UNIT 9 ERRORS IN VARIABLES

Structure

9.0 Objectives
9.1 Introduction

i'. 9.2 Consequences of Errors in Variables

9.2.1 Measurement Errors in Y
I
9.2.2 Measurement Errors i- X
9.2.3 Measurement Errors 11, both X and Y
a
:
9.3 Instrumental Variables Method
9.4 Test of Measurement Errors
g. l
I
% , 9.5 Inverse Regression
d 1
ri:, 9.6 Let Us SurnL.Jp
-1
i
...
@
61
2 .
1 9.7

9.8
9.9
Key Words
some Useful BookslReferences
AnswersIHints to Check Your Progress Exercises

9.0 OBJECTIVES.
After going through this unit, you should be in a position to:

I , explain the concept of errors in variables;

indetify the consequences of errors in variables;
explain the concept of instrumental variables;
I
, test for the presence of measurement errors; and
take remedial measures to solve for the problem of measurement error.

9.1 INTRODUCTION - -

In ordinary least squares model we assume that sample observations are measured
accurately. All our formulae are based upon the presumption ttat variables (both
explained and explanatory) are measured without error. The only form of error
admitted to our model is in the form of disturbance term. Here the error term
represents the influence of variohs explanatory variables that have not accurately
been included in the model. However, the assumption may not be realistic,
particularly in the case of secondary data. \

Variables, both dependent and independent, are measured subject to error. In

particular, the available data may not refer to the variable as specified, as in the
case of proxy variable, or there may be systematic biases in the collection or
publication of data. If the measurement errors are systematic, in general, auxiliary
equations can be specified to capture these errors. This unit will focus only on the
impact of random measurement errors on the regression model.
Treatment of Violations of
Basic Assumptions 9.2 CONSEQUENCES OF ERRORS IN VARIABLES
Most of the published data or summary information contain errors of summsrizing
or misrepresenting by informer. When these data are used, one of the assumptions
of classical least-squares method is violated. In this case, the classical least-squares
estimator will be biased even when sample size increases, which is alternatively
known as asymptotic biases or inconsistency.

Under the classical assumptions the ordinary least squares (OLS) estimators are best
linear unbiased. One of the major underlying assumptions is the interdependence of
regressors fiom the disturbanceterm. If this condition does not hold, OLS estimators
are biased and inconsistent. This statement may be illustrated by simple errors in
variables.

We discuss about the consequences in the following, if the error appears in the
measurement of dependent variable, independent variable or both.

9.2.1 Measurement Error in Y

Let us assume that only the dependent variable contains error of measurement.
Assume that the true regression model (written in deviation forn~)is

Y, = Px,+E,- ...(9.1)

where E, represents errors associated with the specification of the model (the effects
. of omitted variables, etc.).

Assume that the variable y e , instead.of y, is obtained in the measurement process

such that ,

The measurement error ul is not associated with the regressor. Thus we have

c o v (u,, x,) = 0

The regression model is estimated with y*as the dependent variable, with no accoun
being taken of the fact that y' is not an accurate measure ofy. Therefore, instead o
estimating Eq. (9. I), we estimate

or, y*.= P x, + + ~1) ,/

or, y*=/3xl+el . ...(9.3)

where el is ? composite error term, containing the population disturbance term E,

(which may be called the error term in the equation) and the measurement error
term ul.

For simplicity let us assume that E(E,)= E(u ,) = 0, COV(x, ) = 0 (which is the
assumption of classical linear regression), and Cov (x,, u , ) = 0, i.e., the errors of
 Errors in Variables
measurement in y' are uncorrelated with x , and Cov (c, , u, ) = 0, the equation
error and measurement error are uncorrelated.
With these assumptions, it can be shown that p estimated from either (9.1) or (9.3)
will be an unbiased estimator of the true P. Thus, the errors of measurement in
dependent variable y' do not destroy the unbiased property of the OLS estimators.
However, the variances and standard errors of P estimated from (9.1) and (9.3)
will be differentbecause, employing the usual formula, we obtain

Qbviously, the varianoe given at 19.5) is larger than the variance given at (9.4).
Therefore, although the errors of measurement in the dependent variable still give
unbiased estimates of the parameters and their variables, the estimated variances
are now larger than the case where there are no. such errors of measurement.

9.2.2 Measurement Error in X

We have taken the true regreasion model deviation form to be

Now let us assume that explanatory variable xl is measured with error and the
observed value becomes x,' such that

or, Xi = Xi - vi
Putting the value of x, in (9.6), we have

where w,= E, - fll.We have assumed earlier that the measuiement error in x is
n o d b distributed with zero mean, it has no serial conelati&, and it is uneornlated
with E, . However, we can no longer assume that the composite error term w, is
independent of the explanatory variable x,*.
Treatment of Violations of
Basic Assumptions Cov (w,, xi') = E [ w , - E(w,)] [ x,*- E( xi*)]

The composite error term wl has mean zero as E (wl) = E[E,- @,=I E(q>- PE(v)=O.
Thus, the explanatory variable and the error term are correlated, which violates the
cpcial assumption of the classical linear regression model that the explanatory
variable is uncorrelated with the stochastic disturbance term. If this assumption is
violated, then OLS estimates are not only biased but also inconsistent; they remain I

biased asymptotically. Let us now show this.

I
Instead of being able to estimate true relationship ( 9 .l), we are forced to replace xl I
I
by x,*. Now the OLS estimator is I

C ( x , + v1)(Bx1+&/I
-
( x , + v,)?

From above we find that B is a biased estimateas E( p) t P. Let us see the asymptotic
properties of b.
Since v, and E, are stochastic and are uncorrelated with each other as well as with x,,
we can say that

Py-
XI2
plim b =plim - c v,?
X12 +
 Errors in Variables
- PVflr(x>
Var(x) + a,,2 (Since x, = X - Fby definition)
-
- Pax2
2 2
0, +a,,

Thus, b is biased even for ah infinite sample and b is an inconsistent estimator of

0.
9.2.3 Measurement Errors in both X and Y
Now let us assume that both X and Y have errors of measurement. The true model
is as before

Since X and Y have errors of measurement, we observe x,' k d y,' instead of xl and
yl, such that
* .

where u, and vl present the errors in the values ofyl and xlrespectively. We make the
following assumptions about the error terms:

(i) There is no correlation between the error term and correspondingvariable, i.e.,

(i) There is no correlation between error of one variable and measurement of the
other variable, i.e.,

(i) There is no correlation between errors in measurement of both the variables,

i.e.,

On the basis of the above assumptions, our estimated regression equation will be

' Yl = P XI

or, Y,* - u, = p (x,'- v,)

or, yl' = p xl*- P vl+ u,

~reatmentof Violations of
Basic Assumptions or, yl* =$x: + (u,- pvl)

Equation (9.8) shows that if we model y* as a function of x', the transformed

disturbance contains the measurament error in P . We then write the OLS estimate.

By substiuing the values of x* and y* in terms of x and y we find that

From above we observe that E( p) + 8.

plim 4 = plim c.:a x+cv: X:

-
(x)
~ a(x)r + cv2

a
Thus, will not be a consistent estimator of p. The presence of measurement error
ofthe type in question will lead to an underestimate of the true regressiorrparameter
if ordinary least squares are used.

A specific illustration of this problem in economics is thepermanent income theory

of the consumptionfirnction, which is an errors-in-variables model. According to
this model meaqured income. YI' . has a nermanent (true) cnmnnnent Y and n
1 transitory component, YEsuch that Errors in variables'

Y '= Y + YE ...(9.10)
I

I Measured consumptionexpenditure, C,similarly, has a permanent (true) component,

I C, and a transitory component, CEsuch that .

As you know form MEC-002, Block 4, the permanent consunlption (C) is

proportional to permanent income (Y). Due to tr'ansitary components, however, the
average propensity to consume declines as income of a household increases. Let us
analyse this issue through errors in variables.
The transitory componentsrepresent accidental or chance factors, including cyclical
variations. They can be treated as the errors in measuring income and consumption.
It is usually assumed that errors, on the average, vanish:

It is also usually assumed that the errors are correlated neither with each other nor
with the permanent components:

Cov (Y, Yc) = Cov (C, CE)= Cov (YE)C) =0 ...(9.13)

'i I
L I According to permanent-income hypothesis,

C = p l y + p,+ u ...(9.14)
that is, permanent consumption is a linear function of permanent income, where
p, and p, are constant parameters and u is a stochastic disturbance term. Notice
that the parameter p, in (9.14) represents marginal propensity to consume (MPC).
Combining (9.14) with (9.10) and (9.1 1) yields

C = p l Y + / ? 2 + ( ~ - P I Y & ) where
, v = u+C
= .
...(9.15)
Equation (9.15) is a simple linear regression model of the same form as (9.8). Then,
r

plim B ,< PI (as we know from (9.9))

*
Since the ratio PIis less than unity, we observe that the above term is less than the
true parameter.

Thus,.the estimated marginal propensity to consume, ,, even in the probability

limit, systematicallyunderestimates the true PI.The greater the variance of transitory
1 income, the greater will be the underestimation. Only when the variance of transitory
(error) component of income is zero, the least squares estimator of PIis a cobistent

i
!
estimator. To solve this problem, it is necessary to construct a measure of permanent
income Y before estimating the consumption function.
Check Your Progress 1
i 1) Explain the concept of errors in variable. What are its consequences?
I
Treatment of Violations of .................................................................................................................
Basic Assumptions

2) State whether the following statements are true or false.

a) Errors in variables lead to estimates of the regression coefficients that are
biased.
b) The presence of measurement error in an equation leads to an underestimate
of the true regression parameter if ordinary least squares are used.

3) Explain briefly why measurement eiror in the explanatory variables leads to

biased and inconsistent parameter estimates while measurement error in the
dependent variable does not.
I

9.3 INSTRUMENTAL VARIABLES MI *'ElOD

The problem of errors in the measurement of variables in regression models is quite
I
important, yet econometricians do not have much to offer in the way of qseful ,I
solutions. As ageneral rule, we tend to pass over the problem of measurement error,
hoping that the errors are small enough which will not destroy the validity of the
1
estimation procedure. One technique which is available and can solve the 1
measurement error problem is the technique of instrumental variablss estimation. I
We briefly outline the concept of instrumental variables because it is likely to b e 1.
useful with measurement errors.

The method of instrumental variables involves the search for a new variable Zwhich
II
is highly correlated with the indepe ' - . variableXand at the same time uncorrelated
with the error term in the equation ':: M 11.as the errors of measurement of both
variables). In practice, we are conceined , ith the consistency of parameter estimates
and therefore concentrate on the relationship between the variable Z and the
remaining variables in the model when the sample size gets large. We define the
random variabl, Z to be an instrument if the following conditions are met.

1) The correlations between Z and E, u, and v, respectively, in the regression

approach zero, as the sample size gets large.
2) The correlation bet&een Z and Xis nonzero, as the sample size gets large.
We simply select one instrument (or combination of instruments)that has the highest Errors in Variables
correlation with the X variable.

Assuming for the moment that such a variable can be found, we can alter the l e ~ s t
squares regression procedure to obtain estimated parameters that are consistent.
Unfortunately, there is no guarantee that the estimation process will yield unbiased
parameter estimates.To simpliG the matter, let us consider the case of measurement
errors in the independeqtvariable such thaty, = px, + ci an.! only x ismeasured with
error (as x*=x + v). In order to solve the problem we take the regression equation
y, = pz, + E, Where Z is the instrumental variable. The instrumental variables estimator
of the regression slope in the above model is

The choice of this particular slope formula is made so that the resulting estimator
will be consistent. To see this, we can derive the relationship between the
instrumental-variablesestimator and the true slope parameter as follows:

p
Clearly, the choice of Z as an instrument guarantees that will approach P as the
sample size gets large [Cov (z, E*) approaches 03 and will therefore be a consistent
estimator of 0. Remember that the variable x,* in (9.16) was not replaced by z, in

the denominator of instrumental variables estimator, as the estimator

C ~i.1
does
L" 1
not yield a consistent estimator of P. Let us show this.
 --
Treatment of Violations of
Basic Assumptions

-PCx,z,+ PCz,vl+ C z~&,*

C zi2
plim jj = plim
PCx1z.
C2
.
'

The above shows that

C Y I ~ I

C.will not yield a consistent estimator of P.

2

The technique of i n s k e n t a l variables appears to provide a simple solution to a

difficult problem. We have defined an estimation technique, which yields consistent
estimates if we can find an appropriate instrument.
A few concluding comments may be instructive.

i) First, the OLS. estimation technique is actually a special case of instrumental

variables, This follows because in the classical regression model X is
uncorrelated with the error term and becauseXis perfectly correlated with itself.
ii) Second, if we generalize the measurement error problem to errors in more than
one independent variable, one instrument is needed to replace each of the
designated independent variables.
iii) Finally, we repeat that instrumental-vdables estimation guarantees consistent ,

estimation but does not guarantee unbiased estimation.

- - -- --- - -- -- - -

9.4 TEST OF MEASUREMENT ERRORS

Suppose we are estimating the two-variable regression model

Y, =Pxl+&, ...(9.1)
There is a possibility that x might be measured with error.

If x, = xle- vl ,then the actual least squares regression would be

 Errors in vgriables
where E,* = - Pv,.

If x is measured with error, we have seen that consisent estimator of p can be obtained
by using an instrument z which is correlated with x' but uncorrelated with E And v.
Suppose the relationship between z and x* is given by

When (9.18) is estimated using least squares, we obtain

'
where ?I, are the regression msidwls. Substituting the value of eq. (9.19) into eq.
(9.17) we have

Whether or not there is measwemsnt error, the coefficient of il*

will be consistently
estimated by ordinary laast squares, since

plim 2 il*el' - plim iC zl (cl- PV, ) = 0.

N N
I
In fact, the least-squares estimator of the coefficient of il*
in eq. (9.20) is identically
equal to the instrumental-variablesestimator, which is given by

To look at the coefficientof the variable G,,note that

= plim
- P z v ; ( x 1 +a)
N
= -pay2

When there is no melwrewnt error, av2

= 0, so that OLS applied to eq. (9.20)

will generate a consistent estimator of the coefficient of GI.However, when there

is measurement error, the coefficient will be estimated inconsistently. '
ll& suggests a relatively easy measurement error test. Let Srepresent the coefficient
of the variable GI in eq. (9.20).
Treatment of Violations of
Basic Assumptions
Substituting ii'= x,* - Gi, we get

With no measurement error, S= P, so that the coefficient of G, should equal zero.

However, with measurement error, S+ p, and the coefficient will (ingeneral) be
different from zero. We can test for measurement error by doing a simple two-stage
procedure. First, we regress x* on z to obtain the residuals 6. Then, we regress y
onx* and ii, and perform a t test on the coefficient of the $ variable. If we are
concerned with measurement error in more than one variable of a multiple regression
model, in equivalent F test could be applied.
The test just described is a special case of a more general test for specification error
proposed by Hausman. The Hausman specification test relies on the fact that under
the null hypothesis the ordinary least-squares estimator of the parameters of the
original eq. (9.17) is consistent and (for large samples) efficient, but is inconsistent
if the alternative hypothesis is true. However, the instrumental-variablesestimator
[the least-squares estimator of eq. (9.20)] is consistent whether or not the null
hypothesis is true, although it is inefficient if the null hypothesis is not valid.

8,
The Hausman specificationtest is as follows: If there are two estimators and $,
that converge to the true value p under the null but converge to different values
under the alternative, the null hypothesis can be verified by testing whether the
probability limit of the difference of the two estimators is zero,
Example 9.2

The expenditure of states and local governments (EXP) in US vary substantially by

state tind region. Among the important variables that explain differences in spending
levels are federal grants-in-aid (AID), the income of states (MC), and the population 1I
of states (POP). When a model that relates the dependent variable EXP to the
independent variablesAID, INC, and POP was estimated by ordinary least squares, 4
using census data for 50 states, the following results were obtained (with t statistics
in parentheses):

\ EXP = -46.81 + 0.00324 AID + 0.00019 INC - 0.597 POP

(-0.56) (13.64) (8.12) (-5.71)
R2 = 0.993 F = 2190
There is an important possible source of measurement error in the AID variable.
This component can be subjedt to substantial error as the state aid programmes
involve fixed sums of money, and the money involved are easy to measure even before
stak and local budgets are set.
We can test the presence of measurement error by using a Hausman specification
. test. To perform the test, we use the population of primary and secondary school
.children (PS) as an instrument. The test proceeds in two stages. In the first stage
AID is regressed on PS, and the residual variable 6, is calculated as follows:
 Errors in Variables
GI = AID - 77.95 + 0.845 PS

I1 h the second stage % is added to the original regression to correct for measurement
error. The resulting eqpatidrl is
I
EXP = -138.51 + 0.00174 AID + 0.00018INC - 0.275 POP + 1.372 it,

A two-tailed t test of the null hypothesis that there is no measurement error would
be accepted at the 5 per cent level, since 1.73 < 1.96. However, measurement error
would be important if we were using either a one-tail test, or a two-tailed test at 10
per cent significance level. Note that cokrecting for the possibility of measurement
error has substantiall) lowered the coefficient on the AID variable, suggesting that
measurement error causes the effect of AID on spending to be overstated.

9.5 INVERSE REGRESSION

-

When we have the variablesy and x both measured with errors (the observed values
being y* and x*),we consider two regreesion equations:

1) 'Regression of y' on x', which is called the "direct" regression.

2) Regression of x* on y*,which is called the "inverse" or "reverse" regression.
Inverse regression bas been frequently advocated in the case of analysisbf salary
discrimination. Since the problem is one of usual errors in variables, both regressions
need to be computed and whefier inverse regression alone gives the correct estimates
depends on the assumptions we make.
The usual model, in its simplest form, is that of two explanatory variables, one of
which is measured with error

Y P J ~ +, E ...(9.22)
-
= PIX, +

where y salary
x, =true qulifications
x, - gender

What we are interested in is the coefficient ofx2.The problem is that x , is measured

with error.

Let xi*= measured qualifiations

xi*= x i + vi

Suppose we adopt the notion that

I, for men
-
X2 -
0, for women LC) I
 -

Treatment of Violations of Thgh A > 0 implies that men are paid mdre than women'wth the same qwifi~ations
Basic Assumptions and thus there is gender discrimination. Adirect least squares estimation of eq. (9.22)
with x,' substituted for x,, and j2> 0 has been frequently used as eviden& of gender
. discrimination.
In the inverse regression we take

x,' = Y,Y + Y p 2+ W ..,(9.23)

We are asking whether mep at.e m e o r less qualified than women h a ~ thesame~ g
salaries. The proponents of inverse qgygssion argue that to estarbllsh discrimination,
one has to have 3,< 0 in eq. (9.23);that is, among men and women receiving equal
salaries, men possess lower qualifications.
The usual errots in variables results however show that we shoul$ not make
inferences on the basis of B2
and f 2 but obtain bounds for p,, from the direct
regression and the inverse regression estimates.

Check Y ~ u Progress
r 2
..
Cy1z,
1) . Show that in the two-variable model B =7
Czl
.(where z is an instrument)
will not yield a consistent estimate of the true slope parameter,
.................................................................................................................

..............%t...........................................r.....................................................
2) What is,theHausmw specification test? How would you carry out a Hausman
-
specification test to evaluate the presence or absence of measurement error?
...................................................................................................................

LET US SUM UP
In ordinary least squares model we assume that sample observations are measured
without error, which is always not true. When this assumption does not hold, OLS
estimators are biased and inconsistent. Errors may appear in the measurement of
dependeit variable, indepehdent vhable or both. When there is error in dependent
variable, this does not destroy the unbiased property of the OLS estimators but the
estimated variances are larger than the case where there is nQ such errors of
measurement. On the other hand, when there is error in the independent variable,
or both dependent and independent variables, then OLS estimates are not only biqed
hut also inconsistent.
 We can test for the presence of measurement error by using Hausman specification Errors in Variables
test. One technique which is available arid can solve the measurement error problem
is the technique of instrumental variables estimation.

9.4 KEY WORDS

Asymptotic bias : When the classical least-squares estimator is biased even
with the increase in sample size, it is known as asymptotic
bias or inconsistency.
- 1

Equation error : It represents the influence of various explanatory

variables that have not accurately been included in the
model and is in the form of disturbance term.
Errors in Variables : When errors are found in the measurement of variables,
we say that there are errors in variables. These errors
can substantially alter the properties of the estimated
regression parameters.
Explanatory variable : It is the independent varkble .which explains the
dependent variable.
Instrumental variable : It is a random variable, which is highly correlated with
the independent variable and at the same time
uncorrelated with the error term in the equation.
, Inverse regression : It is the reverse of direct regressioh. If we consider
regression of Y on X as direct regression, then regression
of X on Y is inverse regression.
heasurement error :When either dependent variable, or independent variable
- .I
or both, are measured subject to error, we say that there
is measurement error.
Parameter : It is a measure of some characteristicsofthe population.
Proxy variable : It is a variable, which is proxy for the true variable.
Specification error : Errors found in the specification of the model.
Stochastic disturbance : It is the random error term whose\value is based on an
term underlying probability distribution.

9.8 SOME USEFUL BOOKSIREFERENCES

Intriligator, Michael, Ronald Bodkin and Cheng Hsiao (1996); Econometric Models,
Techniques, and Applications; Prentice-Hall International Inc., New Jersey.
Johnston, Jack and John Dinardo (1997); Econbmetric Methoh, The McGraw-Hill
Companies Inc., Singapore.
Maddala, GS. (2002); Introduction to Econometrics, John WiIey & Sons Ltd., Delhi.
t and Daniel L. Rubinfeld (1998): Econometric Models and
Pindyck, ~ o t e r S.
Economic Forecasts, IrwinMcGraw-Hill, singapo&. .
Treatment of Violations of
Basic Assumptions 9.9 ANSWERS/HINT$ TO CHECK YOUR PROGRESS
EXERCISES
Check Your Progress 1

2) a) true b) true.

3) Go through Section 9.3

i

Check Your Progress 2

1) Gd through Section 9.4.

2) Go through Section 9.5 and Example 9.2.