MerQuR+Statistics Glossary

Statistics Glossary → Cointegration (Johansen / Engle-Granger)

Cointegration (Johansen / Engle-Granger)analysis

A method (Johansen, Engle-Granger) that tests whether two or more series, each non-stationary and drifting freely over time, move together in the long run, that is, whether there is an equilibrium relationship between them.

Concept

Many economic series such as prices, income and consumption are non-stationary: they drift in one direction over time and shocks leave permanent marks (a unit root, I(1)). Relating two such series directly by regression carries the danger of 'spurious regression': a high R² and significant coefficients can appear when there is no link at all. Yet some series, though each drifts, cannot move far apart; their difference, or some linear combination of them, stays stationary. This is called cointegration, and it shows a long-run equilibrium between the series.

The Engle-Granger method works in two steps: first a regression between the series is fitted, then a unit-root test checks whether the residuals of that regression are stationary; if they are, the series are cointegrated. The Johansen method instead treats all the series together within a VAR model and determines how many independent equilibrium relationships (the cointegration rank r) exist using the trace and maximum eigenvalue tests. When cointegration is found, the result is turned into an error correction model (VECM): the adjustment coefficients (α) show how quickly each series returns to equilibrium after a departure.

The method assumes the series are integrated of the same order (usually I(1)), so unit-root tests should be run first. The results are sensitive to the VAR lag order and to how the deterministic terms (constant, trend) are specified. Engle-Granger can find only a single equilibrium relationship and has low power; with more than two series the Johansen method is preferred.

When to use it

  • To test whether there is a long-run equilibrium relationship between non-stationary economic or financial series (prices, interest rates, exchange rates, income and consumption).
  • For two series and a single relationship, use Engle-Granger; with three or more series, or if more than one equilibrium relationship is possible, use the Johansen method.
  • If the series are already stationary, cointegration is not needed and a VAR is fitted directly; if they are non-stationary but not cointegrated, use a VAR on the differenced series.

An everyday example

Suppose the prices of flour and bread in a province have been recorded every month for fifteen years. Both rise steadily with inflation, and neither is stationary on its own. Still, because bakers pass the price of flour on to bread, the price of bread cannot drift away from the price of flour for long; if flour becomes dearer one month while bread stays the same, bread goes up within a few months. A cointegration test checks this relationship, and the VECM's adjustment coefficient might show that about 20% of any departure from equilibrium is closed each month through changes in the price of bread.

Parameters in MerQur

Series (≥ 2, I(1))

Concept: Cointegration is sought between series integrated of the same order (usually I(1), meaning they become stationary after differencing once). The series are entered in levels, not differenced, because what is tested is the long-run relationship between the levels.

Everyday example: The flour and bread prices are entered as the prices themselves (per kg), not as monthly changes. It should first be confirmed with unit-root tests that both series have a unit root but that their monthly changes are stationary.

What it does: Level (I(1)) series. The Johansen eigenvectors and β are normalised on the first series or on the selected normalisation variable.

Effect on the result: The selected set of series and their order.

On the validation data: Canada (JSS order prod, e, U, rw), denmark (LRM, LRY, IBO, IDE) and finland.

Date (optional)

Concept: Because the model relies on past values, the observations must be arranged in time order. The date column puts the rows in chronological order.

Everyday example: If the price table was recorded from the newest month to the oldest and no date is selected, the lags are built the wrong way round and the test gives wrong results. Selecting the date fixes the order of the months.

What it does: Sorts the rows by the date column.

Effect on the result: In a reverse-ordered file the result is built backwards in time.

On the validation data: In the reversed Canada file, selecting the date gives a result equal to that of the ordered file; without it the trace r = 0 statistic differs from 105.844.

Method (Johansen + Engle-Granger / Johansen / Engle-Granger)

Concept: The Johansen method treats all the series as a system and can find more than one equilibrium relationship; Engle-Granger rests on a single regression and its residuals, can find only one relationship and has low power. Running both shows how much the conclusion depends on the method.

Everyday example: With only flour and bread prices, the two methods often reach the same conclusion. If yeast and electricity prices are added, Johansen may find, say, two separate equilibrium relationships, whereas Engle-Granger looks for only one of them.

What it does: Multivariate maximum likelihood rank test and VECM (Johansen) and/or the two-step Engle-Granger.

Effect on the result: Only produces the relevant section; the numbers do not change.

On the validation data: The default run gives both sections together (Johansen rank 2, EG τ = -2.307).

→ Johansen trace and maximum-eigenvalue tests

VAR lag K (in levels)

Concept: The lag order in levels of the VAR model on which the Johansen test rests; in the error correction model it corresponds to K − 1 lagged differences. Too few lags leave autocorrelation in the residuals, too many reduce the test's power, and the rank decision is sensitive to this choice.

Everyday example: If bakers usually pass a flour price rise on to bread within one or two months, K = 2 or 3 is reasonable. If K = 2 finds two relationships and K = 3 finds one, the result should be interpreted carefully alongside the residual diagnostics.

What it does: VAR lag K in levels; K - 1 difference lags in the VECM.

Effect on the result: Trace statistics and the rank decision.

On the validation data: Canada, restricted constant, trace r = 0: 105.84 for K = 2, 100.94 for 3, 64.79 for 6; 5% rank decision 2, 1, 1. In the Pfaff (2008) Table 4 model (trend, K = 3) trace 84.92, 36.42, 18.72, 3.85.

→ Lag length selection

Deterministic case (Restricted constant (in relation) / Unrestricted constant (trend in data) / Restricted trend + unrestricted constant)

Concept: Sets where the constant and trend enter the model: only inside the equilibrium relationship (restricted constant), as a linear trend in the series themselves (unrestricted constant), or as a trend within the equilibrium relationship. Because the critical values change accordingly, the choice should rest on the plot of the data and on theory.

Everyday example: If flour and bread prices rise steadily with inflation, the series themselves have a trend and an unrestricted constant may be suitable. If the prices fluctuate around a flat line, the constant is placed only in the equilibrium relationship.

What it does: Three deterministic cases: restricted constant in the cointegrating relation, unrestricted constant (linear trend in the data), restricted trend + unrestricted constant. The critical values are chosen accordingly from the Osterwald-Lenum (1992) table.

Effect on the result: Trace statistics, critical values and the rank.

On the validation data: K = 2, trace r = 0: restricted constant 105.84 (5% critical 53.12), unrestricted constant 75.93 (48.28), trend 86.12 (62.99); rank 2, 1, 1.

→ Deterministic terms (constant, trend)

VECM rank r (-1 = %5 decision of the trace test)

Concept: The rank r is the number of independent equilibrium relationships between the series; for n I(1) series it lies between 0 and n − 1. The error correction model is fitted with this many relationships, and adjustment coefficients (α) are estimated for each.

Everyday example: For flour and bread, r can be at most 1: 'the price of bread stays around a certain multiple of the price of flour'. A VECM with r = 1 shows, after a departure from equilibrium, how much of the gap closes each month and in which price.

What it does: Number of cointegrating relations used in the VECM.

Effect on the result: α, β, Γ and the VECM equations. When r = 0 or r = number of series, the VECM is not estimated and the reason is stated.

On the validation data: Canada α_prod: 0.0004 for r = 1, 0.0386 for r = 2, 0.0475 for r = 3. Pfaff (2008) Table 5 (rw order, trend, K = 3, r = 1): β = (1, 0.545, -0.013, 1.727, -0.709), α_rw = -0.085; MerQur is the same.

→ Johansen trace and maximum-eigenvalue tests

Normalization variable

Concept: An equilibrium relationship can be multiplied by any constant and so is not identified on its own; normalisation sets the chosen series' coefficient to 1 so that the relationship is read in terms of that series. It is an aid to interpretation and does not change the relationship itself.

Everyday example: Normalised on the bread price, the relationship reads something like 'bread = 0.8 × flour + constant'. Normalised on flour, the same relationship is written in terms of the flour price; which is clearer depends on the research question.

What it does: The Johansen eigenvectors, β and the VECM equations are normalised on this series; the default is the first series.

Effect on the result: Scale of β and α; the trace and maximum eigenvalue statistics do not change.

On the validation data: Selecting rw in the JSS order reproduces the Pfaff (2008) Table 5 β and α values.

→ Johansen trace and maximum-eigenvalue tests

Seasonal dummies (Johansen) (None / Quarterly (4) / Monthly (12))

Concept: Centred seasonal dummies add the regularly recurring effect of each month or quarter to the model. Seasonal fluctuation is thus not confused with the long-run relationship.

Everyday example: If certain months, such as the start of the school year or a religious holiday, affect bread demand and prices in a similar way every year, monthly seasonal dummies separate out this regular effect. The equilibrium relationship is then estimated from prices cleaned of it.

What it does: Adds centred seasonal dummies to the model.

Effect on the result: Trace statistics, β and α.

On the validation data: The Denmark (K = 2, restricted constant) and Finland (unrestricted constant) models of Johansen and Juselius (1990) and the Canada model with monthly seasonal dummies are estimated.

→ Seasonality and seasonal period

Engle-Granger dependent variable

Concept: Which series is the dependent variable in Engle-Granger's first-step regression. In finite samples the test result can depend on this choice, which is one of Engle-Granger's weaknesses compared with Johansen.

Everyday example: Regressing bread on flour may give one result, and regressing flour on bread a slightly different one. If cointegration is found in both directions, the conclusion is more reassuring.

What it does: Dependent variable of the first-step regression. Affects only Engle-Granger; the Johansen order does not change.

Effect on the result: First-step coefficients and the residual ADF τ.

On the validation data: Canada τ: prod (first) -2.307, e -3.293, U -3.419.

→ Engle-Granger two-step method

EG regression terms (Constant / Constant + trend)

Concept: Sets whether only a constant, or a constant and linear trend, is added to the first-step regression. With a trend, the hypothesis tested is 'the series are in equilibrium once a common time trend is removed', and the critical values become stricter.

Everyday example: If the bread price also carries a trend unrelated to flour, because bakeries' labour costs have risen over the years, constant + trend is chosen. That extra trend is then not mistaken for a departure from the equilibrium relationship.

What it does: A constant or a constant + trend is added to the first-step regression; the MacKinnon (2010) c and ct critical values are used.

Effect on the result: τ and the critical values.

On the validation data: Constant + trend: τ = -2.845, 5% critical value -4.609 (constant: -4.235).

→ Deterministic terms (constant, trend)

ADF lag selection (Fixed / AIC / BIC) and ADF lag k

Concept: Sets how many lagged differences are added to the ADF test applied to the residuals in Engle-Granger's second step. The lags remove autocorrelation from the residuals; a fixed number can be given or it can be chosen by AIC/BIC.

Everyday example: If the residuals of the flour-bread regression tend to stay high the month after a high month, lags must be added to the ADF test. The τ statistic and the decision may change with the number of lags, so the choice should be stated in the report.

What it does: Lag of the residual ADF regression (no constant): fixed k or chosen from 1..k by AIC/BIC (on a common sample). Under AIC/BIC with k ≤ 1, a maximum lag of 8 is taken and a note is written.

Effect on the result: τ and the number of observations used.

On the validation data: Canada: τ = -1.905 for k = 0 (n = 83), -2.307 for k = 1 (82), -2.396 for k = 4 (79); AIC/BIC select k = 1, τ = -2.052 (common sample n = 75).

→ Lag length selection

Example on real data

Data: Canada, Breitung et al. 2004, R vars (Pfaff 2008 ordering)

Variables: series prod (productivity), e (employment), U (unemployment), rw (real wage)

  • Question: Is there a long-run equilibrium (cointegration) relationship among Canada's four macro series (prod, e, U, rw)?
  • Johansen trace test: K = 2, restricted constant, T = 82; λ_trace(r = 0) = 105.84 (5% critical 53.12), λ_trace(r ≤ 1) = 42.31 (34.91), λ_trace(r ≤ 2) = 17.39 (19.96); rank r = 2 at the 5% level.
  • Max. eigenvalue test: 63.54, 24.92, 10.51, 6.88; the same decision (r = 2).
  • VECM: Adjustment coefficients (ect1) prod = 0.039, e = 0.026, U = -0.019 (significant), rw = -0.003 (not significant); rw behaves like a variable that does not adjust to equilibrium (weakly exogenous).
  • Engle-Granger: In the residuals of the prod ~ e + U + rw regression, ADF τ = -2.31 (5% critical -4.23, p = .762); this single-equation test finds no cointegration.
  • Limitation: The conflict between the two methods arises because Engle-Granger is limited to one vector and has low power; Johansen is preferred in the multivariate case.
  • APA: Johansen trace test, λ_trace(r = 0) = 105.84 (5% critical value = 53.12); r = 2. Engle-Granger τ = -2.31, p = .762.

Related concepts

Johansen trace and maximum-eigenvalue testsEngle-Granger two-step methodLag length selectionVector error correction model (VECM)Deterministic terms (constant, trend)Unit root tests (ADF, Phillips-Perron)Seasonality and seasonal periodEigenvalue and eigenvector

The numbers in the examples are MerQur's output on independent, publicly available validation data; see the MerQur Technical Note for details.