Error correction model

The first step of this method is to pretest the individual time series one uses in order to confirm that they are non-stationary in the first place. This can be done by standard unit root DF testing and ADF test (to resolve the problem of serially correlated errors). Take the case of two different series ${\displaystyle x_{t}}$ and ${\displaystyle y_{t}}$. If both are I(0), standard regression analysis will be valid. If they are integrated of a different order, e.g. one being I(1) and the other being I(0), one has to transform the model.

If they are both integrated to the same order (commonly I(1)), we can estimate an ECM model of the form

${\displaystyle A(L)\,\Delta y_{t}=\gamma +B(L)\,\Delta x_{t}+\alpha (y_{t-1}-\beta _{0}-\beta _{1}x_{t-1})+\nu _{t}.}$

If both variables are integrated and this ECM exists, they are cointegrated by the Engle–Granger representation theorem.

The second step is then to estimate the model using ordinary least squares: ${\displaystyle y_{t}=\beta _{0}+\beta _{1}x_{t}+\varepsilon _{t}}$ If the regression is not spurious as determined by test criteria described above, Ordinary least squares will not only be valid, but also consistent (Stock, 1987). Then the predicted residuals ${\displaystyle {\hat {\varepsilon _{t}}}=y_{t}-\beta _{0}-\beta _{1}x_{t}}$ from this regression are saved and used in a regression of differenced variables plus a lagged error term

${\displaystyle A(L)\,\Delta y_{t}=\gamma +B(L)\,\Delta x_{t}+\alpha {\hat {\varepsilon }}_{t-1}+\nu _{t}.}$

One can then test for cointegration using a standard t-statistic on ${\displaystyle \alpha }$. While this approach is easy to apply, there are numerous problems:

• The univariate unit root tests used in the first stage have low statistical power
• The choice of dependent variable in the first stage influences test results, i.e. we need weak exogeneity for ${\displaystyle x_{t}}$as determined by Granger causality
• One can potentially have a small sample bias
• The cointegration test on ${\displaystyle \alpha }$ does not follow a standard distribution
• The validity of the long-run parameters in the first regression stage where one obtains the residuals cannot be verified because the distribution of the OLS estimator of the cointegrating vector is highly complicated and non-normal
• At most one cointegrating relationship can be examined.

The Engle–Granger approach as described above suffers from a number of weaknesses. Namely it is restricted to only a single equation with one variable designated as the dependent variable, explained by another variable that is assumed to be weakly exogeneous for the parameters of interest. It also relies on pretesting the time series to find out whether variables are I(0) or I(1). These weaknesses can be addressed through the use of Johansen's procedure. Its advantages include that pretesting is not necessary, there can be numerous cointegrating relationships, all variables are treated as endogenous and tests relating to the long-run parameters are possible. The resulting model is known as a vector error correction model (VECM), as it adds error correction features to a multi-factor model known as vector autoregression (VAR). The procedure is done as follows:

• Step 1: estimate an unrestricted VAR involving potentially non-stationary variables
• Step 2: Test for cointegration using Johansen test
• Step 3: Form and analyse the VECM.

The idea of cointegration may be demonstrated in a simple macroeconomic setting. Suppose, consumption ${\displaystyle C_{t}}$ and disposable income ${\displaystyle Y_{t}}$ are macroeconomic time series that are related in the long run (see Permanent income hypothesis). Specifically, let average propensity to consume be 90%, that is, in the long run ${\displaystyle C_{t}=0.9Y_{t}}$. From the econometrician's point of view, this long run relationship (aka cointegration) exists if errors from the regression ${\displaystyle C_{t}=\beta Y_{t}+\varepsilon _{t}}$ are a stationary series, although ${\displaystyle Y_{t}}$ and ${\displaystyle C_{t}}$ are non-stationary. Suppose also that if ${\displaystyle Y_{t}}$ suddenly changes by ${\displaystyle \Delta Y_{t}}$, then ${\displaystyle C_{t}}$ changes by ${\displaystyle \Delta C_{t}=0.5\,\Delta Y_{t}}$, that is, marginal propensity to consume equals 50%. Our final assumption is that the gap between current and equilibrium consumption decreases each period by 20%.

In this setting a change ${\displaystyle \Delta C_{t}=C_{t}-C_{t-1}}$ in consumption level can be modelled as ${\displaystyle \Delta C_{t}=0.5\,\Delta Y_{t}-0.2(C_{t-1}-0.9Y_{t-1})+\varepsilon _{t}}$. The first term in the RHS describes short-run impact of change in ${\displaystyle Y_{t}}$ on ${\displaystyle C_{t}}$, the second term explains long-run gravitation towards the equilibrium relationship between the variables, and the third term reflects random shocks that the system receives (e.g. shocks of consumer confidence that affect consumption). To see how the model works, consider two kinds of shocks: permanent and transitory (temporary). For simplicity, let ${\displaystyle \varepsilon _{t}}$ be zero for all t. Suppose in period t − 1 the system is in equilibrium, i.e. ${\displaystyle C_{t-1}=0.9Y_{t-1}}$. Suppose that in the period t, disposable income ${\displaystyle Y_{t}}$ increases by 10 and then returns to its previous level. Then ${\displaystyle C_{t}}$ first (in period t) increases by 5 (half of 10), but after the second period ${\displaystyle C_{t}}$ begins to decrease and converges to its initial level. In contrast, if the shock to ${\displaystyle Y_{t}}$ is permanent, then ${\displaystyle C_{t}}$ slowly converges to a value that exceeds the initial ${\displaystyle C_{t-1}}$ by 9.

This structure is common to all ECM models. In practice, econometricians often first estimate the cointegration relationship (equation in levels), and then insert it into the main model (equation in differences).