Economics•Chapter 6•4 min read•Updated September 27, 2026

Econometrics — Stationarity and Spurious Regression

O
OiyoContributor
6/6

Econometrics — Before regressing over time, ask whether the mean stays put

Cross-sectional OLS commonly assumes observations are close to independent. When this year’s inflation depends on last year’s, that assumption fails. Chapter 5 concerns individual heterogeneity in panels; this chapter concerns persistence over time.

1. Nonstationarity can inflate a t-statistic

Weak stationarity means that the mean and variance do not change over time and that the covariance between two observations depends on their lag, not their calendar dates. A series that becomes stationary after removing a deterministic trend differs from a unit-root series in which shocks accumulate permanently. The random walk Yt=Yt−1+εtY_t=Y_{t-1}+\varepsilon_t is a standard example of the latter. Regressing the levels of two independent random walks can produce apparently large R2R^2 and conventional tt-statistics, especially as the sample grows. This is a spurious regression.

Unit root
Yt=ρYt−1+εt,ρ=1Y_t = \rho Y_{t-1} + ε_t,\quad \rho=1
When ρ=1, shocks accumulate. The first difference ΔYₜ=εₜ may be stationary. Failure to reject a test is not proof of a unit root.

As a hypothetical example, regressing 100 observations of two independent random walks might yield t=4.2t=4.2 and R2=38%R^2=38\% despite there being no relationship. If the differenced regression yields t=0.4t=0.4, the original statistic was an illusion created by the trends.

The usual null hypothesis of an augmented Dickey–Fuller (ADF) test is that a unit root is present. Failure to reject does not prove it. Whether to include an intercept and deterministic trend, and how many lagged differences are needed to address residual autocorrelation, depends on the data and model. Short samples and structural breaks can make the test weak.

2. Cointegration is a long-run relation that indiscriminate differencing can erase

Two I(1)I(1) series YtY_t and XtX_t are cointegrated if some linear combination zt=Yt−α−βXtz_t=Y_t-\alpha-\beta X_t is I(0)I(0). Their levels may then contain a long-run statistical relation that indiscriminate differencing would discard. But residual stationarity must be tested using cointegration-specific critical values; ordinary level-OLS tt-statistics cannot be interpreted as if this were a stationary regression. Cointegration alone does not identify a causal effect of XX on YY.

An error-correction model relates last period’s deviation from the long-run relation to this period’s change. For example, if ΔYt=λzt−1+γΔXt+et\Delta Y_t=\lambda z_{t-1}+\gamma\Delta X_t+e_t, λ=−0.2\lambda=-0.2, and zt−1=10z_{t-1}=10, then this period’s change in YY is −2-2 when the other terms are zero. This is a hypothetical calculation of adjustment speed, not an estimate for a particular economy. Check the sign and stability of λ\lambda and diagnose the residuals.

Terms that are easily overinterpreted
TermWhat it establishesWhat it does not establish
Granger causalityPast X helps predict YAn intervention or causal identification
Failure to reject a unit rootThe test could not reject ρ=1Proof of a unit root; power may be low
CointegrationA stationary long-run residualA relation immune to structural breaks

Granger causality asks whether past XX helps predict YY. It does not establish that a policy changing XX would change YY. A leading indicator and a cause are not synonyms.

The unpredictability of asset prices also depends on the information set available to investors, discussed in Financial Economics, Chapter 3. Econometrics first asks whether the regression itself is spurious.

Check your understanding

Question. Annual consumption and GDP rise together in 50 observations, and the conventional level-regression statistic is t=6t=6. Both series might have unit roots. A differenced regression yields t=1.1t=1.1, while a residual-based cointegration test also fails to reject its null. What can you conclude?

Solution. The level statistic alone supports neither a relationship nor a causal effect. In this example the differenced result gives weak evidence of a short-run linear association. Failure to reject the cointegration test’s null is not proof that cointegration is absent. Examine the trend and lag specification, sample length, and structural breaks. If an appropriate test later supports cointegration, an error-correction model can separate the long-run relation from short-run adjustment, but a causal interpretation still requires a separate identification design.

References

O

Oiyo

Editorial Desk

The OIYO editorial desk researches money, law, lifestyle, and self-understanding topics against primary sources and public statistics. Every piece carries source notes and is reviewed on a regular cycle for accuracy and usefulness.