How the Unit Root Test Prevents Spurious Correlation in Non-Stationary Time Series

Before an econometrician can trust a relationship between two variables over time, they must first prove the underlying data does not wander endlessly. The tool they deploy to make this decision is the unit root test, which determines whether a time series has a constant mean and variance, or if it must be mathematically differenced before any regression model is run.[2]

The danger of skipping this step is a mathematical trap known as spurious regression. When two completely independent datasets—such as cumulative rainfall in London and the closing price of the S&P 500—both trend upward over time, standard regression models will falsely declare them highly correlated. The model interprets the shared upward trajectory as a causal relationship, generating artificially high confidence metrics that deceive the analyst.[5][6]

This phenomenon was formally diagnosed in 1974 by statisticians Clive Granger and Paul Newbold. Through Monte Carlo simulations of 120 independent random walks, they demonstrated that standard Ordinary Least Squares (OLS) regressions would report a statistically significant relationship 76% of the time, despite the variables having absolutely zero actual connection. They found that these false regressions consistently produced an R-squared value greater than 0.90 and t-statistics well above the standard 2.0 threshold.[5]

The root cause of this deception is non-stationarity. A stationary time series reverts to a constant mean over time, with a stable variance that makes future states predictable. A non-stationary series, however, possesses a "stochastic trend" where each new value is simply the previous value plus a random shock. Because these shocks accumulate permanently, the variance of the series grows infinitely over time, violating the core assumptions of classical regression.[1][3]

In mathematical terms, this accumulating behavior is called a "unit root." The name derives from the characteristic equation of the autoregressive model; if one of the roots of this equation equals exactly 1.0, the series is non-stationary. As the WallStreetMojo documentation explains, detecting this condition is the primary gateway in time-series forecasting, ensuring that "the statistical properties of the system do not change over time."[4][7]

To catch a unit root before it ruins a model, analysts rely on the Augmented Dickey-Fuller (ADF) test, developed in 1979. The ADF test sets a null hypothesis that a unit root is present. The analyst calculates a test statistic and compares it to a critical value; if the statistic is more negative than the critical value (typically at a 0.05 alpha level), the null hypothesis is rejected, confirming the data is stationary.[7]

The "Augmented" portion of the ADF test allows it to handle more complex time-series dynamics by including lagged differences of the data. This ensures that serial correlation in the error terms does not artificially inflate the test statistic. However, selecting the correct number of lags is crucial; too few lags leave the test biased, while too many drain its statistical power.[2]

Because the ADF test assumes the error terms have a constant variance, it can struggle with real-world financial data that exhibits volatility clustering. In 1988, Peter Phillips and Pierre Perron introduced the Phillips-Perron (PP) test to solve this. Instead of adding lagged differences, the PP test applies a non-parametric correction to the standard error of the test statistic, making it robust to unspecified autocorrelation and heteroskedasticity.[1]

Both the ADF and PP tests suffer from low statistical power in small samples, meaning they often fail to reject the null hypothesis when a series is actually stationary but highly persistent. To counter this, econometricians use the KPSS test, which flips the framework entirely. The KPSS test sets its null hypothesis to assume the data is stationary. By running the ADF and KPSS tests concurrently, analysts can cross-verify the results and isolate series that are merely trend-stationary.[2]

When a unit root test confirms the data is non-stationary, the analyst cannot proceed with a standard regression. Instead, they must transform the data through "first differencing"—subtracting the previous observation from the current one. As MathWorks documentation notes, "A unit root process is a data-generating process whose first difference is stationary."[2][3]

Differencing shifts the analysis from absolute values to rates of change. Instead of regressing the total GDP against total carbon emissions, the econometrician regresses the month-over-month change in GDP against the month-over-month change in emissions. This strips away the shared stochastic trend, revealing whether the two variables actually move together on a granular level.[6]

While unit root tests are foundational to modern econometrics, they are not infallible. Structural breaks—sudden shifts in a time series caused by external shocks like a pandemic or a policy change—can trick both the ADF and PP tests into falsely identifying a unit root. In these edge cases, analysts must deploy specialized algorithms, such as the Zivot-Andrews test, which account for structural breaks while evaluating stationarity.[1]