Statistics Glossary → Serial correlation tests (Breusch-Godfrey)
Serial correlation tests (Breusch-Godfrey)concept
Tests of whether successive error terms over time or sequence are correlated with one another; the Breusch-Godfrey test is one of the most general of these.
Regression models usually assume that the error terms are independent of one another. In time-series and panel data, however, one period's error is often related to the next period's; this is serial correlation. In the Breusch-Godfrey test the model residuals are regressed, in an auxiliary regression, on the predictors and on p lagged values of the residuals; under the null hypothesis the n·R² statistic follows a chi-square distribution with p degrees of freedom. Unlike the Durbin-Watson test, it can test for higher-order serial correlation and remains valid when the model includes a lagged dependent variable.
Serial correlation does not bias the coefficient estimates (as long as exogeneity holds), but it makes ordinary standard errors misleading; with positive serial correlation they are usually too small and p values too optimistic. If the model includes a lagged dependent variable, the coefficients can also become inconsistent. Remedies include Newey-West or clustered (by unit, in panels) robust standard errors, adding lagged terms to the model, or generalized least squares. In panel data, Wooldridge's test, based on the residuals of first differences, is a dedicated option.
An everyday example
You build a model explaining an ice-cream seller's daily sales by temperature. Throughout a hot week the model keeps underpredicting sales, and during the cool week that follows it keeps overpredicting. The errors do not occur at random but follow one another in waves: that is serial correlation, and it leads us to trust the model's standard errors more than we should.
In MerQur
In Panel Data Regression the 'Serial correlation and BG test lag order' additional tests provide the Breusch-Godfrey/Wooldridge test and Wooldridge's first-difference test. In VAR, the residual diagnostics include a Portmanteau test of autocorrelation in the residuals.
Analyses where this concept appears
Panel Data RegressionVAR (Vector Autoregression)
Related concepts
Autocorrelation (ACF) and partial autocorrelation (PACF) functionsLjung-Box (portmanteau) test