Augmented Dickey–Fuller Test

A regression-based time-series hypothesis test whose null is a unit root, augmenting the Dickey–Fuller equation with lagged differences to absorb serial correlation under a declared deterministic specification and lag order.

Version

v1 · 2026-09-28 · History

Domain-specific #

8073

Domain group

Formal Sciences

Origin domain

Experimental Design & Statistics

Subdomains

Time Series Analysis, Unit Root Testing → Experimental Design & Statistics

Aliases

ADF test, Augmented Dickey Fuller test

Core Idea

ADF turns a question about stochastic persistence into a carefully specified autoregression. The coefficient on the lagged level carries the unit-root null, while lagged differences protect the regression from residual dynamics that the basic Dickey–Fuller equation cannot absorb.

A result is inseparable from its intercept, trend, lag order, sample, and critical-value convention. Failure to reject is not evidence that a unit root certainly exists, and rejection supports only the stated stationary or trend-stationary alternative.

How would you explain it like I'm…

Does It Drift Back?

Some things, after a bump, drift back toward their usual spot, like a swing settling down. Others just wander wherever the bumps push them, like a balloon blowing around. This test looks at a long record of something to see if there is good proof that it drifts back. If it can't find that proof, that still doesn't prove it is a wanderer.

The Wander-or-Return Test

The Augmented Dickey–Fuller test is used on numbers measured over time, like prices each day. It asks whether the numbers have a pull back toward a usual level or trend, or whether they wander with nothing pulling them back, where every bump sticks around forever. The test starts by assuming the numbers wander with no pull back, and looks for enough evidence to say otherwise. It uses a formula that also includes the last few changes, so that ordinary ups and downs from step to step don't fool it. If the test can't find evidence of a pull back, that doesn't prove there isn't one. The answer depends on choices like whether you include a trend and how many past changes you use.

Unit-Root Testing With Lags

The Augmented Dickey–Fuller (ADF) test checks whether a time series has a "unit root," meaning shocks to it persist forever, like a random walk, versus being stationary, meaning it tends to return to a mean or trend. It runs a regression of the change in the series on its previous level, plus lagged changes. The coefficient on the lagged level carries the test: the null hypothesis is a unit root, and a significantly negative coefficient points toward stationarity. The lagged differences are the "augmentation"; they soak up extra short-term dynamics that the basic Dickey–Fuller regression can't handle, so the errors behave properly. Results depend on whether you include an intercept or trend, how many lags you use, the sample and which critical values apply. Failing to reject doesn't prove a unit root exists, and rejecting supports only the specific stationary or trend-stationary alternative you tested.

 

The Augmented Dickey–Fuller test converts a question about stochastic persistence into a specified autoregression, typically Δy_t = α + βt + γ y_{t−1} + Σ_{i=1}^{p} δ_i Δy_{t−i} + ε_t. The unit-root null is carried by the coefficient on the lagged level, γ = 0, against the alternative γ < 0, which indicates a stationary or trend-stationary process depending on the deterministic terms included. The lagged differences augment the basic Dickey–Fuller equation so that residual serial correlation from short-run dynamics does not invalidate the test. Under the null, the test statistic does not follow the usual t distribution, so nonstandard Dickey–Fuller critical values that depend on the deterministic specification are used. A result is therefore inseparable from its intercept and trend choice, lag order p, sample and critical-value convention. Failing to reject is not evidence that a unit root certainly exists, and rejection supports only the specific stationary or trend-stationary alternative that was stated.

Structural Signature

Sig role-phrases:

• Time-series sample — Supplies ordered observations whose persistence is assessed. It is data carrier. Counterfactual: Cross-sectional ordering cannot establish a temporal unit root.
• Lagged level — Carries the coefficient that distinguishes unit-root from stationary dynamics. It is hypothesis term. Counterfactual: Removing it eliminates the Dickey–Fuller restriction.
• Lagged differences — Model short-run autocorrelation in regression errors. It is augmentation. Counterfactual: Too few lags invalidate residual assumptions; too many reduce power.
• Deterministic terms — Specify no constant, drift, or time-trend cases. It is model frame. Counterfactual: Critical values and alternative change with this choice.
• Test statistic and critical law — Convert the fitted level coefficient into a rejection decision. It is inference rule. Counterfactual: A standard Student t table gives the wrong calibration.
• Lag-selection rule — Fixes augmentation without outcome-driven searching. It is tuning rule. Counterfactual: Post hoc lag changes distort reported size.

What It Is Not

• Failure to reject does not prove a unit root.
• Rejection does not guarantee every form of stationarity.
• Ordinary t critical values do not calibrate the null.
• Lag and trend choices are not interchangeable preprocessing details.
• Closest near-miss. A Phillips–Perron test retains a simpler regression and adjusts inference nonparametrically for serial correlation; ADF handles it parametrically through augmentation.

Scope of Application

• Macroeconomics. Tests persistence before modeling levels or differences.
• Finance. Assesses unit-root behavior in prices, rates, or spreads.
• Forecasting. Guides transformations while preserving uncertainty.
• Model diagnostics. Compares persistence conclusions across specifications and break tests.

Clarity

Report series definition, sampling frequency, transformation, sample window, missing-data treatment, deterministic terms, maximum and selected lag, selection rule, statistic, critical values or p-value method, residual checks, and alternative hypothesis.

Manages Complexity

One statistic compresses choices about deterministic structure, short-run dynamics, breaks, and finite-sample calibration. The test is useful because those choices are explicit, not because it makes stationarity a one-button fact.

Abstract Reasoning

1. Plot and define the series, transformations, sample, and plausible deterministic structure.
2. Choose the no-constant, intercept, or intercept-plus-trend regression before reading the outcome.
3. Select enough lagged differences by a declared criterion and diagnostic checks.
4. Estimate the level coefficient statistic and use matching Dickey–Fuller critical values.
5. Interpret rejection or nonrejection conditionally, then test robustness to lag, breaks, and complementary procedures.

Knowledge Transfer

The regression logic transfers among regularly sampled series only when deterministic terms, dependence, breaks, and critical values are rebuilt for the new data. The generic lesson—test persistence under an explicit null—travels farther than ADF's exact time-series machinery.

Examples

Canonical

For a quarterly series, an analyst preregisters an intercept-plus-trend ADF regression, selects lag order by a stated criterion, checks residual autocorrelation, and compares the statistic with matching unit-root critical values.

Mapped back: series → quarterly observations; hypothesis → unit root; augmentation → lagged differences; frame → intercept and trend; decision → ADF critical law.

Applied / In Practice

Regressing the level on time and applying an ordinary t threshold to its slope is not an ADF test because no unit-root regression or nonstandard reference distribution is used.

Mapped back: model → trend regression; lagged level restriction → absent; critical law → ordinary t; verdict → not ADF.

Structural Tensions

T1 — Serial-Correlation Control versus Finite-Sample Power. Additional difference lags can whiten errors while consuming degrees of freedom and weakening rejection.

Diagnostic: Does the chosen order pass residual checks without gratuitous augmentation?

T2 — Simple Unit-Root Null versus Structural Change. A level shift or trend break can mimic persistent nonstationarity and reduce the test's power.

Diagnostic: Were breaks and regime changes examined with an appropriate alternative test?

Structural–Framed Character

Augmented Dickey–Fuller Test is structural as a unit-root restriction in an augmented difference regression and framed by time-series inference. Calibration depends on specification rather than ordinary regression tables.

Structural Core vs. Domain Accent

The reusable core is null-restricted model comparison under nuisance dynamics. Econometrics contributes stochastic trends, deterministic terms, lag selection, and nonstandard asymptotics.

Instantiates / Related Primes

This entry is a kind of Hypothesis Testing (Null vs. Alternative).

• Approved unparented root. The live catalog lacks a test node whose identity entails ADF's unit-root null, augmented difference terms, and Dickey–Fuller critical law.

• Related — Dickey–Fuller, Phillips–Perron, KPSS, and stationarity. They supply the base test, alternate serial-correlation correction, reversed null, and target property.

Relationships to Other Abstractions

Current abstraction Augmented Dickey–Fuller Test Domain-specific

Parents (1) — more general patterns this builds on

• Augmented Dickey–Fuller Test is a kind of Hypothesis Testing (Null vs. Alternative) Prime

  Augmented Dickey–Fuller Test is a strict kind of Hypothesis Testing (Null vs. Alternative): it tests a unit-root null against stationarity alternatives using an augmented regression.

Hierarchy paths (5) — routes to 5 parentless roots

• Augmented Dickey–Fuller Test → Hypothesis Testing (Null vs. Alternative) → Statistical Inference → Inductive Reasoning

Show alternative paths (4)

Neighborhood in Abstraction Space

Augmented Dickey–Fuller Test sits in a crowded region of the domain-specific corpus (36^th percentile for distinctiveness): several abstractions share nearly its structure, so a description that fits it tends to fit its neighbors too.

Family — Statistical Hypothesis Tests & Diagnostics (9 abstractions)

Nearest neighbors

• Ljung–Box Test — 0.91
• First-Hitting-Time Model — 0.88
• Ecosystem Model — 0.88
• CUSUM — 0.87
• Welfare Cost of Business Cycles — 0.87

Computed from structural-signature embeddings · 2026-10-08

Not to Be Confused With

• Dickey–Fuller test. Tell: Omits the extra lagged differences in its basic form.
• Phillips–Perron test. Tell: Uses nonparametric correction rather than autoregressive augmentation.
• KPSS test. Tell: Takes stationarity or trend-stationarity as its null.
• Cointegration test. Tell: Studies stationary combinations among multiple nonstationary series.

References

• Frozen Wikipedia discovery revision: https://en.wikipedia.org/wiki/Augmented_Dickey%E2%80%93Fuller_test (revision 1328097950).
• Preserved source candidate: http://econterms.com/glossary.cgi?action%3D++Search++%26query%3Daugmented+dickey-fuller
• Preserved source candidate: https://web.archive.org/web/20090302082540/http://econterms.com/glossary.cgi?action=++Search++&query=augmented+dickey-fuller
• Preserved source candidate: http://www.nber.org/papers/t0130.pdf
• Preserved source candidate: http://www.inside-r.org/packages/cran/forecast/docs/ndiffs
• Preserved source candidate: https://web.archive.org/web/20160717021256/http://www.inside-r.org/packages/cran/forecast/docs/ndiffs
• Preserved source candidate: http://finzi.psych.upenn.edu/R/library/tseries/html/adf.test.html
• Preserved source candidate: https://fabian-kostadinov.github.io/2015/01/27/comparing-adf-test-functions-in-r/
• Preserved source candidate: https://cran.r-project.org/web/packages/urca/urca.pdf

The frozen Wikipedia revision is discovery provenance. The retained source set was reviewed for identity, formal or operational relation, and scope. The encyclopedia's structural synthesis is bounded to those claims; a thin authority surface is recorded as a nonblocking source-strengthening repair rather than concealed.