome or propensity score models are misspecified.

  2. Implement an equilibrium reference outcome approach for inequality decomposition, which is based on the weighted average of potential outcomes, as this method avoids the need for common support assumptions and trimming observations. This provides a more flexible framework for analyzing distributional differences across many characteristics.

  3. Design AI systems capable of performing doubly robust estimation of the unexplained part of mean difference using Augmented Inverse Probability Weighting estimators, which are consistent if either the outcome model or propensity score model is correctly specified. These systems can handle cases where common support is not required, such as when using the equilibrium reference outcome for decomposition.

  4. Create an algorithm based on Algorithm 1 that integrates cross-fitting and machine learning to estimate the unexplained part of the observed difference in outcomes for all three reference outcomes (r=0, r=1, r=2). This provides a comprehensive toolkit for researchers to compare different decomposition strategies empirically.

  5. Implement calibration correction methods to improve the performance of propensity score models used within inequality decomposition frameworks. This addresses the issue where machine learning estimates of probabilities may not match empirical probabilities, enhancing the robustness of estimators like AIPW.

Abstract

Kitagawa-Oaxaca-Blinder decompositions split a gap in mean outcomes between two groups into a part explained by characteristics and an unexplained part. The unexplained part depends on the counterfactual "reference outcome" that individuals would receive if characteristics were valued equally. Several competing references have been proposed, mostly within the linear regression model. We show that the main proposals, namely using the outcome of either group as the reference and the references of Cotton (1988), Reimers (1983) and Neumark (1988), all belong to a single family of mixtures of the two groups' potential outcomes. This separates the choice of target from the choice of estimator, and allows nonparametric estimation by double machine learning. We find that the appropriate orthogonal score depends on the reference: Neumark's propensity-weighted reference requires a pooled regression, while all the others require group-specific regressions. We then show that Neumark's reference avoids both extrapolation and extreme weights under limited overlap. Finally, we discuss in a general setting the shrinkage effect of the unexplained part associated with the choice of Neumark's reference.

Sources

Related papers