1 Background and purpose
The Breusch-Godfrey test is a diagnostic procedure in regression analysis for detecting serial correlation in the disturbance terms. It is widely used in time series and other econometric settings, where residual dependence can distort standard errors, test statistics, and confidence intervals. The method helps assess whether a fitted model adequately captures the data’s dynamic structure.
1.1 Definition of serial correlation
Serial correlation, also called autocorrelation, occurs when regression errors are correlated across observations rather than being independently distributed. In a time-ordered sequence, a positive error in one period may be associated with a positive error in the next. Such dependence can arise from omitted dynamics, unmodeled trends, or other sources of systematic structure in the residuals.
1.2 Why autocorrelated errors matter
When errors are autocorrelated, ordinary least squares estimates may remain unbiased under some conditions, but conventional inference often becomes unreliable. Standard errors can be understated or overstated, leading to misleading significance tests. Serial correlation may also indicate that the model is incomplete, since important lagged effects or explanatory variables may be missing.
1.3 Relationship to regression diagnostics
The Breusch-Godfrey test is one of several regression diagnostics used to evaluate model adequacy. It is typically considered alongside checks for heteroskedasticity, functional form, and outliers. Because it can detect more general forms of autocorrelation than some older methods, it is often favored in applied econometric work.
2 Historical development
The test was developed in the context of classical econometrics, where researchers needed tools to diagnose violations of regression assumptions in dynamic data. Its formulation provided a flexible alternative to earlier tests that were more limited in scope.
2.1 Breusch and Godfrey
Trevor Breusch and Leslie Godfrey independently contributed to the development of the procedure, which came to bear both names. Their work formalized a test based on an auxiliary regression, allowing researchers to examine residual dependence at specified lag lengths. The resulting method became a standard diagnostic for linear regression models.
2.2 Econometric context
The test emerged during a period when econometric practice increasingly dealt with time series models, lagged variables, and more complex error structures. As empirical work expanded beyond simple cross-sectional regression, the need for tools that handled dynamic dependence became more pronounced. The Breusch-Godfrey test answered that need with a general and practical framework.
2.3 Comparison with earlier tests
Earlier procedures, such as the Durbin-Watson statistic, were useful but had restrictions, especially in models containing lagged dependent variables. The Breusch-Godfrey approach was more adaptable because it could test for higher-order autocorrelation and remain applicable in broader regression settings. This flexibility made it especially valuable in applied econometrics.
3 Statistical framework
The test is built around the idea that, under the null hypothesis, the regression disturbances are not serially correlated. Under the alternative, past error terms help explain the current disturbance.
3.1 Regression error structure
Consider a regression model in which the observed dependent variable is represented as a function of explanatory variables plus an unobserved error term. If the error term follows a pattern across time or ordered observations, the residuals from the fitted model may inherit that dependence. The Breusch-Godfrey test probes for this structure by examining whether lagged residuals contribute explanatory power.
3.2 Null and alternative hypotheses
The null hypothesis states that the regression errors are not autocorrelated up to a chosen lag order. The alternative states that one or more lagged error terms are correlated with the current error. In practice, the test is often framed as a joint hypothesis that all autocorrelation coefficients through the selected order are zero.
3.3 Order of autocorrelation
The procedure allows the analyst to test for first-order, second-order, or higher-order serial correlation. The chosen order determines how many lagged residuals are included in the auxiliary regression. A higher order can detect more complex dependence patterns, though it may reduce power in small samples.
4 Test formulation
The Breusch-Godfrey test is typically implemented by estimating an auxiliary regression that augments the original model with lagged residuals. The significance of those added terms forms the basis of the test.
4.1 Auxiliary regression approach
First, the original regression model is estimated and residuals are obtained. Then these residuals are regressed on the original explanatory variables together with a selected number of their own lagged values. If the lagged residuals are jointly significant, that suggests serial correlation in the error process.
4.2 Construction of the test statistic
The test statistic can be derived from the auxiliary regression’s goodness of fit. Common versions use the sample size multiplied by the auxiliary regression’s coefficient of determination, which under the null follows an approximate chi-square distribution. This provides a straightforward way to assess whether the residual lags add explanatory content.
4.2.1 Lagrange multiplier interpretation
The Breusch-Godfrey test is often described as a Lagrange multiplier, or score, test. This means it evaluates whether adding autocorrelation parameters would significantly improve the model without requiring full estimation of the more general alternative. The approach is attractive because it relies on the restricted model and an auxiliary check rather than fitting a separate unrestricted dynamic error model.
4.2.2 Chi-square approximation
Under standard large-sample arguments, the test statistic is approximately chi-square distributed with degrees of freedom equal to the number of lagged residuals included. This approximation allows a p-value to be computed directly. In many software packages, this is the default reported version of the test.
4.3 F-version of the test
An F-form of the Breusch-Godfrey test is also used in practice. It is based on the same auxiliary regression but expresses the result as an F statistic rather than a chi-square statistic. This version is often convenient in finite samples and is especially familiar to users of standard linear-model output.
5 Assumptions and applicability
The test is most useful when applied within a linear regression framework and when the analyst seeks evidence of serial dependence in the residuals. Its applicability extends across several empirical settings.
5.1 Linearity and model specification
The usual formulation assumes a linear model with a correctly specified set of regressors, apart from possible autocorrelation in the errors. If the model is badly misspecified, the test may detect residual patterns that reflect omitted variables, nonlinearities, or structural changes rather than pure serial correlation. For this reason, interpretation should always be paired with broader diagnostic checking.
5.2 Time series and panel settings
The test is especially common in time series analysis, where sequential dependence is a natural concern. It can also be adapted to panel-data contexts, though the implementation may differ depending on whether the panel is balanced, the time dimension is short, or the error structure is more complicated. In all cases, the basic goal is to identify correlation across ordered observations.
5.3 Presence of lagged dependent variables
A major advantage of the test is that it remains valid when lagged dependent variables appear among the regressors. This distinguishes it from methods that rely on stronger restrictions. As a result, it is particularly useful in dynamic models where past outcomes are included as explanatory variables.
6 Implementation
In applied work, the test is straightforward to carry out using standard regression software. The main choices involve selecting the lag order and interpreting the resulting auxiliary regression output.
6.1 Choosing the number of lags
The selected lag length should reflect the data’s time scale, sample size, and plausible dynamics. Too few lags may miss existing autocorrelation, while too many can dilute power and consume degrees of freedom. Analysts often examine several orders or use domain knowledge to guide the choice.
6.2 Step-by-step procedure
A typical implementation proceeds as follows. First, estimate the original regression model and save the residuals. Second, construct an auxiliary regression with the original regressors and the chosen lagged residuals. Third, test whether the lagged residual coefficients are jointly zero using either the chi-square or F form. Finally, use the p-value to assess whether serial correlation is present.
6.3 Software implementations
The test is widely available in statistical software and econometric packages. Implementations usually automate the auxiliary regression and report the relevant test statistic and significance level.
6.3.1 R
In R, the test is commonly available through econometrics-oriented packages. Users typically fit a linear model and then call a diagnostic function specifying the number of lags to test. The output usually includes a test statistic, degrees of freedom, and a p-value.
6.3.2 Stata
In Stata, the procedure is commonly invoked after regression estimation through a postestimation command. The software reports both the auxiliary-regression-based statistic and an associated significance measure. This makes it easy to integrate the test into standard diagnostic workflows.
6.3.3 Python
In Python, econometric libraries provide functions for serial-correlation testing in regression models. The user generally supplies the fitted model and the desired lag order. The result is returned in a form suitable for quick interpretation and reporting.
7 Interpretation
The test output indicates whether the data provide evidence of residual autocorrelation at the selected lag order. Interpretation should be grounded in both statistical significance and the practical setting of the model.
7.1 Reading p-values
A small p-value suggests that the lagged residuals jointly contribute to explaining the current residual, which is evidence against the null hypothesis of no serial correlation. A large p-value indicates that the data do not provide strong evidence of autocorrelation at the tested order. As with other hypothesis tests, the p-value should be interpreted in context rather than mechanically.
7.2 Rejecting or failing to reject the null
Rejecting the null implies that the regression errors exhibit dependence at one or more of the tested lags. This usually signals that the model may need dynamic terms, different error assumptions, or additional explanatory variables. Failing to reject the null does not prove that the residuals are independent; it only means the test did not detect significant autocorrelation at the chosen order.
7.3 Practical significance
Statistical significance does not always imply serious practical consequences. Mild autocorrelation in a large sample may produce a significant result even if the effect on substantive conclusions is limited. Conversely, a non-significant result in a small sample may fail to uncover meaningful dependence. Analysts often consider the size and pattern of residual correlation alongside the formal test.
8 Limitations
Although widely used, the test has limitations that affect how results should be interpreted. These concerns are especially relevant in small samples or in models with broader specification problems.
8.1 Sensitivity to model misspecification
The test can respond to errors that are not purely serial correlation, such as omitted variables, incorrect functional form, or neglected seasonality. In such cases, a rejection may reveal that the model is inadequate more generally rather than indicating a single identifiable defect. Diagnostic follow-up is therefore essential.
8.2 Finite-sample considerations
The asymptotic justification for the chi-square approximation may be less accurate in small samples. Finite-sample performance can also depend on the number of lags tested and the complexity of the original model. For that reason, analysts may compare the chi-square and F versions or supplement the test with other evidence.
8.3 Comparison with alternative tests
No single test is best in every setting. Some alternatives are better suited to simple models, while others focus on residual patterns in pure time-series contexts. The Breusch-Godfrey test is valued for flexibility, but it is often used together with complementary diagnostics rather than in isolation.
9 Related methods
Several other procedures are commonly discussed alongside the Breusch-Godfrey test because they address serial dependence in regression or time-series residuals.
9.1 Durbin-Watson test
The Durbin-Watson test is an earlier diagnostic for first-order serial correlation in regression errors. It is easy to compute and widely known, but it has restrictions in models that include lagged dependent variables. The Breusch-Godfrey test is more general in such settings.
9.2 Ljung-Box test
The Ljung-Box test is a portmanteau test used to examine whether a sequence of autocorrelations is jointly zero. It is often applied to residuals from time-series models rather than general regressions. Its purpose overlaps with serial-correlation diagnostics, though its formulation differs from the Breusch-Godfrey approach.
9.3 Portmanteau tests
Portmanteau tests assess whether residual autocorrelations up to a chosen lag are jointly insignificant. They are useful in examining overall whiteness of residuals. Compared with the Breusch-Godfrey test, they are often more common in pure time-series model checking than in general regression analysis.
9.4 Wooldridge test for serial correlation
The Wooldridge test is commonly used in panel-data contexts to assess serial correlation in idiosyncratic errors. It is tailored to the structure of longitudinal data and may be preferable in some panel applications. The Breusch-Godfrey framework, by contrast, is more general in standard linear regression and time-series settings.
10 Applications
The test is widely used in empirical work where ordered observations may exhibit persistence. Its appeal lies in its simplicity and broad applicability.
10.1 Macroeconomic time series
In macroeconomics, variables such as output, inflation, and interest rates often display temporal dependence. Regression models for these series may generate autocorrelated residuals if important lagged effects are omitted. The Breusch-Godfrey test helps detect such problems before results are interpreted.
10.2 Financial econometrics
Financial return and volatility models frequently involve time dependence in residual behavior. Although specialized models are often used, the Breusch-Godfrey test can still serve as a diagnostic in simpler regressions. It helps determine whether the error structure is sufficiently captured by the chosen specification.
10.3 Applied regression modeling
Beyond economics, the test appears in any setting where data are ordered and residual dependence is plausible. Researchers may use it in epidemiology, transportation studies, or engineering applications involving repeated measurements over time. In these contexts, it functions as a general check on model adequacy.
11 Extensions and variants
The basic logic of the test has been extended in several ways to address richer forms of dependence and different data structures.
11.1 Higher-order autocorrelation
The standard procedure naturally extends to autocorrelation of order greater than one by including more lagged residuals in the auxiliary regression. This allows the analyst to test for dependence that persists across several periods. Higher-order versions are useful when residual patterns are not well described by a simple first-lag process.
11.2 Heteroskedasticity-robust approaches
In some applications, serial correlation may coexist with changing error variance. Robust variants and related corrections can be used when heteroskedasticity is also a concern. These approaches aim to improve inference while preserving the diagnostic role of the original test.
11.3 Panel-data adaptations
Panel-data versions adapt the core idea to repeated observations on multiple units over time. Such methods account for within-unit dependence and the structure of longitudinal data. They are especially useful when the data combine cross-sectional and temporal dimensions, and when standard time-series assumptions do not directly apply.