Skip to content
Back to search results

Theil Inequality Decomposition

Introduced by
Henri Theil (1967); decomposition class by Anthony Shorrocks (1980)
Year
1967
Type
Decomposable inequality measure
Family
Income inequality measurement
Font size

What does it do?

The Theil index, introduced by Henri Theil in 1967 by importing Shannon's information theory into economics, measures income inequality as the divergence between each unit's income share and its population share.

Its defining advantage is exact additive decomposability: total inequality splits cleanly into a within-group component (inequality inside each subgroup) and a between-group component (inequality between subgroup means). Theil's T and its companion L (mean log deviation) are the two best-known members of the generalized-entropy class, which Anthony Shorrocks showed in 1980 to be the only inequality measures that are additively decomposable in this way.

The intuition behind the method

Imagine partitioning a population into groups — say regions or education levels — and asking: how much of total income inequality comes from differences between the average incomes of the groups, and how much from inequality among individuals within groups? The Theil index answers this exactly. Borrowing from information theory, it treats inequality as the 'surprise' in how far each person's income share departs from an equal share. Because entropy is additive, the total surprise decomposes into surprise about which group you are in (between) plus surprise about your position within your group (within). That clean accounting — with the two parts summing exactly to the total — is what makes Theil the tool of choice when you want to attribute inequality to structure.

History

Henri Theil introduced his inequality index in the 1967 book Economics and Information Theory, drawing a direct analogy between income inequality and the entropy of an information source: the more concentrated the income shares, the lower the entropy and the higher the inequality. For more than a decade the decomposability of inequality measures was studied case by case until Anthony Shorrocks's 1980 Econometrica paper proved that the generalized-entropy class — of which Theil's T (GE(1)) and the mean log deviation (GE(0)) are members — is, up to transformation, the unique family that is additively decomposable into within- and between-group components. This result made Theil and the GE class the standard apparatus for structural analyses of inequality.

When to use it

Use Theil decomposition when you have individual or household income (or consumption) data, a meaningful partition of the population into groups, and you want to attribute total inequality exactly to within-group and between-group sources. It is the standard tool for questions like 'how much of national inequality is between regions versus within them?' or 'how much is explained by education?' Theil's T is appropriate when top-end sensitivity matters; the mean log deviation L is preferred when you want population-share weights, a clean residual-free decomposition, and consistency with Shapley factor decompositions. Report which GE member you use, since the between/within split and its interpretation depend on beta. For non-nested factor attribution (multiple overlapping characteristics) combine GE decomposition with regression-based or Shapley methods.

Limitations and considerations

  • Undefined for non-positive incomes (T and L use logarithms), requiring censoring or use of consumption data.
  • Less intuitive to communicate than the Gini coefficient because it is in entropy units with no direct geometric reading.
  • The within-group weights for general beta are neither pure population nor pure income shares, complicating interpretation away from beta = 0.
  • Sensitive to the top of the distribution (for T) where survey data are least reliable and top incomes are often under-captured.

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

Strengths

  • Exactly additively decomposable into within-group and between-group inequality with no residual, the defining advantage over the Gini coefficient.
  • Grounded in information theory and characterized axiomatically (Shorrocks 1980) as essentially the unique decomposable inequality class.
  • A tunable sensitivity parameter beta lets the analyst emphasize the bottom (low beta) or top (high beta) of the distribution.
  • Satisfies the standard inequality axioms — scale invariance, the transfer principle, population replication, and symmetry.

Applications

  • Decomposing national income inequality into between-region and within-region components for regional policy.
  • Quantifying how much inequality is associated with education, occupation, gender, ethnicity, or sector by between-group shares.
  • Tracking how the within/between split of inequality evolves with development, urbanization, or globalization.
  • Comparing inequality across countries using a decomposable measure consistent with axiomatic foundations.
  • Serving as the inequality measure in nested and Shapley-based factor decompositions of distributional change.

How it works

How it works

1

Theil's T weights each person's log income ratio by their income share, T = (1/N) sum of (y_i/mean) ln(y_i/mean), so it is more sensitive to changes at the top of the distribution. The mean log deviation L (Theil's second measure, sometimes called Theil-L or GE(0)) instead averages the log of the inverse income ratio, L = (1/N) sum of ln(mean/y_i), making it more sensitive to the bottom. Both equal zero under perfect equality and rise with inequality, and both are unbounded above, but they weight the distribution differently.

Formula
T = \frac{1}{N}\sum_{i=1}^{N} \frac{y_i}{\bar{y}} \ln\!\frac{y_i}{\bar{y}}, \qquad L = \frac{1}{N}\sum_{i=1}^{N} \ln\!\frac{\bar{y}}{y_i}
2

T and L are special cases of the generalized-entropy class GE(beta), indexed by a sensitivity parameter beta: GE(1) is Theil's T, GE(0) is the mean log deviation L, and GE(2) is half the squared coefficient of variation. Lower beta makes the measure more sensitive to inequality at the bottom of the distribution, higher beta to the top. Shorrocks (1980) proved that, up to monotonic transformation, the GE class is the unique family of inequality measures that is additively decomposable with population-share-and-income-share weights.

Formula
GE(\beta) = \frac{1}{\beta(\beta-1)}\left[\frac{1}{N}\sum_{i=1}^{N}\left(\frac{y_i}{\bar{y}}\right)^{\beta} - 1\right], \quad GE(1)=T,\; GE(0)=L
3

For Theil's T the decomposition is especially clean: total inequality equals the income-share-weighted sum of the within-group Theil indices plus a between-group term that is itself a Theil index computed on group mean incomes. The weights s_k are each group's share of total income. The within part measures how unequal incomes are inside groups; the between part measures how unequal the group means are. Their sum reproduces the overall index exactly, with no residual or interaction term.

Formula
T = \underbrace{\sum_{k=1}^{K} s_k\, T_k}_{\text{within}} + \underbrace{\sum_{k=1}^{K} s_k \ln\!\frac{\bar{y}_k}{\bar{y}}}_{\text{between}}, \qquad s_k = \frac{N_k \bar{y}_k}{N \bar{y}}
4

Every member of the generalized-entropy class decomposes into a weighted sum of within-group indices plus a between-group index, but the within-group weights depend on beta — they mix each group's population share f_k and income share s_k as f_k^(1−beta) s_k^beta. Only at beta = 0 (the mean log deviation) are the weights the pure population shares, which is why L is often preferred when one wants the within-group weights to sum to one and the decomposition to be 'path independent' in the Shapley sense.

Formula
GE(\beta) = \sum_{k} w_k(\beta)\, GE_k(\beta) + GE_{\text{between}}(\beta), \quad w_k(\beta) = f_k^{1-\beta}\, s_k^{\beta}

Common pitfalls

  • Comparing a Theil T decomposition with an L decomposition as if they were the same; the within-group weights and the between/within balance differ by construction.
  • Interpreting a small between-group share as evidence that group differences are unimportant, when the result depends heavily on how finely the population is partitioned.
  • Applying the index to data with zero or negative incomes without censoring, which makes T and L undefined.
  • Reading the between-group term as a causal effect of group membership; it is an accounting decomposition, not a structural one.

Frequently asked questions

What is the difference between Theil's T and Theil's L?

Both are generalized-entropy measures but with different sensitivity. Theil's T (GE(1)) weights log income ratios by income shares, making it more responsive to changes at the top of the distribution. The mean log deviation L (GE(0)) averages log inverse income ratios and is more responsive to the bottom. In decompositions, L is often preferred because its within-group weights are the pure population shares (summing to one), giving a cleaner, path-independent split, whereas T uses income-share weights.

Why is Theil decomposable when the Gini coefficient is not?

The Gini is based on pairwise differences across the whole distribution, so when you split the population into groups the between- and within-group parts do not add up to the total — there is a residual that depends on how much the group income ranges overlap. Theil and the broader generalized-entropy class are built from sums of individual entropy-like terms, which are additive by construction, so the within and between components sum exactly to the total with no overlap residual. Shorrocks (1980) showed the GE class is essentially the only one with this property.

How should I choose the sensitivity parameter beta?

Choose beta to match where in the distribution you care most about inequality. Low or negative beta (including GE(0), the mean log deviation) emphasizes the bottom of the distribution; beta = 1 (Theil's T) is balanced toward the top; beta = 2 (half the squared coefficient of variation) is very top-sensitive and is the only member defined for zero incomes. Because the within/between split depends on beta, report results for more than one value to show robustness, and prefer GE(0) when you want population-share weights.

Related methods

Sources

  1. Theil, H. (1967). Economics and Information Theory. Amsterdam: North-Holland.

    ISBN: 9780444814630
    Copied.
  2. Shorrocks, A. F. (1980). The class of additively decomposable inequality measures. Econometrica, 48(3), 613–625.

    Open DOI reference
    Copied.