Statistics Glossary → Unit root tests (ADF, Phillips-Perron)
Unit root tests (ADF, Phillips-Perron)concept
Tests of whether a time series contains a unit root, i.e. whether it is a non-stationary process (e.g. a random walk) in which shocks have permanent effects.
In a series with a unit root (e.g. yₜ = yₜ₋₁ + εₜ) the effect of a shock never fades, the variance grows over time, and the series does not return to a fixed mean. The augmented Dickey-Fuller (ADF) test fits the regression Δyₜ = α + βt + γyₜ₋₁ + Σ δⱼΔyₜ₋ⱼ + εₜ; the null hypothesis is γ = 0 (a unit root is present). The lagged difference terms absorb autocorrelation in the errors, and their number can be chosen with criteria such as AIC or SIC. The test statistic follows not the ordinary t distribution but the Dickey-Fuller distribution, and the critical values (MacKinnon) depend on whether a constant and trend are included. The Phillips-Perron test fits the same regression without lagged differences and handles autocorrelation and heteroscedasticity with a nonparametric Newey-West type correction.
Unit root tests have low power, especially for short series and for stationary series with a root very close to 1, so they are recommended alongside the KPSS test, whose null hypothesis is stationarity. A wrong choice of constant and trend terms changes the result, and structural breaks can make a stationary series look as if it had a unit root. If a unit root is found, the series is usually made stationary by differencing; but if several unit-root series move together, cointegration analysis is needed. In panel data, the tests for individual units are combined (Im-Pesaran-Shin, Maddala-Wu Fisher).
An everyday example
The water level of a lake changes a little every day. If the level fluctuates around an average from season to season, it is stationary and the effect of a dry week fades over time. But if each day's level drifts by adding a random change to the previous day's (a unit root), the drop caused by the dry week is permanent and the lake does not return to its old level by itself. The ADF test tries to tell these two situations apart.
In MerQur
Available in panel data regression as 'Panel unit root tests' with the 'Unit root: exogenous terms', 'lag selection' and 'lag / maximum lag' options, and in the Time Series analysis as the ADF and KPSS tests plus the 'Phillips-Perron unit-root test' and 'Differencing recommendation' outputs. In cointegration analysis it also appears as the Engle-Granger residual ADF test.
Analyses where this concept appears
Panel Data RegressionTime SeriesCointegration (Johansen / Engle-Granger)