Why Panel Cointegration Tests Matter
When two variables are nonstationary, a regression in levels can look convincing even if it reflects unrelated trends. Cointegration is a more specific possibility: a linear combination of nonstationary variables is stationary, so their levels do not drift apart without bound in the long run. Panel cointegration tests ask whether this relationship is supported across multiple units, such as regions, firms, or countries, while allowing some forms of heterogeneity across units.
This tutorial considers tests that assume cross-sectional independence: after accounting for the specified deterministic terms, shocks for one unit are independent of shocks for another. Serial dependence within a unit can still be allowed through lagged differences or long-run variance adjustments. If common shocks induce cross-unit dependence, the reference distributions and resulting inferences can be unreliable; the independence assumption is not a minor technical detail.
Here \(I(1)\) means that a series is nonstationary in levels but its first difference is stationary; \(I(0)\) means stationary in levels under the chosen specification. The relevant variables should be assessed as \(I(1)\) before a cointegration test is used. The earlier material on panel unit-root testing covers tests for that prerequisite.
Specify the Long-Run Relation First
A basic heterogeneous panel cointegrating regression is
The intercept \(\alpha_i\), trend coefficient \(\delta_i\), and long-run slope vector \(\beta_i\) may be restricted or allowed to differ across units, depending on the test and specification. A specification without a trend omits \(\delta_i t\); one with an intercept but no trend allows unit-specific level differences without a deterministic linear trend. The residual \(u_{it}\) is the candidate stationary disequilibrium. The null hypothesis of no cointegration says it retains a unit root.
Deterministic terms must reflect the proposed relationship. Omitting a genuine trend can make a stationary trend-adjusted residual appear nonstationary; adding an unnecessary trend can reduce power. Also distinguish a deterministic trend in the cointegrating relation from a time trend in a short-run equation: they have different roles and should not be included mechanically in the same way.
Worked Example: Choosing and Interpreting a Cointegrating Regression
Suppose a hypothetical panel tracks an outcome \(y_{it}\) and a single regressor \(x_{it}\), both measured in index points, for 24 regions over 60 years. Economic reasoning suggests the regions may have different baseline levels but no deterministic linear trend in their long-run relation. Use \(y_{it}=\alpha_i+\beta_i x_{it}+u_{it}\). For one region, suppose the fitted intercept is 4.0 and the fitted slope is 1.25. At \(x_{it}=10\), the fitted level is
If its observed value is 17.1 index points, the fitted residual is \(17.1-16.5=0.6\) index points. This single residual does not establish cointegration. The test uses the residual sequence over time and asks whether it is stationary after allowing for the selected deterministic terms. A trend should be added only if the long-run specification warrants it, not merely because the variables rise over time.
Residual-Based Tests
Residual-based methods first estimate the long-run regression and then test the estimated residuals for a unit root. A generic residual test equation is
The lagged differences accommodate serial correlation in the residual process. Under no cointegration, the residual has a unit root, corresponding to \(\rho_i=0\) in this representation. Evidence for stationarity typically corresponds to a negative \(\rho_i\). The residuals are estimated rather than observed, so the test statistics do not generally have the ordinary single-series ADF reference distribution. Panel procedures use test-specific corrections for the residual estimation, deterministic terms, and dependence over time.
Pedroni-style tests provide two broad ways to combine evidence. Within-dimension (panel) statistics pool information across units and impose a common autoregressive component in the residual test under the alternative. Between-dimension (group-mean) statistics average unit-specific evidence and allow the residual adjustment to differ across units. Both approaches can allow heterogeneous short-run dynamics; their distinguishing feature is the restriction on the residual adjustment and the resulting alternative, not whether every aspect of the panel is identical.
Commonly reported Pedroni statistics include panel \(v\), panel rho, panel Phillips–Perron (PP), panel ADF, group rho, group PP, and group ADF. The panel \(v\) statistic has a right-tailed rejection region; the other listed statistics are generally left-tailed. Reported standardized statistics use corrections and reference distributions appropriate to the chosen deterministic specification. A standard normal approximation may be used for those corrected statistics in suitable large panels, but raw residual ADF statistics must not be compared casually with standard normal critical values.
Worked Example: Reading a Residual-Based Test
Consider hypothetical output from a Pedroni group-ADF test for the 24-region panel, using the intercept-only cointegrating regression above. Suppose the corrected, standardized statistic is \(-2.41\). For a left-tailed standard normal approximation, the p-value is
At \(\alpha=0.05\), \(0.0080<0.05\), so reject the null of no cointegration for the group-mean test. This supports an alternative under which cointegration may hold for some units, with heterogeneous residual adjustment. It does not show that every region is cointegrated, nor does it establish a causal effect of \(x_{it}\) on \(y_{it}\). The conclusion is conditional on the deterministic specification, the test’s assumptions, and cross-sectional independence.
Error-Correction Tests
Error-correction tests examine whether changes in the outcome respond to the previous period’s disequilibrium. A representative conditional error-correction equation is
The expression in parentheses is the lagged deviation from the proposed long-run relation. The constants \(c_i\) and \(d_i(t-1)\) represent the chosen level and trend terms; they may be omitted or restricted to match the intended deterministic specification. The differenced regressors and lagged differences represent short-run dynamics. In this parameterization, \(\phi_i=0\) means that \(y_{it}\) does not adjust to the lagged disequilibrium through this channel. A negative \(\phi_i\) gives a downward adjustment in \(\Delta y_{it}\) when the lagged disequilibrium is positive, but negativity alone does not guarantee convergence.
Westerlund-style error-correction tests test the joint null of no error correction, often expressed as \(H_0:\phi_i=0\) for every unit. The group-mean statistics \(G_\tau\) and \(G_\alpha\) allow unit-specific adjustment and can reject when there is evidence of error correction for at least some units. The panel statistics \(P_\tau\) and \(P_\alpha\) pool evidence under a more restrictive panel alternative in which adjustment is present across units. The \(\tau\) statistics are based on t-type evidence about adjustment; the \(\alpha\) statistics use an adjustment-coefficient form. All four are typically left-tailed. Exact finite-sample behavior and reported p-values depend on the implementation and deterministic choices.
These tests focus directly on adjustment toward a long-run relation rather than first constructing residuals and then applying a residual unit-root test. They still require a correctly specified long-run relation and suitable treatment of short-run dynamics. They do not remove the need to assess cross-sectional independence.
Worked Example: Checking the Sign and Size of Adjustment
In a simplified hypothetical error-correction equation, suppose \(\Delta y_{it}=\phi_i e_{i,t-1}\), where \(e_{i,t-1}\) is the previous disequilibrium. Given \(e_{i,t-1}=2.0\) index points and \(\widehat{\phi}_i=-0.4\), the fitted change attributable to adjustment is
In the simplified scalar recursion with no other changes, \(e_{it}=e_{i,t-1}+\Delta y_{it}=(1+\phi_i)e_{i,t-1}\). Thus the next disequilibrium is \(2.0-0.8=1.2\) index points, or 60% of its prior value. The magnitude declines here because \(|1+\phi_i|=0.6<1\). In this scalar recursion, convergence requires \(|1+\phi_i|<1\), equivalently \(-2<\phi_i<0\). In the full panel model, other regressors and dynamics affect stability, so this scalar check is not a complete stability analysis.
For an illustrative individual adjustment estimate, suppose the standard error is 0.15. Its t-ratio is \(-0.4/0.15=-2.67\). That ratio alone is not a panel cointegration conclusion: a panel procedure combines unit information and uses its own reference distribution. The example shows why the sign, magnitude, specification, and panel test result all matter.
Comparing the Approaches
| Feature | Residual-based tests | Error-correction tests |
|---|---|---|
| Quantity tested | Stationarity of estimated long-run residuals | Adjustment to lagged disequilibrium |
| Typical null | No cointegration; residuals have a unit root | No error correction; adjustment coefficients are zero |
| Heterogeneity | Within-dimension tests pool residual adjustment; group-mean tests allow unit-specific adjustment | Group-mean tests allow heterogeneous adjustment; panel tests use a more restrictive pooled alternative |
| Deterministic terms | Specified in the cointegrating regression and relevant test correction | Specified in the long-run relation and consistently represented in the error-correction equation |
| Main interpretation | Evidence that a long-run combination is stationary | Evidence that deviations from a long-run relation trigger adjustment |
Neither approach is universally preferable. Residual-based tests make the residual-stationarity logic explicit. Error-correction tests make the adjustment mechanism explicit and can be useful when the response to disequilibrium is central to the question. Their alternatives differ, so a group-mean rejection and a pooled-panel rejection do not support identical claims. In either case, cointegration indicates a stable long-run statistical relation, not a causal effect.
Cross-Sectional Independence and Inference
The tests described here rely on cross-sectional independence for their standard panel inference. This means that, conditional on included terms, shocks are not correlated across units. Independence does not rule out serial correlation within a unit; lag augmentation or long-run variance corrections address that different feature. A common macroeconomic shock, shared policy change, or spatial spillover can violate cross-sectional independence even if each unit’s own time series is modeled carefully.
For illustration, suppose hypothetical Westerlund group-\(\tau\) output is a corrected standardized statistic of \(-2.36\), with a left-tailed normal approximation. Then
At the 5% level, this rejects no error correction under that test’s group-mean alternative. It supports adjustment in at least some units under the maintained assumptions; it does not prove that all units share one long-run slope. If cross-unit dependence is present, this p-value need not have its stated calibration. The earlier material on testing for cross-sectional dependence and common-factor error models is relevant when deciding whether independence is credible.
Common Pitfalls
- Using the wrong deterministic terms. A regression with an intercept but no trend does not test the same specification as one with unit-specific trends. Choose terms based on the long-run model and interpret the test accordingly.
- Reading a group-mean rejection as universal evidence. Group tests can be driven by a subset of units. They do not by themselves establish cointegration for every unit.
- Assuming any negative adjustment coefficient guarantees convergence. In the scalar recursion \(e_{it}=(1+\phi_i)e_{i,t-1}\), the magnitude declines only if \(|1+\phi_i|<1\), or \(-2<\phi_i<0\). For example, \(\phi_i=-3\) gives \(e_{it}=-2e_{i,t-1}\), which increases the magnitude rather than reducing it.
- Treating cointegration as causation. A stationary long-run combination describes a persistent statistical relationship; it does not identify a causal effect.
- Ignoring cross-unit dependence. Common shocks can invalidate the independence-based reference distribution, even when the test statistic has been computed correctly.
Check Your Understanding
Use the model assumptions and test alternatives, not the statistic’s sign alone, to answer these questions.
- What does the null hypothesis of a residual-based panel cointegration test assert about the estimated long-run residuals?
- How does a group-mean alternative differ from a pooled panel alternative in the scope of its conclusion?
- Why can omitting a deterministic trend alter the interpretation of a cointegration test?
- In the scalar recursion \(e_{it}=(1+\phi_i)e_{i,t-1}\), what condition on \(\phi_i\) ensures the disequilibrium magnitude declines?
- Why can common shocks make a reported independence-based panel p-value unreliable?