Applied Macro and Financial

Econometrics

Multivariate Time Series

Vector Autoregressive Model

Course Coordinator:
Dr. Devasmita Jena
Spurious Regression
Consider following independent RWM:

and are white noise and are independent of each other. Consider regressing on :

What will the estimated value of ?

 Since and are independent of each other,
 But it may so happen that comes out to be statistically significant
• T-test indicating meaningful relationship between and
• R2 is also moderately high
 So, statistically speaking, there is a relationship between but we may not be able support it by economic theory
 This is called spurious regression: meaningless regression
 Example
• Crime rate in India and Life expectancy in Nigeria
• India’s export intensity and literacy rate in Iceland
Spurious Regression: Why it happens?
 RWM have equal chances of moving up and/or down
• may show tendency of moving together in the same direction 50% of the times
• Higher R2 when you regress on
Even though the general movements in one series may be explained “very well” by another
series but if a period by period analysis is carried out the relationship may hold “good” in one
period but “not so good” in another period
• This could be a cross check to see whether you are estimating a spurious regression
What are the other cross checks?
• If relationship holds in levels, then it should hold in 1st differences by and large (exceptions are there) : do the
regression of on in 1st differences
o It may be the case that 1st difference
• Residual analysis
o Residuals in the case of spurious regression will not be stationary
o Residuals with stochastic trend => the error in one period stays permanently => thus model is spurious
 R2 > D-W test statistic
 However, there are situations when the residuals from regression model of on could decay quickly, i.e., they are
white noise with mean zero and constant variance then such a regression is useful for forecasting
Vector Autoregressive Models
 VARs are generalization of univariate autoregressive models
It is a systems regression model, i.e., there are more than one dependent variables
Hybrid between univariate TS model and SEM: alternative to large-scale simultaneous equation
structural models
Bivariate VAR: we’ve two variables, each of whose current values depend on different
combinations of the previous k values of both the variables, and error terms:

VARMA: VAR extended to include MA errors

VARX: with exogenous component
 You could extend VAR to include more than two variables or to include 1 st differences and
cointegrating relationships
VAR Models: Advantages
 Flexibility
No need to specify which variables are endogenous – all are endogenous
AR models are restricted VAR: VARs allow the value of a variable to depend on more than just
its own lags or combinations of white noise terms
Provided that there are no contemporaneous terms on the RHS of the equations, it is possible to
simply use OLS separately on each equation
• All variables on the RHS are pre-determined: at time t, they are known
• no possibility for feedback from any of the LHS variables to any of the RHS variables
• Pre-determined variables include all exogenous variables and lagged values of the endogenous variables
The forecasts generated by VARs are often better than ‘traditional structural’ models
• Structural models impose ad-hoc nature of restrictions for identification
VAR Models: Disadvantages
 VARs are a-theoretical in nature (as are ARMA models)
Determination of lag length : trade-off between loss of df and model mis-specification
• For g equations - g variables, k lags of each variable in each equation - (g+kg 2) parameters are to be estimated
• VAR models should be as unrestricted as possible
o A VAR with different lag lengths for each equation could be viewed as a restricted VAR
o For example, a VAR with 3 lags of both variables in one equation and 4 lags of each variable in the other equation is a
restricted model
 the coefficient on the fourth lags of each variable in the first equation have been set to zero
 There are ways to choose optimal length (check Brooks for discussion)
• Information criteria is still the best!
Should all of the components of the VAR be stationary? Yes
• Stationarity is essential for hypothesis testing (single or joint hypotheses) and to examine the statistical
significance of the coefficients
• Differencing of variables to ensure stationarity throws away information of any long run relationships between
series
o VAR estimation is done purely to examine short run relationship
o Exception : if variables are I(0) to begin with, VAR estimation can be used to infer about long run relationship
o It is also possible to combine levels and first differenced terms in a vector error correction model (we’ll see later)
Does VAR include Contemporaneous Terms?
 Consider following B-VAR(*)

No contemporaneous terms => no term on the RHS of the equation and no term on the RHS
of the equation
What if the equation has contemporaneous feedback?

In matrix notation:

----(**)
The above system is VAR in primitive form – akin to structural form of SEM
It can be proved: The VAR (*) is reduced for of generalized structural VAR (**)
From Structural to Reduced form VAR
Writing (**) as:
----(**)

Pre-multiplying both sides by A-1:

This the standard form of VAR that we work with which is similar to reduced form from a set of
simultaneous equations
 This VAR contains only predetermined values on the RHS
 No contemporaneous feedback or endogeneity issue => OLS works well
What is the stability condition of the VAR system of equations?
 Eigen vales of matrix A1 have modulus less than one
 Det. |A1 – λI| = 0
 All the roots of matrix equation: [ I – A1z] = 0 are greater than 1 in modulus
 The condition can be generalized for VAR(q)
Block Significance Test/Granger Causality Test
 Test to see which set of variables in a VAR system have significant effects on each of the
dependent variables and which do not
In other words, the test checks which of the variables in the system are exogenous
Joint tests are usually conducted that restricts all the lags of a particular variable to zero
Consider the following VAR(3) model:

The granger causality test hypotheses:

 Lags of don’t explain current
 Lags of don’t explain current
 Lags of don’t explain current
 Lags of don’t explain current
Under the assumption of stationarity, the joint hypothesis is tested using F-test
Granger Causality Test
 Points:
• Unidirectional causality and exogeneity
• Bi-directional causality/ bi-direction feedback and endogeneity
• Independent variables
• Granger causality => correlation between the current value of one variable and past values of the other
variables
Granger causality doesn’t imply causation, i.e. movement of one variable causes the movement
of the other.
Impulse Response Function
 Granger causality test is silent on the sign of relationship
Granger causality is silent on how long the effects require to take place
IRF: traces out the responsiveness of the dependent variables in the VAR to shocks to each of
the variables
A unit shock is applied to the error, and the effects on the VAR system is noted over time
For g variables, there is possibility of g2 impulse responses
VAR stability => shocks should gradually dies down
Theoretically, impulses are calculated by expressing VAR as VMA
Variance Decomposition Function
 Variance decompositions give the proportion of the movements in the dependent variables that
are due to their “own” shocks Vs. shocks to other variables
How much of the s-step ahead forecast error variance of a given variable is explained by the
innovations to each explanatory variables for s=1,2,3,…
Usually, own shock matters the most
Note:
• For calculating IRFs and variance decompositions, the ordering of variables important
o in practice the error terms of the equations in the VAR will be correlated
o so the result will be dependent on the order in which the equations are estimated in the model
• Theory or evidence may guide on ordering
Taking theory to praxis- steps
 Check stationarity of individual time series
All elements of VAR system should be I(d)
Information criteria to obtain lag lengths
Estimate VAR and check the significance of coefficients
VAR stability
Granger Causality, IRF, Variance Decomposition Function