A Guide to the Two-Stage Least Squares Estimator in Econometrics

If you’ve ever read a news article that breathlessly reports “People who drink coffee live longer!” or “College graduates earn $1 million more!” you’ve probably had a nagging feeling. Is it really the coffee? Or do coffee drinkers also happen to walk more? Is it really the diploma? Or do the kind of people who go to college already have the drive and intelligence to earn more anyway? This “chicken-and-egg” confusion is the single biggest challenge in data analysis. In econometrics, we have a name for this messy, real-world problem: endogeneity.

Our most common statistical tool, Ordinary Least Squares (OLS) regression, has a simple job: to draw the best possible straight line through a cloud of data points. But it only works if a crucial assumption is met: the explanatory variable we’re studying (like education) must be independent of all the other unmeasured factors (like “innate ability” or “family connections”) that also affect the outcome. When this assumption is violated-when our explanatory variable is “tainted” by being correlated with these unmeasured factors-OLS breaks down. The results are biased, inconsistent, and, in a word, wrong. This is where a more clever, powerful tool is required: the Two-Stage Least Squares (2SLS) Estimator.

The big problem: What is endogeneity? 🔗

To understand why we need 2SLS, we first have to appreciate the problem it solves. Endogeneity is the boogeyman of econometrics, the technical term for when our “cause” variable (let’s call it x) is tangled up with the “error term” (let’s call it u). The error term is just a statistical catch-all for every other factor influencing our outcome (y) that we don’t have in our model. This tangling can happen for a few key reasons:

• Omitted Variable Bias: This is the most common culprit. We want to measure the effect of Education (x) on Wages (y). But we can’t measure Innate Ability (which is in the error term u). Since ability likely affects *both* how much education a person gets and how high their wages are, our x and u are correlated. OLS will mistakenly give Education the credit for work that Ability was doing, biasing our results.
• Simultaneity: This is the classic “chicken-and-egg” problem. We want to know how Price affects Demand for a product. But doesn’t Demand also affect Price? The variables are determined at the same time. OLS can’t untangle this feedback loop and will produce meaningless estimates.
• Measurement Error: Sometimes, we just can’t measure our x variable perfectly. If the way we measure x is flawed in a certain way, this “error in variables” can also create endogeneity and bias our results.

When any of these conditions are present, OLS fails. The estimates it produces are not just wrong; they are inconsistent, which means that even with an infinite amount of data, they would *still* be wrong. We need a way to isolate only the “clean” part of our x variable-the part that is *not* correlated with the error term.

The solution: Finding an ‘instrument’ 🔗

To solve this, economists look for something called an Instrumental Variable (IV). An IV, let’s call it z, is a third variable that acts as a kind of “natural experiment.” It provides a clever workaround to the endogeneity problem. To be a valid instrument, z must follow two golden rules:

Rule 1: The instrument must be relevant 🔗

The instrument z must be strongly correlated with our “problem” variable x. If we’re trying to find a clean path to Education (x), our instrument z must have a real, demonstrable effect on Education. This condition is testable; we can run a regression and check if the correlation is strong.

Rule 2: The instrument must be exogenous 🔗

This is the tricky part. The instrument z must be completely uncorrelated with the error term u. This means the instrument can *only* affect our outcome y (Wages) by working *through* our problem variable x (Education). It cannot have any independent, secret path to the outcome. This rule, known as the “exclusion restriction,” cannot be statistically tested. It must be argued based on economic theory, common sense, and a deep understanding of the real-world situation.

A famous example from economists Joshua Angrist and Alan Krueger helps make this concrete. They wanted to know the causal effect of education on earnings. To get around the “ability” problem, they used a clever instrument: a person’s quarter of birth. The logic was:

1. Relevance: At the time, compulsory schooling laws forced students to stay in school until their 16th birthday. Students born in the fourth quarter (Oct-Dec) started school younger and were therefore forced to complete more schooling before they could legally drop out. Thus, “quarter of birth” was correlated with “years of education.”
2. Exogeneity: It’s highly plausible that the *season* you are born in has no direct effect on your adult wages, *except* for the fact that it changed how much schooling you got. It’s not correlated with innate ability, motivation, or family connections.

By finding this instrument, they could isolate the *causal* effect of education on earnings, free from the bias of endogeneity.

So, what is Two-Stage Least Squares (2SLS)? 🔗

The IV method is brilliant, but what happens if you don’t just find *one* good instrument, but two, or three, or more? This is called the “overidentified” case. We might have “quarter of birth” (z₁) and “distance from nearest college” (z₂) as two plausible instruments for education. We *could* just pick one and ignore the other, but that’s wasteful. We’d be throwing away valuable information.

This is precisely where 2SLS comes in. The Two-Stage Least Squares estimator is an extension of the Instrumental Variable method specifically designed for the common scenario where we have *multiple* instruments for a single endogenous variable. Its goal is to combine all available instruments in the most efficient way possible to generate a single, powerful estimate.

Instead of forcing the researcher to arbitrarily choose one instrument over another, 2SLS automatically calculates the best linear combination of all the available instruments to use in the estimation process.

The 2SLS ‘recipe’: A two-stage process 🔗

The name “Two-Stage Least Squares” is not just creative-it’s a literal description of the process. It’s an algorithm that runs in two distinct steps.

The first stage: Isolating exogenous variation 🔗

The first stage is all about “cleaning” our problem variable. Here’s what we do:

1. We take our endogenous (problem) variable (let’s call it x_k, or “Education”).
2. We regress it on all the exogenous variables in the model. This includes any other “control” variables we already trusted (like Age) and, crucially, all of our available instruments (z₁, z₂, …, z_M).
3. The OLS regression for this stage looks like this:
   x_k = β₀ + β₁(Age) + ... + γ₁(z₁) + γ₂(z₂) + ... + γ_M(z_M) + error
4. After running this regression, we are *not* interested in the coefficients. Instead, we save the fitted values (the predicted values) from this equation. We’ll call this new variable x̂_k (“x-hat-k”).

Think about what x̂_k represents. It is a new, “purified” version of our education variable. It contains *only* the variation in education that can be predicted by our clean, exogenous instruments. It has effectively been “scrubbed” of the problematic part that was correlated with the unmeasured error term (like ability). This x̂_k *is* the “best linear combination” of our instruments.

The second stage: Consistent parameter estimation 🔗

Now that we have our “clean” variable, the second stage is surprisingly simple.

1. We return to our original equation of interest, the one that predicts Wages.
   y = β₀ + β₁(x_k) + β₂(Age) + ... + error
2. We were afraid to run OLS on this equation because x_k was endogenous.
3. Now, we simply substitute our “scrubbed” variable x̂_k in place of the original, “dirty” variable x_k.
4. The new equation looks like this:
   y = β₀ + β₁(x̂_k) + β₂(Age) + ... + new_error
5. We run a simple OLS regression on *this* second-stage equation.

The resulting coefficients from this second regression are the Two-Stage Least Squares estimators. Because x̂_k was constructed to be uncorrelated with the error term, our OLS assumption is now satisfied. The estimates we get for β₁ (the effect of education on wages) are now consistent and asymptotically normal, allowing us to finally make valid, causal inferences.

Why 2SLS is the most efficient IV estimator 🔗

This brings us to a key point. Why go to all this trouble? Why is 2SLS considered the *best* IV estimator when you have multiple instruments?

In econometrics, “efficient” has a precise meaning: it means the estimator with the smallest asymptotic variance. In simpler terms, it’s the most “precise” or “reliable” estimate we can get. The 2SLS estimator is the most efficient IV estimator because it uses *all* available exogenous information in the smartest way.

Remember that “best linear combination” we created in the first stage? That x̂_k is the linear projection of the endogenous variable on all the exogenous variables and instruments. This means that, by construction, x̂_k is the instrument combination that is most highly correlated with the original endogenous variable x_k. This strong “first-stage” relationship is what minimizes the variance (increases the precision) of the “second-stage” estimates.

By using the optimal combination of *all* instruments, 2SLS provides the smallest asymptotic variance among all IV estimators that use linear combinations of those instruments. It’s simply better than arbitrarily picking one instrument and throwing the others away.

A quick note: When 2SLS and IV are the same 🔗

It’s important to clarify that 2SLS *is* an IV estimator. The terms are often used interchangeably, but there’s a slight distinction.

• If you have one endogenous variable and one instrument (what’s called a “just identified” model), the 2SLS procedure and the standard IV formula will give you the exact same result.
• Where 2SLS becomes a distinct and essential tool is in the “overidentified” case: when you have more instruments than endogenous variables (e.g., three instruments for one variable).

In modern econometrics, 2SLS is the standard, go-to method for implementing almost all instrumental variable regressions, as it handles all these cases (just-identified and overidentified) perfectly.

A final word of warning: Weak instruments 🔗

The entire 2SLS procedure, and indeed all IV estimation, hinges on the “Relevance” condition-that our instruments are strongly correlated with our problem variable. But what if they are only *weakly* correlated? This is known as the “weak instrument problem,” and it’s a serious danger.

If our instruments are weak, the first-stage regression does a poor job of “scrubbing” the endogenous variable. In this case, 2SLS estimators, while technically still “consistent,” can be severely biased in a finite sample-sometimes even *more* biased than the OLS estimates we were trying to fix! For this reason, applied economists must always test the “strength” of their first stage, often by looking at a statistic called the “First-Stage F-Statistic.” A common rule of thumb is that this F-statistic should be 10 or higher to be confident that the instruments are strong enough to be reliable.

What do you think? Can you think of another real-world example where a simple correlation (like “cities with more police have more crime”) might be misleading due to simultaneity? Given the ‘untestable’ nature of the exclusion restriction, do you think the search for clever instruments is more of an art or a science?

References

1. https://www.ibm.com/docs/en/spss-statistics/cd?topic=regression-two-stage-least-squares
2. https://en.wikipedia.org/wiki/Instrumental_variables_estimation
3. https://www.nber.org/system/files/working_papers/w25691/w25691.pdf