Quantitative Methods for Finance

Topic 7: Specification Errors in Regression

Models

Dr. Mark Hallam

University of York

Academic Year 2023/24

Overview

Introduction and outline for the topic

Heteroskedasticity

Autocorrelation

Incorrect functional form

Omitting relevant or including irrelevant variables

Measurement error
Overview

Introduction and outline for the topic

Heteroskedasticity

Autocorrelation

Incorrect functional form

Omitting relevant or including irrelevant variables

Measurement error
The OLS Assumptions

▶ We made certain assumptions about the error term, ut , to allow us

to derive the desirable properties of the OLS estimator and
statistical tests.

OLS Assumptions
1. E(ut ) = 0
2. var(ut ) = σ 2 < ∞
3. cov(ui ,uj ) = 0 for i ̸= j
4. cov(ut ,xt ) = 0
5. ut ∼ N(0, σ 2 )

1 / 64
Outline for the topic

▶ We will now look at what happens when these assumptions do not

hold - this is often the case with real financial and economic data

▶ In each case, we will look at:

▶ What may cause the assumption to fail and how to test for violations
of the assumption.
▶ What the consequences for the OLS estimator and statistical tests
are if the assumption does not hold but the researcher does not take
this into account.
▶ Whether there are any solutions or alternative techniques that can be
used when the assumption does not hold.

▶ Note that because the actual error term ut is not observable, the
tests or checks we use rely on the residuals ût from the estimated
sample regression function

2 / 64
Overview

Introduction and outline for the topic

Heteroskedasticity

Autocorrelation

Incorrect functional form

Omitting relevant or including irrelevant variables

Measurement error
Heteroskedasticity and Homoskedasticity

▶ We have previously assumed that the variance of the errors were

finite and constant across observations:

var(ut ) = σ 2 < ∞

which is known formally as the assumption of homoskedastic errors

(homoskedastic/homoscedastic = equal variance).

▶ This is not always realistic - the error variance may vary across units
(for cross-sectional data) or over time (for time series data).
▶ In this case, we say that the errors are heteroskedastic - formally this
can be expressed as:

var(ut ) = σt2 or var(ui ) = σi2

for the cases of time series and cross sectional notation respectively.

3 / 64
Heteroskedasticity
▶ We will focus in this topic on the standard form of heteroskedasticity
that is often found in cross-sectional data.
▶ In a later topic we will also look at a specific form of time series
heteroskedasticity that is important in financial data, where the
current error variance depends on the past error variance.

▶ To see a simple real-world example of heteroskedasticity, let’s start

with a model with a single explanatory variable in which we are
regressing hourly wages on years of education:

wagei = β0 + β1 educi + ui

▶ Estimating this model with historical 1970s US data from the

Current Population Survey we obtain (t-stats in parentheses):

[ = −0.905 + 0.541educ
wage
(−1.321) (10.166)

4 / 64
Heteroskedasticity
▶ Plotting this fitted regression line together with the data:

40

30
Wage ($/hr)

20

10

-10

-2 0 2 4 6 8 10 12 14 16 18 20

Education (years)

▶ We can see that the variability of wages around the fitted/predicted

value of the sample regression line increases with the level of
education.

5 / 64
Heteroskedasticity
▶ Or alternatively, plotting the residuals themselves against the value
of the explanatory variable education:
30

20

10
Residual

-10

-20
-2 0 2 4 6 8 10 12 14 16 18 20

Education (years)

▶ We can see that the variance of the residuals is increasing with the
level of education.

6 / 64
Testing for heteroskedasticity
▶ Therefore, in simple cases with a single explanatory variable, it is in
principle possible to check for heteroskedasticity graphically.
▶ However:
▶ Even if applicable, graphical checks are often imprecise and/or
inconclusive - how strong does the pattern need to be?
▶ For more general regression models, the size of the error variance
may vary with the values of multiple explanatory variables, or vary
over time - not easy to detect graphically.

▶ Therefore, to reliably detect heteroskedasticity we need more formal

procedures - there are numerous statistical tests developed for this.
▶ An early example was the Goldfeld-Quandt (GQ) test (see Section
5.4 of Brooks for details), but we focus here on the later and more
flexible White’s test.
▶ An important advantage of White’s test is that it makes few
assumptions about the form of the heteroscedasticity.

7 / 64
Testing for heteroskedasticity: White’s test
▶ White’s test is conducted as follows:
1. Assume that the regression model of interest is:
yt = β0 + β1 x1t + β2 x2t + ut
and we want to test Var(ut ) = σ 2 (i.e. homoskedasticity). We
estimate the model, obtaining the residuals, uˆt .
2. Then run the auxiliary regression:
ût2 = α0 + α1 x1t + α2 x2t + α3 x1t
2 2
+ α4 x2t + α5 x1t x2t + vt
and test the null hypothesis H0 : α1 = · · · = α5 = 0
(homoskedasticity) vs. H1 : at least one αj ̸= 0 (heteroskedasticity).
3. We can either do this using a standard F-test, or more simply using
an LM test. It can be shown that under H0 :
TR 2 ∼ χ2 (m)
where, R 2 is from the auxiliary regression, T is the number of
observations, and m is the number of regressors in the auxiliary
regression excluding the constant (i.e. m = 5 above).
4. If the χ2 test statistic from step 3 is greater than the corresponding
value from the statistical table then reject the null hypothesis that
the disturbances are homoscedastic.
8 / 64
Testing for heteroskedasticity: White’s test
▶ Let’s look a little more at the logic underlying White’s test.

▶ Use of squared residuals: as previously noted, the errors ut are

unobservable and so we generally base tests about the errors on the
residuals ût instead, which are the observable sample equivalent.
▶ Why squared residuals? The variance of a RV can be written as:
Var(ut ) = E[(ut − E(ut ))2 ]. Because of the assumption that
E(ut ) = 0, this simplifies to Var(ut ) = E[ut2 ].

▶ Structure of auxiliary regression: the auxiliary regression contains

the levels, squares and cross-products of all explanatory variables.
▶ If heteroskedasticity is present, the error variance will most
commonly vary with the level, square, or the product of one or more
explanatory variables.
▶ White’s test allows for any/all of these forms of heteroskedasticity
without forcing the test user to make any assumptions.

9 / 64
Testing for heteroskedasticity: tests in EViews
▶ If you wish to implement White’s test, or other heteroskedasticity
tests in EViews, first estimate your model in the usual way.
▶ Once you have estimated your model, from the estimated equation
object go to: View -> Residual Diagnostics ->
Heteroskedasticity Tests and choose the test to perform.
▶ For the previous wage and education example:

10 / 64
Testing for heteroskedasticity: tests in EViews
▶ If we were to expand the model to include years of experience:

wagei = β0 + β1 educi + β2 experi + ui

then our estimation output and White test output become:

11 / 64
Consequences of ignoring heteroskedasticity
▶ Suppose that the errors for our model of interest are heteroskedastic,
but we do not realise this and apply the methods studied so far.
▶ What are the consequences?
▶ OLS estimator: the standard OLS estimator remains unbiased and
consistent - these properties do not depend on the error variance,
and so heteroskedasticity is irrelevant.
▶ However, the OLS estimator is no longer efficient i.e. it is no longer
the estimator (linear or otherwise) with the minimum variance.
▶ Intuitively, by using information about the form of heteroskedasticity
in the errors (if it were available), then we could derive a more
efficient estimator.
▶ Statistical inference: the standard t-tests and F-tests are no longer
valid - the test statistics depend on the standard errors of the
coefficients, which in turn depend on σ 2 .
▶ With heteroskedasticity the usual SEs are invalid - true SEs may be
larger or smaller, depending on form of heteroskedasticity.
12 / 64
Dealing with heteroskedasticity

▶ What can we do in situations where we have, or think we have,

heteroskedasticity?

▶ If the cause or form of heteroskedasticity is known we can use an

estimation method which takes this into account called weighted
least squares (WLS), which is a form of generalised least squares
(GLS).
▶ Section 5.4.3. of Brooks has additional details, but intuitively WLS
uses the information on the form of heteroskedasticity to transform
the data such that it satisfies the OLS assumptions.
▶ However, in reality, the cause/form of heteroskedasticity is generally
not known and the validity of WLS depends on knowing this, so it is
usually difficult to apply.

13 / 64
Dealing with heteroskedasticity
▶ What about if the cause/form of heteroskedasticity is unknown?
Here there are two common strategies.

▶ Transforming the data into logs: natural logarithm (or similar

transformations) pulls in extreme values and may be appropriate if
SD is proportional to mean value.
▶ Cannot be used for variables that may take zero or negative values,
since log is not defined.

▶ Use heteroskedasticity consistent/robust SEs: these are SE

estimators that have been adjusted to try to account for the effects
of heteroskedasticity - most well-known is White’s approach
▶ Essentially if error variance is +ve related to squared value of an
explanatory variable, the SE for that slope coefficient is increased.
▶ Can be calculated by all standard econometric software packages, so
straightforward to use in practice.

14 / 64
Dealing with heteroskedasticity
▶ To use White’s heteroskedasticity robust SEs in EViews, you need to
manually select them from the estimation options - the EViews
default is the standard OLS SEs.
▶ From the Equation Estimation box, click on the Options tab
and then select Huber-White for the Covariance method option:

15 / 64
Overview

Introduction and outline for the topic

Heteroskedasticity

Autocorrelation

Incorrect functional form

Omitting relevant or including irrelevant variables

Measurement error
Autocorrelation

▶ Autocorrelation (or serial correlation) is a often observed in

regression models fitted to time series data.
▶ In this case, the CLRM assumption that the errors are independent
Cov(ui , uj ) = 0 for i ̸= j is the same as saying that there is no
systematic relationship between the errors in different time periods.
▶ Autocorrelation of the errors occurs when this assumption is violated
and Cov(ui , uj ) ̸= 0 for some i ̸= j.

▶ Again, we never have the actual errors ut , so when testing for

autocorrelation we use their sample counterpart, the residuals uˆt .
▶ As an informal graphical check for autocorrelation, we can plot the
residuals against their lagged values i.e. ût against ût−1 .
▶ Autocorrelation of the errors results in patterns that are visible from
this type of residual plot - common example are given on the next 3
slides.

16 / 64
Testing for autocorrelation - graphical checks

+
ût
ût

– +
û t–1
time

– –

▶ Positive autocorrelation in the residuals is indicated by a cyclical

residual plot over time.
▶ A positive (negative) residual in one time period is likely to be
followed by another positive (negative) residual in the next period.

17 / 64
Testing for autocorrelation - graphical checks

+
ût

ût

– +
û t–1

time

– –

▶ Negative autocorrelation is indicated by an alternating pattern where

the residuals cross the time axis more frequently than if they were
distributed. randomly.
▶ A positive (negative) residual in one time period is likely to be
followed by a negative (positive) residual in the next period.

18 / 64
Testing for autocorrelation - graphical checks

+
ût

ût

+
– +
û t–1

time

– –

▶ No pattern in the residual plot at all: this is what we would like to

see.

19 / 64
Testing for autocorrelation - Durbin Watson test
▶ Graphical checks can be useful, but as always they can be imprecise
and/or inconclusive and so various formal statistical tests for
autocorrelation were developed

▶ A well-known example is the Durbin-Watson (DW) test - this is a

test for first order autocorrelation i.e. it assumes that the
relationship is between an error and the previous one.

ut = ρut−1 + vt

where vt ∼ N(0, σv2 ).

▶ The DW test statistic actually tests:

H0 : ρ = 0 and H1 : ρ ̸= 0

and so under the null, the errors in consecutive time periods are
uncorrelated i.e. no first order autocorrelation. The alternative
implies the presence of first order autocorrelation.

20 / 64
Testing for autocorrelation - Durbin Watson test
▶ We can test this null from the residuals of the original model
without having to estimate the auxiliary model above.
▶ Specifically, the DW test statistic is calculated as:

T
X 2
(ût − ût−1 )
t=2
DW = T
X
ût2
t=2

▶ The above form of the DW statistic is not very intuitive, but it can
be shown (see Section 5.5.3 of Brooks) that as T grows, the DW
test statistic satisfies the approximation:
 
cov(ût , ût−1 )
DW ≈ 2 1 − = 2 (1 − corr(ût , ût−1 ))
var(ût )

21 / 64
Testing for autocorrelation - Durbin Watson test

▶ From the alternative form:

DW ≈ 2 (1 − corr(ût , ût−1 ))

we can see that there are three possibilities for the DW statistic,
depending on the autocorrelation in the errors:
1. The null of no first order autocorrelation =⇒ corr(ût , ût−1 ) = 0
=⇒ DW = 2.
2. Under the alternative, positive first order autocorrelation =⇒
1 ≥ corr(ût , ût−1 ) > 0 =⇒ 2 > DW ≥ 0.
3. Under the alternative, negative first order autocorrelation =⇒
0 > corr(ût , ût−1 ) ≥ −1 =⇒ 4 > DW ≥ 2.

▶ Therefore we fail to reject the null of no AC for values of DW close

to 2, and we reject for values closer to 0 (positive AC), or towards 4
(negative AC).

22 / 64
Testing for autocorrelation - Durbin Watson test
▶ However, the DW test statistic does not follow a standard statistical
distribution - instead, it has two critical values: an upper CV,
denoted dL , and a lower CV, denoted dL .
▶ We also have regions in which the test outcome is inconclusive and
we cannot reject or fail to reject the null.

▶ Critical values dL and dU depend on sample size T , and the number

of explanatory variables in the model K .
▶ Econometrics software like EViews will compute these CVs for you,
or they can be obtained from tables as in appendix of Brooks.

23 / 64
Testing for autocorrelation - Durbin Watson test

▶ Suppose that we estimate a model with 3 explanatory variables:

yt = β0 + β1 x1t + β2 x2t + β3 x3t + ut

using a sample size of T = 100, and the sample value of the DW

test statistic is 1.23.
▶ From the table of CVs for the DW test with K = 3 and T = 100,
we see that dL = 1.48 and dU = 1.60.
▶ Given that the value of DW is lower than dL and so we would reject
the null of no autocorrelation in favour of the alternative of positive
autocorrelation.

24 / 64
Testing for autocorrelation - Breusch-Godfrey test

▶ The DW test has several limitations - see Brooks for a full list, but
the most important from a practical perspective are:
▶ DW only tests for first order autocorrelation, but in reality AC may
be higher order too e.g. may have corr(ût , ût−1 ) = 0 (no first order
AC), but corr(ût , ût−2 ) ̸= 0 (2nd order AC).
▶ Test outcome may be inconclusive.
▶ There cannot be any lags of the dependent variable in the model -
rules out AR and ARDL models, which we will see later on.

▶ An alternative test for AC that addresses these issues is the

Breusch-Godfrey (BG) test, which is a more general test for r th
order autocorrelation:

ut = ρ1 ut−1 + ρ2 ut−2 + ρ3 ut−3 + · · · + ρr ut−r + vt

where vt ∼ N(0, σv2 )

25 / 64
Testing for autocorrelation - Breusch-Godfrey test
▶ The null and alternative hypotheses are:

H0 : ρ1 = · · · = ρr = 0 (no AC)
H1 : at least one ρj ̸= 0 (AC)

▶ Test user must choose value of r , but data frequency can often
provide a guide e.g. r = 4 for quarterly data.
▶ The BG test is carried out as follows:
1. Estimate the original linear regression using OLS and obtain the
residuals, ût .
2. Regress ût on all of the regressors from stage 1 (the xs) plus ût−1 ,
ût−2 , . . . , ût−r
Obtain R 2 from this regression.
3. It can be shown that under the null of no autocorrelation:

(T − r )R 2 ∼ χ2r

▶ If the test statistic exceeds the critical value from the statistical
tables, reject the null hypothesis of no autocorrelation.
26 / 64
Testing for autocorrelation - EViews
▶ The value of the DW test statistic is computed automatically by
EViews when estimating any regression model - see below the
regression output under the information criteria:

▶ EViews does not however give DW critical values or a p-value, so

you will need to compute these yourself to obtain the test outcome.
▶ Here the relevant critical values would be
27 / 64
Testing for autocorrelation - EViews
▶ The BG test is not computed automatically, but is accessed easily in
a similar way to White’s test for heteroskedasticity earlier.
▶ Once you estimate your model, go to View -> Residual
Diagnostics -> Serial Correlation LM Test and choose the
number of lags.

28 / 64
Consequences of ignoring autocorrelation

▶ If the errors are autocorrelated, but we ignore this fact then the
consequences are similar to those we saw for previously
heteroskedastic errors.

▶ The coefficient estimates derived using OLS are still unbiased and
consistent, since these properties do not depend on the lack of
autocorrelation.
▶ However, they are no longer efficient i.e. they do not have lowest
variance out of the class of linear estimaters (they are not BLUE),
even in large sample sizes.

▶ The standard OLS standard error estimates are inappropriate, and so

standard statistical inference such as the usual t-tests and F-tests
may be misleading.

29 / 64
Dealing with autocorrelation

▶ If the form of the autocorrelation in the errors is known (e.g. first

order AC), we could use a GLS procedure that allows for
autocorrelated residuals.
▶ The aim is to estimate the autocorrelation structure, transform the
model, and arrive at a model where the error term is no longer
autocorrelated and OLS is again appropriate.

▶ However, we need to make assumptions about the form of the

autocorrelation and if the assumptions made are inappropriate, then
GLS will be invalid.
▶ In practice it is unlikely that the form of the autocorrelation is
known, and a more ‘modern’ view is that residual autocorrelation
presents an opportunity to modify the regression.
▶ We will discuss dynamic models like autoregressive distributed lag
(ADL) models that would be appropriate for this in later topics.

30 / 64
Dealing with autocorrelation
▶ As with heteroskedasticity, we could also use modified standard
errors that are valid in the presence of autocorrelation.
▶ The White SEs we saw previously only allow for heteroskedastic
errors, but would not be appropriate for autocorrelated errors.

▶ Newey and West developed an alternative that is valid for both

heteroskedastic and autocorrelated errors - known as
heteroskedasticity and autocorrelation consistent (HAC) SEs.
▶ Using these HAC standard errors, we can perform the usual t-tests,
F-tests, etc. in the presence of heteroskedasticity and/or
autocorrelation.

▶ Unlike White’s SEs, the NW HAC standard errors require the user to
select the number of lagged residuals to evaluate the autocorrelation.
▶ Software such as EViews will automatically select this for you based
on the sample size.

31 / 64
Dealing with autocorrelation
▶ NW standard errors can be computed easily in EViews in a similar
way to White’s heteroskedasticity robust SEs we saw previously.
▶ From the Equation Estimation box, click on the Options tab and
then select HAC (Newey-West) for the Covariance method option:

32 / 64
Overview

Introduction and outline for the topic

Heteroskedasticity

Autocorrelation

Incorrect functional form

Omitting relevant or including irrelevant variables

Measurement error
Incorrect Functional Form

▶ In the first half of the module, you saw different functional forms
that were all still linear in the parameters e.g. log-log.
▶ Each implies a different relationship between the dependent variable
and explanatory variable(s), but financial or economic theory will
often not suggest a unique ‘correct’ functional form to describe a
given relationship of interest.

▶ Although there may not be one unique functional form that is always
‘correct’ for modelling a specific relationship between variables, there
will usually be some functional forms that are incorrect or unsuitable.
▶ Therefore, it is important to know what problems will be created if
we use an incorrect functional form for our regression model.
▶ The short answer is that the coefficient estimators will be biased and
inconsistent and the standard t-tests and F-tests will be invalid.

33 / 64
Incorrect Functional Form
▶ To discuss the problem of choosing an incorrect functional form we
will use the sample of data plotted below containing 50 observations
for X and Y :
28

26

24

22
Y

20

18

16

14
0 10 20 30 40 50 60 70 80 90 100 110
X

34 / 64
Incorrect Functional Form

▶ To ensure we know what the true functional form is, these data were
artificially generated and the true relationship between X and Y is
correctly described by the following regression model:

yt = β0 + β1 ln xt + ut

▶ This would not be possible with real-world financial or economic

variables, where the true functional form is usually unknown.

▶ Suppose that we used these data to estimate a regression model

with the correct linear-log functional form, and another model with
an incorrect linear functional form.

35 / 64
Incorrect Functional Form
▶ Plots of estimated (incorrect) linear and (correct) lin-log models:
28

Linear Model
26 Lin-log Model

24

22
Y

20

18

16

14
0 10 20 30 40 50 60 70 80 90 100 110
X

36 / 64
Incorrect Functional Form

▶ As expected, the incorrect linear functional form fits the data worse
than the correct linear-log functional form
▶ Looking at the previous graph, the predicted values of y for the
linear model differ quite substantially from the true values
▶ In addition, the incorrect linear functional form implies that the
predicted change in y for a one unit increase in x is always the same
▶ However, the true effect on y of an increase in x varies for different
values of x, because the slope of the linear-log model form is not
constant
▶ It is obvious therefore that the estimated effect of x on y is not a
reliable estimate of the true effect - as we said before, the coefficient
estimator will be biased (and inconsistent), so this is not surprising.
▶ Essentially, the inappropriate functional form means that the errors
are correlated with the explanatory variable(s) i.e. cov(ut ,xt ) ̸= 0.

37 / 64
Detecting Incorrect Functional Form: Residual Plots
▶ One possible method for detecting incorrect functional form is to use
residual plots in which we plot the fitted values Ybi against the
residuals ûi .
▶ Incorrect functional form will often lead to patterns in the residual
values, which can be visible on such plots - e.g. for the previous
incorrect linear functional form:
4

2
Residuals

-1

-2

16 18 20 22 24 26
Fitted Values for Y 38 / 64
Detecting Incorrect Functional Form: Residual Plots

▶ Like all graphical checks, residual plots are informal and may be
inconclusive and/or subjective.

▶ Therefore, formal statistical methods have been developed to test

for incorrect functional form - most well-known example is the
regression error specification test or RESET developed by Ramsey.
▶ RESET provides a general test for incorrect functional form, that
doesn’t require any detailed assumptions about what the alternative
functional form might be.
▶ This is important because even if we have a relatively small number
of explanatory variables, it would be very time consuming to check
every possible alternative functional form.

39 / 64
Detecting Incorrect Functional Form: RESET

▶ The basic idea of RESET is that if the model:

Yt = β0 + β1 X1,t + . . . + βk Xk,t + ut

is correctly specified, then nonlinear functions of X1 , X2 , . . . Xk

should be statistically insignificant if we add them to the model as
additional explanatory variables.

▶ If we have a small number of explanatory variables then we could

directly add nonlinear functions of them (e.g. their logs or squared
values) and test their statistical significance.
▶ However, this will be problematic if we have many explanatory
variables, because we will end up with many explanatory variables
relative to the sample size (i.e. low degrees of freedom).

40 / 64
Detecting Incorrect Functional Form: RESET

▶ Instead, RESET uses the fact that the powers of the fitted values Yb ,
such as Yb 2 , Yb 3 etc. are nonlinear functions of the original
explanatory variables X1 , X2 , . . . Xk
▶ It then uses these powers of the fitted values as additional
explanatory variables and tests to see if they are statistically
significant

▶ After estimating our original (possibly misspecified) model,

weestimate an auxiliary regression that adds powers of Yb as
additional explanatory variables.
▶ We decide how many powers to include, but Yb 2 and Yb 3 (i.e. two
powers) is usually enough to detect incorrect functional form:

Yt = β0 + β1 X1,t + . . . + βk Xk,t + α1 Ybt2 + α2 Ybt3 + vt

41 / 64
Detecting Incorrect Functional Form: RESET

▶ If the original model was correctly specified then the true values of
α1 and α2 should be zero in the new model.
▶ Therefore, we can use the F-test to test the null hypothesis
H0 : α1 = α2 = 0, which corresponds to no specification error

▶ If the value of the F-statistic is statistically significant and we reject

the null hypothesis, then we we conclude that the functional form of
our original regression model is misspecified.
▶ If the value is statistically insignificant then we conclude there is no
evidence of functional form misspecification.

42 / 64
Some Important Points to Note About RESET

▶ If the test outcome for RESET implies that there is functional form
misspecification, RESET does not provide any guidance about what
the true functional form might be
▶ In this case, we can test different possible functional forms (selected
based on theory, intuition, etc.) until we hopefully find one that
passes RESET
▶ It is also possible for more than one different functional form to pass
RESET - in this case, one option is to select the model with the most
statistically significant RESET F-test (equivalently lowest p-value)
▶ Note that RESET can also be used to choose between models with
different dependent variable forms e.g. between a log-log model with
ln Y and a linear model with Y )

43 / 64
Overview

Introduction and outline for the topic

Heteroskedasticity

Autocorrelation

Incorrect functional form

Omitting relevant or including irrelevant variables

Measurement error
Omitting relevant or including irrelevant variables
Omission of a relevant variable
▶ By relevant we mean the variable has a systematic and significant
effect on the dependent variable.
▶ The effects of any excluded variables are contained in the error term.
If the variable is relevant, it is likely that it will be correlated with at
least some of the included explanatory variables.
▶ In this case, cov(ut ,xjt ) ̸= 0 for some j and the estimated
coefficients on all the explanatory variables will be biased and
inconsistent - often known as omitted variables bias (OVB).
▶ The standard errors will also be biased and so standard statistical
tests will be invalid.

Inclusion of an Irrelevant Variable

▶ Coefficient estimates will still be consistent and unbiased, but the
estimators will be inefficient.

44 / 64
Omitting relevant or including irrelevant variables

▶ Inclusion of irrelevant variables is a less serious problem and solution

is more straightforward: if theoretical basis for including variable is
not clear and variable is not statistically significant, then can drop it
from the model.

▶ Omission of relevant variable(s) is a bigger problem in practice:

consequences much more serious, and testing and solutions are more
complex.
▶ For testing, we can sometimes use residual plots like those we saw
for incorrect functional form to test for omitted variables
▶ Logic is again that the residuals will be correlated with fitted values
of dependent variable, and these patterns may be visible graphically.

45 / 64
Testing for Omitted Variables: Residual Plots
Some examples of residual plots from correctly specified models:
800

600

400
Residuals
200

-200

-400

-600
-1000 -500 0 500 1000 1500 2000 2500 3000
Fitted Values for Y
300

200

100
Residuals

-100

-200

-300

-400
-200 0 200 400 600 800 1000 1200
Fitted Values for Y

46 / 64
Testing for Omitted Variables: Residual Plots
Some examples of residual plots from incorrectly specified models:
25

20

15

Residuals 10

-5

-10

-15

-20
-500 -400 -300 -200 -100 0 100 200 300 400
Fitted Values for Y
100

50

0
Residuals

-50

-100

-150

-200

-250
-5 0 5 10 15 20 25 30
Fitted Values for Y

47 / 64
Testing for Omitted Variables

▶ As always, graphical checks may be imprecise or subjective, so what

other options are available?
▶ The previous RESET test can sometimes be used, but it is primarily
a test of functional form, and will not pick up many forms of
omitted variables.

▶ Often the best approach is to compare the results obtained from the
model to what would be predicted by relevant financial or economic
theory.
▶ If the results from the model are inconsistent with those implied by
theory, this could be because something is wrong with the model we
have used - omitted variables could be one reason.
▶ For example, this may result in estimated coefficients with the
opposite sign to that implied by theory, or variables that are
theoretically important having statistically insignificant coefficients.

48 / 64
Omission of Relevant Variables: Solutions

▶ Suppose that (either based on theory or tests) we think that there is

a problem with omitted variable bias in our model - what solutions
are available to correct the problem?

▶ The ideal solution is to obtain data on the omitted variables and

then include them in the regression model as explanatory variables
▶ We may need to refer to theory, empirical evidence or other
knowledge to identify what these additional explanatory variables
should be

▶ Assuming that there are no additional omitted variables that are

correlated with the explanatory variables, the estimation and testing
methods we learned should again be reliable

49 / 64
Omission of Relevant Variables: Solutions
▶ However, this ideal solution can only be used if we know what the
omitted variables are, and can obtain data for them.
▶ Generally the reason a variable was omitted was either lack of
information (not knowing the variable was relevant) or a lack of data
for the variable.
▶ If we cannot obtain data, then we can instead try to identify one or
more suitable control variables that we can obtain data for.
▶ The key requirement for a control variable is that it must be
correlated with the omitted variable (although it may or may not
have a direct effect on the dependent variable)
▶ Including suitable control variables as explanatory variables will
reduce the correlation between the error term and the explanatory
variables and therefore reduce or eliminate the omitted variable bias

▶ We won’t be able to directly estimate the effect of the omitted

variable on the dependent variable, but we can at least obtain more
reliable estimates and tests for the coefficients on the other variables
50 / 64
Omitted Variables: An Example

▶ Let’s now look at an example of omitted variables in a regression

model, to illustrate the consequences and detection of the problem
▶ Here our aim is to forecast the GBP/USD forward discount rate and
the model we use is:

fwdisct = β0 + β1 UK 3mt−1 + β2 US3mt−1 + ut

where fwdisc is the 6-month GBP/USD forward discount rate

relative to the spot exchange rate, UK 3m and US3m are the UK
and US 3-month interest rates

51 / 64
Omitted Variables: An Example

▶ These two explanatory variables have been chosen based on financial

theory, which suggests that the forward discount for a particular
exchange rate should be determined primarily by the interest rates in
the two countries
▶ When we estimate this model using monthly data from 1979 to 2015
we obtained the following results:

fwdisct = 0.12 − 0.75UK 3mt−1 + 0.83US3mt−1

(1.74) (−42.78) (40.05)
2
R = 0.82

▶ Both coefficients are highly statistically significant and the model R 2

of 0.82 is relatively high

52 / 64
Omitted Variables: An Example

▶ Suppose now that we omitted the UK interest rate from our model:

fwdisct = β0 + β2 US3mt−1 + ut

▶ Estimating this model using the same data we obtain:

fwdisct = − 1.43 + 0.05US3mt−1

(−10.04) (2.25)
2
R = 0.01

▶ Although the estimated coefficient on the US interest rate is still

statistically significant at the 5% significance level (t-stat 2.25), it’s
significance has dropped substantially (previous t-stat was 40.05)
▶ In addition, the estimated value has changed substantially from 0.83
to 0.05 and the R 2 has dropped massively to 0.01

53 / 64
Omitted Variables: An Example

▶ Theory suggests that this second model will be misspecified because

it omits a key explanatory variable that is known to be relevant
(UK 3m)
▶ This will lead to omitted variable bias if the omitted variable UK 3m
is correlated with the included explanatory variable US3m
▶ Calculating the correlation coefficient for our sample of data, we
obtain a value of 0.88 - remember that 1 implies perfect positive
correlation and zero implies no correlation
▶ Therefore, the omitted variable is indeed highly correlated with the
explanatory variable and so we would expect OVB to be a problem
▶ This explains the large change in results between the two models
and suggests that estimated coefficient values and hypothesis tests
using the second model are likely to be unreliable

54 / 64
Omitted Variables: An Example

▶ In this case the results from the estimated model suggest

immediately that something may be wrong
▶ Furthermore, when we compare our model with theory we also see
that we have omitted a variable that is expected to be an important
explanatory factor
▶ In this case therefore it is relatively certain that we have a problem
with omitted variable bias

▶ However, if we wanted to check further for OVB we could do

additional testing
▶ For example, the residual plots showing the fitted values against the
residuals are shown on the following two slides

55 / 64
Omitted Variables: An Example

Below is the residual plot (using fitted values of Y ) for the first model
that includes both UK 3m and US3m:
5

2
Residuals

-1

-2

-3

-4
-8 -6 -4 -2 0 2 4 6
Fitted Values for Y

56 / 64
Omitted Variables: An Example

Here is the residual plot (using fitted values of Y ) for the second model
that omits UK 3m:
6

2
Residuals

-2

-4

-6
-1.5 -1.4 -1.3 -1.2 -1.1 -1 -0.9 -0.8 -0.7 -0.6 -0.5
Fitted Values for Y

57 / 64
Omitted Variables: An Example

▶ Finally It is important to note too that the model that includes both
UK 3m and US3m may still omit other relevant variables
▶ As we discuss next week, the model may also have forms of
misspecification, such as incorrect functional form that we need to
check for
▶ Therefore, whenever we are interpreting regression results we should
always be considering and checking for a range of different forms of
specification error

58 / 64
Overview

Introduction and outline for the topic

Heteroskedasticity

Autocorrelation

Incorrect functional form

Omitting relevant or including irrelevant variables

Measurement error
Measurement error
▶ Often financial and economic data are measured with some error
and so we do not observe the true values of the variables.
▶ This is known as measurement error or errors-in-variables and
can occur in a variety of circumstances, e.g.
▶ Macroeconomic variables are almost always estimated quantities
(GDP, inflation, and so on), as is most information contained in
company accounts.
▶ Sometimes we cannot observe or obtain data on a variable we require
and so we need to use a proxy variable - for instance, many models
include expected. quantities (e.g., expected inflation) but we cannot
typically measure expectations.
▶ In survey data, respondents may intentionally or unintentionally give
incorrect or incomplete responses.

▶ If there is measurement error in one or more of the explanatory

variables, this will violate the assumption that the explanatory
variables are non-stochastic
▶ The OLS estimator will generally be biased and inconsistent.

59 / 64
Measurement error in the explanatory variable(s)
▶ Suppose that we wish to estimate a model containing just one
explanatory variable, xt :

yt = β0 + β1 xt + ut

where ut is a disturbance term.

▶ Suppose further that xt is measured with error so that instead of
observing its true value, we observe a noisy version, x̃, that
comprises the actual xt plus some additional noise, vt that is
independent of xt and ut :

x̃t = xt + vt
▶ Taking the first equation and substituting in for xt from the second:

yt = β0 + β1 (x̃t − vt ) + ut
▶ We can rewrite this equation by separately expressing the composite
error term, (ut − β1 vt )

yt = β0 + β1 x̃t + (ut − β1 vt )
60 / 64
Measurement error in the explanatory variable(s)
▶ It is clear from this equation and the definition of x̃t , that x̃t and the
composite error term, (ut − β2 vt ), are correlated since both depend
on the measurement error vt .
▶ Therefore, the requirement that the explanatory variables are
non-stochastic does not hold, causing the OLS estimator for the
coefficients to be biased and inconsistent.

▶ The size of the bias in the estimates will be a function of the

variance of the noise in xt (i.e. the variance of vt ) as a proportion of
the overall disturbance variance.
▶ For the simple case of one explanatory variable, the parameter
estimate will always be biased towards zero as a result of the
measurement noise.
▶ With multiple explanatory variables, the direction of the bias is
usually difficult to determine.
▶ Furthermore, with multiple explanatory variables, measurement error
in a single variable affects the coefficients for all variables.

61 / 64
Measurement error in the dependent variable

▶ What about the case where all explanatory variables are measured
correctly, but the dependent variable Y is measured with error?
▶ In this case the effects are different from measurement error in the
explanatory variable(s) and are usually less serious.
▶ Using the same notation as before, let’s assume that we cannot
measure and record Y exactly, but instead measure Ỹ .
▶ Therefore, instead of estimating the true model:

Yi = β0 + β1 Xi + ui

we estimate:
Ỹi = β0 + β1 Xi + vi

62 / 64
Measurement error in the dependent variable

▶ As long as the measurement error in Y (i.e. Ỹi − Yi ) is

uncorrelated with the explanatory variables, then the error term
vi will also be uncorrelated with the explanatory variable X .
▶ This would be the true for example if we had classical measurement
error in Y so that Ỹi = yi + wi , where w is completely random.

▶ Therefore, our usual coefficient estimators are reliable in the

sense that they are unbiased (and consistent).
▶ However, the coefficient standard errors will be higher than if we
could measure true values Y and so measurement error in the
dependent variable results in less precise estimates of the
coefficients.

▶ Note: more generally, if the form of measurement error in Y is such

that w is correlated with X then our estimators will again be biased.

63 / 64
Solutions to measurement error

▶ The best solution is to obtain more accurate data for the

explanatory and/or dependent variable, but this may be difficult,
costly or even impossible.

▶ If better data are not available then more complicated solution

methods are available, but we do not look at these approaches in the
current module.
▶ One option is instrumental variables regression - here we can use an
additional variable that is correlated with the true value of X , but
uncorrelated with the measurement error, to separate these two
effects.
▶ Alternatively, it is sometimes possible to develop a model of the
measurement error and can then estimate the bias and adjust the
coefficient estimates accordingly.

64 / 64