Remedial Measures for Heteroscedasticity in Econometrics

You’ve done it. You’ve collected your data, specified your economic model, and run your Ordinary Least Squares (OLS) regression. You feel great. But then, you run a diagnostic test-maybe the Breusch-Pagan test or the White test-and the result is not what you wanted: heteroscedasticity. This intimidating term simply means that the variance of your model’s errors (the “guesswork” or residuals) is not constant. For some of your data, the model’s predictions are very precise (low error variance), while for other parts, the predictions are all over the map (high error variance).

Why is this a problem? While heteroscedasticity doesn’t bias your coefficients (your $\beta$ estimates are still correct on average), it throws your standard errors completely out of whack. If your standard errors are wrong, you can’t trust your t-statistics or your p-values. You might conclude a variable is statistically significant when it isn’t, or vice versa. It undermines your ability to test hypotheses. Fortunately, this is a common problem in econometrics, and there is a well-stocked toolkit of remedial measures to fix it. This post will walk you through the five most common strategies, from the precision tools to the practical workarounds.

The precision tool: Weighted Least Squares (WLS) 🔗

The Weighted Least Squares (WLS) method is the theoretically ideal solution, but it comes with one very large “if”: it works perfectly *if* you know the exact pattern or form of the heteroscedasticity. The core idea is brilliantly intuitive. If OLS gives equal weight to every observation, WLS gives *unequal* weight. It strategically gives less weight to observations that come from the high-variance part of your data (the “noisy” ones) and more weight to observations from the low-variance, “cleaner” part.

Think of it like a teacher grading a set of exams from two different schools. They know that students from School A are incredibly consistent, with scores always close to their true ability. Students from School B, however, are highly erratic; their scores can swing wildly. To get the fairest estimate of the overall student ability, the teacher wouldn’t just average all the scores together. They would give *more weight* to each score from the reliable School A and *less weight* to each score from the “noisy” School B. WLS does exactly this with your data.

How WLS transforms the model 🔗

Technically, WLS “fixes” the model by transforming it. Let’s say your original, heteroscedastic model is:

$ Y_i = \beta_0 + \beta_1 X_i + u_i $

You suspect that the variance of the error, $ Var(u_i) $, is not constant. Let’s say you believe the variance is proportional to some variable, $X_i$. This means the standard deviation of the error is proportional to $ \sqrt{X_i} $. WLS instructs you to divide your *entire* equation by this standard deviation term, $ \sqrt{X_i} $:

$ \frac{Y_i}{\sqrt{X_i}} = \beta_0 \left( \frac{1}{\sqrt{X_i}} \right) + \beta_1 \left( \frac{X_i}{\sqrt{X_i}} \right) + \left( \frac{u_i}{\sqrt{X_i}} \right) $

This looks complicated, but the magic happens with the new error term, $ \left( \frac{u_i}{\sqrt{X_i}} \right) $. By dividing the original error by its own standard deviation, the new error term becomes homoscedastic-it now has a constant variance. You can now run a simple OLS regression on this *transformed* model (using the transformed Y, transformed X, and transformed intercept term). The resulting estimates will be both unbiased and, unlike your original OLS, efficient (meaning they are the most precise estimates you can get).

The big catch, of course, is that we rarely know the true form of the heteroscedasticity. We often have to *assume* what the weights should be (e.g., variance is proportional to $X_i$, or $X_i^2$). When we use an *estimated* or assumed form of the variance to perform WLS, it’s technically called Feasible Weighted Least Squares (FWLS).

The heavyweight champion: Generalised Least Squares (GLS) 🔗

If WLS is a specific tool, Generalised Least Squares (GLS) is the entire professional toolkit. GLS is a much broader and more formal procedure that can handle a wider range of issues with the error terms. Specifically, GLS is designed to produce efficient estimates (the Best Linear Unbiased Estimator, or BLUE) in the presence of *either* heteroscedasticity (non-constant variance) or autocorrelation (errors that are correlated with each other, common in time-series data), or both.

In fact, WLS is just a special case of GLS. GLS works by transforming the original model using the full variance-covariance matrix of the errors. This is a large grid that describes not only the variance of each error term (the diagonal elements) but also how each error term is related to every other error term (the off-diagonal elements).

• When the errors are heteroscedastic but not autocorrelated, this matrix is “diagonal” (all off-diagonal elements are zero), and the GLS procedure simplifies to become WLS.
• When the errors are homoscedastic *and* not autocorrelated (the classic assumptions), the variance-covariance matrix is even simpler, and GLS simplifies to become plain old OLS.

For the purpose of fixing heteroscedasticity, you can think of GLS as the underlying theory and WLS as the practical application. The goal of GLS is to find the most efficient, Best Linear Unbiased Estimator (BLUE). OLS is only BLUE when all its classical assumptions (including homoscedasticity) hold. When heteroscedasticity is present, GLS (or its WLS form) becomes the new BLUE.

The “what if I have no idea?” solution: White’s robust standard errors 🔗

So, WLS/GLS is great, but it requires us to *know* or make a very good guess about the form of the heteroscedasticity. What if we have no idea what’s causing it? What if we guess the weights wrong? In that case, our “fix” could actually make our estimates *worse* than they were with OLS.

This is where a brilliant and now-famous workaround comes in, developed by economist Robert White in 1980. This solution is known as White’s Heteroscedasticity-Corrected Standard Errors, or more simply, robust standard errors.

The philosophy here is completely different. Instead of trying to “fix” the model by transforming it (like WLS), White’s method says: “Let’s just accept that the model is heteroscedastic and our OLS coefficients are what they are. The coefficients themselves are still unbiased, after all. The *only* thing that’s broken is the formula we use to *calculate* the standard errors. So, let’s just use a different, more ‘robust’ formula.”

It’s like having a car with an unreliable speedometer. The WLS approach is to try and rebuild the engine and transmission to make the speedometer accurate. The White’s error approach is to leave the engine alone and just buy a high-quality GPS that tells you your correct speed. You’re still using the same car (the OLS coefficients), but you’re getting your crucial information (the standard errors) from a more reliable source.

White’s method uses the squared residuals from your original OLS regression as an estimate of the non-constant variance, and then uses this information to build a new, corrected variance-covariance matrix. From this, it computes new standard errors that are “robust” to the presence of heteroscedasticity, even when its form is unknown. Today, this is by far the most common remedy used in applied economics. Almost every statistical software package (Stata, R, Python) has a simple command (like `.robust`) that you can add to your regression to get these corrected standard errors. This makes hypothesis testing (t-stats and p-values) valid again.

The only real downside is that this is a “large sample” property. It works best when you have a lot of data. In very small samples, it might not be perfectly accurate, and your OLS coefficients, while unbiased, are still not the most *efficient* (WLS/GLS estimates would be, if you knew the right weights).

The intuitive fix: Logarithmic transformation of variables 🔗

Sometimes, heteroscedasticity isn’t a deep, complex statistical problem but a simple problem of *scale*. Economic data is often right-skewed. Think about income. Most people are clustered around a median income, with a few individuals earning astronomical amounts. If you model food expenditure as a function of income, you’d expect a lot of heteroscedasticity.

Why? A person earning $30,000 a year might vary their weekly food budget by $\pm$$20. A person earning $3,000,000 a year might vary their budget by $\pm$$1,000. The *absolute* variance in spending is much larger for the high-income group. This is a classic case of variance increasing with the level of an independent variable.

A simple and elegant solution is to apply a logarithmic transformation. Instead of modeling $Expenditure = f(Income)$, you model $log(Expenditure) = f(log(Income))$.

This transformation does two things:

1. It compresses the scale. The $log(30,000)$ is $\approx 10.3$, while the $log(3,000,000)$ is $\approx 14.9$. The absolute difference between the logged values is much smaller than the $2,970,000 difference in the original data.
2. It often stabilizes the variance. By pulling in the extreme values and making the distribution more symmetric, a log transformation often makes the error variance much more constant, thus solving the heteroscedasticity problem.

The best part is that this is often strongly supported by economic theory. A log-log model is the standard way to estimate a Cobb-Douglas production function. Even better, the $\beta$ coefficients in a log-log model have a wonderful, direct interpretation: elasticity. A coefficient of 0.8 would mean “a 1% increase in $X$ is associated with a 0.8% increase in $Y$.”

The “common sense” fix: Using deflators to solve heteroscedasticity 🔗

This final method is another type of transformation, similar in spirit to both WLS and the log transformation. It’s about re-framing your research question in *relative* terms rather than *absolute* terms. This is done by deflating your variables.

Let’s go back to the example of modeling a company’s profits. If you regress $Profits$ on $Company\_Size$ (e.g., number of employees), you will almost certainly find heteroscedasticity. Huge multinational corporations will have profits in the billions, with a very large variance, while small businesses will have profits in the thousands, with a small variance. The variance of the error term is clearly related to the size of the company.

A “common sense” fix is to stop modeling absolute profits. Instead, why not model profits per employee? You would do this by dividing your dependent variable (and possibly your other independent variables) by the $Company\_Size$ variable.

This is a “deflator.” You are “deflating” the variables by a scaling factor that you believe is the source of the heteroscedasticity. In fact, you might recognize this: this is just a specific, intuitive application of Weighted Least Squares (WLS), where you have *assumed* that the error variance is proportional to $Company\_Size^2$ (and thus the standard deviation is proportional to $Company\_Size$).

This is extremely common in economics.

• Economists rarely model a country’s total $GDP$. They model per capita GDP (deflating by population).
• They often don’t use nominal $income$. They use real income (deflating by a price index like the Consumer Price Index).

Using deflators like population or a price index is a powerful two-for-one. It often solves the problem of heteroscedasticity by putting all observations on a comparable scale, *and* it makes the model more economically meaningful by focusing on relative measures (like per capita wealth) rather than absolute ones.

What do you think? Which of these methods do you find most appealing? Do you prefer a “pure” fix like WLS, or a “practical” workaround like White’s robust standard errors that is more common in modern research? Can you think of another economic model where you would expect to find heteroscedasticity right from the start?

References

1. https://online.stat.psu.edu/stat462/node/189/
2. https://www.econometrics-with-r.org/5-4-heteroskedasticity-robust-standard-errors.html
3. https://online.stat.psu.edu/stat501/lesson/9/9.1
4. https://www.rbi.org.in/Scripts/BS_PressReleaseDisplay.aspx?prid=54049