Is this the same as sequentially adding controls?

No. Sequential changes depend on entry order; these contributions use the fixed full model.

Can a contribution exceed the total change?

Yes, when another contribution offsets it.

Does the remaining coefficient identify an unexplained causal effect?

Not automatically. Omitted variables, measurement errors, and design limitations still matter.

AI Glossary / Statistics and probability / Advanced

Gelbach decomposition

Definition

Gelbach decomposition is a regression method for dividing the change in a selected coefficient, between a base model and a fuller model, into additive contributions from added covariates or groups of covariates. In ordinary least squares, its accounting follows the omitted-variable formula. For a fixed sample and full specification, the contributions do not depend on the order in which those added covariates are listed.

What it means

A coefficient may change when analysts adjust for experience, location, or other measured characteristics. Adding these controls one at a time can produce different stories depending on which enters first, particularly when controls are correlated.

Gelbach decomposition answers a more precise question: given this base model and this full model, how much of their coefficient difference is associated with each added component? Contributions can be positive or negative and can offset one another.

This is conditional statistical accounting. Calling a contribution “explained” does not establish that changing that characteristic would cause the corresponding outcome change. The chosen controls, measurements, and research design still determine what interpretations are justified.

How it works

Fit the base and full regressions on exactly the same observations, with consistent weights and base controls. Let the coefficient of interest be b_short in the base fit and b_full in the full fit. Use the sign convention difference = b_short − b_full.

For each added covariate z_j, regress z_j on the entire base design matrix. Take the auxiliary coefficient gamma_j corresponding to the variable of interest. Multiply it by z_j’s coefficient beta_j in the full outcome regression: contribution_j = gamma_j × beta_j. Sum contributions within meaningful groups if desired.

Under the usual full-rank least-squares setup, these contributions sum to the coefficient difference. Uncertainty requires appropriate inference, such as a bootstrap that respects clustering and repeats the whole procedure. Order invariance does not mean invariance to changing the full model.

Three practical examples

Workforce analysis: examine how occupation, experience, and location account statistically for a difference between an unadjusted and adjusted pay coefficient, without treating the residual as a complete measure of discrimination.
Customer research: separate the contributions of customer tenure and product mix to a regional spending coefficient’s change after adjustment.
AI evaluation: investigate how case difficulty and language composition relate to a performance gap’s change across regression specifications, while separately assessing data quality and fairness.

Think of it this way

Think of reconciling two account balances using signed ledger entries. Each component explains part of the difference under a declared accounting system. A negative entry can offset a positive one, and the reconciliation says nothing by itself about which intervention would improve the balance.

Worked example

Construct six observations with x = (−1, −1, 1, 1, 0, 0), z1 = (−1, −3, 3, 1, 0, 0), and z2 = (2, 0, 0, −2, 1, −1). Define y = 3x + 4z1 + 2z2, giving y = (−3, −15, 15, 3, 2, −2). Include an intercept in every regression.

The base regression of y on x has slope 9. The full regression has slopes 3, 4, and 2 for x, z1, and z2. Auxiliary regressions give slopes 2 for z1 on x and −1 for z2 on x. Contributions are therefore 2 × 4 = 8 and −1 × 2 = −2. They sum to 6, exactly 9 − 3.

Adding z1 first instead makes x’s intermediate slope −3: the sequential changes become 12 and −6. Gelbach’s contributions remain 8 and −2 for the fixed full model. These constructed fits were independently checked by least squares; they illustrate accounting, not a causal experiment.

Illustrative case study

Fictional case: ClearPath Services studies why a regional satisfaction coefficient shrinks after adding customer tenure and support-channel mix. Two analysts initially tell different stories because they add the controls in opposite orders.

The team freezes the common sample and applies Gelbach decomposition with predeclared covariate groups. It reports signed contributions, uncertainty intervals, and the base and full coefficients together. It also checks alternative reasonable specifications. Management uses the analysis to identify questions for further research, rather than assuming that changing a support channel will reproduce its statistical contribution.

Common misconceptions and mistakes

Treating order independence as model independence. Different controls or samples can produce different decompositions.
Interpreting negative contributions as calculation errors. They can reflect offsetting associations.
Calling contributions causal mechanisms without a defensible causal design.
Comparing regressions with different missing-data exclusions and attributing the entire difference to added controls.
Reporting percentages without caution when the total coefficient change is small, zero, or composed of large offsetting terms.

Business use

For business reporting, use this method when stakeholders need a reproducible explanation of how adjustment changes a regression coefficient. Define the question, common sample, base controls, full controls, and groupings before interpreting results.

Present coefficients and contributions in their original units first. Keep model sensitivity, uncertainty, and data coverage visible. A useful deliverable is an auditable reconciliation that supports discussion; it is not an automatic ranking of investment opportunities or an explanation of every source of an observed disparity.

Check your understanding

Answer: 1. A coefficient changes from 7 to 4. Contributions are 5 and −2.

Do they reconcile?

Answer: Yes: 5 − 2 = 3 = 7 − 4. 2.

Should auxiliary regressions predict the added covariates from the base design?

Answer: Yes, for the formula used here. 3.

Does reordering unchanged covariate groups change their contributions?

Answer: No, apart from numerical rounding.

Related terms

Omitted-variable bias
Confounding
Bootstrap
Regression coefficients
multiple linear regression
covariate adjustment
Oaxaca–Blinder decomposition
regression coefficient
suppression effect

Learn next

Study the omitted-variable formula, then reproduce a decomposition on a common sample. Examine how grouping controls, changing the full specification, and allowing clustered uncertainty alter the report. Learn causal identification separately before interpreting any contribution as a mechanism.

Sources

Gelbach, “When Do Covariates Matter? And Which Ones, and How Much?”: https://www.journals.uchicago.edu/doi/abs/10.1086/683668 — original publisher abstract checked; full paper was unavailable.

PyFixest, regression decomposition documentation: https://pyfixest.org/how-to/regression_decomposition.html — mathematical formula and implementation checked. This entry follows the displayed matrix formula for auxiliary-regression direction and explicitly declares its sign convention. Examples are original.

Previous: Gazetteer Next: GELU Function