Two-Stage Least Squares (2SLS) Estimation

Updated 19 January 2026

Two-stage least squares (2SLS) estimation is a method that addresses endogeneity by using valid instrumental variables to isolate exogenous variation in endogenous regressors.
The technique involves a first stage where endogenous variables are regressed on instruments, followed by a second stage that estimates structural parameters using the predicted values.
Advanced extensions like regularization, Bayesian model averaging, and robust variance estimation make 2SLS adaptable to weak instruments, high-dimensional contexts, and heterogeneous treatment effects.

Two-stage least squares (2SLS) estimation is a foundational technique in econometrics and statistics for addressing endogeneity in linear models with endogenous regressors and valid instrumental variables (IVs). It extends ordinary least squares (OLS) by using instruments to isolate exogenous variation in the endogenous regressors, thereby enabling consistent estimation of structural parameters. This article describes the theoretical formulation, properties, contemporary developments, and application domains of 2SLS, as well as extensions to high-dimensional, Bayesian, and semi-parametric frameworks.

1. Structural Framework and Identification

The canonical 2SLS model is specified as

$y = X \beta + \varepsilon,\quad \varepsilon \sim \text{mean } 0, \ E[\varepsilon|X] \neq 0,$

where $y$ is an $n \times 1$ vector of outcomes, $X$ an $n \times k$ matrix of endogenous regressors, $\beta \in \mathbb{R}^k$ the parameters of interest, and $\varepsilon$ an $n \times 1$ error term. The presence of endogeneity— $E[\varepsilon|X] \neq 0$ —renders OLS estimators inconsistent.

Identification relies on the availability of $l \geq k$ instruments $Z$ ( $n \times l$ ) satisfying:

Exogeneity: $E[Z'\varepsilon] = 0$ (no direct effect on $y$ )
Relevance: $\operatorname{rank}(E[Z'X]) = k$ (instruments predict $X$ )
Full rank: $\operatorname{rank}(E[Z'Z]) = l$

These are the minimal Fisher-type and order conditions for valid IV identification, and their empirical adequacy must be established in applied research (Ginestet et al., 2015).

2. Estimator Construction: Stages and Algebraic Form

The 2SLS algorithm proceeds as:

Stage 1: Regress $X$ on $Z$ (and optionally exogenous controls), obtaining fitted values:

$\hat X = P_Z X, \quad P_Z = Z (Z'Z)^{-1} Z'$

Stage 2: Regress $y$ on $\hat X$ to obtain

$\hat\beta_{\text{2SLS}} = (\hat X' \hat X)^{-1}\hat X' y = (X' P_Z X)^{-1} X' P_Z y$

Under the standard IV regularity, $\hat\beta_{\text{2SLS}} \to_p \beta$ (Ginestet et al., 2015).

For time-varying or panel-dependent settings, the first and second stages generalize to accommodate evolving $Z_t$ and $X_t$ as well as spatial or temporal lags, as seen in spatial panels and dynamic treatment effect models (Tian et al., 2024, Tompsett et al., 2024).

3. Properties: Bias, Variance, and Mean Squared Error

OLS estimators, in the presence of endogeneity ( $E[X'\varepsilon] \neq 0$ ), are biased: $\operatorname*{plim} \hat\beta_{\text{OLS}} = \beta + (E[X'X])^{-1} E[X'\varepsilon]$ while 2SLS is asymptotically unbiased. However, OLS is always at least as efficient in variance:

$\operatorname{Var}(\hat\beta_{\text{OLS}}) \preceq \operatorname{Var}(\hat\beta_{\text{2SLS}})$
$\operatorname{Bias}^2(\hat\beta_{\text{2SLS}}) \preceq \operatorname{Bias}^2(\hat\beta_{\text{OLS}})$

The risk implications can be captured via mean squared error (MSE) decomposition: $\operatorname{MSE}(\hat\beta) = \operatorname{Var}(\hat\beta) + \operatorname{Bias}(\hat\beta) \operatorname{Bias}(\hat\beta)'$

This trade-off motivates convex combinations of OLS and 2SLS to minimize finite-sample MSE, as in the convex least squares (CLS) estimator (Ginestet et al., 2015).

4. Advanced Extensions and Robustification

a) Weak Instrumentation: When instrument strength is weak ( $\pi_1 \to 0$ ), standard 2SLS estimators become unstable, unbounded, and their sampling distributions can be Cauchy. Regularization, such as Ridge IV

$\hat\beta_{\text{ridgeIV}} = (X' P_Z X + \lambda I_k)^{-1} X' P_Z y$

restores boundedness of both bias and variance, uniformly dominating 2SLS in MSE under weak-instrument asymptotics (Rajkumar, 2019).

b) Shrinkage and High-Dimensionality: High-dimensional and large-scale systems deploy shrinkage methods (James–Stein for low-dimensional, ridge or lasso for high-dimensional) at the first or both stages. Two-stage lasso or ridge-lasso hybrids achieve estimation consistency and variable-selection consistency under restricted eigenvalue and mutual incoherence conditions, with finite-sample error rates at the order $(\text{sparsity})^2 \log (\text{dimensions}) / n$ (Zhu, 2013, Chen et al., 2015, Spiess, 2017).

c) Time-Varying, Panel, and Censoring Cases: 2SLS is adapted for complex longitudinal designs—time-varying IVs and treatments, right-censoring, and spatial autocorrelation. For right-censored $Y$ , all OLS regressions are replaced by inverse-probability-of-censoring weighted least squares, with consistency and asymptotic normality preserved under standard IV and independent censoring assumptions. This is operationalized by incorporating Kaplan–Meier weights in both stages (Beyhum, 2021).

5. Model Averaging, Bayesian, and Latent Structure Approaches

Model uncertainty over instrument or covariate sets is addressed via Bayesian Model Averaging (BMA) extensions. Two-stage BMA procedures sequentially average over first-stage and second-stage models, using unit information priors such that the posterior mode matches the 2SLS estimator. Posterior inclusion probabilities quantify instrument relevance; Bayesian Sargan tests provide model-averaged overidentification diagnostics with considerably higher power to detect invalid instruments than classical frequentist tests (Lenkoski et al., 2012, Henry et al., 2018).

In structural equation models (SEM), the model-implied instrumental variable 2SLS (MIIV-2SLS) extends 2SLS to instrument constructs or latent variables using algebraic implications from the model. BMA over all MIIV subsets yields robust estimation and high-powered instrument-specific overidentification and weak instrument tests (Henry et al., 2018).

6. Variance Estimation under Treatment Effect Heterogeneity

Standard variance formulas for 2SLS are inconsistent when instruments identify different local average treatment effects (LATEs); in this case, 2SLS estimates a weighted average of instrument-specific LATEs, and the GMM moment conditions are misspecified. The correct asymptotic variance involves a Hall–Inoue–type "misspecification-robust" sandwich estimator, which must replace conventional robust formulas to ensure valid inference (Lee, 2018). Empirically, standard errors from the conventional formula can severely underestimate true uncertainty, affecting hypothesis testing validity even with strong instruments.

7. Practical Considerations and Applications

Practical data-driven choices include:

Combining OLS and 2SLS: The CLS estimator adaptively chooses the convex weight minimizing estimated MSE from the data; in empirical work (e.g., returns to education with weak instruments), the data-driven CLS can strictly decrease MSE relative to pure 2SLS (Ginestet et al., 2015).
Instrument and Model Selection: Model averaging and selection procedures outperform single-model 2SLS in terms of MSE, coverage, and power to detect misspecification.
ML Integration: ML prediction methods for the first stage do not in general satisfy the orthogonality and exclusion requirements of 2SLS unless explicitly constrained (e.g., post-lasso, split-sample IV). Nonlinear and highly flexible learners (random forests, neural nets) can induce bias exceeding OLS benchmarks due to forbidden regression bias and regularization shrinkage (Lennon et al., 19 May 2025).

Table: 2SLS Extensions and Contexts

Extension/Variant	Main Goal or Feature	Key Reference
Ridge-regularized IV	Stable estimation w/ weak IVs	(Rajkumar, 2019)
High-dimensional penalty	Sparse structural estimation	(Zhu, 2013, Chen et al., 2015)
Bayesian Model Averaging (2SBMA)	Instrument/covariate uncertainty	(Lenkoski et al., 2012, Henry et al., 2018)
Model-implied IVs (MIIV-2SLS)	SEM/latent structure IVs	(Henry et al., 2018)
Misspecification-robust SE	Heterogeneous LATE robust errors	(Lee, 2018)
Time-varying 2SLS	Longitudinal/causal inference	(Tompsett et al., 2024, Tian et al., 2024)
ML-augmented first stage	Flexible first-stage prediction	(Lennon et al., 19 May 2025)

Implementation in empirical studies (e.g., education returns; panel carbon emissions; medical treatment strategies) demonstrates the need to carefully match 2SLS variants to the presence of weak/invalid instruments, high-dimensionality, complex time- or space-dependence, and model uncertainty.

2SLS estimation remains central in causal inference with endogeneity, but rigorous application in modern, complex designs requires advanced extensions: regularization for weak/high-dimensional instrument sets, model averaging for selection uncertainty, misspecification-robust inference for heterogeneous treatment effects, and precise diagnostics to verify the adequacy of instruments and specifications (Ginestet et al., 2015, Rajkumar, 2019, Lenkoski et al., 2012, Lee, 2018, Lennon et al., 19 May 2025).