Is a control function always a raw residual?

No; generalized residuals and more flexible functions can be required.

Is it identical to two-stage least squares?

Certain linear specifications produce the same structural coefficient estimates, but the equivalence is not universal.

Can software verify the exclusion restriction?

No. Diagnostics can inform assessment, but substantive reasoning is indispensable.

AI Glossary / Statistics and probability / Advanced

Control function (econometrics)

Definition

In econometrics, a control function is a modeled term added to an outcome equation to account for dependence between an explanatory variable and unobserved influences on the outcome. It is often constructed from a first-stage model of an endogenous explanatory variable. Under suitable identification and modeling assumptions, conditioning on this term can support estimation that ordinary regression would not justify.

What it means

Suppose firms choose advertising intensity partly because they anticipate unusually strong demand. A regression of sales on advertising may then mix the effect of advertising with unobserved demand expectations. Advertising is endogenous because it is associated with the outcome equation's unobserved component.
A control-function approach explicitly models that dependence rather than assuming it away. The control function is not simply another observed demographic covariate, and adding an arbitrary residual does not automatically remove bias. The method needs a defensible model and a source of identifying information, commonly valid instruments.

How it works

Specify the outcome relationship and why an explanatory variable may be endogenous. Identify instruments or other justified structure that can help separate its variation from unobserved outcome determinants. Instrument relevance and exogeneity are substantive requirements; a strong statistical relationship alone does not establish validity.
In a simple linear setup, model the endogenous variable X using instruments Z and appropriate exogenous covariates. Retain the first-stage residual, which measures the part of X not explained by that model. Include X, the other outcome covariates, and the residual in the outcome regression.
The central assumption is that the chosen residual-based function adequately represents the relevant conditional dependence in the outcome error. More complex settings may require generalized residuals, nonlinear functions, or interactions. A residual formula suitable for a continuous linear first stage should not automatically be transferred to binary, count, or other nonlinear models.
In a standard linear setting with matching instruments and specifications, residual inclusion reproduces the structural coefficient estimates from two-stage least squares. This equivalence does not extend without qualification to arbitrary nonlinear models. Control-function approaches can require stronger modeling assumptions than a basic moment-condition approach.
Use inference that accounts for estimating the control functions. Suitable joint procedures, analytical corrections, or a properly designed bootstrap may be needed. Assess sensitivity to instruments, functional forms, and identifying assumptions before interpreting the result causally.

Three practical examples

An economist studies an input's relationship with output when businesses choose the input using private information.
An analyst models a binary outcome with an endogenous explanatory variable using an appropriate control-function specification.
A research team compares a linear residual-inclusion estimate with a corresponding instrumental-variable estimate as an implementation check.

Think of it this way

Think of the control function as a carefully constructed indicator of the hidden selection component that accompanies a chosen input. It helps the model account for that component under explicit assumptions. It is not a universal filter that can turn any observational comparison into an experiment.

Worked example

Consider an artificial four-row illustration. Let Z and V take all four combinations of -1 and 1 with equal weight. Define X = 2Z + V and Y = 3X + 2V. The rows (Z, V, X, Y) are (-1, -1, -3, -11), (-1, 1, -1, -1), (1, -1, 1, 1), and (1, 1, 3, 11).
All variables have mean zero. Regressing X on Z with an intercept gives fitted X = 2Z, so the residual is exactly V. A regression of Y on X alone has slope sum(XY)/sum(X squared) = 68/20 = 3.4. It mixes the coefficient three with the association between X and V.
Including X and the recovered residual V gives the exact relationship Y = 3X + 2V, recovering coefficient three on X. Z is constructed to be unrelated to V and to influence Y only through X in this toy system.
These are algebraic demonstrations, not estimates from a credible field study. The exact fit leaves no realistic residual-noise analysis, and the construction supplies assumptions that must be defended rather than presumed in applications. If an instrument directly affects Y or the dependence is misspecified, the same recipe need not recover the desired coefficient.

Illustrative case study

Fictional case: Orchard Research studies whether optional training changes employee output. It suspects that unobserved motivation affects both enrollment and output. Analysts investigate a proposed instrument related to training availability, then examine whether it could also influence output through scheduling or management attention.
Those alternative pathways matter before any residual is computed. The team documents the model, uses uncertainty calculations that include the first stage, and compares plausible specifications. Its report presents conclusions as conditional on the identification assumptions. A significant control-function term is not treated as proof that the proposed instrument is valid or that every source of selection has been resolved.

Common misconceptions and mistakes

Do not confuse this method with merely controlling for observed variables. First-stage residuals are generated regressors with a specific role. High first-stage predictive accuracy does not prove exogeneity, and a nonsignificant residual coefficient does not prove absence of endogeneity. Ordinary second-stage standard errors can be inappropriate. Blindly substituting fitted values into a nonlinear outcome model is also not generally equivalent to a valid control-function estimator.

Business use

The method can help answer causal business questions involving self-selected prices, inputs, participation, or other choices. Its usefulness depends more on credible identification than on software convenience. Explain why instruments or structural restrictions are plausible, what behavior the model represents, and which alternative explanations remain. Report sensitivity and uncertainty rather than presenting an adjusted coefficient as an assumption-free effect.

Check your understanding

What creates the naive slope of 3.4 in the example?

Answer: X contains V, which also affects Y directly.

What does the first stage recover here?

Answer: The residual V.

Would adding that residual necessarily work after changing the data-generating relationships?

Answer: No; the validity of the control-function specification must be reconsidered.

Related terms

Instrumental variables
Generated regressor
Endogeneity
instrumental variable
two-stage least squares
residual inclusion
exclusion restriction
structural equation
causal identification

Learn next

Study the linear equivalence with instrumental-variable estimation, then compare the assumptions needed for nonlinear control-function models. Practice explaining an instrument's relevance and exclusion restriction before fitting an application.

Sources

Stata, cfregress: Control-function linear regression, methods and formulas: https://www.stata.com/manuals/rcfregress.pdf Stata, Control-function linear and probit models: https://www.stata.com/new-in-stata/control-function-linear-probit-models/

Previous: Control-flow diagram Next: Control group