From Unit Root To Cointegration

Putting Economics into Econometrics

Many (perhaps most) macroeconomic variables are non stationary Many of these are difference stationary, I(1), variables.

Economics - many variables have stable long-run relationships.

Consumption - Income Prices - Wages Prices at home - prices abroad We can use the unit root testing techniques to identify variables which have stable long-run relationships with one another (as opposed to spurious regression)

Example: Are Y (disposable income) and X (consumption) cointegrated? First be satisfied that the two time series are I(1). E.g. apply unit root tests to X and Y in turn.
55 X 50 45 40 35 30 25 20 15 10 0 10 20 30 40 50 60 70 80 90 100 Y

They are clearly not stationary But they seem to move together

Economic theory tells us that there should be (at least in the long run) a relation like
Ct b0 b1Yt

A regression between these variables gives:

EQ( 1) Modelling C by OLS (using Lecture6a.in7) The estimation sample is: 1 to 99 Coefficient 4.85755 1.00792 0.564679 0.997541 -82.8864 99 Std.Error 0.1375 0.005081 t-value 35.3 198. t-prob Part.R^2 0.000 0.9279 0.000 0.9975

Constant Y

sigma R^2 log-likelihood no. of observations

RSS 30.9296673 F(1,97) = 3.935e+004 [0.000]** DW 2.28 no. of parameters 2

But is this regression spurious? Or is there a genuine long run relationship?

COINTEGRATION Any linear combination of I(1) variables is typically spurious. However if there is a long-run relationship, errors have tendency to disappear and return to zero i.e. are I(0). If there exists a relationship between two non stationary I(1) series, Y and X , such that the residuals of the regression

Yt 0 1 X t ut
are stationary, then the variables in question are said to be cointegrated There is a Long Run Relationship towards which they always come back

Single Equations Errors

No tendency to return to zero

ut

0 -

Error rarely drifts from zero

time

Disequilibrium errors (i.e. ut = Yt - 0 - 1Xt)

Stationary Errors

If we have two independent non-stationary series, then we may find evidence of a relationship when none exists (i.e. spurious regression problem). One way to test if there is a relationship between nonstationary data is if disequilibrium errors return to zero. If long run relationship exists then errors should be a stationary series and have a zero mean.

ut
0
time

Our regression between consumption and income:

Ct b0 b1Yt

How can we distinguish between a genuine long-run relationship and a spurious regression? We need a test.

Testing for Cointegration

After estimating the model save residuals from static regression. (In PcGive after running regression click on Test and Store Residuals) Informally consider whether stationary.
1.0
residuals

0.5 0.0 -0.5

0 1.0 0.5 0.0 -0.5

10

20

30

40

50

60

70

80

90

100

ACF-residuals

10

11

12

(1) Cointegrating Regression Durbin Watson (CRDW) Test

At 0.05 per cent significance level with sample size of 100, the critical value is equal to 0.38.

Ho: DW = 0 => no cointegration (i.e. DW stat. is less than 0.38) Ha: DW > 0 => cointegration (i.e. DW stat. is greater than 0.38)
Ho: ut = ut-1 + et Ha: ut = ut-1 + et

<1

N.B. Assumes that the disequilibrium errors ut can be modelled by a first order AR process. Is this a valid assumption? May require a more complicated model.

First Test: Cointegrating Regression Durbin Watson Test (CRDW)

If the residual are non stationary, DW will go to 0 as the sample size increases. So large values of DW are taken as evidence for rejection of the null hypothesis of NO COINTEGRATION
EQ( 1) Modelling Y by OLS (using Lecture The estimation sample is: 1 to 99 Coefficient 4.85755 1.00792 0.564679 0.997541 -82.8864 99 Std.Error 0.1375 0.005081 6a.in7)

Constant X

t-value 35.3 198.

t-prob Part.R^2 0.000 0.9279 0.000 0.9975

sigma R^2 log-likelihood no. of observations

RSS 30.9296673 F(1,97) = 3.935e+004 [0.000]** DW 2.28 no. of parameters 2

CRDW test statistic = 2.28 >> 0.38 = 5% critical value. This suggests cointegration - assumes residuals follow AR(1) model.

Second Test: Cointegrating Regression DF test (CRDF)

1 - Perform the cointegrating regression

Yt b0 b1 X t ut

2 Save the residuals, ut

3 Run auxiliary regression:

, t 1,..., T

ut

ut 1 et

t 2,..., T

Cointegrating Regression Dickey Fuller (CRDF) Test

ut = ut-1 + et

Critical Values (CV) are from MacKinnon (1991) Ho: = 0 Ha: < 0 => no cointegration (i.e. TS is greater than CV) => cointegration (i.e. TS is less than CV)

Cointegrating Regression Dickey Fuller (CRDF) Test

BETTER: Use lagged differenced terms to avoid serial correlation.

ut = ut-1 + 1ut-1 + 2ut-2 + 3ut-3 + 4ut-4 + et

Use F-test of model reduction and also minimize Schwarz Information Criteria.

Critical Values (CV) are from MacKinnon (1991) Ho: = 0 Ha: < 0 => no cointegration (i.e. TS is greater than CV) => cointegration (i.e. TS is less than CV)

Testing for Cointegration

Using CRDF we incorporate lagged dependent variables into our regression

ut = ut-1 + 1ut-1 + 2ut-2 + 3ut-3 + 4ut-4 + et

And then assess which lags should be incorporated using model reduction tests and Information Criteria.
Progress to date Model T p log-likelihood EQ( 2) 94 5 OLS -74.238306 EQ( 3) 94 4 OLS -74.793519 EQ( 4) 94 3 OLS -74.797849 EQ( 5) 94 2 OLS -74.948145 EQ( 6) 94 1 OLS -75.845305 Tests of model reduction (please ensure models are EQ( 2) --> EQ( 6): F(4,89) = 0.77392 [0.5450] EQ( 3) --> EQ( 6): F(3,90) = 0.67892 [0.5672] EQ( 4) --> EQ( 6): F(2,91) = 1.0254 [0.3628] EQ( 5) --> EQ( 6): F(1,92) = 1.7730 [0.1863]

SC 1.8212 1.7847 1.7364 1.6913 1.6621 nested for

HQ AIC 1.7406 1.6859 1.7202 1.6765 1.6881 1.6553 1.6591 1.6372 1.6459 1.6350 test validity)

All model reduction tests are accepted hence move to most simple model

Consequently we choose

ut = ut-1 + et

Testing for Cointegration using CRADF test

Step 1: Estimate cointegrating regression

Yt = 0 + 1Xt + ut
EQ( 1) Modelling Y by OLS (using Lecture The estimation sample is: 1 to 99 Coefficient 4.85755 1.00792 0.564679 0.997541 -82.8864 99 Std.Error 0.1375 0.005081 6a.in7)

Constant X

t-value 35.3 198.

t-prob Part.R^2 0.000 0.9279 0.000 0.9975

sigma R^2 log-likelihood no. of observations

RSS 30.9296673 F(1,97) = 3.935e+004 [0.000]** DW 2.28 no. of parameters 2

And save residuals.

Testing for Cointegration

Step 2: Use estimated in the auxiliary regression model ut = ut-1 + et
EQ( 6) Modelling dresiduals by OLS (using Lecture6a.in7) The estimation sample is: 6 to 99 Coefficient -1.16140 0.545133 -75.8453 Std.Error 0.1024 RSS DW t-value -11.3 t-prob Part.R^2 0.000 0.5805 27.6367834 1.95

residuals_1 sigma log-likelihood

Which means

ut = -1.161 ut-1 + et
(-11.3)

CRDF test statistic = -11.3 << -3.39 = 5% Critical Value from MacKinnon. Hence we reject null of no cointegration between X and Y.

Testing for Cointegration

Advantages of (CRDF) Test Engle and Granger (1987) compared alternative methods for testing for cointegration. (1) Critical values depend on the model used to simulated the data. CRDF was least model sensitive.

(2) Also CRDF has greater power (i.e. most likely to reject a false null) compared to the CRDW test.

Testing for Cointegration

Disadvantage of (CRDF) Test - Although the test performs well relative to CRDW test

there is still evidence that CRDF have absolutely low power.

Hence we should show caution in interpreting the results.

Cointegration and Consistency

OLS estimates with I(0) variables are said to be consistent. As the sample size increases they converge on their true value. However if the true relationship between variables includes dynamic terms

Yt = 0 + 1Xt + 2Yt-1 + 3Xt-1 + ut

Static models estimated by OLS will be bias or inconsistent.

Yt = 0 + 1Xt + ut
Stock (1987) found that if Yt and Xt are cointegrated then OLS estimates of 0 and 1 will be consistent.

Cointegration and Superconsistency

Indeed, Stock went further and suggested that estimated coefficients from cointegrated regressions will converge at a faster rate than normal. i.e. super consistent. Coefficients from a cointegrated regression are super consistent. => (i) simple static regression dont necessarily give spurious results. (ii) dynamic misspecification is not necessarily a problem.

Consequently we can estimate simple regression

Yt = 0 + 1Xt + ut
even if there are important dynamic terms

Yt = 0 + 1Xt + 2Yt-1 + 3Xt-1 + ut

Cointegration and Superconsistency

However, superconsistency is a large sample result. Coefficients may be biased in finite samples (i.e. typical sample periods) due to omitted lagged values of Yt and Xt Bias in static regressions is related to R2 . A high R2 indicates that the bias will be smaller.

Testing for Cointegration: Summary

To test whether two I(1) series are cointegrated we examine whether the residuals are I(0).

(a) We firstly use informal methods to see if they are stationary (1) plot time series of residuals (2) plot correlogram of residuals

(b) Two formal means of testing for cointegration. (1) CRDW - Cointegrating Regression Durbin Watson Test (2) CRDF - Cointegrating Regression Dickey Fuller Test