1 What the Breusch–Pagan Test Assesses

1.1 Heteroskedasticity in regression models

In ordinary least squares (OLS) linear regression, a standard assumption is that the error terms have constant conditional variance across observations. When this conditional variance changes with the explanatory variables, the model exhibits heteroskedasticity. The Breusch–Pagan test is designed to detect such non-constant variance by checking whether variability in the residuals appears systematically related to predictors.

The test does not measure the “size” of heteroskedasticity directly; rather, it evaluates whether there is statistical evidence that the error variance depends on included regressors, according to the test’s chosen specification.

1.2 Null and alternative hypotheses

The Breusch–Pagan test is formulated around a variance-constant null hypothesis. Under the null, the conditional variance of the error term is constant (homoskedasticity), so the squared residuals should show no systematic relationship with the regressors beyond what sampling noise would produce.

Under the alternative, the conditional variance depends on one or more regressors through a chosen functional form (typically linear in parameters). Evidence against the null is reflected in a test statistic large enough to be unlikely under homoskedasticity.

1.3 When the test is appropriate (practical use cases)

The Breusch–Pagan test is most commonly used as a diagnostic in linear regression workflows where researchers suspect that spread in outcomes increases or decreases with predictors. Typical practical contexts include:

  • Cross-sectional data where measurement error may vary with a covariate magnitude.
  • Models where theory suggests variance changes with income, size, distance, or similar scaling variables.
  • Situations where residual plots show patterns, and formal evidence is desired to support model adjustments.

Because the test depends on a specified link between variance and regressors, it is best viewed as a targeted check aligned with the modeling intent.

2 Model Setup and Assumptions

2.1 Linear regression framework

Consider a linear regression model \[ y_i = x_i^\top \beta + u_i,\quad i=1,\dots,n, \] where \(x_i\) is the vector of regressors (including an intercept if present), \(\beta\) are coefficients, and \(u_i\) are error terms. OLS estimates \(\beta\) by minimizing the sum of squared residuals, and the Breusch–Pagan test uses the fitted residuals to assess whether \(\operatorname{Var}(u_i\mid x_i)\) is constant.

The procedure is usually introduced in the setting of classical linear regression assumptions for inference about coefficients, even though the test itself is used primarily for diagnosing variance structure.

2.2 Error structure and variance modeling

Under homoskedasticity, the conditional variance satisfies \[ \operatorname{Var}(u_i\mid x_i)=\sigma^2, \] a constant independent of \(x_i\). Under heteroskedasticity allowed by the Breusch–Pagan framework, the conditional variance is modeled as varying with regressors, often using a parametrization such as \[ \operatorname{Var}(u_i\mid x_i)=\sigma^2 h(z_i), \] where \(z_i\) is a vector of variables (commonly chosen from among the regressors \(x_i\)) and \(h(\cdot)\) is specified so that the constant-variance case corresponds to a set of parameter restrictions (for example, a slope vector of zeros in a linear approximation).

2.3 Independence and exogeneity considerations

The diagnostic is generally presented with the understanding that the regression residuals behave in a way consistent with the model’s assumptions about exogeneity: regressors should be independent of the error term in an appropriate conditional sense. The test’s validity relies on asymptotic arguments that presume a form of regularity, such as sufficient sample size and stable design matrices.

If exogeneity fails, the residual patterns can arise from misspecification beyond variance instability, and test results may be hard to interpret purely as heteroskedasticity.

2.4 Relationship to residuals and fitted values

The Breusch–Pagan test leverages the idea that when variance changes across observations, the magnitude of residuals (in particular their squares) tends to shift in tandem with predictors. The test constructs an auxiliary relationship involving squared residuals and selected regressors or functions thereof. Since residuals depend on both the underlying variance and the correctness of the mean specification, the test implicitly assumes the mean model is reasonably adequate.

3 Core Test Construction

3.1 Computing residuals from the fitted model

The first step is to fit the baseline regression by OLS and compute residuals \[ \hat{u}_i = y_i - x_i^\top \hat{\beta}. \] These residuals approximate the unobserved errors. Squaring them produces quantities that, under the homoskedastic null, fluctuate around a constant mean related to the error variance, with no systematic dependence on the regressors.

In practice, residuals are often scaled (e.g., divided by an estimate of \(\sigma^2\)) within the auxiliary regression, which can affect numerical behavior but not the conceptual structure of the test.

3.2 Forming the auxiliary regression

The canonical auxiliary regression regresses a function of squared residuals on the regressors implicated in the variance model. A common formulation sets \[ \frac{\hat{u}_i^2}{\hat{\sigma}^2} = z_i^\top \gamma + e_i, \] where \(z_i\) is the vector of variables used to explain heteroskedasticity (often including the intercept and possibly a subset of \(x_i\)), and \(\gamma\) are auxiliary parameters.

Under the null of constant variance, the coefficients on the variance-explaining variables (other than the intercept term) should be zero.

3.3 Test statistic derivation (Lagrange multiplier form)

The Breusch–Pagan statistic is frequently presented as a Lagrange multiplier (LM) test. In the simplest version, the LM statistic is computed from the auxiliary regression’s fit, typically expressed as

  • \(LM = n \cdot R^2\) from regressing the scaled squared residuals on \(z_i\) (with appropriate degrees of freedom),

or an equivalent expression using the number of restrictions under the null.

The underlying derivation connects the auxiliary regression’s explanatory power to how likely the observed residual pattern is under the homoskedastic variance structure, asymptotically.

3.4 Reference distribution and p-values

Under the null hypothesis and regularity conditions, the test statistic follows (approximately) a chi-squared distribution with degrees of freedom equal to the number of independent restrictions in the variance model (for example, the number of non-intercept regressors included in the auxiliary regression).

A p-value is obtained by comparing the computed statistic to this reference distribution. Small p-values indicate evidence against homoskedasticity as specified by the model underlying the test.

4 Variants and Extensions

4.1 Breusch–Pagan test with multiple regressors

The multi-regressor variant uses a vector \(z_i\) containing several predictors believed to influence the conditional variance. The auxiliary regression then includes these variables, and the degrees of freedom for the chi-squared reference distribution increase with the number of slope parameters being tested.

This version is useful when heteroskedasticity may depend on multiple aspects of the design, such as both a scale variable and a quality indicator. It also provides a way to test joint effects on variance rather than focusing on a single predictor.

4.2 Testing specific functional forms of the variance

Breusch–Pagan is not fully agnostic about how variance might vary. Instead, it tests heteroskedasticity relative to a chosen structure—commonly linear in the regressors in the auxiliary equation. Researchers can tailor the auxiliary regressors to reflect plausible variance mechanisms, such as including:

  • Selected covariates thought to scale measurement error,
  • Polynomial terms (e.g., \(x_i^2\)) if variance plausibly grows nonlinearly,
  • Interactions if theory suggests variance depends on combined effects.

This tailoring improves relevance but also means the test may fail to detect heteroskedasticity when the true variance pattern differs substantially from the specified form.

Generalized least squares (GLS) is a modeling approach that accounts for known or estimated heteroskedastic variance structures to achieve efficient estimation. While the Breusch–Pagan test is not itself a GLS estimation method, it provides a diagnostic basis for considering GLS (or related heteroskedasticity-aware methods).

Conceptually, the test can inform whether estimating a variance model is warranted, especially when a variance structure is motivated by theory or when downstream inference depends on correctly specified error variance.

4.4 Relation to other heteroskedasticity diagnostics

Breusch–Pagan is part of a family of diagnostics aimed at heteroskedasticity. Other procedures differ in how flexibly they relate residual variability to predictors and what residual transformation they use. Compared with broader checks, Breusch–Pagan typically corresponds to a more structured alternative, making it sensitive to the specified dependence while potentially less sensitive to unexpected patterns.

A practical approach in applied work often combines formal tests with graphical residual analysis and comparison across multiple diagnostic methods.

5 Interpretation and Reporting

5.1 Interpreting p-values and effect on inference

A statistically significant Breusch–Pagan test p-value (relative to a chosen significance level) signals evidence that the conditional variance is unlikely to be constant under the test’s variance specification. In regression analysis, this affects the reliability of conventional standard errors that assume homoskedasticity.

Importantly, coefficient point estimates from OLS generally remain unbiased under many heteroskedasticity forms, but their standard errors—and therefore t-tests and confidence intervals—may be invalid unless heteroskedasticity-consistent inference is used.

5.2 Distinguishing symptoms from causes of heteroskedasticity

The test detects a pattern in residual variance; it does not identify why the pattern exists. Variance changes can arise from omitted variables affecting noise, nonlinear mean misspecification, data scaling issues, measurement heterogeneity, or model form that fails to capture systematic structure.

Consequently, a prudent interpretation treats the result as a symptom that motivates model checking. Researchers often follow up by examining residual plots, verifying functional form, and considering transformations or alternative variance modeling.

5.3 Common reporting conventions in papers

Reporting typically includes:

  • The test name (Breusch–Pagan) and version (e.g., which variables were used in the auxiliary regression),
  • The test statistic and degrees of freedom,
  • The p-value,
  • A brief statement about implication for standard errors or model re-estimation.

In many research reports, authors also mention whether they subsequently use heteroskedasticity-robust standard errors, even when the test indicates heteroskedasticity.

6 Alternatives and Robust Approaches

6.1 Robust standard errors (conceptual comparison)

Robust standard errors, often called heteroskedasticity-consistent estimators, modify the estimated variance of regression coefficients without requiring a correct specification of the heteroskedasticity form. In contrast, the Breusch–Pagan test is a diagnostic that assumes a particular structure for how variance depends on predictors.

As a result, robust standard errors can be viewed as a safeguard for inference, while the Breusch–Pagan test helps evaluate whether an assumption of constant variance is inconsistent with the data.

6.2 White’s test vs. Breusch–Pagan (high-level contrast)

White’s test is commonly described as being more general in that it can detect a wider set of heteroskedastic alternatives, often by including cross-products and squares of regressors in the auxiliary regression. The Breusch–Pagan test, by comparison, typically uses a simpler variance specification linked to the original regressors (or a chosen subset).

This difference implies trade-offs: the more general approach may have higher sensitivity to various patterns but can also be less parsimonious, affecting degrees of freedom and power under some circumstances.

6.3 Choosing among heteroskedasticity tests

Selection depends on modeling goals and assumptions about variance. Factors that influence choice include:

  • The plausibility of a structured variance mechanism,
  • Sample size and model complexity,
  • Whether the researcher wants evidence tied to specific covariates or a broader screening procedure,
  • The desire to minimize reliance on a particular variance functional form.

In applied research, it is common to pair test outcomes with robustness checks and to ensure inferential methods remain valid under heteroskedasticity.

7 Implementation Details

7.1 Data requirements and preprocessing

The test requires an estimated OLS regression and residuals from that model. No specialized preprocessing is required beyond standard data cleaning and ensuring that regressors are properly defined. However, practical issues can influence stability:

  • Handling missing data consistently across model estimation and auxiliary regression,
  • Scaling very large or tiny predictors to improve numerical conditioning,
  • Ensuring sufficient variation in regressors to avoid singular auxiliary regressions.

If predictors include categorical indicators or transformed variables, they should be included in \(z_i\) in a manner consistent with the intended variance explanation.

7.2 Software and implementation workflows

Many statistical software packages provide built-in routines for the Breusch–Pagan test, sometimes allowing users to specify the variables included in the auxiliary regression. A typical workflow is:

  1. Fit the baseline linear regression by OLS.
  2. Obtain residuals and fitted values internally.
  3. Run the auxiliary regression of scaled squared residuals on chosen \(z_i\).
  4. Compute the LM statistic and retrieve the p-value using the chi-squared approximation.

Some implementations also permit formula interfaces (e.g., specifying a “variance model”) that define \(z_i\) directly.

7.3 Numerical considerations (scaling, leverage, stability)

Residual-based procedures can be sensitive to numerical scaling. Squared residuals can become large in magnitude when the outcome is not scaled, potentially affecting regression stability in the auxiliary step. Scaling residuals by an estimate of error variance is one common remedy.

Additionally, extreme leverage points—observations with unusual regressor values—can influence both the fitted baseline model and the auxiliary regression, potentially overstating evidence of heteroskedasticity. Diagnostic plots and influence measures can help determine whether results are driven by a small subset of observations.

8 Worked Example (Method Walkthrough)

8.1 Fitting the baseline regression

Suppose a researcher models an outcome \(y\) using predictors \(x_1, x_2\) via \[ y_i = \beta_0 + \beta_1 x_{1i} + \beta_2 x_{2i} + u_i. \] They fit the model using OLS and extract residuals \(\hat{u}_i\). At this stage, it is typical to also inspect residual plots to see whether variance appears to expand or contract with fitted values or with individual predictors.

8.2 Running the Breusch–Pagan test steps

Next, the researcher defines the variance specification. A common choice is to let the auxiliary regression use the same set of regressors as the mean model: \[ \frac{\hat{u}_i^2}{\hat{\sigma}^2} = \gamma_0 + \gamma_1 x_{1i} + \gamma_2 x_{2i} + e_i. \] They then estimate this auxiliary regression and compute \(R^2\). The LM test statistic is obtained as \(LM = nR^2\) (with degrees of freedom equal to the number of slope terms tested, here typically 2 if only \(\gamma_1\) and \(\gamma_2\) are restricted to zero).

Finally, they calculate the p-value using the chi-squared reference distribution.

8.3 Verifying assumptions and checking diagnostics

Because the Breusch–Pagan test relies on residuals from a particular mean specification, the researcher checks whether the baseline model appears misspecified in ways unrelated to variance. For example, if residuals show systematic curvature, it may indicate a nonlinear mean structure rather than pure heteroskedasticity. In such cases, adjusting the mean model (e.g., adding nonlinear terms or transforming the dependent variable) can change both residual behavior and the test outcome.

They may also examine whether results are robust to using alternative \(z_i\) choices in the auxiliary regression, such as using only \(x_1\) if theory suggests variance depends mainly on that predictor.

8.4 Presenting results in a research-friendly way

A typical reporting statement might specify:

  • the auxiliary regression variables used,
  • the test statistic and p-value,
  • the follow-up inference choice (e.g., using heteroskedasticity-robust standard errors).

For example: “The Breusch–Pagan test for heteroskedasticity using \(x_1\) and \(x_2\) in the variance specification yielded \(LM=\dots\), df=\dots, p=\dots\). Given evidence of heteroskedasticity, we report heteroskedasticity-consistent standard errors in subsequent inference.”

This framing communicates both the diagnostic evidence and the practical modeling response.