UNIT 8 MULTIPLE LINEAR REGRESSION

MODEL: INFERENCES 
Structure
8.0 Objectives
8.1 Introduction
8.2 Assumptions of Multiple Regression Models
8.2.1 Classical Assumptions
8.2.1 Test for Normality of the Error Term
8.3 Testing of Single Parameter
8.3.1 Test of Significance Approach
8.3.2 Confidence Interval Approach
8.4 Testing of Overall Significance
8.5 Test of Equality between Two Parameters
8.6 Test of Linear Restrictions on Parameters
8.6.1 The t-Test Approach
8.6.2 Restricted Least Squares
8.7 Structural Stability of a Model: Chow Test
8.8 Prediction
8.8.1 Mean Prediction
8.8.2 Individual Prediction
8.9 Let Us Sum Up
8.10 Answers/ Hints to Check Your Progress Exercises

8.0 OBJECTIVES
After going through this unit, you should be able to
 explain the need for the assumption of normality in the case of multiple
regression;
 describe the procedure of testing of hypothesis on individual estimators;
 test the overall significance of a regression model;
 test for the equality of two regression coefficients;
 explain the procedure of applying the Chow test;
 make prediction on the basis of multiple regression model;
 interpret the results obtained from the testing of hypothesis, both individual
and joint; and
 apply various tests such as likelihood ratio (LR), Wald (W) and Lagrange
Multiplier Test (LM).


Dr. Pooja Sharma, Assistant Professor, Daulat Ram College, University of Delhi
Multiple Regression
Models
8.1 INTRODUCTION
In the previous unit we discussed about the interpretation and estimation of
multiple regression models. We looked at the assumptions that are required for
the ordinary least squares (OLS) and maximum likelihood (ML) estimation. In
the present Unit we look at the methods of hypothesis testing in multiple
regression models.
Recall that in Unit 3 of this course we mentioned the procedure of hypothesis
testing. Further, in Unit 5 we explained the procedure of hypothesis testing in the
case of two variable regression models. Now let us extend the procedure of
hypothesis testing to multiple regression models. There could be two scenarios in
multiple regression models so far as hypothesis testing is concerned: (i) testing of
individual coefficients, and (ii) joint testing of some of the parameters. We
discuss the method of testing for structural stability of regression model by
applying the Chow test. Further, we discuss three important tests, viz.,
Likelihood Ratio test, Wald test, and Lagrange Multiplier test. Finally, we deal
with the issue of prediction on the basis of multiple regression equation.
One of the assumptions in hypothesis testing is that the error variable 𝑢 follows
normal distribution. Is there a method to test for the normality of a variable? We
will discuss this issue also. However, let us begin with an overview of the basic
assumptions of multiple regression models.

8.2 ASSUMPTIONS OF MULTIPLE REGRESSION

MODELS
In Unit 7 we considered the multiple regression model with two explanatory
variables 𝑋 and 𝑋 . The stochastic error term is 𝑢 .
𝑌 =𝛽 +𝛽 𝑋 +𝛽 𝑋 +𝑢 … (8.1)

8.2.1 Classical Assumptions

There are seven assumptions regarding the multiple regression model. Most of
these assumptions are regarding the error term. We discussed about these
assumptions in the previous Unit. Let us briefly mention those assumptions
again.
a) The regression model is linear in parameters and variables.
b) The mean of error terms is zero. In other words, the expected value of
error term conditional upon the explanatory variables X2i and X3i is zero.
𝐸 (𝑢 )= 0 or 𝐸 (𝑢 |𝑋 , 𝑋 ) = 0
c) There is no serial correlation (or autocorrelation) among the error terms.
The error terms are not correlated. It implies that the covariance between
the error term associated with ith observation 𝑢 and the error term
associated with jth observation, 𝑢 is zero.
100
 cov 𝑢 , 𝑢 =0 Multiple Linear Regression
Model: Inferences
d) Homoscedasticity: The assumption of homoscedasticity states that the
error variance is constant throughout the population. The variance of the
error term associated at each observation has the same variance.
𝑣𝑎𝑟 (𝑢 ) = 𝜎
e) Exogeneity of explanatory variables: There is no correlation between the
explanatory variables and the error term. This assumption is also called
exogeneity, because the explanatory variables are assumed to be
exogenous (given from outside; X is not determined inside the model). In
contrast, Y is determined within the model. When the explanatory
variable is correlated with the error term, it is called endogeneity problem.
In order to avoid this problem, we assume that the explanatory variables
are kept fixed across samples.
f) Independent variables are not linear combination of one another. If there
is perfect linear relationship among the independent variables, the
explanatory variables move in harmony and it is not possible to estimate
the parameters. It is also called multicollinearity problem.
g) The error variable is normally distributed. This assumption is not
necessary in OLS method for estimation of parameters. It is required for
construction of confidence interval and hypothesis testing. In the
maximum likelihood method discussed in the previous Unit, in order to
estimate the parameters we assumed that the error term follows normal
distribution.
8.2.2 Test for Normality of the Error Term
As pointed out earlier, we look into the assumption of normality of the error
term. In order to test for normality of the error term we apply the Jarque-Bera test
(often called the JB test). It is an asymptotic or large sample test. We do not
know the error terms in a regression model; we know the residuals. Therefore,
the JB test is based on the OLS residuals. Recall two concepts from statistics:
skewness and kurtosis. A skewed curve (i.e., asymmetric) is different from a
normal curve. A leptokurtic or platykurtic curve (i.e., tall or short in height) is
different from a normal curve. The JB test utilises the measures of skewness and
kurtosis.
We know that for a normal distribution S = 0 and K = 3. A signification deviation
from these two values will confirm that the variable is not normally distributed.
Jarque and Bera constructed the J-statistic given by
( )
𝐽𝐵 = 𝑆 + … (8.2)

where
n = sample size
101
Multiple Regression S = measure of skewness ( )
Models
K = measure of kurtosis ( )

Skewness and kurtosis are measured in terms of the moments of a variable. As

you know from BECC 107, Unit 4, the formula for calculating the rth moment of
variable 𝑋 is
μ = ∑ 𝑓 (𝑋 − 𝑋) … (8.4)

Variance is the second moment μ .

In equation (8.2) the JB statistic follows chi-square distribution with 2 degrees of
freedom, ~𝜒( ) .
Let us find out the value of the JB statistic if a variable follows normal
distribution. For the normal distribution, as mentioned above S = 0 and K = 3. By
asubstitutingthese values in equation (8.2) we obtain
𝐽𝐵 = [0 + 0] = × 0 = 0 … (8.3)

For a variable not normally distributed JB statistics will assume increasingly

large values. The null hypothesis is
H0: The random variable follows normal distribution.
We draw inferences from the JB statistic as follows:
a) If the calculated value of JB statistic is greater than the tabulated value of
𝜒 for 2 degrees of freedom, we reject the null hypothesis. We infer that
the random variable is not normally distributed.
b) If the calculated value of the JB statistic is less than the tabulated value of
𝜒 for 2 degrees of freedom, we do not reject the null hypothesis. We
infer that the random variable is normally distributed.
Check Your Progress 1
1) List the assumptions of multiple regression models.
.......................................................................................................................
.......................................................................................................................
.......................................................................................................................
.......................................................................................................................
2) State the Jarque-Bera test for normality.
.......................................................................................................................
.......................................................................................................................
.......................................................................................................................
.......................................................................................................................

102
8.3 TESTING OF SINGLE PARAMETER Multiple Linear Regression
Model: Inferences

The population regression function is not known to us. We estimate the

parameters on the basis of sample data. Since we do not know the error variance
𝜎 , we should apply t-test instead of z-test (based on normal distribution).
Let us consider the population regression line given at equation (8.1).
𝑌 =𝛽 +𝛽 𝑋 +𝛽 𝑋 +𝑢
The sample regression line estimated by ordinary least squares (OLS) method is
𝑌 =𝑏 +𝑏 𝑋 +𝑏 𝑋 … (8.4)
where b1, b2 and b3 are estimators of 𝛽 , 𝛽 and 𝛽 respectively. The estimator of
error variance 𝜎 is given by 𝜎 = .

There are two approaches to hypothesis testing: (i) test of significance approach,
and (ii) confidence interval approach. We discuss both the approaches below.
8.3.1 Test of Significance Approach
In this approach we proceed as follows:
(i) Take the point estimate of the parameter that we want test, viz., b1, or
b2 or b3.
(ii) Set the null hypothesis. Suppose we expect that variable 𝑋 has no
influence on Y. It implies that 𝛽 should be zero. Thus, null hypothesis
is 𝐻 : 𝛽 = 0. In this case what should be alternative hypothesis? The
alternative hypothesis is 𝐻 : 𝛽 ≠ 0.
(iii) If 𝛽 ≠ 0, then 𝛽 could be either positive or negative. Thus we have
to apply two-tail test. Accordingly, the critical value of the t-ratio has
to be decided.
(iv) Let us consider another scenario. Suppose we expect that 𝛽 should be
positive. It implies that our null hypothesis is 𝐻 : 𝛽 > 0 . The
alternative hypothesis is 𝐻 : 𝛽 ≤ 0.
(v) If 𝛽 > 0, then 𝛽 could be either zero or negative. Thus the critical
region or rejection region lies on one side of the t probability curve.
Therefore, we have to apply one-tail test. Accordingly the critical
value of t-ratio is to be decided.
(vi) Remember that the null hypothesis depends on economic theory or
logic. Therefore, you have to set the null hypothesis according to
some logic. If you expect that the explanatory variable should have no
effect on the dependent variable, then set the parameter as zero in the
null hypothesis.
(vii) Decide on the level of significance. It represents extent of error you
want to tolerate. If the level of significance is 5 per cent (α = 0.05),
103
Multiple Regression your decision on the null hypothesis will go be wrong 5 per cent
Models
times. If you take 1 per cent level of significance (α = 0.01), then your
decision on the null hypothesis will be wrong 1 per cent times (i.e., it
will be correct 99 per cent times).
(viii) Compute the t-ratio. Here the standard error is the positive square root
of the variance of the estimator. The formula for the variance of the
OLS estimators in multiple regression models is given in Unit 7.
𝑡= … (8.5)
( )

(ix) Compare the computed value of the t-ratio with the tabulated value of
the t-ratio. Be careful about the two issues while reading the t-table:
(i) level of significance, and (ii) degree of freedom. Level of
significance we have mentioned above. Degree of freedom is (n–k), as
you know from the previous Unit.
(x) If the computed value of t-ratio is greater than the tabulated value of t-
ratio, reject the null hypothesis. If computed value of t-ratio is less
than the tabulated value of t-ratio, do not reject the null hypothesis
and accept the alternative null hypothesis.
8.3.2 Confidence Interval Approach
We have discussed about interval estimation in Unit 3 and Unit 5. Thus, here we
bring out the essential points only.
(i) Remember that confidence interval (CI) is created individually for
each parameter. There cannot be a single confidence interval for a
group of parameters.
(ii) Confidence interval is build on the basis of the logic described above
in the test of significance approach.
(iii) Suppose we have the null hypothesis 𝐻 : 𝛽 = 0 and the alternative
hypothesis is 𝐻 : 𝛽 ≠ 0. The estimator of 𝛽 is 𝑏 . We know the
standard error of 𝑏 .
(iv) Here also we decide on the level of significance (α). We refer to the t-
table and find out the t-ratio for desired level of significance.
(v) The degree of freedom is known to us, i.e., (n–k).
(vi) Since the above is case of two-tailed test, we take 𝛼⁄2 on each side of
the t probability curve. Therefore, we take the t-ratio corresponding to
the probability 𝛼⁄2 and the degrees of freedom applicable.
(vii) Remember that confidence interval is created with the help of the
estimator and its standard error. We test whether the parameter lies
within the confidence interval or not.
(viii) Construct the confidence interval as follows:
104
 𝑏 −𝑡 ⁄ 𝑆𝐸 (𝑏 ) ≤ 𝛽 ≤ 𝑏 + 𝑡 ⁄ 𝑆𝐸 (𝑏 ) . … (8.6) Multiple Linear Regression
Model: Inferences
(ix) The probability of the parameter remaining in the confidence interval
is (1 − 𝛼). If we have taken the confidence interval as 5 per cent, then
the probability that 𝛽 will remain in the confidence interval is 95 per
cent.
𝑃 𝑏 −𝑡 ⁄ 𝑆𝐸 (𝑏 ) ≤ 𝛽 ≤ 𝑏 + 𝑡 ⁄ 𝑆𝐸 (𝑏 ) = (1 − 𝛼) … (8.7)
(x) If the parameter (in this case, 𝛽 ) remains in the confidence interval,
do not reject the null hypothesis.
(xi) If the parameter does not remain within the confidence interval, reject
the null hypothesis, and accept the alternative null hypothesis.
Check Your Progress 2
1) Describe the steps you would follow in testing the hypothesis that 𝛽 < 0.
.......................................................................................................................
.......................................................................................................................
.......................................................................................................................
.......................................................................................................................
........................................................................................................................

2) Create a confidence interval for the population parameter of the partial

slope coefficient.
...........................................................................................................................
...........................................................................................................................
...........................................................................................................................
...........................................................................................................................
..........................................................................................................................

8.4 TEST OF OVERALL SIGNIFICANCE

The overall test of significance of a multiple regression model is carried out by
applying F-test. We have discussed about the F-test in Unit 5 of this course in the
context of two variable models. For testing of the overall significance of a
multiple regression model we proceed as follows:
(i) Set the null hypothesis. The null hypothesis for testing the overall
significance of a multiple regression model is given as follows:
𝐻 : 𝛽 = 𝛽 =. . . 𝛽 = 0 … (8.8)
(ii) Set the corresponding alternative hypothesis.
𝐻 : 𝛽 =. . . = 𝛽 ≠ 0 … (8.9)

105
Multiple Regression (iii) Decide on the level of significance. It has the same connotation as in
Models
the case of t-test described above.
(iv) For multiple regression model the F-statistic is given by
/( )
𝐹= ( )
… (8.10)

(v) Find out the degrees of freedom. The F-statistic mentioned in

equation (8.10) follows F distribution with degrees of freedom (k–1,
n–k).
(vi) Find out the computed value of F on the basis of equation (8.10).
Compare it with the tabulated value of F (given at the end of the
book). Read the tabulated F value for desired level of significance and
applicable degrees of freedom.
(vii) If the computed value of F is greater than the tabulated value, then
reject the null hypothesis.
(viii) If the computed value is less than the tabulated value, do not reject the
null hypothesis.

8.5 TESTOF EQUALITY BETWEEN TWO

PARAMETERS
We can compare between the parameters of a multiple regression model.
Particularly, we can test whether two parameters are equal in a regression model.
For this purpose we apply the same procedure as we have learnt in the course
BECC 107.
Let us take the following regression model:
𝑌 = 𝛽 +𝛽 𝑋 +𝛽 𝑋 +𝛽 𝑋 +𝑢 … (8.11)
Recall that we do not know the variance of the parameters. Thus, for comparison
of the parameters we apply the t-test. Secondly, we do not the parameters.
Therefore, we take their OLS estimators for comparison purposes.
Our null hypothesis and alternative hypothesis are as follows:
𝐻 :𝛽 = 𝛽 or (𝛽 − 𝛽 ) = 0 … (8.12)
𝐻 :𝛽 ≠ 𝛽 or (𝛽 − 𝛽 ) ≠ 0 … (8.13)
For testing of the above hypothesis, the t -statistic is given as follows:
( ) ( )
𝑡= ( )
… (8.14)

The above follows t-distribution with (n – k) degrees of freedom.

Since 𝛽 = 𝛽 under the null hypothesis, we can re-arrange equation (8.14) as
follows:
𝑡= ( ) ( ) ( )
… (8.15)
,
106
The computed value of t-statistic is obtained by equation (8.15). We compare the Multiple Linear Regression
Model: Inferences
computed value of t-ratio with the tabulated value of t-ratio. We read the t-table
for desired level of significance and applicable degrees of freedom.
If the computed value of t-ratio is greater than the tabulated value, then we reject
the null hypothesis. If the computed value of t-ratio is less than the tabulated
value, then we do not reject the null hypothesis and accept the alternative
hypothesis.
We need to interpret our results. If we reject the null hypothesis we conclude that
the partial slope coefficients  3 and  4 are statistically significantly different. If
we do not reject the null hypothesis, we conclude that there is no statistically
significant difference between the slope coefficients  3 and  4 .

Check Your Progress 3

1) Mention the steps of carrying out a test of the overall significance a multiple
regression model.
...........................................................................................................................
...........................................................................................................................
...........................................................................................................................
...........................................................................................................................
...........................................................................................................................

2) State how the equality between two parameters can be tested.

...........................................................................................................................
...........................................................................................................................
...........................................................................................................................
...........................................................................................................................
...........................................................................................................................

8.6 TEST OF LINEAR RESTRICTIONS ON

PARAMETERS
Many times we come across situations where we have to test for linear
restrictions on parameters. For example, let us consider the Cobb-Douglas
production function.
𝑌 =𝛽𝑋 𝑋 𝑒 … (8.16)
where 𝑌 is output, 𝑋 is capital and 𝑋 is labour. The parameters are 𝛽 and
𝛽 . The stochastic error term is 𝑢 . The subscript ‘i’ indicates the ith observation.
The Cobb-Douglas production function exhibits constant returns to scale if the
parameters fulfil the following condition:
𝛽 +𝛽 =1 … (8.17) 107
Multiple Regression As we have discussed in Unit 6, by taking natural log, the Cobb-Douglas
Models
production function can be expressed in linear form as
𝑙𝑛 𝑌 = 𝑙𝑛 𝛽 + 𝛽 𝑙𝑛𝑋 + 𝛽 𝑙𝑛𝑋 + 𝑢 … (8.18)
Suppose have collected data on a sample of firms; our sample size is n. The
production function is Cobb-Douglas as given above. We want to test whether
the production function exhibits constant returns to scale. For this purpose we
need to apply the F-test. We can follow two approaches as discussed below.
8.6.1 The t-Test Approach
We will discuss two procedures for testing the hypothesis.
(a) For the In this case our null hypothesis and alternative hypothesis are as
follows:
𝐻 :𝛽 + 𝛽 = 1 … (8.19)
𝐻 :𝛽 + 𝛽 ≠ 1 … (8.20)
For testing of the above hypothesis, the t -statistic is given as follows:
( ) ( )
𝑡= ( )
… (8.21)

The above follows t-distribution with (n – k) degrees of freedom.

We can re-arrange equation (8.21) as follows:
t= ( ) ( ) ( )
… (8.22)
,

The computed value of t-statistic is obtained by equation (8.22). We compare

the computed value of t-ratio with the tabulated value of t-ratio. We read the
t-table for desired level of significance and applicable degrees of freedom.
If the computed value of t-ratio is greater than the tabulated value, then we
reject the null hypothesis. If the computed value of t-ratio is less than the
tabulated value, then we do not reject the null hypothesis and accept the
alternative hypothesis.
We need to interpret our results. If we reject the null hypothesis we conclude
that the firms do not exhibit constant returns to scale. If we do not reject the
null hypothesis, we conclude that the firms exhibit constant returns to scale.
(b) Let us look again at the null hypothesis given at (8.19).
𝐻 :𝛽 + 𝛽 = 1
If the above restriction holds, then we should have
𝛽 = (1 − 𝛽 )
Let us substitute the above relationship in the Cobb-Douglas production
function
𝑙𝑛 𝑌 = 𝑙𝑛 𝛽 + (1 − 𝛽 )𝑙𝑛𝑋 + 𝛽 𝑙𝑛𝑋 + 𝑢 … (8.23)
108
We can re-arrange terms in equation (8.23) to obtain Multiple Linear Regression
Model: Inferences
𝑙𝑛 𝑌 − 𝑙𝑛𝑋 = 𝑙𝑛 𝛽 − 𝛽 𝑙𝑛𝑋 + 𝛽 𝑙𝑛𝑋 + 𝑢
Or,
𝑙𝑛(𝑌 ⁄𝑋 ) = 𝛽 + 𝛽 𝑙𝑛(𝑋 ⁄𝑋 ) + 𝑢 … (8.24)
Note that the dependent variable in the above regression model is output-labour
ratio and the explanatory variable is capital-labour ratio. We can estimate the
regression model given at equation (8.24) and find the OLS estimator of 𝛽 .
If 𝛽 = 1, then the Cobb-Douglas production will exhibit constant returns to
scale.
Therefore, we set the null hypothesis and alternative hypothesis as
𝐻 : 𝛽 = 1 and 𝐻 : 𝛽 ≠ 1
We apply t-test for individual parameters as mentioned in sub-section 8.3.1. If the
null hypothesis is rejected we conclude that the firms do not exhibit constant
returns to scale.
8.6.2 Restricted Least Squares
The t-test approach mentioned above may not be suitable in all cases. There may
be situations where we have more than two parameters to be tested. In such
circumstances we apply the F-test. This approach is called the restricted least
squares.
Let us consider the multiple regression model given at equation (8.11).
𝑌 = 𝛽 +𝛽 𝑋 +𝛽 𝑋 +𝛽 𝑋 +𝑢
Suppose we have to test the hypothesis that X3 and X4 do not influence the
dependent variable Y. In such a case, the parameters 𝛽 and 𝛽 should be zero.
Recall that if we increase the number of explanatory variables in a regression
model, there is an increase 𝑅 . Recall further that 𝑅 = . Thus, if two of the
explanatory variables in equation (8.11) are dropped (i.e., their coefficients are
zero), there will be a decrease in the value 𝑅 . If the variables that X3 and X4 are
relevant, there will be a significant decline the value of 𝑅 . On the other hand, if
the variables X3 and X4 are not relevant for the regression model, then the decline
in the value of 𝑅 will be insignificant. We use this property of the regression
model to test hypotheses on a group of parameters. Therefore, while applying F-
test in restricted least squares we estimated the regression model twice: (i) the
unrestricted model, and (ii) the restricted model.
We proceed as follows:

(i) Suppose there are k explanatory variables in the regression model.

(ii) Out of these k explanatory variables, suppose the first m explanatory

variables are not relevant. 109
Multiple Regression (iii) Thus our null hypothesis will be as follows:
Models

𝐻 :𝛽 = 𝛽 = ⋯ = 𝛽 = 0 … (8.25)

(iv) The corresponding alternative hypothesis will be that the 𝛽s are not
zero.

(v) Estimate the unrestricted regression model given at (8.11). Obtain the
residual sum of squares (RRS) on the basis of the estimated regression
equation. Denote it as RSSUR.

(vi) Estimate the restricted regression model by excluding the explanatory

variables for which the parameters are zero. Obtain the residual sum
of squares (RRS) from this restricted model. Denote it as RSSR.

(vii) Our F-statistic is

/
𝐹= … (8.26)
/( )

The F-statistic at (8.26) follows F-distribution with degrees of

freedom (m, n–k).

(ix) Find out the computed value of F on the basis of equation (8.10).
Compare it with the tabulated value of F (given at the end of the
book). Read the tabulated F value for desired level of significance and
applicable degrees of freedom.

(x) If the computed value of F is greater than the tabulated value, then
reject the null hypothesis.

(xi) If the computed value is less than the tabulated value, do not reject the
null hypothesis.
As mentioned earlier, the residual sum of squares (RSS) and the coefficient of
determination (𝑅 ) are related. Therefore, it is possible to carry out the F-test on
the basis of 𝑅 also. If we have the coefficient of determination for the
unrestricted model (𝑅 ) and the coefficient of determination for the restricted
model (𝑅 ), then we can test the joint hypothesis about the set of parameters.
The F-statistic will be
/
𝐹= … (8.27)
/( )

which follows F-distribution with degrees of freedom (m, n–k).

The conclusion to be drawn and interpretation of results will be the same as
described in points (x) and (xi) above.

110
8.7 STRUCTURAL STABILITY OF A MODEL: Multiple Linear Regression
Model: Inferences
CHOW TEST
Many times we come across situations where there is a change in the pattern of
data. The dependent and independent variables may not remain the same
throughout the sample. For example, saving behaviour of poor and rich
households may be different. The production of an industry may be different after
a policy change. In such situations it may not be appropriate to run a single
regression for the entire dataset. There is a need to check for structural stability of
the econometric model.
There are various procedures to bring in structural breaks in a regression model.
We will discuss about the dummy variable cases in unit 9. In this Unit we discuss
a very simple and specific case.
Suppose we have data on n observations. We suspect that the first 𝑛
observations are different from the remaining 𝑛 observations (we have 𝑛 +
𝑛 = 𝑛). In this case run the following three regression equations:
𝑌 =𝜆 +𝜆 𝑋 +𝑢 (number of observations: 𝑛 ) … (8.28)
𝑌 =𝑟 +𝑟 𝑋 +𝑣 (number of observations: 𝑛 ) … (8.29)
𝑌 = 𝛼 + 𝛼 𝑋 + 𝑤 (number of observations: 𝑛 = 𝑛 + 𝑛 ) … (8.30)
If both the sub-samples are the same, then we should have 𝜆 = 𝑟 = 𝛼
and 𝜆 = 𝑟 = 𝛼 . If both the sub-samples are different then there will be a
structural break in the sample. It implies the parameters of equations (8.28) and
(8.29) are different. In order to test for the structural stability of the regression
model we apply Chow test.
We process as follows:
(i) Run the regression model (8.28). Obtain residual sum of squares RSS1.
(ii) Run regression model (8.29). Obtain residual sum of squares RSS2.
(iii) Run regression model (8.30). Obtain residual sum of squares RSS3.
(iv) In regression model (8.30) we are forcing the model to have the same
parameters in both the sub-samples. Therefore, let us call the residual
sum of squares obtained from this model RSSR.
(v) Since regression models given at (8.28) and (8.29) are independent, let
us call this the unrestricted model. Therefore, 𝑅𝑆𝑆 = 𝑅𝑆𝑆 + 𝑅𝑆𝑆
(vi) Suppose both the sub-samples are the same. In that case there should not
be any difference between 𝑅𝑆𝑆 and 𝑅𝑆𝑆 . Our null hypothesis in that
case is H0: There is not structural change (or, there is parameter
stability).
(vii) Test the above by the following test statistic:
111
Multiple Regression )/
𝐹= … (8.31)
Models /

It follows F-distribution with degrees of freedom 𝑘, (𝑛 + 𝑛 − 2𝑘) ,

where k is the number of explanatory variables in the regression model.
(viii) Check the F-distribution table given at the end of the book for desired
level of significance and applicable degrees of freedom.
(ix) Draw the inference on the basis of computed value of the F-statistic
obtained at step(vii).
(x) If the computed value of F is greater than the tabulated value, then reject
the null hypothesis.
(xi) If the computed value is less than the tabulated value, do not reject the
null hypothesis.
The Chow test helps us in testing for parameter stability. Note that there are three
limitations of the Chow test.
(i) We assume that the error variance 𝜎 is constant throughout the
sample. There is no difference in the error variance between the sub-
samples.
(ii) The point of structural break is not known to us. We assume that point
of structural change.
(iii) We cannot apply Chow test if there are more than one structural
break.
8.8 PREDICTION
In Unit 5 we explained how prediction is made on the basis of simple regression
model. We extend the same procedure to multiple regression models. As in the
case of simple regression models, there are two types of prediction in multiple
regression models.
If we predict an individual value of the dependent variable corresponding to
particular values of the explanatory variables, we obtain the ‘individual
prediction’. When we predict the expected value of Y corresponding to particular
values of the explanatory variables, it is called ‘mean prediction’. The expected
of Y in both the cases (individual prediction and mean prediction) is the same.
The difference between mean and individual predictions lies in their variances.
8.8.1 Mean Prediction
Let
1
⎡𝑋 ⎤
⎢ ⎥
X = ⎢𝑋 ⎥ (8.32)
⎢ ⋮ ⎥
⎣𝑋 ⎦
112
be the vector of values of the X variables for which we wish to predict 𝑌 . Multiple Linear Regression
Model: Inferences
The estimated multiple regression equation, in scalar form, is
𝑌 = 𝛽 + 𝛽 𝑋 + 𝛽 𝑋 + ⋯𝛽 𝑋 + 𝑢 ... (8.33)
which in matrix notation can be written compactly as
𝑌 = 𝑋′ 𝛽 ... (8.34)
where
𝑋 ′ = [1 𝑋 𝑋 … 𝑋 ] ... (8.35)
and
𝛽
⎡ ⎤
𝛽 = ⎢𝛽 ⎥ ... (8.36)
⎢⋮ ⎥
⎣𝛽 ⎦

Equation (8.34) is the mean predication of Yi corresponding to given Xi’.

If Xi’ is as given in (8.35), then (8.34) becomes

𝑌 𝑋′ = 𝑋′ 𝛽 ... (8.37)

where the values of x0 are fixed. You should note that (8.36) gives an unbiased
prediction of 𝐸 𝑌 𝑋 ′ , since 𝐸 𝑋 ′ 𝛽 = 𝑋 ′ 𝛽 .

Variance of Mean Prediction

The formula to estimate the variance of 𝑌 𝑋 ′ is as follows:

var 𝑌 𝑋 ′ = 𝜎 𝑋 (𝑋 𝑋) 𝑋 ... (8.38)

where 𝜎 is the variance of 𝑢

𝑋 are the X variables for which we wish to predict, and

since we do not know the error variance ( 𝜎 ), we replace it by its unbiased

estimator 𝜎 .
8.8.2 Individual Prediction
As mentioned earlier, expected value of individual prediction is the same as that
of individual prediction, i.e., 𝑌 . The variance of the individual prediction is

var(𝑌 |𝑋 ) = 𝜎 [1 + 𝑋 (𝑋 𝑋) 𝑋 ] ... (8.39)


where var(𝑌 |𝑋 ) stands for E Y0  Ŷ0 | X 2 . In practice we replace 2 by its
unbiased estimator ̂2 .
113
Multiple Regression Check Your Progress 4
Models
1) Consider a Cobb-Douglas production. Write down the steps of testing the
hypothesis that it exhibits constant returns to scale.
.........................................................................................................................
.........................................................................................................................
.........................................................................................................................
.........................................................................................................................

2) Write down the steps of carrying out Chow test.

.........................................................................................................................
.........................................................................................................................
.........................................................................................................................
.........................................................................................................................

3) Point out why individual prediction has higher variance than mean
prediction.
.........................................................................................................................
.........................................................................................................................
.........................................................................................................................
.........................................................................................................................

8.9 LET US SUM UP

This unit described the assumptions of classical multiple regression that fortifies
normality of error term also tested by Jarque-Bera Test (J-Test for Normality).
The testing of hypothesis about individual coefficients is distinguished from the
overall significance test in the unit. The unit also describes the testing of equality
of two regression coefficients. Later the structural stability is tested using Chow
test. The multiple regression is also used for prediction of dependent variables for
given values of independent variables. Both individual and joint hypothesis
testing is described in the unit. Various tests such as likelihood ratio (LR), Wald
(W) and Lagrange Multiplier Test (LM) are explained in the unit

8.10 ANSWERS/ HINTS TO CHECK YOUR

PORGRESS EXERCISES
Check Your Progress 1
1) Refer to Sub-Section 8.2.1 and answer.
2) The Jarque-Bera test statistic is given at equation (8.2). Describe how
thetest is carried out.
114
Check Your Progress 2 Multiple Linear Regression
Model: Inferences
1) Refer to Sub-Section 8.3.1 and answer. Decide on the null and alternative
hypotheses. Describe the staps you would follow.
2) Refer to Sub-Section 8.3.2 and answer.
Check Your Progress 3
1) It can be tested by F-test. See Section 8.4 for details.
2) Refer to Sub-Section 8.5 and answer.
Check Your Progress 4
1) We have explained in Sub-Section 8.6.1. Refer to it.
2) Refer to Sub-Section 8.7 and answer.
3) Refer to Sub-Section 8.8 and answer. It has the same logic as in the case of
two variable models discussed in Section 5.7 of Unit 5.

115