Introductory Econometrics

Review

Monash Econometrics and Business Statistics

2019

1 / 15
Learning goals of this unit
I Understand and derive the properties of ordinary least squares in
summation and matrix notation
I Interpret, evaluate and apply inferential methods to multiple linear
regression
I Understand the use and implications of data scaling, functional
form, dummy variables in regression modelling
I Understand the consequences of heteroscedasticity in regression
models, and know how to test for it and the practical ways of
dealing with it
I Understand the consequences of serial correlation in errors in
regression models, and know how to test for it and the practical
ways of dealing with it
I Understand that some time series in finance and economics are
non-stationary, and be introduced to how to deal with
non-stationary time series
I Be introduced to the large sample properties of estimators (namely
consistency and asymptotic normality)
2 / 15
The OLS estimator and its properties

I The OLS estimator βb produces a linear combination of columns of

X that are closest to y
I That is, ŷ = Xβb is the closest vector in the column space of X to y
I That is, û = y − ŷ is perpendicular to the column space of X,
implying that X0 û = 0.
I This leads to the formula βb = (X0 X)−1 X0 y

3 / 15
The OLS estimator and its properties
I We have studied the multiple regression model and learnt that when:
1. model is linear in parameters: y = Xβ + u
2. conditional mean of errors is zero: E (u | X) = 0
3. columns of X are linearly independent
⇒ then the OLS estimator βb is an unbiased estimator of β

4 / 15
The OLS estimator and its properties
I We have studied the multiple regression model and learnt that when:
1. model is linear in parameters: y = Xβ + u
2. conditional mean of errors is zero: E (u | X) = 0
3. columns of X are linearly independent
⇒ then the OLS estimator βb is an unbiased estimator of β
I if in addition,
4. sample is random and errors are homoskedastic: Var (u | X) = σ 2 In ,
b = σ 2 (X0 X)−1
⇒ then βb is the BLUE and Var (β)

4 / 15
The OLS estimator and its properties
I We have studied the multiple regression model and learnt that when:
1. model is linear in parameters: y = Xβ + u
2. conditional mean of errors is zero: E (u | X) = 0
3. columns of X are linearly independent
⇒ then the OLS estimator βb is an unbiased estimator of β
I if in addition,
4. sample is random and errors are homoskedastic: Var (u | X) = σ 2 In ,
b = σ 2 (X0 X)−1
⇒ then βb is the BLUE and Var (β)
I If, in addition to the above,
5. errors are normally distributed,
⇒ then conditional on X, βb is normally distributed, and we can use the
usual t and F tests to make inferences based on the OLS estimator

4 / 15
Interpretation of the estimated parameters

I We learnt how to interpret the estimated parameters in multiple

regression, including
I in level-level, log-level, level-log and log-log models
I in quadratic models
I in models with dummy variables
I in dynamic models
I Almost all tutorials, several moodle lessons and all past exams
have examples of this
I We learnt how to use dummy variables to bring in categorical or
qualitative information, with applications to investigating differences
between groups, or seasons, or before and after a certain time period

5 / 15
Inference based on the OLS estimator

I Formulating a real world question in terms of a restriction on one or

more parameters in a multiple regression model
I The tests:
I Using a t-test to test hypothesis about a single parameter
against a one-sided or two-sided alternatives
I Using an F -test to test joint hypotheses involving several
parameters
I Using a t-test in a reparameterised model to test a single
hypothesis involving more than one parameter against a
one-sided or two-sided alternatives

6 / 15
Inference based on the OLS estimator

I Mathematical statement: Formal statement of all steps of

hypothesis testing using precise notation (the null, the alternative,
the test statistic and its distribution under the null, computing the
test statistic, obtaining the critical value, comparison and
conclusion)
I Stating the conclusion in the context of the problem
I Special applications:
I Using F -test to test the overall significance of a regression
I Using t-test or F -test in conjunction with dummy variables to
test difference between groups, seasons, or structural breaks

7 / 15
Heteroskedasticity (HTSK)
Consequences and remedies

I The assumption of equal conditional variance for all u may be not a

reasonable assumption is some applications.
I Since constant conditional variances is not required for unbiasedness,
the OLS estimator will remain unbiased even with HTSK
I HTSK violates assumption 4, therefore the OLS estimator will no
longer be BLUE and Var (β) b 6= σ 2 (X0 X)−1 . This means that the
default standard errors reported by the statistical package for the
OLS estimator will be incorrect
I Even if errors are normally distributed, the t and F tests based on
the default OLS standard errors will be unreliable
I Fortunately, if we detect HTSK, we have ways to conduct reliable
inference based on the OLS estimator using robust standard errors,
or use weighted least squares, which is a more efficient estimator
than the OLS estimator when we have HTSK

8 / 15
Heteroskedasticity (HTSK)
Consequences and remedies

I Detecting HTSK: We learnt that HTSK can be detected using tests

that are based on n × R 2 of an auxiliary regression of û 2 on a set of
appropriate regressors. These tests are:
1. Breusch-Pagan (BP) test, which is used when we know (or we
are told) that under the alternative the variance of the error is
a linear function of particular set of variables
2. White test, which is used when the only thing that we suspect
is that the variance of the error is a smooth function of all
explanatory variables
I These tests have a χ2q distribution under the null of no HTSK, where
q is the number of regressors (not counting the constant) in the
auxiliary regression. These tests are only accurate in large samples.
I We learnt (and we can show) that when Var (ui ) = σ 2 hi , multiplying
the entire model by √1h leads us to a model in which OLS is BLUE
i
and its standard errors are correct.

9 / 15
Serial Correlation
Consequences for OLS

I Time series data are temporally dependent, so time series samples

are not random samples. Therefore, it is quite possible that the
errors are not white noise (a white noise process is a serially
uncorrelated process with mean zero and finite variance). This
means that Var (u | X) 6= σ 2 In .
I However, even though the sample is not random, we may still have
E (u | X) = 0. If that is the case and columns of X are linearly
independent, then the OLS estimator will still be unbiased.
I But, the OLS estimator is no longer BLUE, and the usual OLS
standard errors are incorrect.
I This means that the usual t-tests and F -tests will be unreliable.

10 / 15
Serial Correlation
Detection

I We learnt to use
1. the time plot and the correlogram of the residuals to detect
serial correlation visually
2. to use Breusch-Godfrey test to detect serial correlation
formally. This tests requires stating an AR(q) alternative for
the errors, where q is determined from the context or the
frequency of the data. The test statistic is n × R 2 of an
auxiliary regression of ût on all independent variables
(including the constant) and q lags of ût , and has a χ2q
distribution under the null.

11 / 15
Serial Correlation
Remedies

I If we conclude that errors are serially correlated, we can

1. use OLS estimates with HAC standard errors; or
2. augment the model with an AR(q) model for errors, i.e.,

yt = β0 + β1 xt1 + · · · + βk xtk + ut
ut = ρ1 ut−1 + · · · + ρq ut−q + et

and use FGLS instead of OLS; or

3. specify a more general dynamic (ARDL(p,q)) model, i.e.,

yt = β0 +β1 yt−1 +· · ·+βp yt−p +γ0 xt +γ1 xt−1 +· · ·+γq xt−q +et

and estimate that using OLS.

I The FGLS estimator and the OLS estimator in ARDL model are not
unbiased, but they are consistent and asymptotically normal

12 / 15
Modelling Dynamics
I The OLS estimators of the parameters of the ARDL(p, q) model

yt = β0 + β1 yt−1 + · · · + βp yt−p + γ0 xt + γ1 xt−1 + · · · + γq xt−q + et

are consistent and asymptotically normal provided that yt and xt are

stationary, and et is white noise.
I This means that the OLS estimates are reliable and we can use the
usual tools for hypothesis testing as long as we have a large sample.
I We went over the definitions of a consistent estimator and an
asymptotically normal estimator.
I We learnt the definition of covariance stationarity, which is the only
notion of stationarity that is adequate for our purposes.
I We also learnt that even if yt and xt have a linear trend, but their
fluctuations around their trend is stationary, we can add a time
trend to the right hand side of the ARDL model and use OLS.

13 / 15
Modelling Dynamics
Interpretation of parameters
I In the estimated ARDL(p, q) model
ŷt = β̂0 + β̂1 yt−1 + · · · + β̂p yt−p + γ̂0 xt + γ̂1 xt−1 + · · · + γ̂q xt−q
when x increases by one unit at time t, it immediately increases the
predicted y by γ̂0 at time t. This is the immediate or the impact
effect of a change in x on predicted y .
I However, because of the dynamics, a change in x at time t will
continue to affect y in t + 1, t + 2, . . .
I The overall or the long-run effect of a unit increase in x at time t
on predicted y is given by
sum of the coefficients of xt and its lags
1 − sum of the coefficients of lags of yt
which in the above equation would be
γ̂0 + γ̂1 + · · · + γ̂q
1 − (β̂1 + · · · + β̂p )

14 / 15
Modelling Dynamics
Nonstationarity and unit roots
I A covariance stationary process is a time series which at all t has a
finite mean and finite variance that do not depend on t, and the
covariance between two elements of this series depend on the time
difference, not on the time itself
I A stationary time series reverts to its mean often
I If a time series does not satisfy any of these requirements, it is
non-stationary
I A prevalent type of non-stationarity in business and economics data
is unit root non-stationarity. This is when the time series its
autoregressive, but the sum of the coefficients of lag of yt is 1.
I Unit root processes are not mean reverting and have high
persistence
I We identify them by the plot of the time series, by their
correlogram: their first order autocorrelation coefficient will be very
close to 1, and the autocorrelations decay very slowly
I If yt has a unit root, ∆yt will not have a unit root, and we can use
∆yt in our econometric models 15 / 15