Endogeneity: The Reason Your Regression Coefficients Are Arguing With Themselves

Suppose you're trying to figure out whether more police officers reduce crime. You collect city-level data, run a regression, and find a troubling result: cities with more officers have higher crime rates. Does hiring cops cause crime?

Photo by Brett Jordan on Pexels.

No. What you've found is endogeneity. Cities hire more officers because they already have more crime. Your explanatory variable and your outcome are tangled up in a feedback loop, and your regression has no idea which direction causality runs. The coefficient you're reading as an answer is actually a confession that the model is confused.

Endogeneity is what happens when one of your predictor variables is correlated with the error term in your regression. That sounds technical, so here's the plain version: your model assumes the variable you're using to explain something is independent of all the stuff you didn't measure. When that assumption breaks, your coefficients are biased. Often severely. Sometimes in the wrong direction entirely.

Three Ways Endogeneity Gets Into Your Model

The police-crime example illustrates one source: reverse causality. The outcome influences the predictor as much as the predictor influences the outcome. This is common in economics, medicine, and basically any domain where the world adapts to itself over time.

Omitted variable bias is another entry point. If you leave out a variable that affects both your predictor and your outcome, that missing influence leaks into the error term. Your included variable picks up credit (or blame) it doesn't deserve. A regression of ice cream sales on drowning deaths looks alarming until you realize both are driven by hot weather, which you forgot to include.

Measurement error is the third culprit. If your predictor is measured with noise, the true signal and the error term become correlated. This typically biases coefficients toward zero, a problem called attenuation bias. You think your variable has a weak effect. The truth is your measure of it is garbage.

graph TD
    A[Reverse Causality] --> D{Endogeneity}
    B[Omitted Variable] --> D
    C[Measurement Error] --> D
    D --> E(Biased Coefficients)
    E --> F[Wrong Conclusions]

Why OLS Can't Save You

Ordinary least squares is built on a set of assumptions, and one of the load-bearing ones is exogeneity: your predictors are uncorrelated with the error term. Violate that, and OLS is no longer the best unbiased estimator. It's just an estimator. A biased one.

What makes endogeneity particularly treacherous is that nothing in your regression output will flag it directly. Your R-squared won't drop. Your standard errors won't balloon. The model will produce tidy coefficients with stars next to them, and those coefficients will be wrong. The diagnostic is theoretical, not statistical. You have to think about the data-generating process before you trust any number the software hands you.

What You Can Actually Do About It

Instrumental variables (IV) estimation is the classic fix. The idea: find a variable (the instrument) that affects your endogenous predictor but has no direct effect on your outcome except through that predictor. In the police example, a valid instrument might be a city's budget cycle quirks that cause officer hiring to fluctuate independently of crime trends.

Good instruments are rare. A weak instrument, one only loosely correlated with your predictor, makes things worse by amplifying bias and inflating variance. The literature is full of papers that used technically valid but practically useless instruments to produce estimates nobody can replicate.

Randomized controlled trials sidestep the problem by design. Random assignment breaks the feedback loop between treatment and outcome. When randomization isn't possible, difference-in-differences and regression discontinuity designs can help by exploiting natural experiments where the assignment to treatment is plausibly exogenous.

None of these are magic. Each comes with its own assumptions, and those assumptions deserve scrutiny that often doesn't happen.

The Bigger Point

Regression is a tool for measuring associations. Turning those associations into causal claims requires extra work, extra assumptions, and a lot more honesty about what you actually know versus what you're inferring.

Endogeneity is the reminder that correlation structure alone can't tell you which way the arrows point. Before you trust a coefficient, ask yourself: could the outcome be causing the predictor? Did I leave something important out of the model? Is my measurement of this variable reliable?

If the answer to any of those is "maybe," the answer to "can I trust this coefficient?" is "not yet."