Instrumental_variables

What Is Instrumental Variables?

Instrumental variables (IV) is a statistical technique used in Econometrics and Causal Inference to estimate causal relationships between variables when standard regression methods would yield biased results. This method is particularly valuable when an explanatory variable, known as the Independent Variable, is correlated with the error term in a regression model, a condition referred to as Endogeneity. Instrumental variables address issues such as Omitted Variable Bias, Measurement Error, and Simultaneity, which can prevent accurate estimation of a causal effect. An instrumental variable, or simply an "instrument," is a third variable that influences the endogenous independent variable but does not directly affect the Dependent Variable and is uncorrelated with the error term.

History and Origin

The concept of instrumental variables first emerged in the field of econometrics in the late 1920s. American economist Philip G. Wright is credited with proposing the use of instrumental variables in his 1928 book, "The Tariff on Animal and Vegetable Oils," as a solution to the identification problem in supply and demand models.¹⁶ Wright recognized that simply observing prices and quantities in a market wouldn't reveal the true supply or demand curves because both are simultaneously determined. He sought a variable that would shift one curve without affecting the other, thereby allowing him to isolate and estimate the elasticity of the curve of interest. His innovative approach laid the theoretical groundwork for instrumental variables estimation.¹⁴, ¹⁵

While Wright's initial contribution was largely overlooked for decades, the method was independently rediscovered and formalized by Olav Reiersøl in 1945, who coined the term "instrumental variables." The technique gained significant prominence in applied economics with the work of economists like Joshua Angrist and Guido Imbens. In 2021, Angrist and Imbens, along with David Card, were awarded the Nobel Memorial Prize in Economic Sciences for their methodological contributions to the analysis of causal relationships, particularly their work in formalizing the instrumental variables framework and its application in "natural experiments." ¹², ¹³Their work clarified how instrumental variables could be used to estimate well-defined causal effects even in settings with heterogeneous treatment effects.
¹¹

Key Takeaways

• Instrumental variables (IV) methods are used in Regression Analysis to obtain consistent estimates when explanatory variables are endogenous.
• IV helps overcome biases caused by omitted variables, measurement error, or simultaneous relationships between variables.
• A valid instrument must be strongly correlated with the endogenous explanatory variable but uncorrelated with the error term and not directly affect the dependent variable.
• The technique allows researchers to isolate the causal effect of interest by leveraging variation in the endogenous variable that is driven solely by the instrument.
• A key challenge in applying instrumental variables is finding valid and strong instruments.

Formula and Calculation

For a simple linear regression model with an endogenous regressor (X) and an instrumental variable (Z), the basic idea of instrumental variables can be understood through a Two-Stage Least Squares (2SLS) approach.

Consider the structural equation:

where (Y_i) is the dependent variable, (X_i) is the endogenous independent variable, (\beta_0) is the intercept, (\beta_1) is the coefficient of interest, and (\epsilon_i) is the error term. Since (X_i) is endogenous, (\text{Cov}(X_i, \epsilon_i) \neq 0).

The instrumental variable (Z_i) satisfies two conditions:

1. Relevance: (\text{Cov}(Z_i, X_i) \neq 0) (The instrument is correlated with the endogenous variable).
2. Exclusion Restriction: (\text{Cov}(Z_i, \epsilon_i) = 0) (The instrument is uncorrelated with the error term and affects (Y_i) only through (X_i)).

The 2SLS estimation proceeds in two stages:

First Stage: Regress the endogenous variable (X_i) on the instrument (Z_i) and any other Exogenous Variables (controls) included in the structural equation:

From this regression, we obtain the predicted values of (X_i), denoted as (\hat{X}_i). These predicted values capture the variation in (X_i) that is explained by the instrument (Z_i) and is, by construction, uncorrelated with (\epsilon_i).

Second Stage: Regress the dependent variable (Y_i) on the predicted values (\hat{X}_i) (and any other exogenous controls):

The coefficient (\hat{\beta}_1) from this second stage regression is the instrumental variables estimator for (\beta_1).

In the case of a single endogenous regressor and a single instrumental variable, the IV estimator for (\beta_1) can also be expressed as:

where (\text{Cov}(Z, Y)) is the sample covariance between the instrument and the dependent variable, and (\text{Cov}(Z, X)) is the sample covariance between the instrument and the endogenous independent variable. This formula highlights how IV uses the variation in (X) that is driven by (Z) to identify the causal effect.

Interpreting the Instrumental Variables Estimate

Interpreting the estimate derived from instrumental variables requires careful consideration. Unlike Ordinary Least Squares (OLS) regression, which estimates the average effect of an independent variable on a dependent variable for the entire sample, the IV estimate often captures a "local average treatment effect" (LATE). This means the estimate is specific to the subpopulation whose behavior is influenced by the instrument. For instance, if an instrument affects the educational attainment of individuals who would otherwise not pursue higher education, the IV estimate for the return to schooling would reflect the causal effect for this "complier" group, not necessarily the entire population. Understanding the mechanism by which the instrumental variable affects the endogenous variable is crucial for proper interpretation and Statistical Inference. Researchers must ensure the instrument's validity by carefully considering the exclusion restriction—that the instrument influences the outcome only through its effect on the endogenous variable.

Hypothetical Example

Consider a hypothetical scenario where an investor wants to understand the causal effect of Company ESG (Environmental, Social, Governance) Score on Stock Returns. A direct OLS regression might be biased because companies with higher stock returns might also have more resources to invest in improving their ESG scores (reverse causality), or unobserved factors like strong management quality could simultaneously drive both high ESG scores and high stock returns (omitted variable bias).

To address this, an analyst might seek an instrumental variable. A potential instrument could be the geographical proximity of a company's headquarters to major environmental disasters in the recent past (e.g., severe floods or wildfires).

1. Relevance: Companies located closer to recent environmental disasters might face increased pressure from local communities, regulators, and employees to improve their environmental and social governance, thus influencing their ESG scores. This proximity is plausibly correlated with changes in ESG scores.
2. Exclusion Restriction: It's assumed that the mere geographical proximity to past environmental disasters, after controlling for initial company characteristics and industry factors, does not directly impact future stock returns other than through its influence on the company's ESG efforts and subsequent ESG score improvements.

Step-by-step application:

• Data Collection: Gather data on company stock returns, ESG scores, and headquarters' proximity to major environmental disasters.
• First Stage: Regress ESG scores on proximity to environmental disasters (and other control variables). This identifies the portion of ESG score variation that is driven by the environmental disaster proximity.
• Second Stage: Use the predicted ESG scores (from the first stage) to regress stock returns. The coefficient on the predicted ESG score would provide an instrumental variables estimate of the causal effect of ESG scores on stock returns, attempting to remove the bias from endogeneity.

This hypothetical example illustrates how instrumental variables can be used to attempt to uncover a causal link in complex financial relationships where direct measurement is difficult due to confounding factors.

Practical Applications

Instrumental variables methods are widely applied across economics and finance, particularly when researchers seek to establish causal links in complex systems. In labor economics, IV has been used to estimate the returns to education, where factors like an individual's innate ability might influence both their schooling and future earnings. Seminal work by Angrist and Krueger used quarter-of-birth as an instrumental variable for schooling, leveraging compulsory schooling laws, to estimate its causal effect on earnings.

I⁹, ¹⁰n finance, instrumental variables can be used to analyze the impact of various corporate decisions or market regulations on firm performance, investment behavior, or asset prices, where endogeneity is often a concern. For example, researchers might use IV to estimate the causal effect of corporate governance structures on firm valuation, or the impact of financial regulations on bank risk-taking. However, finding valid instruments in financial markets can be challenging due to their dynamic and interconnected nature. A critical survey by Angrist and Krueger highlights the mechanics and qualities of a good instrument, including those derived from "natural experiments," which are situations where external events create conditions akin to a Randomized Controlled Trials.

#⁷, ⁸# Limitations and Criticisms

Despite their utility, instrumental variables methods are not without limitations and criticisms. A primary concern is the "weak instruments" problem. This occurs when the instrument chosen has a weak correlation with the endogenous independent variable in the first stage of estimation. We⁶ak instruments can lead to biased and inconsistent IV estimates, inflated standard errors, and poor finite sample properties, meaning the estimates may not be reliable even with a large sample size. Th⁴, ⁵e bias can be substantial, pushing the IV estimator closer to the biased OLS estimator. Re³searchers often check for weak instruments using statistics like the first-stage F-statistic, where a low value (e.g., below 10) indicates a potential problem. Fo²r further reading on this topic, a detailed discussion on the theory and practice of weak instruments is available.

A¹nother critical assumption is the exclusion restriction, which states that the instrument must not directly affect the dependent variable other than through the endogenous variable. Violating this assumption can lead to biased results, as the instrument would then be capturing more than just its intended effect. The validity of this assumption often relies on strong theoretical arguments and institutional knowledge rather than purely statistical tests, making the selection of a credible instrument a demanding task. The precision of IV estimates typically requires larger sample sizes than OLS, and the finite sample properties can be poor, meaning the desirable statistical properties of instrumental variables generally hold only asymptotically (in very large samples).

Instrumental Variables vs. Ordinary Least Squares

Instrumental variables (IV) and Ordinary Least Squares (OLS) are both techniques used in Regression Analysis to estimate relationships between variables, but they are applied in different contexts and address different problems.

The confusion between IV and OLS often arises because OLS is a more straightforward and commonly taught method. However, when the underlying assumptions of OLS are violated, specifically regarding the exogeneity of the independent variables, the estimates from OLS regression are unreliable for inferring causality. Instrumental variables provide a methodological solution to these challenges, allowing researchers to potentially uncover true causal relationships even with observational Panel Data.

FAQs

What is endogeneity, and why does instrumental variables address it?

Endogeneity occurs when an explanatory variable in a statistical model is correlated with the error term. This correlation can be due to omitted variables, Measurement Error, or Simultaneity (where variables mutually influence each other). Instrumental variables address endogeneity by finding an "instrument"—a variable that affects the problematic explanatory variable but is independent of the error term. This allows the IV method to isolate the truly exogenous variation in the explanatory variable, leading to unbiased estimates of its causal effect.

How do I know if an instrumental variable is "valid"?

A valid instrumental variable must satisfy two key conditions: it must be correlated with the endogenous explanatory variable (relevance), and it must not be correlated with the error term of the main equation (exclusion restriction). The relevance condition can typically be tested statistically (e.g., using an F-statistic in the first stage of 2SLS). However, the exclusion restriction cannot be directly tested and largely relies on strong theoretical arguments and domain knowledge. Researchers must demonstrate that the instrument affects the outcome only through the endogenous variable, not through any other unobserved channels.

What happens if the instrumental variables are "weak"?

If instrumental variables are "weak," meaning they have a low correlation with the endogenous explanatory variable, the IV estimator can perform poorly. This leads to several issues: the IV estimates can be biased, often towards the inconsistent OLS estimates; the standard errors can be very large, making Statistical Inference unreliable; and hypothesis tests may have incorrect size (rejecting true null hypotheses too often). Detecting and addressing weak instruments is a critical step in applied Econometrics and there are specific diagnostic tests and alternative estimation methods developed for this problem.