Autocorrelation, the correlation of a signal with a delayed copy of itself, is a critical diagnostic concern across time series analysis, econometrics, and signal processing. When present in regression residuals, it violates the assumption of independence underlying many statistical models, potentially invalidating standard errors and t-statistics. Detecting and quantifying this dependency is therefore essential for robust inference and accurate forecasting. A robust suite of tests for autocorrelation exists, each designed for specific data characteristics, model structures, and types of dependence.
Understanding the Nature of Autocorrelation
Before selecting a testing methodology, it is vital to clarify the form of autocorrelation under investigation. The most common form is first-order serial correlation, where the current error term depends linearly on the previous period's error. Higher-order autoregressive structures involve dependencies on multiple past lags. Furthermore, the context dictates the appropriate test; one must distinguish between autocorrelation in the raw time series and, more critically, in the residuals of a fitted regression model. The latter scenario is the primary focus of diagnostic testing, as it indicates model misspecification in capturing dynamic relationships.
The Durbin-Watson Statistic: The Classic Frontier
Arguably the most recognized test for autocorrelation is the Durbin-Watson statistic, specifically designed for detecting first-order autocorrelation in the residuals of an ordinary least squares (OLS) regression. The statistic ranges from 0 to 4, with a value of 2 indicating no autocorrelation. Values significantly below 2 suggest positive autocorrelation, while values above 2 point to negative autocorrelation. Despite its popularity, the test has limitations, including an indeterminate region in its distribution and an inability to directly test for higher-order autoregressive processes.
Limitations and the Durbin's h Test
A significant drawback of the Durbin-Watson test is its invalidity when the model includes lagged dependent variables as regressors, a common scenario in dynamic models. In such cases, the test statistic tends to cluster around zero, rendering interpretation impossible. To address this specific issue, economists rely on Durbin's h test. This test adjusts the standard Durbin-Watson statistic by approximating the distribution of the statistic under the null hypothesis of no autocorrelation, providing a valid t-test for the significance of the lagged dependent variable.
Testing Higher-Order and General Dependence
For scenarios involving higher-order autocorrelation or the need to test multiple lags simultaneously, the Breusch-Godfrey serial correlation LM test stands as the most flexible and powerful general-purpose tool. Unlike the Durbin-Watson test, it is valid in models with lagged dependent variables and can assess autocorrelation up to a specified order p. The test involves regressing the residuals on the original regressors and lagged residuals, with the test statistic following a chi-squared distribution based on the explained sum of squares from this auxiliary regression.
Parsimony and the Ljung-Box Q-Test
When the primary goal is to determine whether any autocorrelation exists in the raw time series itself—rather than in the residuals of a specific model—the Ljung-Box Q-test is the standard instrument. This test examines the null hypothesis that the data are independently distributed (i.e., all autocorrelations up to a certain lag m are zero) against the alternative that at least one is non-zero. It is particularly valuable for checking the adequacy of white noise models in time series forecasting, ensuring that the model has successfully extracted all relevant linear information from the data.
