Heteroscedasticity: How to Detect It and Fix It in Regression

What Is Heteroscedasticity?

The name comes from the Greek roots hetero (different) and skedasis (dispersion). When dispersion is the same across all observations, the condition is called homoscedasticity. When it differs, it is heteroscedasticity. Both spellings — heteroscedasticity and heteroskedasticity — refer to exactly the same concept; different textbooks and disciplines simply favour one over the other.

The clearest mental image: plot the regression residuals on the vertical axis and fitted values on the horizontal axis. Under constant variance, residuals form a band of roughly equal width from left to right. Under heteroscedasticity, that band fans out, narrows, or shifts in some systematic way. The pattern is a diagnostic clue, not an automatic verdict.

The Constant-Variance Assumption

In a standard simple linear regression model:

The classical ordinary least squares framework assumes, among other things, that the error variance is the same for every observation:

Heteroscedasticity replaces that fixed σ² with a value that can differ across observations:

The subscript on σᵢ² signals that the variance is now observation-specific. It may grow with X, shrink with X, follow a nonlinear variance relationship, or change in some other structured way. Residuals — calculated as eᵢ = yᵢ − ŷᵢ — are the observable proxies for these unobserved errors, so analysts examine residuals to diagnose the problem. They are not identical to the true errors, but they carry the signal needed for practical diagnostics. For a full treatment of what regression residuals are and how to interpret them, see the regression residuals guide.

Homoscedasticity vs Heteroscedasticity

The two conditions sit at opposite ends of the variance-assumption spectrum. Both are properties of the conditional error distribution — how spread out errors are given a particular predictor value — not simply how spread out the raw Y values are.

What Does Heteroscedasticity Look Like?

The residual-versus-fitted-values plot is the primary visual diagnostic. The question to ask is simple: does the vertical spread of residuals remain roughly stable from left to right, or does it change in some systematic way?

Four Common Variance Patterns

Heteroscedasticity is not only an increasing fan. The defining feature is a systematic change in spread, regardless of the direction. Here are the patterns analysts most often encounter.

Heteroscedasticity vs Nonlinearity

A curved band of residuals — where the center of the residual cloud bows above or below zero — points toward a misspecified conditional mean, not changing variance. Changing residual width is the variance signal. Changing residual center is the mean-specification signal. Both can appear simultaneously, which is why careful visual inspection of the full residual plot matters before any remedy is chosen.

Why Does Heteroscedasticity Occur?

There is no single cause. Understanding what produced the changing variance in a particular dataset is the first step toward choosing a sensible response. Common origins include the following.

Identifying the likely cause matters because it shapes the remedy. A scale effect suggests a response transformation. Omitted variables suggest model respecification. Unknown variance structure suggests robust standard errors. Do not reach for a fix before understanding the source.

Why Is Heteroscedasticity a Problem?

Effect on OLS Coefficients

This is the most widely misunderstood consequence. Under a correctly specified linear conditional mean model with the appropriate error-predictor independence conditions, heteroscedasticity alone does not make OLS coefficient estimates biased. The OLS regression line can still estimate the conditional mean appropriately. The slope β̂₁ and intercept β̂₀ are determined by the first-order conditions that minimize squared residuals, and those conditions do not require constant variance to produce an unbiased result.

Effect on Standard Errors

This is where heteroscedasticity creates real inferential damage. The conventional OLS standard-error formula was derived under the assumption of constant variance. When error variance changes across observations, that formula uses the wrong variance information, and the resulting standard errors can be inconsistent — meaning they do not converge to the right value even in large samples.

Effect on P-Values and Confidence Intervals

Because t-statistics divide coefficient estimates by their standard errors, distorted standard errors flow through to distorted t-statistics, distorted p-values, and confidence intervals with incorrect nominal coverage. A slope that appears statistically significant at the 5% level under conventional standard errors may not be significant under a correct variance estimate — or vice versa. This is why the inferential chain matters:

Effect on OLS Efficiency

Under the Gauss-Markov conditions — which include the constant-variance assumption — OLS has the familiar property of being the best (minimum variance) linear unbiased estimator. When heteroscedasticity is present, OLS no longer necessarily holds that efficiency property. Alternative estimators such as weighted least squares, properly specified, can be more efficient. Whether this matters practically depends on how severe the variance imbalance is and what precision is required for the analysis.

Effect on Prediction Uncertainty

When conditional error variance changes across X, prediction intervals of constant width are inadequate. A single RMSE figure summarises average prediction-error magnitude across all observations, but it cannot capture that predictions at some fitted values are inherently more uncertain than predictions at others. Understanding the variance structure allows better-calibrated prediction intervals. For more on how prediction accuracy is measured overall, see the RMSE guide, bearing in mind that a single RMSE does not diagnose whether variance is constant.

How to Detect Heteroscedasticity

Step-by-Step: Residual-vs-Fitted Plot

Breusch-Pagan Test

The Breusch-Pagan test is a formal procedure for evaluating whether residual variance is systematically related to the explanatory variables. Rather than relying on visual judgment alone, it produces a test statistic and p-value under clearly stated assumptions.

Null hypothesis: Var(εᵢ | Xᵢ) = σ² — error variance is constant.

Alternative: Error variance depends systematically on the specified explanatory variables.

The core logic: if the squared residuals (which proxy for error variance at each observation) are predictable from the explanatory variables, the null of constant variance is undermined. At a high level, the test auxiliary regression relates squared residuals to the predictors and evaluates how much explained variation exists. The resulting LM test statistic follows an approximate chi-squared distribution under the null, with degrees of freedom corresponding to the number of predictors in the auxiliary regression.

White Test

The White test takes a more general approach. Rather than regressing squared residuals only on the original predictors, it also includes squares and cross-products of those predictors, allowing it to detect forms of heteroscedasticity that do not follow a simple linear pattern with the original X variables.

This generality is both an advantage and a limitation. The White test can detect more complex variance relationships, but in models with several predictors the auxiliary regression accumulates many terms, which consumes degrees of freedom and can reduce statistical power in smaller samples. The chi-squared test statistic follows the same distributional reference as Breusch-Pagan under the null, but with a larger degrees-of-freedom parameter.

In practice, plotting residuals first gives direction and shape; then a formal test adds inferential grounding. The choice between Breusch-Pagan and White depends on the model size and whether there is prior theory about the variance structure.

How to Fix Heteroscedasticity

No single remedy applies universally. The right choice depends on what caused the changing variance, what the goal of the analysis is, and how well the variance structure can be modelled. Avoid choosing a remedy solely because it makes a diagnostic-test p-value non-significant.

Heteroscedasticity-Robust Standard Errors

Robust standard errors — sometimes called heteroscedasticity-consistent (HC) standard errors or sandwich standard errors — recompute the covariance matrix of the OLS estimator using a formula that does not assume constant variance. Crucially, they do not change the OLS coefficient estimates. The same β̂₀ and β̂₁ remain; only the estimated covariance matrix, and therefore the standard errors, t-statistics, and confidence intervals, changes.

OLS coefficients:   β̂ (unchanged — same OLS point estimates)
Covariance matrix:  V̂ (changed — uses HC-corrected formula)
Standard errors:    SE(β̂) = sqrt(diag(V̂))  (changed)
t-statistics:       β̂ / SE(β̂)              (changed)
p-values:           from corrected t         (changed)
Confidence intervals: β̂ ± t* · SE(β̂)      (changed)

HC0, HC1, HC2, HC3 — Which Variant?

Several versions of the heteroscedasticity-consistent covariance estimator exist, each applying different small-sample corrections.

HC3 is often the default recommendation in applied econometrics for small to moderate samples because its stronger leverage correction tends to produce better-calibrated inference. No single variant is universally optimal, and software packages use different defaults — always check which is being applied.

Response Transformation

A logarithmic transformation of the response variable may stabilise variance when variability is roughly proportional to the mean — a pattern common in positive, right-skewed outcomes. After taking logs, the model becomes multiplicative on the original scale, which changes both the variance structure and the interpretation of every coefficient.

Weighted Least Squares

Weighted least squares assigns each observation a weight inversely related to its estimated error variance. Observations with larger error variance receive lower weight, reducing their influence on the estimated coefficients. When the variance structure is correctly modelled, WLS can be more efficient than OLS under heteroscedasticity.

The challenge is that the variance structure Var(εᵢ | Xᵢ) is rarely known precisely. Incorrect weights introduce their own problems. Feasible generalised least squares (FGLS) estimates the variance parameters first, then uses those estimates as weights. If the variance model is misspecified, the resulting estimator may be worse than OLS with robust standard errors. WLS requires substantive justification — do not weight observations simply to make residual plots look flat.

Heteroscedasticity vs Other Assumption Violations

Heteroscedasticity vs Non-Normality

They are separate assumptions. Heteroscedasticity concerns whether the conditional variance is constant across observations. Non-normality concerns whether the conditional error distribution follows a normal distribution — relevant for exact finite-sample inference in classical regression. A model can have constant variance with non-normal errors, or changing variance with approximately normal conditional errors, or both, or neither. A normality test (such as a Shapiro-Wilk test on residuals) does not diagnose heteroscedasticity, and a heteroscedasticity test does not diagnose non-normality.

Heteroscedasticity vs Autocorrelation

Autocorrelation concerns dependence among errors across observations — a pattern common in time series or spatial data where nearby observations share information. Heteroscedasticity concerns changing variance magnitude, not error dependence. The two can coexist. Standard heteroscedasticity-robust standard errors do not correct for serial correlation; cluster-robust or HAC (heteroscedasticity and autocorrelation consistent) standard errors are needed when both are present.

Heteroscedasticity vs Outliers

A single extreme residual can locally distort the apparent spread in a residual plot, mimicking a heteroscedastic pattern. Before concluding that variance is systematically non-constant, identify whether any specific observations have unusually high leverage or influence on the regression. The influential points guide covers leverage, Cook's distance, and how to assess whether individual observations are driving apparent diagnostic patterns.

A Numeric Illustration

The table below uses two small hypothetical datasets to show the contrast between constant and increasing variance. The regression relationship (a slope of approximately 2) is similar in both; only the residual spread differs.

Dataset A — Roughly Constant Variance

Dataset B — Increasing Variance

In both datasets the fitted slope is approximately 2. The OLS coefficient is not obviously wrong in Dataset B. What is wrong is that if you apply the constant-variance standard-error formula to Dataset B, you get a single variance estimate that blends the small errors at X=1 with the large errors at X=5, producing standard errors that do not accurately reflect precision at any particular X value.

Frequently Asked Questions

Quick Reference