Heteroskedasticity is a statistical condition in which the variance of a variable changes across observations. In regression analysis, it means the spread of residuals (the differences between predicted and actual values) is not constant, which violates a core assumption of ordinary least squares (OLS) estimation. While OLS estimates remain unbiased, their standard errors become unreliable, leading to invalid hypothesis tests and confidence intervals.
Key Takeaways
- Heteroskedasticity means the variance of errors changes across observations.
- It violates the OLS assumption of constant variance (homoskedasticity).
- OLS estimates remain unbiased but standard errors become wrong.
- Detection tests include Breusch-Pagan, White, and visual inspection of residual plots.
- Robust standard errors (Huber-White)correct standard errors without fixing the estimator.
What is Heteroskedasticity?
Heteroskedasticity occurs when the variability of a variable is not the same across all values of another variable. For instance, household consumption varies more among high-income families than low-income families. If you regress consumption on income, the residuals (vertical distance from the regression line) fan out as income rises. This outward-spread pattern is a visual signature of heteroskedasticity.
In notation, homoskedasticity means the variance of the error term is constant: Var(epsilon) equals sigma squared. Heteroskedasticity means Var(epsilon_i) equals sigma squared_i, which can differ for each observation. The error variance might increase with income, decrease with firm size, or follow a cycle over time.
How Does Heteroskedasticity Work?
Heteroskedasticity arises in three common ways. First, scale effects: when a variable scales with another, the residual variance scales too. Income vs. expenditure is a classic example. Second, outliers at specific levels: if extreme observations cluster at high values of the predictor, variance is larger there. Third, structural breaks: if the data generating process changes at a known point (e.g., a policy change), variance can shift accordingly.
Detection usually begins with a residual plot. If the residuals look like a horizontal band (constant width), homoskedasticity holds. If they fan out or form a pattern, heteroskedasticity is likely. Formal tests include the Breusch-Pagan test, which regresses squared residuals on the predictors and tests if the R-squared is significant. The White test is more general, including squares and cross products. A significant test statistic indicates heteroskedasticity.
Treatment options include robust standard errors (Huber-White), which adjust the standard error estimates without changing the OLS coefficients. For students, this is often the first line of defense, because it is simple and does not require knowing the variance structure. Alternatively, you can use weighted least squares (WLS), where each observation is weighted by the inverse of its estimated variance. This requires a model of how variance changes, which is not always known.
Why Does Heteroskedasticity Matter?
Heteroskedasticity matters because it produces incorrect standard errors, which lead to wrong conclusions. A coefficient might appear statistically significant when it is not, or vice versa. For a researcher, this means p-values and confidence intervals cannot be trusted without correction.
In finance, heteroskedasticity is the rule, not the exception. Stock return volatility clusters in time (high-volatility days follow high-volatility days). This motivated ARCH and GARCH models (Engle 1982, 2003 Nobel Prize), which explicitly model time-varying variance. The insight that volatility changes over time became central to options pricing (implied volatility surfaces), value-at-risk (VaR), and portfolio optimization.
In policy evaluation, heteroskedasticity affects treatment-effect estimates. If outcomes have different variances across treatment and control groups, unweighted estimates can overstate or understate the treatment effect. Empirical economists routinely report both conventional and heteroskedasticity-robust standard errors to address this.
What Are the Limitations in Handling Heteroskedasticity?
- Robust errors fix only standard errors, not efficiency - OLS remains inefficient (higher variance than necessary) when errors are heteroskedastic. WLS is more efficient but requires specifying the variance function.
- Sample size matters - robust standard errors are valid in large samples but can be unreliable in small samples. In samples under 50, bootstrap methods may be more trustworthy.
- Identification of the variance structure is difficult - choosing the right weights for WLS requires knowing how variance scales. A misspecified weight can make things worse, not better.
- Conditional vs. unconditional heteroskedasticity - ARCH/GARCH handle time-varying conditional variance. Cross-sectional heteroskedasticity requires different tools.
- Multicollinearity can mask heteroskedasticity - when predictors are highly correlated, residual patterns may not reveal heteroskedasticity cleanly.
Frequently Asked Questions
Does heteroskedasticity cause bias?
No. OLS coefficients remain unbiased. Heteroskedasticity affects the standard errors, not the point estimates. However, tests based on those standard errors (t-tests, F-tests) become invalid, so significance conclusions may be wrong.
When should I use robust standard errors?
As a default practice in applied regression with cross-sectional data, robust standard errors are a good insurance policy. They do no harm under homoskedasticity (they converge to the standard OLS errors) and correct under heteroskedasticity. Most statistical packages support them (Stata, Python statsmodels, R sandwich package).
What is the difference between ARCH and GARCH?
ARCH (Autoregressive Conditional Heteroskedasticity) models variance as a function of past squared errors. GARCH (Generalized ARCH) adds past variances to the equation, similar to adding an AR term to an MA model. GARCH is more parsimonious and is the standard model in financial econometrics for time-varying volatility.
This article is for educational purposes only and does not constitute financial advice.
