When Some Errors Are Bigger Than Others: White's Robust Standard Errors

Ordinary least squares makes an assumption so quiet that most people who use it have never consciously noticed it. The assumption is: every observation is equally noisy.

The error on a tiny illiquid micro-cap in a panic is assumed to have the same typical size as the error on Apple on a sleepy Tuesday. The error on a rich household's spending is assumed to be as predictable as a poor household's. The error in 2008 is treated exactly like the error in 2017.

This assumption, called homoskedasticity, is essentially never true in economics or finance. Halbert White's 1980 paper showed how to stop pretending, and did it in a way that required almost nothing from the user. It is one of the most-used results in the applied social sciences.

The problem: the estimate survives, the confidence does not

Here is the good news first, and it is important because it is often misunderstood. When errors are unequally sized, your OLS coefficient is still unbiased. The line you fit is still, on average, pointing at the truth. It is not as efficient as it could be, but it is not wrong.

The bad news is that your standard error is wrong, and therefore your t-statistic is wrong, and therefore every conclusion you draw about significance is unreliable. And the error goes in an unpredictable direction: sometimes your standard errors are too small (you become overconfident and find effects that are not there), sometimes too large (you become timid and miss effects that are).

This matters more than it sounds. The coefficient is what you report. The t-statistic is what decides whether anyone believes you. If your t-statistic is unreliable, then the decision your regression exists to inform is unreliable, even though the number in the middle is fine.

The old fix and why it was bad

Before White, the standard approach was: model the heteroskedasticity. Figure out how the error variance depends on the data. Perhaps errors grow with firm size, so weight each observation inversely by firm size and re-run. This is called generalised least squares, and when it works it is genuinely better than OLS, because it also recovers efficiency.

The catch is the word "figure out." You need to know the actual form of the relationship between the data and the error variance. Get it right and you win. Get it wrong, and you have introduced a new misspecification on top of the one you were trying to fix, and your results can be worse than if you had left it alone.

And in practice, nobody knows the form. It is not something economic theory tells you. It is a nuisance property of the data-generating process, and the honest answer to "how does the variance depend on X?" is almost always "I have no idea."

The key idea via analogy: let each point report its own noise

White's insight is a small piece of genius. Ask yourself: what would a good estimate of the error variance at observation number 47 look like?

The obvious answer is: the squared residual at observation 47. Yes, it is a terrible estimate. It is a single draw. It is enormously noisy. As an estimate of the local variance, it is nearly worthless.

White's argument is that it does not need to be a good estimate individually. The standard error formula does not require you to know the variance at any single point. It requires you to know a certain weighted sum across all the points. And when you add up hundreds or thousands of individually terrible, unbiased estimates, the noise averages out. The sum converges to the right answer even though every single term in it is garbage.

The analogy: suppose you want to know the total weight of a thousand crates, and your scale is wildly inaccurate, giving readings that are off by a random amount each time but not biased in either direction. Any single crate's reading is useless. But weigh all thousand and add them up, and the overshoots cancel the undershoots. The total is accurate even though no individual measurement was.

That is the whole idea. You do not model the heteroskedasticity. You do not need to know its form. You just let every residual carry its own weight into the sum, and let the aggregation do the work.

The result is what everyone now calls robust standard errors, or White standard errors, or in software, HC errors. Same coefficient. Honest uncertainty.

The second half: the test

The paper's other contribution, the one that appears in its title but is less famous, is a direct test for heteroskedasticity.

The logic is elegantly self-referential. You now have two ways to compute the standard errors: the ordinary way, which assumes equal-sized errors, and White's way, which does not. If the errors really are equally sized, the two answers should agree. If they disagree substantially, then the equal-size assumption is violated. So the test amounts to formally comparing the two estimates and asking whether they diverge more than chance would allow.

This became known as White's test, and it is still in every econometrics package.

Why it mattered

• It removed an excuse and an obstacle at the same time. You no longer needed to know the form of the heteroskedasticity to get honest inference, so there was no longer any reason not to fix the problem.
• Robust standard errors became the default. In most of applied economics and finance today, reporting non-robust standard errors is treated as a mild embarrassment. That is a direct consequence of this paper. It is hard to overstate how completely the profession adopted this.
• It launched the "robust inference" mindset. The core philosophical move, do not model the nuisance, just correct for it in the aggregate, propagated everywhere. Newey-West extended it to serial correlation. Clustered standard errors extended it to grouped data. Hansen's GMM built a whole estimation framework on the same spirit.
• It is essential in finance specifically. Volatility clustering means financial error variances are wildly unequal by construction. Cross-sectional regressions on stocks have errors that scale with everything from size to liquidity to sector. There is essentially no financial regression where homoskedasticity is a safe assumption.

The honest limitations

• Robust does not mean magic. Robust standard errors fix inference under heteroskedasticity. They do nothing about serial correlation, which in time-series data is usually the more dangerous problem. Running White standard errors on an overlapping-return regression and calling it robust is a common and costly mistake: you need Newey-West for that.
• They are only asymptotically valid, and in small samples they are too small. With a few dozen observations, robust standard errors can be badly downward-biased, making you more overconfident than plain OLS would have. This is a well-documented issue and it is why the small-sample corrections (HC1, HC2, HC3) exist. Using plain HC0 on a small sample is a real error.
• You give up efficiency. If you genuinely do know the form of the heteroskedasticity, properly weighted least squares gives you a more precise estimate than OLS with robust errors. Robustness is bought with information.
• It patches the symptom, not always the disease. Heteroskedasticity can be a signal that your model is misspecified: a missing interaction, a wrong functional form, an omitted regime. Applying a robust correction and moving on can mean silencing a useful alarm.
• It does not help with the biggest problem in most regressions, which is endogeneity. Robust standard errors give you honest uncertainty about a coefficient that may be pointing at entirely the wrong thing.

The one-line takeaway

White showed that you can get honest standard errors under unequal error variances without knowing anything at all about how the variances vary, because a sum of individually useless estimates converges to the right answer, and in doing so he made "robust" the default setting for applied econometrics.