Solving the Simultaneity Bias Problem in Econometrics

Imagine trying to figure out if advertising (X) leads to more sales (Y). You run a simple analysis and find a strong positive link. Great! But what if you work at a company where the advertising budget is *set* as a percentage of last quarter’s sales? Suddenly, sales (Y) are also causing advertising (X). This two-way street, this feedback loop, is at the heart of a tricky problem in econometrics called simultaneity. When this feedback loop exists, our trusty old statistical tool, Ordinary Least Squares (OLS), starts to give us misleading answers. It suffers from what we call simultaneity bias, and it’s a critical issue to understand if we want our models to tell us the truth about cause and effect.

What is the endogeneity problem, exactly? 🔗

In econometrics, we rely on a set of rules to trust our results. One of the most important rules for an OLS regression-think of it as the standard “line of best fit” model-is that our explanatory variable (X) must not be correlated with the error term (u). The error term is a catch-all bucket for everything else that affects our dependent variable (Y) but isn’t in our model. So, this rule just means “the variable we’re using for explanation shouldn’t be related to the *unexplained* part.” This is called the assumption of exogeneity.

Simultaneity breaks this rule. The “endogeneity problem” simply means our explanatory variable is *endogenous*, or “determined from within” the system, making it correlated with that error term. When X causes Y, but Y also causes X, they become a tangled mess. Any random, unobserved shock that affects Y (and is part of the error term) will *also* affect X through the feedback loop. Now, our explanatory variable X is correlated with the error term (u), and the core OLS assumption is violated.

A classic example: consumption and income 🔗

Let’s use the exact example from the prompt, which is a cornerstone of macroeconomic theory. We want to estimate a simple consumption function:

Consumption (C) = β0 + β1 * Income (Y) + u

Here, we want to find β₁, the “marginal propensity to consume” (MPC)-or how much a person’s spending increases for every extra dollar of income. Using OLS seems logical. However, we also know from basic economics that in a simple model, national income (Y) is *defined* by what people do with it:

Income (Y) = Consumption (C) + Investment (I)

Do you see the feedback loop? Income (Y) affects Consumption (C), but Consumption (C) also affects Income (Y). They are jointly determined. This is a classic simultaneous equations model.

Now, think about the error term (u) in the consumption function. This ‘u’ represents unobserved factors that affect spending, like a sudden wave of consumer optimism. If consumer optimism (u) rises, people spend more (C). But because C is a component of Y, this rise in C will *also* cause Y to rise. The result? The error term (u) is positively correlated with the explanatory variable (Y). This violates the OLS assumption, and our model is in trouble.

[Image: A simple diagram showing a feedback loop, with an arrow from Income (Y) to Consumption (C) and another arrow from Consumption (C) back to Income (Y)]

Why ordinary least squares (OLS) fails 🔗

When this core assumption-Cov(X, u) = 0-is violated, the OLS estimator (the formula we use to get our β coefficients) stops being the reliable tool we need. Its desirable properties vanish. Econometricians say that the estimator becomes both biased and inconsistent.

Proof: The OLS estimator is biased 🔗

In statistics, “bias” has a very specific meaning. An estimator is unbiased if, on average, it hits the true population parameter. Think of it like a trustworthy archer: even if their arrows don’t hit the bullseye every time, they are clustered symmetrically around it. On average, they’re perfectly accurate.

A biased estimator is like an archer whose bow is warped. Their arrows might be tightly clustered, but they consistently land to the left of the bullseye. The average of all their shots will *not* be the true center.

In our consumption example, because income (Y) is positively correlated with the error term (u), the OLS formula will mistakenly attribute some of the effect of that unobserved optimism (u) to income (Y). It can’t tell them apart. The result is that our estimate for β₁ (the MPC) will be systematically overestimated. The expectation of our OLS estimator does not equal the true β₁. It’s biased.

Proof: The OLS estimator is not consistent 🔗

This is arguably even worse. “Consistency” is a large-sample property. A consistent estimator is one that gets closer and closer to the true population value as we give it more and more data. If we had an infinite amount of data, a consistent estimator would give us the *exact* true value. This is a fundamental property we need for our estimates to be reliable.

An inconsistent estimator, however, doesn’t get better with more data. Even with an infinite sample size, it will *still* be wrong. The “probability limit” of the OLS estimator (where it converges as the sample size ‘n’ goes to infinity) will not be the true parameter β₁.

Why? Because the root of the problem-the correlation between Y and u-is a structural feature of the system. It doesn’t go away just because we collect more observations. The bias persists no matter how large our sample gets. This means our results aren’t just slightly off; they are fundamentally misleading, and gathering more data won’t fix it.

How do we solve the simultaneity problem? 🔗

If OLS is broken, what do we use instead? We need methods that are designed to handle endogeneity. These methods are clever ways to isolate only the “good,” exogenous variation in our explanatory variable. Here are some of the most common solutions.

Indirect least squares (ILS) 🔗

This method is a bit old-school but provides a great intuition for the solution. It’s used in systems that are “exactly identified” (a technical condition meaning we have just enough information to solve the puzzle). Instead of estimating the “structural equations” (like our consumption function) directly, ILS takes a different path:

1. Estimate the “reduced form” equations: We first use algebra to solve our system of equations so that every endogenous variable (like C and Y) is expressed *only* as a function of all exogenous variables (like Investment (I) and the error terms).
2. Apply OLS to the reduced form: Since the reduced form equations only have exogenous variables on the right-hand side, OLS is perfectly valid! We can estimate these equations without bias.
3. Solve for the structural parameters: After we have the (consistent) estimates from the reduced form, we use algebra to work backward and find the original “structural” parameters we cared about (like β₁).

We get our answer *indirectly*, by solving a different, easier problem first. The main limitation is that this only works for exactly identified systems, which are rare in complex, real-world models.

Instrumental variables (IV) 🔗

This is the conceptual foundation for the most popular solutions. The goal of IV estimation is to find a new variable, called an instrument (Z), that can “cleanse” our problematic endogenous variable (Y).

A valid instrument (Z) must satisfy two strict conditions:

1. Relevance: The instrument (Z) must be strongly correlated with the endogenous variable (Y). It has to be a good predictor. (Cov(Z, Y) ≠ 0)
2. Exogeneity (or the “exclusion restriction”): The instrument (Z) must be completely uncorrelated with the original error term (u). It can *only* affect the final outcome (C) *through* its effect on the endogenous variable (Y). (Cov(Z, u) = 0)

[Image: A diagram showing an instrumental variable (Z) pointing to Income (Y), and Income (Y) pointing to Consumption (C). A crossed-out arrow shows Z cannot directly affect C.]

Finding a good instrument is famously difficult and requires deep economic intuition. In our example, we might use “past rainfall” as an instrument for “agricultural income.” Rainfall (Z) is likely correlated with income (Y) but is *not* correlated with unobserved consumer optimism (u). The IV method then essentially isolates *only* the part of the variation in Income (Y) that is driven by rainfall (Z) and uses that clean variation to estimate the MPC.

Two-stage least squares (2SLS) 🔗

Two-Stage Least Squares (2SLS or TSLS) is the most common and practical way to implement the instrumental variables strategy, especially when we have multiple instruments.

As the name suggests, it works in two stages:

• Stage 1: We run a “cleansing” regression. We regress the endogenous variable (Y) on all the valid instruments (Z) and any other exogenous variables from the original equation. We then save the *predicted values* from this regression. Let’s call them “Y-hat” (Ŷ). This Ŷ represents the portion of Y that is “clean”-the part that is fully explained by the exogenous instruments and free from the taint of the error term (u).
• Stage 2: We run our original equation but with one crucial substitution. We replace the “dirty” endogenous variable (Y) with its “clean” predicted version (Ŷ) from Stage 1.

  Consumption (C) = β0 + β1 * (Ŷ) + v

  Since Ŷ is, by construction, uncorrelated with the original error term, this regression now satisfies the OLS assumptions, and the resulting estimate of β₁ will be a consistent estimator.

Other advanced methods 🔗

While 2SLS is the workhorse, other methods exist, particularly for complex systems. Limited Information Maximum Likelihood (LIML) is another “single-equation” method similar to 2SLS that is known to perform better when instruments are weak. Full Information Maximum Likelihood (FIML) is a “system” method that estimates all the equations in the model at the same time. These are generally more complex, but they use information from the entire system more efficiently.

Ultimately, recognizing simultaneity is the critical first step. When you see a two-way causal street, you now know that OLS will lead you astray and that a new toolkit, built on the logic of instrumental variables, is required to find a consistent and reliable answer.

What do you think? Can you think of another real-world example where two variables might cause each other, leading to simultaneity bias? How does this problem change the way you might interpret a simple correlation or regression you read about in the news?

References

1. https://www.reed.edu/economics/parker/312/notes/Notes11.pdf
2. https://home.iitk.ac.in/~shalab/econometrics/Chapter17-Econometrics-SimultaneousEquationsModels.pdf