Skip to content
Back to search results

Shapley Decomposition of Inequality

Introduced by
Anthony Shorrocks (working paper 1999; published 2013)
Year
2013
Type
Axiomatic decomposition procedure
Family
Distributional decomposition
Font size

What does it do?

The Shapley decomposition, formalized for distributional analysis by Anthony Shorrocks (in a widely circulated 1999 working paper, published in 2013), is a general procedure for attributing an inequality or poverty statistic to its contributing factors — income sources, population subgroups, or determinants.

It borrows the Shapley value from cooperative game theory: each factor's contribution is its average marginal effect on the indicator across all possible orders in which factors could be eliminated. The result is an exact, symmetric, residual-free decomposition that applies to any indicator, even those (like the Gini) that have no natural analytic decomposition of their own.

The intuition behind the method

Suppose total income inequality depends on several income sources — wages, capital, transfers — and you want to know how much each contributes. The obvious idea is to remove one source and see how much inequality falls, but the answer depends on which sources you have already removed, so the order matters and contributions do not add up. The Shapley value solves this fairness problem exactly as it does when splitting the value of a coalition among players: consider every possible order of eliminating the factors, compute each factor's marginal effect in each order, and average over all orders. By symmetry the order-dependence cancels, the contributions sum exactly to the total, and every factor is treated even-handedly.

History

The Shapley value was introduced by Lloyd Shapley in 1953 as the canonical fair allocation of a coalition's worth among its players. Its application to distributional analysis was developed by Anthony Shorrocks, whose 1999 working paper 'Decomposition Procedures for Distributional Analysis: A Unified Framework Based on the Shapley Value' showed that a vast range of inequality and poverty decompositions — by income source, by subgroup, and into growth and redistribution components — are special cases of one Shapley procedure. The paper circulated for over a decade before its 2013 publication in the Journal of Economic Inequality, by which time the Shapley approach had become a standard tool, complementing the Owen value for nested (hierarchical) decompositions.

When to use it

Use the Shapley decomposition when you want to attribute a distributional statistic to several factors that interact, and especially when the indicator has no natural built-in decomposition. Typical uses are decomposing inequality by income source, attributing a poverty or inequality change to determinants (growth, redistribution, prices, demographics), and apportioning inequality to covariates via a regression-based first stage. It is the method of choice for the Gini coefficient, which is not additively decomposable, and for unifying many ad hoc decompositions under one principled framework. The cost is computational: an exact Shapley decomposition requires evaluating the indicator over all 2^K factor subsets, so with many factors one uses sampling approximations. Define the reference state and elimination convention carefully, since they determine what the contributions mean.

Limitations and considerations

  • Computationally expensive: an exact decomposition evaluates the indicator over all 2^K factor subsets, so large K requires Monte Carlo approximation.
  • Results depend on the chosen reference state and on how a factor is 'eliminated' (zeroed, set to its mean, etc.), which is a modeling decision.
  • It is an accounting attribution, not a causal decomposition; contributions reflect correlations and the chosen functional form, not structural effects.
  • When factors are defined via a first-stage regression, the decomposition inherits any misspecification or omitted-variable bias of that model.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Applies to any distributional indicator, including ones like the Gini that have no natural analytic decomposition.
  • Exact and residual-free by the efficiency property, so factor contributions always sum to the total with no leftover interaction term.
  • Symmetric and order-independent: it resolves the path-dependence that plagues naive sequential decompositions.
  • Unifies many previously ad hoc decompositions (income-source, subgroup, growth-redistribution) within one axiomatic framework.

Applications

  • Decomposing the Gini coefficient or other inequality indices by income source (labor, capital, transfers).
  • Attributing a change in poverty or inequality to growth, redistribution, prices, and demographic shifts within one framework.
  • Apportioning inequality to determinants estimated from a first-stage regression (regression-based Shapley decomposition).
  • Decomposing multidimensional poverty and well-being indices by dimension or indicator.
  • Unifying and benchmarking competing ad hoc decompositions of distributional change in applied studies.

How it works

How it works

1

Begin by writing the target indicator — a Gini coefficient, a Theil index, an FGT poverty measure, or a change in any of them — as a function of a set of factors x_1 through x_K. A factor might be an income source, a population characteristic, or a component such as growth versus redistribution. A reference state is fixed at which the indicator (or the eliminated factor's effect) is zero, so the contributions are measured relative to a well-defined baseline.

Formula
I = F(x_1, x_2, \dots, x_K), \qquad I(\varnothing) = 0 \text{ (reference state)}
2

For each subset S of factors already present, the marginal contribution of adding factor k is the change in the indicator, F(S union {k}) − F(S). Equivalently, in the elimination formulation, it is how much the indicator falls when factor k is removed from a given configuration. Because this marginal effect generally depends on which other factors are present (the factors interact), a single removal cannot give an order-free attribution.

Formula
\Delta_k(S) = F(S \cup \{k\}) - F(S), \qquad S \subseteq \{1,\dots,K\}\setminus\{k\}
3

The Shapley value assigns to factor k the weighted average of its marginal contributions over all subsets S, with weights equal to the probability that S is exactly the set of factors preceding k when all K! orderings are equally likely. This is the unique allocation satisfying efficiency (contributions sum to the total), symmetry (factors with identical marginal effects get equal shares), the null-factor property, and additivity, which is why it is the canonical solution to the order-dependence problem.

Formula
C_k = \sum_{S \subseteq N\setminus\{k\}} \frac{|S|!\,(K-|S|-1)!}{K!}\,\bigl[F(S\cup\{k\}) - F(S)\bigr]
4

By the efficiency property of the Shapley value, the factor contributions sum exactly to the value of the full indicator: the decomposition has no residual and no interaction term left over. This holds for any indicator, regardless of whether it possesses a natural closed-form decomposition, which is what makes the Shapley approach a universal decomposition tool — it can decompose the Gini, multidimensional indices, and changes over time that resist analytic splitting.

Formula
\sum_{k=1}^{K} C_k = F(\{1,\dots,K\}) = I

Common pitfalls

  • Treating Shapley contributions as causal effects of the factors rather than as an order-free accounting attribution.
  • Failing to state the reference/elimination convention, which can change the magnitude and even the sign of contributions.
  • Using a sampling approximation with too few permutations and reporting point estimates without acknowledging the approximation error.
  • Decomposing by income source while ignoring that the same exercise on a different indicator (Gini vs. Theil) can apportion shares quite differently.

Frequently asked questions

How does the Shapley decomposition differ from the Theil within/between decomposition?

The Theil decomposition is an exact analytic split that exists only because the generalized-entropy class is additively decomposable; it works for Theil and the mean log deviation but not for the Gini. The Shapley decomposition is a general procedure that works for any indicator by averaging marginal contributions over all elimination orders. For indicators that already decompose analytically the Shapley result usually coincides with the natural decomposition, but Shapley extends to indices with no natural split, which is its main advantage.

Why is the order of eliminating factors a problem, and how does Shapley solve it?

Because factors interact, the marginal effect of removing one factor depends on which others have already been removed, so different elimination orders give different contributions and a leftover interaction term. The Shapley value removes this arbitrariness by averaging each factor's marginal contribution over all possible orders, weighting each order equally. The averaging makes the result symmetric across factors and, by the efficiency axiom, guarantees the contributions sum exactly to the total with no residual.

Is the Shapley decomposition computationally feasible for many factors?

An exact decomposition requires evaluating the indicator over all 2^K subsets of factors, which is fine for a handful of factors but explodes combinatorially. With many factors, practitioners use Monte Carlo sampling of random permutations to approximate the Shapley contributions, trading exactness for tractability. The approximation error shrinks with the number of sampled permutations, and software implementations typically expose the number of replications as a tuning parameter.

Related methods

Sources

  1. Shorrocks, A. F. (2013). Decomposition procedures for distributional analysis: a unified framework based on the Shapley value. Journal of Economic Inequality, 11(1), 99–126.

    Open DOI reference
    Copied.