Does rejection prove every kind of stationarity?

No; the conclusion is limited to the specified unit-root comparison.

Does PP require independent regression errors?

Its correction allows certain dependence, subject to conditions.

Can PP and ADF disagree?

Yes, especially with finite samples and different treatment of dependence.

AI Glossary / Statistics and probability / Intermediate

Phillips - Perron test

Definition

The Phillips - Perron test is a time-series test whose null hypothesis is that a series has a unit root. It modifies a Dickey - Fuller-type statistic using an estimate of the long-run variance of regression errors, accommodating certain forms of serial correlation and heteroskedasticity under its regularity conditions. The alternative and reference distribution depend on the deterministic terms included in the test.

What it means

A unit root describes a form of persistence in which shocks do not decay in the same way as they do in a stationary autoregressive model. Whether this is an appropriate description affects forecasting, differencing, and interpretation of time-series relationships.
The test provides evidence about a specified statistical model. Failing to reject a unit root does not prove that one exists, and rejecting it does not establish that the series is free of structural changes, seasonal complications, or all other forms of nonstationarity.

How it works

Inspect the series, its time spacing, and the substantive context. Decide whether the test regression should contain no deterministic terms, a constant, or a constant and linear trend. These choices represent different possible behaviors under the alternative and affect the critical values.
Estimate a regression of the current observation on its previous value and the selected deterministic terms. The unit-root null corresponds to a coefficient of one on the lagged level; an equivalent differenced formulation tests a zero coefficient on the lagged level.
Calculate residual quantities and estimate the long-run variance using an appropriate heteroskedasticity-and-autocorrelation-consistent procedure. The bandwidth determines how many residual autocovariances contribute and can materially influence the correction.
Phillips - Perron differs from the augmented Dickey - Fuller approach. ADF adds lagged differences to its regression to address serial dependence; PP adjusts the test statistic using a long-run variance estimate. This distinction does not mean one procedure dominates in every sample.
Common PP variants include a corrected t-type statistic and a corrected coefficient-type statistic. Software may call them tau and rho. Their numerical scales differ, so a statistic must be paired with the correct reference calculation.
Compare the result with unit-root critical values or a suitable p-value approximation for the chosen deterministic specification. Ordinary Student-t critical values are inappropriate simply because one variant begins with a regression t statistic.
Assess sensitivity to defensible bandwidth and trend choices, and consider the limitations of the sample. Strong finite-sample dependence, near-unit-root behavior, breaks, and short records can complicate interpretation. Selecting settings only because they produce rejection undermines the analysis.
Use complementary diagnostics where appropriate. A stationarity-null test such as KPSS asks a different question, and agreement or disagreement between tests needs interpretation rather than a majority vote. Forecast validation remains necessary after choosing any transformation.

Three practical examples

Forecasting: analysts investigate whether a measured level series may require differencing before fitting a model.
AI feature preparation: a team examines persistence before using time-series levels in a predictive pipeline.
Operations analysis: researchers assess whether shocks in a monitoring series appear compatible with a unit-root model.

Think of it this way

Think of comparing two kinds of disturbance: a displacement that gradually fades and a displacement that becomes part of the series' ongoing level. The test looks for evidence against a particular persistent-shock model, while allowing the regression errors to have some dependence.

Worked example

Consider stipulated regression summaries solely to illustrate the coefficient-type PP correction. Let the effective regression sample size be n = 100, the lagged-level coefficient be 0.94, its conventional OLS standard error be 0.02, and the residual variance with regression degrees-of-freedom adjustment be s squared = 1. Suppose there are two fitted regression coefficients, giving the unadjusted residual second moment gamma0 = 98/100 = 0.98. Let the estimated long-run variance be 1.44.
Using the coefficient-type convention documented in arch, the statistic is n(rho minus 1) minus one-half times [n squared × SE squared / s squared] times (long-run variance minus gamma0).
The first term is 100 × (0.94 minus 1) = -6. The adjustment is 0.5 × 10000 × 0.0004 × (1.44 minus 0.98) = 0.92. The resulting PP rho statistic is -6.92.
These are invented summary inputs, not fitted observations. The calculation demonstrates the correction only. It does not provide a p-value or a rejection decision; those require the correct coefficient-type unit-root reference and deterministic specification.

Illustrative case study

Fictional case: Pine Forecasting studies a long-running operational demand series. The analysts record the trend specification and bandwidth before evaluating PP and other diagnostics. When plausible specifications disagree, they retain that uncertainty, inspect known operational changes, and compare forecast performance across candidate transformations rather than announcing that one test has proven the series stationary.

Common misconceptions and mistakes

Do not reverse the null hypothesis: PP starts with a unit root, whereas some other tests start with stationarity.
Do not read a nonsignificant result as proof of the null. Unit-root tests can have limited power against persistent alternatives.
The bandwidth in the long-run variance estimator is not the same as the number of lagged-difference regressors in ADF.
Do not use an ordinary regression p-value for testing a coefficient of one as though it were the PP result. The null distribution and correction are specialized.

Business use

The test supports time-series modeling decisions by making persistence assumptions explicit. Business reporting should identify the series transformation, sample period, deterministic terms, bandwidth, statistic variant, result, and sensitivity. It should connect those diagnostics to forecast evaluation and the operational cost of modeling choices.

Check your understanding

Why is the illustrative statistic -6.92 instead of -6?

Answer: The long-run variance correction subtracts 0.92.

Why is no ordinary t-test decision reported?

Answer: The example uses the coefficient-type PP statistic and requires its specialized unit-root reference.

Related terms

Stationarity
Differencing
unit root
Dickey–Fuller test
augmented Dickey–Fuller test
Newey–West estimator
long-run variance
KPSS

Learn next

Learn next how deterministic trends differ from stochastic trends, how bandwidth choices affect long-run variance estimates, and why structural breaks require additional analysis.

Sources

Sources: arch, PhillipsPerron documentation, https://arch.readthedocs.io/en/latest/unitroot/generated/arch.unitroot.PhillipsPerron.html ; arch implementation source, https://arch.readthedocs.io/en/latest/_modules/arch/unitroot/unitroot.html . These primary developer materials document the null, variants, long-run correction, and reference calculations, and identify Phillips and Perron's 1988 paper. The summary-input calculation and fictional case are original.

Previous: Phase of Spectrum Next: Phone