Econometricians seeking to establish cause-and-effect relationships from observational data often face severe obstacles when instrumental variables fail to meet the strict requirement of exogeneity. In a study published in the Journal of Statistical Theory and Applications, researcher Nicolas Ronderos Pulido introduces the Modified Instrumental Variables framework, an iterative estimation methodology designed to reduce or fully eliminate the inconsistency of two-stage least squares estimates when an instrument exhibits partial endogeneity. Rather than discarding imperfect instruments or relying on subjective Bayesian priors, the framework establishes that endogenous instrumental variables can be systematically modified using an iterative search process, provided that the instrument contains a sufficiently strong exogenous component that outweighs its endogenous contamination.
The foundational challenge of instrumental variable estimation lies in isolating exogenous variation within explanatory variables that are otherwise correlated with unobserved disturbances. In classical econometrics, variables such as omitted individual traits, simultaneity, and measurement errors cause explanatory regressors to correlate with the regression error term, rendering ordinary least squares inconsistent. Two-stage least squares resolves this issue only if an instrument correlates with the regressor while remaining completely uncorrelated with the disturbance term. However, identifying an instrument that faces no theoretical counter-arguments is notoriously difficult in empirical practice, as updated theories frequently reveal subtle relationships between proposed instruments and unobserved background factors, undermining their credibility.
To overcome this limitation, the Modified Instrumental Variables framework formalizes a linear system where both the endogenous regressor and the questionable instrument are decomposed into structural components. In this setup, both variables are constructed as linear combinations of an unobserved error term, a vector of structurally weak exogenous elements, and idiosyncratic disturbances. The magnitude of endogeneity entering the instrument is governed by structural weighting parameters. If the parameter weighting the unobserved disturbance equals zero, the instrument remains weakly exogenous; otherwise, it becomes endogenous. The core mechanism of the new estimator operates by subtracting a scaled component of the estimated structural residuals from the first-stage fitted values, thereby systematically filtering out the endogenous error that contaminates the instrument.
The theoretical engine of the estimator is an iterative cycle defined across successive iterations. Beginning with initial values where the modified instrument is set to the first-stage fitted values of standard two-stage least squares, the method estimates the structural parameters and calculates standardized, recentered residuals. In each subsequent iteration, these residuals are multiplied by an optimal hyperparameter matrix and subtracted from the first-stage projections to generate an updated modified instrument. A formal convergence space theorem demonstrates that under random sampling, first-stage full-rank conditions, a balance between exogeneity and endogeneity, and linear control structure, the sequence of estimated coefficients converges in probability to the true structural parameter vector, reducing estimation inconsistency at each step.
The critical condition governing the success of the correction mechanism is Assumption A3, which demands that the exogenous covariance between the instrument and the regressor exceeds their endogenous covariance. When the exogenous variation dominates, the theoretical inconsistency sequence diminishes toward zero as iterations increase, driven by higher-order even moments that cause the exogenous component in the denominator to grow faster than the endogenous component in the numerator. Conversely, if the endogenous component equals or exceeds the exogenous component, the effectiveness of the iterative process declines, and the estimator can only partially mitigate inconsistency rather than fully eliminate it.
Because the true scaling hyperparameter matrix and population disturbances are unknown in empirical practice, the author develops an optimization search algorithm to select the parameter over a bounded convergence space. This parameter search minimizes the sum of squared residuals in an auxiliary regression where standardized residuals enter as regressors alongside the original explanatory variables. To keep computation feasible, the parameter grid is explicitly restricted to regions where the iterative sequence demonstrates numerical convergence. Within this convergence region, the optimal hyperparameter is identified either by maximizing the first-stage F-statistic—indicating optimal instrument relevance—or by minimizing the estimated variance of the structural residuals.
The study also demonstrates that statistical inference cannot rely on standard two-stage least squares formulas, which severely underestimate uncertainty in iterative contexts. Because the iterative estimator involves ratios of dependent random variables that generate heavy-tailed coefficient distributions, the variance of the estimator increases as iterations progress. To resolve this inferential challenge while maintaining consistency, the author formulates a bootstrap subsampling procedure. By drawing random subsamples of size smaller than the full sample, the estimator limits sensitivity to extreme values. The researcher provides a practical heuristic where the optimal number of iterations ranges between 3 and 10, balanced by sample size and residual standard deviation, and recommends decreasing the subsample size across iterations to optimize computational efficiency against precision.
To evaluate performance under controlled settings, the study details extensive finite-sample Monte Carlo simulations with sample sizes of 100,000 individuals across 2,500 replications. In baseline models where endogeneity accounted for one-third of the variation in the regressor and instruments, ordinary least squares yielded an expected coefficient of 1.33 and standard two-stage least squares produced a probability limit of 1.50 against a true parameter of 1.00. The iterative framework successfully identified the optimal hyperparameter at 0.80 across various selection criteria, rapidly reducing the coefficient inconsistency toward the true parameter value. Additional sensitivity analyses varying the exogenous-to-endogenous variance ratio from 0 to 2 showed that when the exogenous variance matched or exceeded the error variance, inconsistency was heavily mitigated.
The practical utility of the Modified Instrumental Variables framework is demonstrated in two contrasting empirical applications drawn from historical econometric literature. In the first application, the author revisits the seminal returns-to-education analysis of Angrist and Krueger, which employed quarter-of-birth indicators as instruments for schooling in a Mincer earnings equation using data from roughly one million individuals across three age cohorts. Addressing long-standing criticisms that birth quarter correlates with family background and health endowments, the algorithm estimated negative hyperparameters across all three cohorts: minus 0.0016 for men aged 20 to 29, minus 0.0051 for those aged 30 to 39, and minus 0.0343 for those aged 40 to 49. These negative values revealed that conventional two-stage least squares systematically underestimated the returns to education, and modifying the instrument produced revised estimates aligned with broader meta-analytic evidence.
In the second application, the methodology is applied as an empirical diagnostic tool to examine the Vietnam War draft lottery study conducted by Angrist. In this setting, draft eligibility assigned by lottery birthdates served as an instrument for military service to estimate impacts on civilian lifelong earnings across FICA, adjusted FICA, and W-2 administrative records. Because lottery numbers were randomly assigned, the eligibility indicator was hypothesized to be strictly exogenous unless procedural flaws had compromised the draw. Applying the search algorithm across wide and fine parameter grids yielded an optimal hyperparameter of exactly zero for all three earnings measures. This result confirmed that the iterative process introduced no modifications, empirically verifying the integrity of the randomization mechanism and validating the draft lottery as an unmodified instrumental variable.
The investigation also clarifies the theoretical boundaries of the framework, noting that the estimator remains vulnerable to specific forms of instrument weakness. An instrument becomes uninformative if both its endogenous and exogenous links to the regressor are absent, or if the two channels perfectly offset one another, causing the first-stage projection to vanish. In such cases, the first-stage parameter collapses to zero and the algorithm produces no modification, signaling underlying weakness. Overall, the study establishes that rather than serving as an unguided black box, the Modified Instrumental Variables framework provides an adaptable diagnostic and corrective foundation, enabling empirical researchers to validate natural experiments and extract consistent causal inferences when strict exogeneity cannot be guaranteed.
Subject of Research: Econometric methodology for correcting inconsistency in endogenous instrumental variable estimation
Article Title: The Modified Instrumental Variable (MIV) for Endogenous Instrumental Variables
Article References: Pulido, N. R. (2025). The Modified Instrumental Variable (MIV) for Endogenous Instrumental Variables. Journal of Statistical Theory and Applications, 24(4), 861-914. https://doi.org/10.1007/s44199-025-00132-z
Image Credits: AI Generated
DOI: 10.1007/s44199-025-00132-z
Keywords: econometrics, instrumental variables, endogeneity, exogeneity, causal inference, Two-Stage Least Squares, Modified Instrumental Variables, iterative estimator, bootstrap subsampling, Mincer earnings equation, quarter of birth, Vietnam War draft lottery, omitted variable bias, statistical consistency, first-stage F-statistic, convergence theorem, Monte Carlo simulation
Cite Scienmag News
Reid Dalton. (November 20, 2026). Modified Instrumental Variable Method Corrects Endogeneity Bias in Econometric Causal Inference. Scienmag. https://scienmag.com/modified-instrumental-variable-method-corrects-endogeneity-bias-in-econometric-causal-inference/
Reid Dalton. "Modified Instrumental Variable Method Corrects Endogeneity Bias in Econometric Causal Inference." Scienmag, 20 November 2026, https://scienmag.com/modified-instrumental-variable-method-corrects-endogeneity-bias-in-econometric-causal-inference/. Accessed 20 November 2026.
Reid Dalton. "Modified Instrumental Variable Method Corrects Endogeneity Bias in Econometric Causal Inference." Scienmag. November 20, 2026. https://scienmag.com/modified-instrumental-variable-method-corrects-endogeneity-bias-in-econometric-causal-inference/








