Two-Stage Least Squares: A Practical Guide to Instrumental Variables and Endogeneity

PortalOwner Misc 9. May 2025 | 0

Two-Stage Least Squares, more formally known as Two-Stage Least Squares (2SLS), is a cornerstone tool in econometrics for untangling causal relationships when key explanatory variables are endogenous. In plain terms, endogeneity occurs when the regressor of interest is correlated with the error term, which can distort ordinary least squares (OLS) estimates. The Two-Stage Least Squares approach uses instruments—variables that influence the endogenous regressor but do not directly affect the outcome except through that regressor—to recover consistent estimates. This article walks through the core ideas, the step‑by‑step procedure, practical testing strategies, and common pitfalls, with an emphasis on clarity and application in the British academic and policy context.

Two-Stage Least Squares: the core idea

The central idea behind Two-Stage Least Squares is simple in principle but powerful in practice. If X is endogenous, we cannot rely on OLS to identify the effect of X on Y. Instead, we replace X with its predicted values based on instruments Z that are correlated with X but uncorrelated with the unobserved determinants of Y. By doing so, the second stage uses these fitted values to estimate the effect of X on Y, yielding estimates that are less biased by endogeneity under the required assumptions.

Two-Stage Least Squares is commonly used in a wide range of settings—from economics to epidemiology—where randomised experiments are unavailable or impractical, and observational data must be relied upon. The method is particularly valuable when simultaneity, omitted variables, or measurement error contaminate the relationship of interest.

When to apply Two-Stage Least Squares

Two-Stage Least Squares should be considered when you suspect endogeneity in the primary regressor(s) and you can plausibly identify valid instruments. Typical scenarios include:

Endogenous regressor: An economics study of how education affects earnings, where unobserved ability may influence both education decisions and earnings.
Simultaneity: In supply and demand analyses, price may be determined within the system along with quantity, creating endogeneity.
Measurement error: If X is measured with error, the observed X deviates from the true underlying variable, inducing bias in OLS estimates.

Instrument selection is a crucial step. A good instrument must be relevant (it should be correlated with the endogenous regressor) and valid (it should not directly affect the outcome Y except through the endogenous regressor, satisfying the exclusion restriction). In practice, researchers often use policy changes, geographical variation, or historical features as instruments, provided these instruments pass a battery of diagnostic tests.

Assumptions behind Two-Stage Least Squares

For Two-Stage Least Squares to identify causal effects consistently, several key assumptions must hold:

Relevance: The instruments must be sufficiently correlated with the endogenous regressor(s). Weak instruments can lead to biased estimates and large standard errors.
Exogeneity: The instruments must be uncorrelated with the structural error term in the outcome equation. Violations here undermine validity.
Exclusion restriction: The instruments affect the outcome solely through their impact on the endogenous regressor(s), not via other channels.
Linearity and additivity: In the standard linear Two-Stage Least Squares model, relationships are assumed to be linear in parameters and additive in error terms.
No perfect multicollinearity: The instruments and exogenous controls should not be perfect linear combinations of each other.

In more complex systems, such as models with several endogenous variables or multiple equations, Two-Stage Least Squares extends to the multiple-instrument, multi-equation framework. The fundamental principles—instrument relevance and validity—still apply, although tests and interpretation become more intricate.

The Two-Stage Least Squares procedure: step by step

Implementing Two-Stage Least Squares involves two principal stages. Below is a clear, practical outline you can apply in typical applied work.

Step 1: First-stage regression

Estimate the endogenous regressor(s) on the set of instruments and exogenous controls. If X is a single endogenous regressor, the first stage is a regression of X on the instruments Z and any exogenous covariates W:

X = π0 + π1 Z + π2 W + η

From this regression, obtain the fitted values X̂ (the predicted values of X given Z and W). The idea is that X̂ captures the portion of X that is explained by the instruments, free from the endogenous portion tied to the error term in the second stage.

Step 2: Second-stage regression

Replace the original endogenous regressor with its fitted values in the outcome equation, and estimate the regression of Y on X̂ and the exogenous controls W:

Y = β0 + β1 X̂ + β2 W + ε

The coefficient β1 is the Two-Stage Least Squares estimate of the causal effect of X on Y, under the standard assumptions. In the multiple endogenous regressor setting, the procedure generalises to a matrix of endogenous variables and a corresponding matrix of instruments, using a projection of the endogenous vector onto the instrument set.

In practice, many econometric packages implement 2SLS under the hood, but understanding the two-stage logic helps in diagnosing weak instruments, interpreting results, and communicating assumptions clearly to readers and reviewers.

Instrument validity and diagnostic tests

Valid instruments are the lifeblood of Two-Stage Least Squares. A failure in instrument validity can produce biased estimates even when the estimation procedure is correctly executed. Below are the key tests and checks commonly employed in empirical work.

Relevance: the first-stage F-statistic

A standard diagnostic is the strength of the relationship between instruments and the endogenous regressor(s), assessed by the F-statistic in the first-stage regression. A commonly used rule of thumb is that a first-stage F-statistic above 10 indicates sufficiently strong instruments. If the statistic is substantially lower, weak instruments may bias 2SLS estimates toward OLS, defeating the purpose of the method.

Exogeneity and overidentification tests

When there are more instruments than endogenous variables (an overidentified model), researchers can test whether the instruments, as a group, are uncorrelated with the error term in the second stage. Classic tests include:

Sargan test for overidentifying restrictions, applicable under homoskedasticity.
Hansen J test for overidentification, which is robust to heteroskedasticity and increasingly preferred in applied work.

If these tests reject the null hypothesis of valid instruments, the researcher must reassess instrument relevance and the exclusion restriction, possibly removing instruments or seeking alternative valid instruments.

Robustness to heteroskedasticity and inference

In the presence of heteroskedasticity, standard errors from the two-stage procedure may be biased. Robust standard errors or bootstrap methods are commonly used to obtain reliable inference. Some practitioners also employ robust 2SLS variants or switch to methods like LIML (Limited Information Maximum Likelihood) when instruments are weak or the model is near the boundary of identification.

Interpreting Two-Stage Least Squares results

The interpretation of the Two-Stage Least Squares coefficient depends on the specification and the instruments used. In the canonical single-equation model with one endogenous regressor, the 2SLS estimate of β1 represents the causal effect of X on Y for the subpopulation affected by the instrument(s)—often referred to as a local average treatment effect (LATE) in certain contexts. When multiple instruments or heterogenous effects are present, the interpretation remains the average causal effect among compliers, contingent on the instrument and control set used.

Key practical considerations for interpretation include:

Clear articulation of the instrument’s mechanism: how the instrument shifts the endogenous regressor and why it does not directly affect Y apart from through that regressor.
Awareness of heterogeneity: treatment effects may vary across groups; the 2SLS estimate is an average effect unless a model explicitly allows for heterogeneity.
Awareness of partial identification: if instruments are weak or many exogenous controls are present, standard errors can become large and point estimates unstable.

Practical examples and applications

Education and earnings: a classic Two-Stage Least Squares example

A well-trodden application is estimating the causal impact of education on earnings. Suppose researchers are concerned that individuals with higher ability both obtain more education and earn more, creating endogeneity. An instrument could be eligibility for a compulsory schooling expansion or proximity to universities, assuming such instruments influence educational attainment but do not directly determine earnings in adulthood aside from education.

First stage: regress years of schooling on the instrument(s) and controls (e.g., family background, parental education).

Second stage: regress earnings on the predicted schooling years from the first stage and controls.

The resulting coefficient on X̂ represents the causal effect of education on earnings for compliers—those whose schooling is affected by the instrument. This estimate helps policy evaluation by indicating potential returns to educational interventions that alter schooling decisions.

Healthcare access and outcomes

In health economics, researchers might study how access to primary care influences health outcomes. If access is correlated with unobserved health status, an instrument such as the distance to the nearest primary care centre or the implementation date of a regional programme can be used. The Two-Stage Least Squares framework helps isolate the causal impact of access on health outcomes, informing policy on where to allocate resources.

Extensions and alternatives to Two-Stage Least Squares

Two-Stage Least Squares is a foundational method, but several extensions and alternatives can be appropriate in different scenarios, especially when instruments are weak, the model is nonlinear, or issues of endogeneity are more intricate.

Limited Information Maximum Likelihood (LIML)

LIML is an alternative instrumental variables estimator that can perform better than 2SLS in the presence of weak instruments. It relies on a different likelihood-based formulation and, in some settings, provides more accurate estimates and smaller biases when instruments are marginally relevant.

Generalised Method of Moments (GMM)

When multiple endogenous variables and instruments are involved, the Generalised Method of Moments offers a flexible framework. Two-Stage Least Squares can be viewed as a special case of IV estimation that aligns with a single-equation GMM approach. In practice, robust or efficient GMM variants can improve inference, particularly under heteroskedasticity or autocorrelation.

Other robust approaches

Researchers may also consider Anderson-Rubin tests, Kleibergen-Paap statistics, or conditional likelihood ratio tests to assess identification and inference under a broader set of assumptions. These tools can be particularly valuable when the standard 2SLS assumptions are in doubt or when the data structure features complexities such as clustering or nonlinearity.

Software implementation: practical notes

Two-Stage Least Squares is implemented in most major econometrics packages. The core idea remains the same, but the syntax varies across platforms. Here are some practical pointers for common tools, focused on achieving clear, replicable results.

R

In R, the ivreg function from the AER or/and the AER::ivreg or the ivpack package provides a straightforward path to 2SLS. For a basic model with endogenous X, instruments Z, and controls W, the workflow is typically:

library(AER)
# Example: Y ~ X + W ; X ~ Z + W
model <- ivreg(Y ~ X + W | Z + W, data = mydata)
summary(model)

For robust standard errors, consider vcovHC from the sandwich package or use the robust option in the summary. When dealing with multiple endogenous variables and an overidentified model, Sargan/Hansen tests can be obtained via the summary output or dedicated test functions.

Stata

Stata users often employ the ivregress command. A typical specification is:

ivregress 2sls Y (X = Z) W
estat firststage
estat overid

The firststage results, overidentification tests, and robust standard errors are readily available post-estimation.

Python (statsmodels)

In Python, statsmodels provides two-stage least squares functionality through the IV2SLS class in the linearmodels package. A typical workflow looks like:

from linearmodels.iv import IV2SLS
est = IV2SLS.from_formula('Y ~ 1 + X + W1 + W2 + ... + Wn + [X ~ Z]', data=mydata).fit()
print(est.summary)

As with other platforms, robust standard errors and overidentification tests can be accessed through the model’s results interface.

Common pitfalls and best practices

Even with a clear two‑stage plan, practical econometrics demands vigilance. Here are some commonly encountered issues and how to address them.

Weak instruments: If the first-stage F-statistic is low, consider seeking stronger instruments, combining instruments (if justifiable), or employing LIML or GMM methods that can offer more reliable inference.
Violation of the exclusion restriction: Reassess the instrument’s pathway to the outcome. When in doubt, discuss theoretical justification and perform sensitivity analyses.
Overfitting in the first stage: Including too many instruments can lead to weak identification and biased second-stage estimates. Prefer parsimonious instrument sets with theoretical justification.
Heteroskedasticity: Use robust standard errors or bootstrap methods to ensure valid inference in the presence of heteroskedastic errors.
Interpretation with multiple instruments: The 2SLS coefficient identifies a local average effect for compliers. When instruments are heterogeneous or the sample is diverse, interpretwith care and transparently report the scope of inference.

A concise checklist for implementing Two-Stage Least Squares

Clearly justify the endogenous regressor(s) and the economic or policy question addressed.
Provide a convincing theoretical or institutional argument for the chosen instruments and discuss the exclusion restriction.
Report the first-stage F-statistic and examine the strength of instruments.
Conduct overidentification tests (Sargan/Hansen) if you have more instruments than endogenous variables.
Use robust standard errors and consider alternative estimators if instruments are weak or the errors are heteroskedastic.
Offer a careful interpretation of the results, noting the population to which the estimates apply (e.g., compliers) and any limitations.

Putting it all together: a small synthesis

Two-Stage Least Squares is a disciplined, transparent approach to causal inference with observational data. By leveraging instruments to disentangle endogenous variation, the method offers a route to credible estimates when randomisation is not feasible. When applied with care—through strong theoretical grounding, careful instrument selection, and robust inference—Two-Stage Least Squares remains a reliable workhorse for empirical research across the social sciences and beyond.

In summary: the value of Two-Stage Least Squares

Two-Stage Least Squares, or Two-Stage Least Squares as a formal estimator, is a foundational technique for addressing endogeneity in econometric modelling. With careful instrument selection, rigorous testing, and robust inference, the method enables researchers to uncover more credible causal relationships in economics, public policy, health analytics, and beyond. By understanding the two-stage logic, researchers can design better studies, communicate more persuasively, and contribute insights that withstand critical scrutiny.

Two-Stage Least Squares: A Practical Guide to Instrumental Variables and Endogeneity