Instrumental Variables

Hey students! 👋 Welcome to one of the most powerful tools in an economist's toolkit - instrumental variables! This lesson will teach you how economists solve one of their biggest headaches: figuring out what actually causes what when simple correlations just aren't enough. By the end of this lesson, you'll understand how instrumental variables help us identify causal relationships, master the two-stage least squares technique, and interpret local average treatment effects like a pro. Get ready to unlock the secrets behind some of the most convincing economic research! 🔍

Understanding the Endogeneity Problem

Imagine you're trying to figure out if more education actually makes people earn more money 💰. You might think, "Easy! Just compare people with different education levels and see who earns more." But here's the catch - smarter, more motivated people tend to get more education AND would probably earn more money anyway, even without the extra schooling. This creates what economists call endogeneity - when our explanatory variable (education) is correlated with unobserved factors (natural ability, motivation) that also affect our outcome (earnings).

In mathematical terms, we have a problem when:

$$E[\epsilon_i | X_i] \neq 0$$

Where $\epsilon_i$ represents all the unobserved factors affecting our outcome, and $X_i$ is our explanatory variable. When this condition is violated, our regular ordinary least squares (OLS) estimates become biased and inconsistent - they don't give us the true causal effect we're looking for.

Real-world examples of endogeneity are everywhere! Does having more police reduce crime, or do high-crime areas simply get more police? Does foreign aid help economic growth, or do struggling countries just receive more aid? These questions plague economists because simple correlations can't tell us the direction of causation or account for confounding factors.

The endogeneity problem shows up in three main ways: omitted variable bias (like the ability example above), reverse causality (does A cause B or does B cause A?), and measurement error (when our data isn't perfectly accurate). Each of these issues can make our estimates unreliable for policy decisions.

The Instrumental Variables Solution

This is where instrumental variables come to the rescue! 🦸‍♂️ An instrumental variable (or "instrument") is a variable that affects your outcome ONLY through its effect on the problematic explanatory variable. Think of it as a "nudge" that randomly pushes some people to get more of the treatment (like education) without directly affecting the outcome (like earnings) through any other channel.

For an instrument to work, it must satisfy two crucial conditions:

Relevance: The instrument must be correlated with the endogenous explanatory variable. Mathematically: $Cov(Z_i, X_i) \neq 0$, where $Z_i$ is our instrument. If your instrument doesn't actually affect the treatment, it's useless!

Exogeneity (or the exclusion restriction): The instrument must only affect the outcome through its effect on the explanatory variable. In other words: $Cov(Z_i, \epsilon_i) = 0$. This means the instrument can't have any direct effect on the outcome - it must work ONLY through the channel we're studying.

Let's use a famous example: Angrist and Krueger's study of education and earnings using quarter of birth as an instrument. In many U.S. states, compulsory schooling laws meant that students born in different quarters of the year had to stay in school for slightly different amounts of time. Your birth quarter is essentially random (relevance condition met - it affects education), and there's no reason why being born in January versus September should directly affect your adult earnings except through education (exogeneity condition met).

Another classic example comes from studying the effect of military service on earnings. Researchers used the Vietnam War draft lottery as an instrument - your draft number was randomly assigned (exogeneity), and it strongly predicted whether you served in the military (relevance), allowing economists to estimate the true causal effect of military service on later earnings.

Two-Stage Least Squares (2SLS)

The most common way to implement instrumental variables is through Two-Stage Least Squares (2SLS) 📊. Don't let the name intimidate you - it's literally just two simple regression steps!

Stage 1: Regress your endogenous explanatory variable on the instrument(s) and any other control variables:

$$X_i = \pi_0 + \pi_1 Z_i + \pi_2 W_i + v_i$$

Where $X_i$ is the endogenous variable, $Z_i$ is the instrument, $W_i$ represents other control variables, and $v_i$ is the error term. From this regression, you get predicted values $\hat{X}_i$.

Stage 2: Regress your outcome variable on the predicted values from Stage 1:

$$Y_i = \beta_0 + \beta_1 \hat{X}_i + \beta_2 W_i + u_i$$

The coefficient $\beta_1$ from Stage 2 is your IV estimate - this gives you the causal effect you're looking for!

Why does this work? In Stage 1, you're extracting only the variation in $X_i$ that comes from the instrument $Z_i$ - the "good" variation that's unrelated to unobserved confounders. In Stage 2, you're using only this clean variation to estimate the effect on your outcome.

The 2SLS estimator can be written as:

$$\hat{\beta}_{IV} = \frac{Cov(Z_i, Y_i)}{Cov(Z_i, X_i)}$$

This shows that IV estimation is essentially comparing how much the outcome changes when the instrument changes, relative to how much the treatment changes when the instrument changes.

Instrument Validity and Testing

Just because you have a potential instrument doesn't mean it's a good one! 🔍 Testing instrument validity is crucial for credible research.

Testing Relevance: This is the easy part - just look at your Stage 1 regression. You want a strong, statistically significant relationship between your instrument and the endogenous variable. A common rule of thumb is that the F-statistic on your instrument(s) in the first stage should be at least 10. If it's weaker than this, you have a "weak instrument" problem that can lead to biased and imprecise estimates.

Testing Exogeneity: This is much harder because the exclusion restriction can't be directly tested - it's fundamentally an assumption based on economic theory and institutional knowledge. However, you can do some indirect checks:

Overidentification tests: If you have more instruments than endogenous variables, you can test whether the extra instruments give consistent results using the Sargan or Hansen J-test.
Placebo tests: Check whether your instrument affects outcomes it shouldn't affect if the exclusion restriction holds.
Balance tests: Verify that your instrument doesn't correlate with observed characteristics that might violate exogeneity.

Real research often involves extensive robustness checks and institutional details to make the case for instrument validity. For example, researchers using rainfall as an instrument for agricultural productivity need to argue convincingly that rainfall only affects economic outcomes through agriculture, not through other channels like transportation or mood.

Local Average Treatment Effects (LATE)

Here's where things get really interesting! 🎯 IV estimates don't give you the average treatment effect for everyone - they give you something called the Local Average Treatment Effect (LATE). This represents the average treatment effect for "compliers" - people whose treatment status is actually affected by the instrument.

Think about the draft lottery example. Not everyone drafted actually served (some got deferments), and some people not drafted still enlisted voluntarily. The IV estimate tells us about the effect of military service specifically for those whose service was determined by the lottery - the "compliers" who served if and only if they were drafted.

Mathematically, we can divide the population into four types:

Always-takers: Would get the treatment regardless of the instrument
Never-takers: Would never get the treatment regardless of the instrument
Compliers: Get the treatment if and only if the instrument encourages it
Defiers: Do the opposite of what the instrument encourages (usually assumed away)

The LATE framework shows that:

$$LATE = E[Y_i(1) - Y_i(0) | \text{Complier}]$$

This is the average treatment effect for compliers only. Whether this generalizes to other populations depends on the specific context and research question.

Understanding LATE is crucial for policy interpretation. If you're using compulsory schooling laws as an instrument for education, your results tell you about the returns to education for students who stayed in school only because of the law - not necessarily for students who would have gotten more education anyway or those who dropped out despite the law.

Conclusion

Instrumental variables represent one of economics' most elegant solutions to the fundamental problem of establishing causation from observational data. By finding variables that create quasi-random variation in treatments, economists can identify causal effects even when controlled experiments aren't possible. The two-stage least squares method provides a straightforward way to implement IV estimation, while the LATE framework helps us understand exactly what population our estimates apply to. Remember, the key to successful IV research lies in finding credible instruments that satisfy both relevance and exogeneity conditions - a challenging but rewarding endeavor that has revolutionized empirical economics! 🎉

Study Notes

• Endogeneity problem: Occurs when $E[\epsilon_i | X_i] \neq 0$, making OLS estimates biased and inconsistent

• Instrumental variable requirements:

Relevance: $Cov(Z_i, X_i) \neq 0$ (instrument affects treatment)
Exogeneity: $Cov(Z_i, \epsilon_i) = 0$ (instrument only affects outcome through treatment)

• Two-Stage Least Squares (2SLS):

Stage 1: $X_i = \pi_0 + \pi_1 Z_i + \pi_2 W_i + v_i$
Stage 2: $Y_i = \beta_0 + \beta_1 \hat{X}_i + \beta_2 W_i + u_i$

• IV estimator formula: $\hat{\beta}_{IV} = \frac{Cov(Z_i, Y_i)}{Cov(Z_i, X_i)}$

• Weak instrument problem: First-stage F-statistic should be ≥ 10

• Local Average Treatment Effect (LATE): IV estimates apply only to "compliers" whose treatment status is affected by the instrument

• Population types in LATE framework: Always-takers, Never-takers, Compliers, Defiers

• Instrument validity tests: Overidentification tests (Sargan/Hansen J), placebo tests, balance tests

• Common instruments: Natural experiments, policy changes, random assignment, geographic variation

• Three sources of endogeneity: Omitted variable bias, reverse causality, measurement error