COLLEGE OF AGRICULTURE AND ENVIRONMENTAL SCIENCE

DEPARTMENT OF AGRICULTURAL ECONOMICS

Econometrics Module for III Year Agricultural Economics students

MODULE PART II

March, 2026
Debre Tabor, Ethiopia
 Debre Tabor University, CAES, Department of Agricultural Economics

Table of Contents
Contents -------------------------------------------------------------------------------------------------------Page
CHAPTER 4: MULTIPLE REGRESSION ANALYSIS ................................................................................... 1

Introduction............................................................................................................................................................ 1

4.1. Model with two Explanatory Variables .................................................................................................... 1

4.2. Notations and Assumptions........................................................................................................................ 1

4.3. Estimation of Partial regression coefficients ............................................................................................ 5

4.4 Variances and Standard Errors of OLS Estimators................................................................................. 7

4.5 Hypothesis testing and Interval Estimation ...................................................................................... 20

4.6. Other functional forms ............................................................................................................................. 36

4.6.1 The Cobb–Douglas Production Function ......................................................................................... 36

4.6.2 Polynomial Regression Models .......................................................................................................... 39

Review Question................................................................................................................................................... 41

CHAPTER FIVE: VIOLATIONS OF THE CLASSICAL ASSUMPTIONS ............................................... 47

5.1 Non-Normality ....................................................................................................................................... 47

5.1.1 Nature of the problem ........................................................................................................................... 47

5.1.2 Source of non-normality ....................................................................................................................... 49

5.1.3 Consequence of non-normality ............................................................................................................. 49

5.1.4 Detecting non –normality ..................................................................................................................... 50

5.1.5 Remedial measure ................................................................................................................................. 50

5.2 MULTICOLLINEARITY ..................................................................................................................... 51

5.2.1 The nature of the problem .............................................................................................................. 51

5.2.2 Sources of Multicollinearity .......................................................................................................... 52

5.2.3 Consequences of Multicollinearity ................................................................................................. 52

5.2.4 Detecting Multicollinearity ............................................................................................................ 54

5.2.5 Remedial Measures ........................................................................................................................ 56

5.3 HETEROSCEDASTICITY ................................................................................................................... 57

ii | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

5.3.1 The Nature of Heteroskedasticity .................................................................................................. 57

5.3.2 Causes of Heteroskedasticity ......................................................................................................... 58

5.3.3 The consequence of Heteroskedasticity......................................................................................... 58

5.3.4 Detecting the problem of Heteroskedasticity: ................................................................................ 59

5.3.5 Remedial Measures ........................................................................................................................ 65

5.4 AUTOCORRELATION .......................................................................................................................... 67

5.4.1 The Nature of Autocorrelation ....................................................................................................... 67

5.4.2 Sources of Autocorrelation............................................................................................................. 68

5.4.3 Consequences of Autocorrelation.................................................................................................. 70

5.4.4 Testing (Detecting) for Autocorrelation......................................................................................... 70

5.4.5 Remedial Measures........................................................................................................................ 74

CHAPTER SIX: NON-LINEAR REGRESSION AND TIME SERIES ECONOMETRICS ...................... 80

6.1. Non-linear regression models Overview .................................................................................................... 80

6.2. Time series Analysis ................................................................................................................................... 82

6.2.1. Univariate Time Series Analysis ......................................................................................................... 82

6.2. 2 Multivariate Analysis ........................................................................................................................... 87

APPENDIX ........................................................................................................................................................... 94

iii | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

iv | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

CHAPTER 4: MULTIPLE REGRESSION ANALYSIS

Introduction

The two-variable model studied extensively in the previous chapters is often inadequate in
practice. In our consumption–income example, for instance, it was assumed implicitly that
only income X affects consumption Y. But economic theory is seldom so simple for, besides
income, a number of other variables are also likely to affect consumption expenditure. An
obvious example is wealth of the consumer. As another example, the demand for a commodity
is likely to depend not only on its own price but also on the prices of other competing or
complementary goods, income of the consumer, social status, etc. Therefore, we need to
extend our simple two-variable regression model to cover models involving more than two
variables. Adding more variables leads us to the discussion of multiple regression models, that
is, models in which the dependent variable, or Regressand, Y depends on two or
more explanatory variables, or regressors. The simplest possible multiple regression model is
three-variable regression, with one dependent variable and two explanatory variables. In this
and the next chapter we shall study this model. Throughout, we are concerned with multiple
linear regression models, that is, models linear in the parameters; they may or may not be
linear in the variables.
4.1. Model with two Explanatory Variables
Generalizing the two-variable population regression function (PRF) (4.1.1), we may write the three-
variable PRF as:

𝑌 = 𝛽 + 𝛽 𝑋 + 𝛽 𝑋 + 𝑢 ---------------------------------------------------------------------------------4.1.1

Where 𝛽 the intercept term; the coefficients are 𝛽 𝑎𝑛𝑑𝛽 are called the partial
regression coefficients and Y is the dependent variable, 𝑋 𝑎𝑛𝑑𝑋 the explanatory
variables (or regressors), u the stochastic disturbance term, and ―i‖ the 𝑖 observation; in
case the data are time series, the subscript t will denote the 𝑡 observation.

4.2. Notations and Assumptions

 Gauss–Markov Theorem: Given the assumptions of the classical linear regression

1 | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

model, the least-squares estimators, in the class of unbiased linear estimators, have
minimum variance, that is, they are Best Linear Unbiased Estimator (BLUE).
 The Gaussian, standard, or classical linear regression model (CLRM),
which is the cornerstone of most econometric theory, makes 10 assumptions.
 We continue to operate within the framework of the classical linear regression model
(CLRM) first introduced in Chapter 3.

Assumptions

Assumption 1: Linear regression model. The regression model is linear in the parameters.
As shown in equation (4.1.1). The model should be linear in parameter but not in explanatory
variables. Keep in mind that the Regressand Y and the regressor X themselves may be
nonlinear. Of the two interpretations of linearity, linearity in the parameters is relevant for the development of
the regression theory to be presented shortly. Therefore, from now on the term “linear” regression
will always mean a regression that is linear in the parameters; the β’s (that is, the parameters
are raised to the first power only). It may or may not be linear in the explanatory variables,
the X’s.

2 | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

FIGURE 4.1 Linear-in-parameter functions.

Assumption 2: X values are fixed in repeated sampling. Values taken by the regressor X are
considered fixed in repeated samples. More technically, X is assumed to be non-stochastic.
What all this means is that our regression analysis is conditional regression analysis, that is,
conditional on the given values of the regressor(s) X.
Assumption 3: Zero mean value of disturbance ui. Given the value of X, the mean, or
expected, value of the random disturbance term ui is zero. Technically, the conditional mean
value of ui is zero. Symbolically, we have
𝑢 𝑥 =0
Assumption 4: Homoscedasticity or equal variance of ui. Given the value of X, the variance
of ui is the same for all observations. That is, the conditional variances of ui are identical.
Symbolically, we have,
𝑣𝑎𝑟 𝑢 𝑥 = 𝑢 𝑢 𝑥 ]
= 𝑢 𝑥 ) because of assumption 3
= Where ―var‖ stands for variance
Assumption 5: No autocorrelation between the disturbances. Given any two X values, Xi
and Xj (= j), the correlation between any two ui and uj (i ≠ j) is zero. Symbolically,
𝑐𝑜𝑣 𝑢 𝑢 = 𝑢 𝑢 𝑥 𝑢 𝑢 𝑥 }

= 𝑢 𝑥 𝑢 𝑥
=
Where i and j are two different observations and where cov means covariance.
In words, (3.2.5) postulates that the disturbances ui and uj are uncorrelated. Technically, this is
the assumption of no serial correlation, or no autocorrelation. This means that, given Xi, the
deviations of any two Y values from their mean value do not exhibit patterns such as those
shown in Figure 3.6a and b. In Figure 4.2a, we see that the u‘s are positively correlated, a
positive u followed by a positive u or a negative u followed by a negative u. In Figure 4.2b,
the u‘s are negatively correlated, a positive u followed by a negative u and vice versa

3 | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

FIGURE 4.2: Patterns of correlation among the disturbances. (a) Positive serial correlation;
(b) negative serial correlation; (c) zero correlation.
Assumption 6: zero covariance between or = .formally,
𝑐𝑜𝑣 𝑢 𝑥 = 𝑢 𝑢 ] 𝑥 𝑥 ]
= 𝑢 𝑥 = 𝑢 𝑥 𝑥 ] since 𝑢 =
= 𝑢 𝑥 𝑢 𝑥 Since 𝑥 non-stochastic
𝑢 𝑥 = 𝑠𝑖𝑛𝑐𝑒 𝑢 =
=
Assumption 7: The number of observations n must be greater than the number of
parameters to be estimated. Alternatively, the number of observations n must be greater than
the number of explanatory variables.
Assumption 8: Variability in X values. The X values in a given sample must not all be the
same. Technically, var (X ) must be a finite positive number.

4 | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

Assumption 9: The regression model is correctly specified. Alternatively, there is no

specification bias or error in the model used in empirical analysis.
Assumption 10: There is no perfect Multicollinearity. That is, there are no perfect linear
relationships among the explanatory variables.
4.3. Estimation of Partial regression coefficients

The Meaning of Partial Regression Coefficients: As mentioned earlier, the regression

coefficients β1 and β2 are known as partial regression or partial slope coefficients. The
meaning of partial regression coefficient is as follows: β1 measures the change in the mean value
of Y, E(Y), per unit change in X1, holding the value of X2 constant. Put differently, it gives the
―direct‖ or the ―net‖ effect of a unit change in X2 on the mean value of Y, net of any effect that X2
may have on mean Y. Likewise, β2 measures the change in the mean value of Y per unit change
in X3, holding the value of X1 constant.4 That is, it gives the ―direct‖ or ―net‖ effect of a unit
change in X2 on the mean value of Y, net of any effect that X1 may have on
mean Y.
In addition, as in Chapter 3, we assume that the multiple regression models is linear in the
parameters, that the values of the regressors are fixed in repeated sampling, and that there is
sufficient variability in the values of the regressors. Assumption (4.2.5), that there be no exact
linear relationship between X2 and X3, technically known as the assumption of no collinearity
or no Multicollinearity if more than one exact linear relationship is involved, is new and needs
some explanation. Informally, no collinearity means none of the regressors can be written as
exact linear combinations of the remaining regressors in the model.

To find the OLS estimators, let us first write the sample regression function (SRF) corresponding to the
PRF of (4.1.1) as follows:
𝑌̂ = 𝛽̂ + 𝛽̂ 𝑋 + 𝛽̂ 𝑋 + 𝑢̂ …………………………………………………4.3.1
where 𝑢̂ i is the residual term, the sample counterpart of the stochastic disturbance term 𝑢

As noted in Chapter 3, the OLS procedure consists in so choosing the values of the unknown
parameters that the residual sum of squares (RSS) ∑ 𝑢̂ is as small as possible. Symbolically,

𝑚𝑖𝑛 ∑ 𝑢̂ =∑ 𝑌̂ 𝛽̂ 𝛽̂ 𝑋 𝛽̂ 𝑋 … … … … … … … … … … … … … … … … … … … 4.3.2

Where the expression for the RSS is obtained by simple algebraic manipulations of (4.3.1).

5 | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

The most straightforward procedure to obtain the estimators that will minimize (4.3.1) is to differentiate
it with respect to the unknowns, set the resulting expressions to zero, and solve them simultaneously.
This procedure gives the following normal equations
• Given in the equation 4.3.3 and 4.3.5: differentiate the equation 4.3.2 partially with respect
to the three unknowns and setting the resulting equations to zero, we obtain:

∑ 𝑢̂
= 2∑ 𝑌 𝛽̂ 𝛽̂ 𝑋 𝛽̂ 𝑋 =
𝛽
∑ 𝑢̂
= 2∑ 𝑌 𝛽̂ 𝛽̂ 𝑋 𝛽̂ 𝑋 𝑋 =
𝛽
∑ 𝑢̂
= 2∑ 𝑌 𝛽̂ 𝛽̂ 𝑋 𝛽̂ 𝑋 𝑋 =
𝛽
• Simplifying these, we obtain Equations. (4.3.3) to (4.3.5).

𝑌̅ = 𝛽̂ + 𝛽̂ 𝑋̅ + 𝛽̂ 𝑋̅ …………………….…………………………………….4.3.3
𝛽̂ = 𝑌̅ 𝛽̂ 𝑋̅ + 𝛽̂ 𝑋̅ )
𝛽̂ = 𝑌̅ 𝛽̂ 𝑋̅ 𝛽̂ 𝑋̅ -----------------------------------------------------------------------4.3.3a
∑ 𝑌 𝑋 = 𝛽̂ ∑ 𝑋 + 𝛽̂ ∑ 𝑋 + 𝛽̂ ∑ 𝑋 𝑋 ………………………………………….…4.3.4
∑ 𝑌 𝑋 = 𝛽̂ ∑ 𝑋 + 𝛽̂ ∑ 𝑋 + 𝛽̂ ∑ 𝑋 𝑋 ………………………………………….…4.3.5
Thus, equation known as normal equation of OLS
Differentiating this expression partially with respect to each of the k unknowns, setting the
resulting equations equal to zero, and rearranging, we obtain the following k normal equations in
the k unknowns:
∑ 𝑌 = 𝑛𝛽̂ + 𝛽̂ ∑ 𝑋 + 𝛽̂ ∑ 𝑋 + + 𝛽̂ ∑ 𝑋 …………………………………..4.3.6
∑ 𝑌 𝑋 = 𝛽̂ ∑ 𝑋 + 𝛽̂ ∑ 𝑋 + 𝛽̂ ∑ 𝑋 𝑋 …………………………………………….4.3.7
∑ 𝑌 𝑋 = 𝛽̂ ∑ 𝑋 + 𝛽̂ ∑ 𝑋 + 𝛽̂ ∑ 𝑋 𝑋 …………………………………………….4.3.8
---------------------------------------------------------------------------------
∑ 𝑌 𝑋 = 𝛽̂ ∑ 𝑋 + 𝛽̂ ∑ 𝑋 𝑋 + 𝛽̂ ∑ 𝑋 𝑋 + 𝛽̂ ∑ 𝑋 +……………..……………4.3.9
Or, switching to small letters, these equations can be expressed as
∑ 𝑦 𝑥 = 𝛽̂ ∑ 𝑥 + 𝛽̂ ∑ 𝑥 + 𝛽̂ ∑ 𝑥 𝑥 ……………………………………………...4.3.10

6 | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

∑ 𝑦 𝑥 = 𝛽̂ ∑ 𝑥 + 𝛽̂ ∑ 𝑥 + 𝛽̂ ∑ 𝑥 𝑥 ……………………………………………4.3.11
---------------------------------------------------------------------------------
∑𝑦 𝑥 = 𝛽̂ ∑ 𝑥 + 𝛽̂ ∑ 𝑥 𝑥 + 𝛽̂ ∑ 𝑥 𝑥 + 𝛽̂ ∑ 𝑥 +…………………………..4.3.12
It should further be noted that the k-variable model also satisfies these equations:

∑ 𝑢̂ = 4.3. 2𝑎

∑ 𝑢̂ 𝑥 = ∑ 𝑢̂ 𝑥 ∑ 𝑢̂ 𝑥 = 4.3. 2𝑏
In the case of three variable regression model we substitute equation (4.3.3a) in equation
(4.3.4)and (4.3.5) finally we get the formula for 𝛽̂ and 𝛽̂
By Appling Cramer's rule of matrix from the normal equation of deviation
∑ 𝑦 𝑥 = 𝛽̂ ∑ 𝑥 + 𝛽̂ ∑ 𝑥 𝑥 ……………………………………………………..4.3.13
∑ 𝑦 𝑥 = 𝛽̂ ∑ 𝑥 𝑥 + 𝛽̂ ∑ 𝑋 ….……………………………………………….…4.3.14

∑ ∑ ∑ ∑
𝛽̂ = ∑ ∑ ∑
------------------------------------------------------------------4.3.15

∑ ∑ ∑ ∑
𝛽̂ = ∑ ∑ ∑
------------------------------------------------------------------4.3.16

Which give the OLS estimators of the population partial regression coefficients β2 and β3,
respectively.
In passing, note the following: (1) Equations (4.3.15) and (4.3.16) are symmetrical in nature
because one can be obtained from the other by interchanging the roles of X1and X2; (2) the
denominators of these two equations are identical; and (3) the three-variable case is a natural
extension of the two-variable case.

4.4 Variances and Standard Errors of OLS Estimators

Having obtained the OLS estimators of the partial regression coefficients, we can derive the
variances and standard errors of these estimators. As in the two-variable case, we need the
standard errors for two main purposes: to establish confidence intervals and to test statistical
hypotheses. The relevant formulas are as follows:

7 | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

̅ ∑ ̅ ∑ ̅ ̅ ∑
• 𝑣𝑎𝑟(𝛽̂ ) = [ + ∑ ∑ ∑
] 4.4.

𝛽̂ = +√𝑣𝑎𝑟 𝛽̂ 4.4.2

∑
• (𝛽̂ ) = ∑ ∑ ∑
4.4.3

• Equivalently (𝛽̂ ) = ∑ 4.4.4

• 𝛽̂ = +√𝑣𝑎𝑟 𝛽̂ 4.4.5

∑
• (𝛽̂ ) = ∑ ∑ ∑
4.4.6

• (𝛽̂ ) = ∑ 4.4.8

• 𝛽̂ = +√𝑣𝑎𝑟 𝛽̂ 4.4.9

In all these formulas ̂ is the (homoscedastic) variance of the population disturbances𝑢 . An

unbiased estimator of ζ 2 is given by:
∑̂
• ̂ = 4.4.8

But 𝑢̂ = 𝑌 𝛽̂ 𝛽̂ 𝑋 𝛽̂ 𝑋 . .

Also substitute the value of 𝛽̂ = 𝑌̅ 𝛽̂ 𝑋̅ 𝛽̂ 𝑋̅ . .

We get 𝑢̂ = 𝑌 𝑌̅ 𝛽̂ 𝑋̅ 𝛽̂ 𝑋̅ 𝛽̂ 𝑋 𝛽̂ 𝑋
=𝑌 𝑌̅ + 𝛽̂ 𝑋̅ + 𝛽̂ 𝑋̅ 𝛽̂ 𝑋 𝛽̂ 𝑋
=𝑦 𝛽̂ 𝑋 𝑋̅ 𝛽̂ 𝑋 𝑋̅
𝑋 𝑋̅ 𝑎𝑛𝑑 𝑋 𝑋̅ 𝑎𝑟𝑒 𝑑𝑒𝑣𝑖𝑎𝑡𝑖𝑜𝑛𝑠 𝑓𝑟𝑜𝑚 𝑟𝑒𝑠𝑝𝑒𝑐𝑡𝑖𝑣𝑒 𝑚𝑒𝑎𝑛.
𝑡𝑕𝑒𝑟𝑒 𝑓𝑜𝑟𝑒 𝑡𝑕𝑒 𝑒𝑞𝑢𝑎𝑡𝑖𝑜𝑛 𝑏𝑒𝑐𝑜𝑚𝑒𝑠

𝑢̂ = 𝑦 𝛽̂ 𝑥 𝛽̂ 𝑥 4.4.8𝑐
Now ∑ 𝑢̂ = ∑ 𝑢̂ 𝑢̂

= ∑ 𝑢̂ 𝑦 𝛽̂ 𝑥 𝛽̂ 𝑥

= ∑ 𝑢̂ 𝑦

𝑏𝑒𝑐𝑎𝑢𝑠𝑒 𝑜𝑓 𝑡𝑕𝑒 𝑓𝑎𝑐𝑡 𝑡𝑕𝑎𝑡 ∑ 𝑢̂ 𝑥 = ∑ 𝑢̂ 𝑥 =

8 | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

Then ∑ 𝑢̂ 𝑦 = ∑ 𝑦 𝑢̂ = ∑ 𝑦 𝑦 𝛽̂ 𝑥 𝛽̂ 𝑥 by plugging equation 4.4.8c

It becomes ∑ 𝑢̂ = ∑𝑦 𝑦 𝛽̂ 𝑥 𝑦 𝛽̂ 𝑥 𝑦

∑ 𝑢̂ = ∑𝑦 𝛽̂ 𝑥 𝑦 𝛽̂ 𝑥 𝑦 4.4.8𝑑

The estimator ̂ can be computed from (4.4.8) once the residuals are available, but it can also be
obtained more readily by using the following relation

∑ 𝑢̂ = ∑𝑦 𝛽̂ ∑ 𝑦 𝑥 𝛽̂ ∑ 𝑦 𝑥 4.4.9

Properties of OLS Estimators

The properties of OLS estimators of the multiple regression models parallel those of the two-
variable model. Specifically:
1. The three-variable regression line (surface) passes through the means 𝑌̅ 𝑋̅ 𝑎𝑛𝑑𝑋̅ . This
property holds generally. Thus in the k-variable linear regression model [a Regressand and
(k - 1) regressors].
𝑌𝑖 = 𝛽 + 𝛽 𝑋 +𝛽 𝑋 + +𝛽 𝑋 +𝑢 4.4.
𝛽̂ = ̅𝑌 𝛽̂ 𝑋̅ + 𝛽̂ 𝑋̅ + + 𝛽̂ 𝑋̅ 4.4.

2. The mean value of the estimated 𝑌 ( =𝑌̂ ) is equal to the mean value of the actual Yi,
3. . ∑ 𝑢̂ = 𝑢̅̂ = , which can be verified from
SRF which can be expressed in the deviation forms as
𝑦 = 𝑦̂ + 𝑢̂ = 𝛽̂ 𝑥 + 𝛽̂ 𝑥 + 𝑢̂ Because 𝑦̂ = 𝛽̂ 𝑥 + 𝛽̂ 𝑥
4. The residuals 𝑢̂ are uncorrelated with 𝑋 and 𝑋 , that is
∑ 𝑢̂ 𝑋 = ∑ 𝑢̂ 𝑋 =
5. The residuals 𝑢̂ are uncorrelated with 𝑌 ; that is ∑ 𝑢̂ 𝑌 =
6. From (4.4.4) and (4.4.9) it is evident that as 𝑟 , the correlation coefficient between 𝑋 and
𝑋 , increases toward 1, the variances of 𝛽̂ 𝑎𝑛𝑑 𝛽̂ increase for given values of and
∑𝑥 or ∑ 𝑥 .
• 𝐼𝑛 𝑡𝑕𝑒 𝑙𝑖𝑚𝑖𝑡, 𝑤𝑕𝑒𝑛 𝑟 = (𝑖.𝑒., 𝑝𝑒𝑟𝑓𝑒𝑐𝑡 𝑐𝑜𝑙𝑙𝑖𝑛𝑒𝑎𝑟𝑖𝑡𝑦), 𝑡𝑕𝑒𝑠𝑒 𝑣𝑎𝑟𝑖𝑎𝑛𝑐𝑒𝑠
𝑏𝑒𝑐𝑜𝑚𝑒 𝑖𝑛𝑓𝑖𝑛𝑖𝑡𝑒.
7. It is also clear from (4.4.4) and (4.4.9) that for given values of 𝑟 and ∑ 𝑥 or ∑ 𝑥 , the

9 | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

variances of the OLS estimators are directly proportional to ; that is, they increase as
increases.
Similarly, for given values of and 𝑟 , the variance of β1 is inversely proportional to ∑ 𝑥 ;
that is, the greater the variation in the sample values of X1, the smaller the variance of 𝛽̂ and
therefore 𝛽̂ can be estimated more precisely. A similar statement can be made about the
variance of 𝛽̂ .
8. Given the assumptions of the classical linear regression model, one can prove that the OLS
estimators of the partial regression coefficients not only are linear and unbiased but also have
minimum variance in the class of all linear unbiased estimators. In short, they are BLUE:
Put differently, they satisfy the Gauss-Markov theorem.
The Multiple Coefficients of Determination and the Multiple Coefficients of Correlation R
In the two-variable case we saw that as defined as measures the goodness of fit of the
regression equation; that is, it gives the proportion or percentage of the total variation in the
dependent variable Y explained by the (single) explanatory variable X. This notation of can
be easily extended to regression models containing more than two variables. Thus, in the three
variable model we would like to know the proportion of the variation in Y explained by the
variables X1 and X2 jointly. The quantity that gives this information is known as the multiple
coefficient of determination and is denoted by ; conceptually it is akin to
To derive we may follow the derivation of Recall that:

𝑌 = 𝛽̂ + 𝛽̂ 𝑋 + 𝛽̂ 𝑋 + 𝑢̂
--------------------4.4.12
𝑌 = 𝑌̂ + 𝑢̂

̂ is the estimated value of Yi from the fitted regression line and is an estimator of true
Where 𝑌
𝑌𝑖 𝑋 𝑖 𝑋2𝑖 . Upon shifting to lowercase letters to indicate deviations from the mean values,
Eq. (4.4.12) may be written as

𝑦 = 𝛽̂ 𝑥 + 𝛽̂ 𝑥 + 𝑢̂
𝑦 = 𝑦̂ + 𝑢̂ --------------------4.4.13

10 | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

Squaring (4.4.13) on both sides and summing over the sample values, we obtain

∑𝑦 =∑ 𝑦̂ +∑ 𝑢̂ + 2∑𝑦
̂𝑖 𝑢
̂𝑖 4.4. 4

= ∑ 𝑦̂ +∑ 𝑢̂ 𝑤𝑕𝑦? 4.4. 4

Verbally, Eq. (4.4. 4) states that the total sum of squares (TSS) equals the explained sum of
squares (ESS) + the residual sum of squares (RSS). Now substituting for ∑ 𝑢̂ from (4.4.9),
we obtain

∑𝑦 = ∑ 𝑦̂ + ∑ 𝑦 𝛽̂ ∑ 𝑦 𝑥 𝛽̂ ∑ 𝑦 𝑥

which, on rearranging, gives

= ∑ 𝑦̂ = 𝛽̂ ∑ 𝑦 𝑥 + 𝛽̂ ∑ 𝑦 𝑥 4.4. 5

∑̂ ̂ ∑ ̂ ∑
By definition = =∑ = ∑
4.4. 6

2
Since the quantities entering (4.4.16) are generally computed routinely, can be computed easily.
2
Note that , like 𝑟2 , lies between 0 and 1. If it is 1, the fitted regression line explains 100 percent of
the variation in Y. On the other hand, if it is 0, the model does not explain any of the variation in Y.
2
Typically, however, lies between these extreme values. The fit of the model is said to be “better’’
the closer is to 1.
Recall that in the two-variable case we defined the quantity r as the coefficient of correlation and
indicated that it measures the degree of (linear) association between two variables. The three-or-more-
variable analogue of r is the coefficient of multiple correlations, denoted by R, and it is a measure of
the degree of association between Y and all the explanatory variables jointly. Although r can be
positive or negative, R is always taken to be positive. In practice, however, R is of little importance.
2
The more meaningful quantity is . Before proceeding further, let us note the following relationship

11 | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

2
between and the variance of a partial regression coefficient in the k-variable multiple regression
model given in (.4.4.10):

𝑣𝑎𝑟(𝛽̂ ) = [ ] 4.4. 7
∑𝑥

Where 𝛽̂ is the partial regression coefficient of regressor 𝑋 and is the R2 in the regression of 𝑋
on the remaining (k − 2) regressors. [Note: There are (k − 1) regressors in the k-variable regression
model.]
Partial Correlation Coefficients: Explanation of Simple and Partial Correlation Coefficients
In Chapter 2 we introduced the coefficient of correlation r as a measure of the degree of linear
association between two variables. For the three-variable regression model we can compute three
correlation coefficients: 𝑟 (correlation between Y and X1), 𝑟 (correlation coefficient between Y and
X2), and 𝑟 (correlation coefficient between X1 and X2); notice that we are letting the subscript 1
represents Y for notational convenience. These correlation coefficients are called gross or simple
correlation coefficients, or correlation coefficients of zero order. These coefficients can be
computed by the definition of correlation coefficient given in chapter 3.
But now consider this question: Does, say, r1 2 in fact measure the ―true‖ degree of (linear) association
between Y and X1 when a third variable X2 may be associated with both of them? This question is
analogous to the following question: Suppose the true regression model is (4.1.1) but we omit from the
model the variable X2 and simply regress Y on X1, obtaining the slope coefficient of, say, b12. Will this
coefficient be equal to the true coefficient β1 if the model (4.1.1) were estimated to begin with? In
general, r12 is not likely to reflect the true degree of association between Y and X1 in the presence of
X2. As a matter of fact, it is likely to give a false impression of the nature of association between Y and
X2, as will be shown shortly. Therefore, what we need is a correlation coefficient that is independent of
the influence, if any, of X2 on X1 and Y. Such a correlation coefficient can be obtained and is known
appropriately as the partial correlation coefficient. Conceptually, it is similar to the partial regression
coefficient. We define
r12.3 = partial correlation coefficient between Y and X1, holding X2 constant
r13.2 = partial correlation coefficient between Y and X2, holding X1 constant
r23.1 = partial correlation coefficient between X1 and X2, holding Y constant
These partial correlations can be easily obtained from the simple or zero order, correlation coefficients
as follows.

12 | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

𝑟 𝑟 𝑟
𝑟 . = 4.4. 8
√ 𝑟 𝑟
𝑟 𝑟 𝑟
𝑟 . = 4.4. 9
√ 𝑟 𝑟
𝑟 𝑟 𝑟
𝑟 . = 4.4.2
√ 𝑟 𝑟
The partial correlations given in Equations (4.4.18) to (4.4.20) are called first order correlation
coefficients. By order we mean the number of secondary subscripts. Thus r1 2.3 4 would be the
correlation coefficient of order two, r1 2.3 4 5 would be the correlation coefficient of order three, and so
on. As noted previously, r12, r13, and so on are called simple or zero-order correlations. The
interpretation of, say, r 1 2.3 4 is that it gives the coefficient of correlation between Y and X1, holding X2
and X3 constant.
and the Adjusted ̅

An important property of R2 is that it is a non-decreasing function of the number of explanatory

variables or regressors present in the model; as the number of regressors increases, almost
invariably increases and never decreases. Stated differently, an additional X variable will not
decrease . To see this, recall the definition of the coefficient of determination:
∑̂
= = =∑ 4.4.26

Now ∑ 𝑦 is independent of the number of X variables in the model because it is simply ∑ 𝑌 𝑌̅ .

The RSS, ∑ 𝑢̂ , however, depends on the number of regressors present in the model. Intuitively, it is
clear that as the number of X variables increases, ∑ 𝑢̂ is likely to decrease (at least it will not
increase); hence as defined in (4.4.26) will increase. In view of this, in comparing two regression
models with the same dependent variable but differing number of X variables, one should be very wary
of choosing the model with the highest .
To compare two terms, one must take into account the number of X variables present in the model.
This can be done readily if we consider an alternative coefficient of determination, which is as follows:

̅ = 𝑛 𝑘 = ∑ 𝑢̂ 𝑛 𝑘
4.4.27
∑𝑦 𝑛
𝑛
Where k = the number of parameters in the model including the intercept term. (In the three-variable

13 | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

regression, k = 3. Why?) The R2 thus defined is known as the adjusted , denoted by ̅ . The term
adjusted means adjusted for the degree of freedom (df) associated with the sums of squares entering
into (4.4.26): ∑ 𝑢̂ has n − k df in a model involving k parameters, which include the intercept term,
and ∑ 𝑦 has n − 1 df. (Why?) For the three-variable case, we know that ∑ 𝑢̂ has n − 3 df. Equation
(4.4.27) can also be written as:
̂
̅ = 4.4.28

Where ̂ is the residual variance, an unbiased estimator of true , and is the sample variance of
Y. It is easy to see that ̅ and are related because, substituting (4.4.26) into (4.4.27), we obtain

𝑛
̅ = 4.4.29
𝑛 𝑘
It is immediately apparent from Eq. (4.4.29) that:
1. For 𝑘 , ̅ which implies that as the number of X variables increases, the adjusted R2
increases less than the unadjusted ; and
2. ̅ can be negative, although R2 is necessarily non-negative. In case ̅ turns out to be negative in
an application, its value is taken as zero.
Which should one use in practice? As Theil notes:
“It is good practice to use ̅ rather than because tends to give an overly optimistic picture
of the fit of the regression, particularly when the number of explanatory variables is not very small
compared with the number of observations”.
But Thiel’s view is not uniformly shared, for he has offered no general theoretical justification for the
“superiority’’ of ̅ . For example, Goldberger argues that the following , call it modified , will
do just as well
𝑘
𝑜𝑑𝑖𝑓𝑖𝑒𝑑 =( ) 4.4.3
𝑛
Note, However, that if = 1, ̅ = = 1. When = 0, ̅ = (1 − k)/ (n− k), in which case
̅ can be negative if k > 1
Besides and adjusted as goodness of fit measures, other criteria are often used to judge the
adequacy of a regression model.
Numerical Example: Child Mortality In Relation To Per Capita GNP and Female Literacy Rate: based

14 | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

on the table 4.1 answer the following question:

a) Estimate and interpret the model of child mortality rate given by
=𝛽 + 𝛽 𝑋 + 𝛽 𝑋 + 𝑢 4.4.3
𝑤𝑕𝑒𝑟𝑒 𝑋 = 𝑎𝑛𝑑 𝑋 = 𝑓𝑒𝑚𝑎𝑙𝑒 𝑙𝑖𝑡𝑟𝑎𝑐𝑦 𝑟𝑎𝑡𝑒
b) Estimate the variance of each regressor and the model?
c) Compute and interpret the R2 and adjusted R2?
Solution a) the coefficients estimation
∑𝑥 𝑦 ∑𝑥 ∑𝑥 𝑦 ∑𝑥 𝑥
𝛽̂ =
∑𝑥 ∑𝑥 ∑𝑥 𝑥

∑𝑥 𝑦 ∑𝑥 == 4.76797 + 3

∑𝑥 𝑦 ∑𝑥 𝑥 = 6.379 9 + 2

= 4.76797 + 3 6.379 9 + 2 = 4. 3 6 + 3

∑𝑥 ∑𝑥 ∑𝑥 𝑥 = .99455 + 3 .43824 + 2 = .85 73 + 3

4. 3 6 + 3
𝛽̂ = = 2.23 585732
.85 73 + 3

∑𝑥 𝑦 ∑𝑥 ∑𝑥 𝑦 ∑𝑥 𝑥
𝛽̂ =
∑𝑥 ∑𝑥 ∑𝑥 𝑥

∑𝑥 𝑦 ∑𝑥 ∑𝑥 𝑦 ∑𝑥 𝑥

2.2667 + .22 67 + = . 45 3 +

∑𝑥 ∑𝑥 ∑𝑥 𝑥 = .85 73 + 3

= .99455 + 3 .43824 + 2 = .85 73 + 3

. 45 3 +
𝛽̂ = = . 5646595
.85 73 + 3

15 | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

Table 4.1: Fertility and other data for 64 countries

Observation Yi X1 X2 X3 Observation Yi X1 X2 X3
CM FLFP PGNP TFR CM FLFP PGNP TFR
1 128 37 1870 6.66 33 142 50 8640 7.17
2 204 22 130 6.15 34 104 62 350 6.6
3 202 16 310 7 35 287 31 230 7
4 197 65 570 6.25 36 41 66 1620 3.91
5 96 76 2050 3.81 37 312 11 190 6.7
6 209 26 200 6.44 38 77 88 2090 4.2
7 170 45 670 6.19 39 142 22 900 5.43
8 240 29 300 5.89 40 262 22 230 6.5
9 241 11 120 5.89 41 215 12 140 6.25
10 55 55 290 2.36 42 246 9 330 7.1
11 75 87 1180 3.93 43 191 31 1010 7.1
12 129 55 900 5.99 44 182 19 300 7
13 24 93 1730 3.5 45 37 88 1730 3.46
14 165 31 1150 7.41 46 103 35 780 5.66
15 94 77 1160 4.21 47 67 85 1300 4.82
16 96 80 1270 5 48 143 78 930 5
17 148 30 580 5.27 49 83 85 690 4.74
18 98 69 660 5.21 50 223 33 200 8.49
19 161 43 420 6.5 51 240 19 450 6.5
20 118 47 1080 6.12 52 312 21 280 6.5
21 269 17 290 6.19 53 12 79 4430 1.69
22 189 35 270 5.05 54 52 83 270 3.25
23 126 58 560 6.16 55 79 43 1340 7.17
24 12 81 4240 1.8 56 61 88 670 3.52
25 167 29 240 4.75 57 168 28 410 6.09
26 135 65 430 4.1 58 28 95 4370 2.86
27 107 87 3020 6.66 59 121 41 1310 4.88
28 72 63 1420 7.28 60 115 62 1470 3.89
29 128 49 420 8.12 61 186 45 300 6.9
30 27 63 19830 5.23 62 47 85 3630 4.1
31 152 84 420 5.79 63 178 45 220 6.09
32 224 23 530 6.5 64 142 67 560 7.2
Note: CM=Child mortality ; the number of death of children under the age of 5 in a year per 1000 live
births FLFP=Female Literacy Rate Percent PGNP= Per capita GNP in 1980

TFR= Total Fertility Rate; the average number of children born to a woman using age specific fertility
rate for a given year

16 | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

Table 4.2: Estimation Procedure of the Parameters /Coefficients of the Child Mortality Regression Model

𝑦 𝑥 𝑥 𝑥 𝑥 x1y x2y
𝑌 𝑌̅ 𝑋 𝑋̅ 𝑋 𝑋̅ 𝑥 𝑥
-13.5 -14.1875 468.75 201.2852 219726.5625 -6650.390625 191.53125 -6328.125
62.5 -29.1875 -1271.25 851.9102 1616076.563 37104.60938 -1824.21875 -79453.125
60.5 -35.1875 -1091.25 1238.16 1190826.563 38398.35938 -2128.84375 -66020.625
55.5 13.8125 -831.25 190.7852 690976.5625 -11481.64063 766.59375 -46134.375
-45.5 24.8125 648.75 615.6602 420876.5625 16097.10938 -1128.96875 -29518.125
67.5 -25.1875 -1201.25 634.4102 1443001.563 30256.48438 -1700.15625 -81084.375
28.5 -6.1875 -731.25 38.28516 534726.5625 4524.609375 -176.34375 -20840.625
98.5 -22.1875 -1101.25 492.2852 1212751.563 24433.98438 -2185.46875 -108473.125
99.5 -40.1875 -1281.25 1615.035 1641601.563 51490.23438 -3998.65625 -127484.375
-86.5 3.8125 -1111.25 14.53516 1234876.563 -4236.640625 -329.78125 96123.125
-66.5 35.8125 -221.25 1282.535 48951.5625 -7923.515625 -2381.53125 14713.125
-12.5 3.8125 -501.25 14.53516 251251.5625 -1911.015625 -47.65625 6265.625
-117.5 41.8125 328.75 1748.285 108076.5625 13745.85938 -4912.96875 -38628.125
23.5 -20.1875 -251.25 407.5352 63126.5625 5072.109375 -474.40625 -5904.375
-47.5 25.8125 -241.25 666.2852 58201.5625 -6227.265625 -1226.09375 11459.375
-45.5 28.8125 -131.25 830.1602 17226.5625 -3781.640625 -1310.96875 5971.875
6.5 -21.1875 -821.25 448.9102 674451.5625 17400.23438 -137.71875 -5338.125
-43.5 17.8125 -741.25 317.2852 549451.5625 -13203.51563 -774.84375 32244.375
19.5 -8.1875 -981.25 67.03516 962851.5625 8033.984375 -159.65625 -19134.375
-23.5 -4.1875 -321.25 17.53516 103201.5625 1345.234375 98.40625 7549.375

17 | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

127.5 -34.1875 -1111.25 1168.785 1234876.563 37990.85938 -4358.90625 -141684.375

47.5 -16.1875 -1131.25 262.0352 1279726.563 18312.10938 -768.90625 -53734.375
-15.5 6.8125 -841.25 46.41016 707701.5625 -5731.015625 -105.59375 13039.375
-129.5 29.8125 2838.75 888.7852 8058501.563 84630.23438 -3860.71875 -367618.125
25.5 -22.1875 -1161.25 492.2852 1348501.563 25765.23438 -565.78125 -29611.875
-6.5 13.8125 -971.25 190.7852 943326.5625 -13415.39063 -89.78125 6313.125
-34.5 35.8125 1618.75 1282.535 2620351.563 57971.48438 -1235.53125 -55846.875
-69.5 11.8125 18.75 139.5352 351.5625 221.484375 -820.96875 -1303.125
-13.5 -2.1875 -981.25 4.785156 962851.5625 2146.484375 29.53125 13246.875
-114.5 11.8125 18428.75 139.5352 339618826.6 217689.6094 -1352.53125 -2110091.875
10.5 32.8125 -981.25 1076.66 962851.5625 -32197.26563 344.53125 -10303.125
82.5 -28.1875 -871.25 794.5352 759076.5625 24558.35938 -2325.46875 -71878.125
0.5 -1.1875 7238.75 1.410156 52399501.56 -8596.015625 -0.59375 3619.375
-37.5 10.8125 -1051.25 116.9102 1105126.563 -11366.64063 -405.46875 39421.875
145.5 -20.1875 -1171.25 407.5352 1371826.563 23644.60938 -2937.28125 -170416.875
-100.5 14.8125 218.75 219.4102 47851.5625 3240.234375 -1488.65625 -21984.375
170.5 -40.1875 -1211.25 1615.035 1467126.563 48677.10938 -6851.96875 -206518.125
-64.5 36.8125 688.75 1355.16 474376.5625 25354.60938 -2374.40625 -44424.375
0.5 -29.1875 -501.25 851.9102 251251.5625 14630.23438 -14.59375 -250.625
42613.75 468053300 1199265 -101868 -5319170

18 | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

𝛽̂ = 𝑌̅ 𝛽̂ 𝑋̅ 𝛽̂ 𝑋̅

= 4 .5 2.23 585732 5 . 875 . 5646595 4 .25

𝛽̂ = 263.64 5856

Now let us bring in female literacy as measured by the female literacy rate (FLR). A priori, we
expect that FLR too will have a negative impact on CM. Now when we introduce both the variables
in our model, we need to net out the influence of each of the regressors. That is, we need to estimate
the (partial) regression coefficients of each regressor. Thus our model is:

̂ =𝛽 + 𝛽 𝑋 + 𝛽 𝑋 + 𝑢

̂ = 263.64 5856 2.23 6 𝑋 . 56 𝑋 + 𝑢

𝑠𝑒 = .5932 . 9 .2 99 = .7 77 ̅ = .698

Where figures in parentheses are the estimated standard errors. Before we interpret this regression,
observe the partial slope coefficient of PGNP, namely, −0.0056.

Let us now interpret these regression coefficients: The coefficient −2.2316 tells us that holding the
influence of PGNP constant, on average, the number of deaths of children under 5 goes down by
about 2.23 per thousand live births as the female literacy rate increases by one percentage point. The
intercept value of about 263, mechanically interpreted, means that if the values of FLR and PGNP
rate were fixed at zero, the mean child mortality would be about 263 deaths per thousand live births.
−0.0056 is the partial regression coefficient of PGNP and tells us that with the influence of FLR held
constant, as PGNP increases, say, by a dollar, on average, child mortality goes down by 0.0056
units. To make it more economically interpretable, if the per capita GNP goes up by a thousand
dollars, on average, the number of deaths of children under age 5 goes down by about 5.6 per
thousand live births. Of course, such an interpretation should be taken with a grain of salt. All one
could infer is that if the two regressors were fixed at zero, child mortality will be quite high, which
makes practical sense. The R2 value of about 0.71 means that about 71 percent of the variation in

19 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

child mortality is explained by FLR and PGNP, a fairly high value considering that the maximum
value of R2 can at most be 1. All told, the regression results make sense.

4.5 Hypothesis testing and Interval Estimation

The Normality Assumption Once Again

We know by now that if our sole objective is point estimation of the parameters of the regression
models, the method of ordinary least squares (OLS), which does not make any assumption about the
probability distribution of the disturbances ui, will suffice. But if our objective is estimation as well
as inference, then, as argued in chapter 3, we need to assume that the ui follow some probability
distribution. For reasons already clearly spelled out, we assumed that the ui follow the
normal distribution with zero mean and constant variance σ2. We continue to make the same
assumption for multiple regression models. With the normality assumption and following the
discussion of Chapters 3, we find that the OLS estimators of the partial regression coefficients,
estimators, are best linear unbiased estimators (BLUE). Moreover, the estimators, 𝛽̂ 𝛽̂ , and 𝛽̂ are
themselves normally distributed with means equal to true β2, β1, and β0 and the variances given in
this Chapter.
As a result and following, one can show that, upon replacing ζ 2 by its unbiased estimator ̂ 2 in the
computation of the standard errors, each of the following variables
 We can derive the t-value of the OLS estimates follows the t distribution with n− 3 df.

̂
𝑡̂ = ̂

-------------------------4.5.1
̂
𝑡̂ = ̂

With n-3/n-k degree of freedom.

̂
𝑡̂ = ̂

 Where: SE = is standard error and k = number of parameters in the model including the
intercept.

20 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

Example 4.6.1: Child Mortality Example Revisited

̂ = 263.64 5856 2.23 6 𝑋 . 56 𝑋 + 𝑢

𝑠𝑒 = .5932 .2 99 . 9
-------4.5.2
𝑡 = 22.74 .6293 2.8 87
𝑝 𝑣𝑎𝑙𝑢𝑒 = . . . 65
= .7 77 ̅ = .698
What about the statistical significance of the observed results? Consider, for example, the of
PGNP of −0.0056. Is this coefficient statistically significant, that is, statistically different
from zero? Likewise, is the coefficient of FLR of −2.2316 statistically significant? Are both
coefficients statistically significant? To answer this and related questions, let us first consider
the kinds of hypothesis testing that one may encounter in the context of a multiple regression
model.
Hypothesis Testing In Multiple Regressions:
General Comments
Once we go beyond the simple world of the two-variable linear regression model, hypothesis
testing assumes several interesting forms, such as the following:
1. Testing hypotheses about an individual partial regression coefficient
2. Testing the overall significance of the estimated multiple regression model, that is, finding
out if all the partial slope coefficients are simultaneously equal to zero
3. Testing that two or more coefficients are equal to one another
4. Testing that the partial regression coefficients satisfy certain restrictions
5. Testing the stability of the estimated regression model over time or in different cross-
sectional units
6. Testing the functional form of regression models. Since testing of one or more of these
types occurs so commonly in empirical analysis, we devote a section to each type.
Hypothesis Testing About Individual Regression Coefficients
If we invoke the assumption that ui ∼ N (0, ζ 2), then, as noted first, we can use the t test to
21 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module
 Debre Tabor University, CAES, Department of Agricultural Economics

test a hypothesis about any individual partial regression coefficient. To illustrate the
mechanics, consider the child mortality regression(4.5.2): Let us postulate that
Example: 𝛽 =
𝛽
The null hypothesis states that, with X1 (female literacy rate) held constant, X2 (PGNP) has
no (linear) influence on Y (child mortality). To test the null hypothesis, we use the t test
given in (4.5.1). Following Chapter 3 if the computed t value exceeds the critical t value at
the chosen level of significance, we may reject the null hypothesis; otherwise, we may not
reject it. For our illustrative example, using (4.5.1) and noting that β2 = 0 under the null
hypothesis, we obtain
̂ .
𝑡̂ = ̂ = = 2.8 87 = 𝑡
.

as shown in Equation 4.6.2

Notice that we have 64 observations. Therefore, the degrees of freedom in this example are
61 (why?). If you refer to the t table given in t-distribution table, we do not have data
corresponding to 61 df. The closest we have are for 60 df. If we use these df, and assume α,
the level of significance (i.e., the probability of committing a Type I error) of 5 percent, the
critical t value is 2.0 for a two-tail test (look up tα/2 for 60 df) or 1.671 for a one-tail test
(look up tα for 60 df). For our example, the alternative hypothesis is two-sided. Therefore,
we use the two-tail t value. Since the computed t value of 2.8187 (in absolute terms) exceeds
the critical t value of 2, we can reject the null hypothesis that PGNP has no effect on child
mortality. To put it more positively, with the female literacy rate held constant, per capita
GNP has a significant (negative) effect on child mortality, as one would expect a priori.
Graphically, the situation is as shown in Figure 4.3

22 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

.
FIGURE 4.3: The 95% confidence interval for t (60 df)
which in the present case is 0.0065. The interpretation of this p value (i.e., the exact level of
significance) is that if the null hypothesis were true, the probability of obtaining a t value of as much
as 2.8187 or greater (in absolute terms) is only 0.0065 or 0.65 percent, which is indeed a small
probability, much smaller than the artificially adopted value of α = 5%.
This example provides us an opportunity to decide whether we want to use a one-tail or a two-tail t
test. Since a priori child mortality and per capita GNP are expected to be negatively related (why?),
we should use the one-tail test. That is, our null and alternative hypothesis should be:
𝛽
𝛽
As the reader knows by now, we can reject the null hypothesis on the basis of the one-tail t test in the
present instance.
In Chapter 3 we saw the intimate connection between hypothesis testing and confidence interval
estimation. For our example, the 95% confidence interval for β2 is:
𝛽̂ 𝑡 𝑠𝑒 (𝛽̂ ) 𝛽2 𝛽̂ + 𝑡 𝑠𝑒 (𝛽̂ ) 4.5.3

. 56 2. . 2 𝛽2 . 56 + 2. . 2
or
𝛽̂ (𝛽̂ )𝑡 = . 56 .85 . 2 2. = . 56 .79
. 96 𝛽 . 6 4.5.4
That is, the interval, -0.0096 to -0.0016 includes the true β2 coefficient with 95% confidence
coefficient. Thus, if 100 samples of size 64 are selected and 100 confidence intervals like (4.5.4) are

23 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

constructed, we expect 95 of them to contain the true population parameter β2. Since the interval
(4.5.4) does not include the null-hypothesized value of zero, we can reject the null hypothesis that
the true β2 is zero with 95% confidence.
Testing the Overall Significance of the Sample Regression
Throughout the previous section we were concerned with testing the significance of the estimated
partial regression coefficients individually, that is, under the separate hypothesis that each true
population partial regression coefficient was zero. But now consider the following hypothesis:
𝛽 =𝛽 = 4.5.5
This null hypothesis is a joint hypothesis that β1 and β2 are jointly or simultaneously equal to zero. A
test of such a hypothesis is called a test of the overall significance of the observed or estimated
regression line, that is, whether Y is linearly related to both X1 and X2.
Can the joint hypothesis in (4.5.5) be tested by testing the significance of 𝛽̂ and 𝛽̂ individually as
in previous? The answer is no, and the reasoning is as follows.
testing a series of single [individual] hypotheses is not equivalent to testing those same hypotheses
jointly. The intuitive reason for this is that in a joint test of several hypotheses any single hypothesis
is ―affected‘‘ by the information in the other hypotheses.
The upshot of the preceding argument is that for a given example (sample) only one confidence
interval or only one test of significance can be obtained. How, then, does one test the simultaneous
null hypothesis that β1 = β2 = 0? The answer follows.
The Analysis of Variance Approach to Testing the Overall Significance of an Observed
Multiple Regression: The F Test
For reasons just explained, we cannot use the usual t test to test the joint hypothesis that the true
partial slope coefficients are zero simultaneously. However, this joint hypothesis can be tested by the
analysis of variance (ANOVA) technique, which can be demonstrated as follows.
Model Sum of square df Mean square F-statistic
Regression/ ∑ 𝑌̂ 𝑌̅ 𝑘
= =
model/SSE
Due to residual/SSR ∑ 𝑌 𝑌̂ 𝑛 𝑘
= = ̂
𝑛 𝑘
Total + 𝑛

24 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

Table 4.3: ANOVA table for three variable regressions

Source of variation SS df MSS

Due to regression 𝛽̂ ∑ 𝑦 𝑥 + 𝛽̂ ∑ 𝑦 𝑥 2 𝛽̂ ∑ 𝑦 𝑥 + 𝛽̂ ∑ 𝑦 𝑥
2
Due to residual ∑ 𝑢̂ 𝑛 3 ∑ 𝑢̂
= ̂
𝑛 3
Total ∑𝑦 𝑛

Recall identity
∑𝑦 = ∑ 𝑦̂ + ∑ 𝑢̂

∑𝑦 = 𝛽̂ ∑ 𝑦 𝑥 + 𝛽̂ ∑ 𝑦 𝑥 + ∑ 𝑢̂
------------4.5.6
= +

Now it can be shown that, under the assumption of normal distribution for ui and the null hypothesis
β1 = β2 = 0,

Therefore we can compute f-statistic as follows:

𝑑𝑓 𝛽̂ ∑ 𝑦 𝑥 + 𝛽̂ ∑ 𝑦 𝑥 2
= = 4.5.7
𝑑𝑓 ∑ 𝑢̂
𝑛 3
is distributed as the F distribution with 2 and n− 3 df.

Table 4.4: ANOVA table for child mortality example

Source of variation SS df MSS

Due to regression 257 362.4 2 2868 .2
Due to residual 6 3 5.6 6 742.88
Total 363 678 63

25 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

.
We obtain F –statistic: = = 73.8325
.

The p value of obtaining an F value of as much as 73.8325 or greater is almost zero, leading to the
rejection of the hypothesis that together PGNP and FLR have no effect on child mortality. If you
were to use the conventional 5 percent level-of-significance value, the critical F value for 2 df in the
numerator and 60 df in the denominator (the actual df, however, are 61) is about 3.15 or about 4.98
if you were to use the 1 percent level of significance. Obviously, the observed F of about 74 far
exceeds any of these critical F values.

Decision Rule. Given the k-variable regression model:

𝑌𝑖 = 𝛽 + 𝛽 𝑋 𝑖 + 𝛽𝛽 𝑖 + ··· + 𝛽𝑘 𝑋𝑘𝑖 + 𝑢𝑖
To test the hypothesis
𝐻 𝛽 = 𝛽 = ··· = 𝛽𝑘 =
(i.e., all slope coefficients are simultaneously zero) versus H1: Not all slope coefficients are
simultaneously zero compute
𝐸𝑆𝑆
𝐸𝑆𝑆 𝑑𝑓 𝑘 𝐸𝑆𝑆 𝑛 𝑘
𝐹= = = 4.5.8
𝑅𝑆𝑆 𝑑𝑓 𝑅𝑆𝑆 𝑅𝑆𝑆 𝑘
𝑛 𝑘
If F > Fα(k - 1, n- k), reject 𝐻 ; otherwise you do not reject it, where Fα(k - 1, n- k) is the critical F
value at the α level of significance and (k - 1) numerator df and (n- k) denominator df. Alternatively,
if the p value of F obtained from (8.5.7) is sufficiently low, one can reject H0.

An Important Relationship between R2 and F

There is an intimate relationship between the coefficient of determination R2 and the F test used in
the analysis of variance. Assuming the normal distribution for the disturbances ui and the null
hypothesis that β1 = β2 = 0, we have seen that

𝑛 𝑘 𝑛 𝑘 𝑘
= = = 4.5.9
𝑘 𝑘 𝑛 𝑘
Thus the F test, which is a measure of the overall significance of the estimated regression, is also a

26 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

test of significance of R2. In other words, testing the null hypothesis (4.5.5) is equivalent to testing
the null hypothesis that (the population) R2 is zero.
Table 4.5: ANOVA table in terms of R2

Source of variation SS df MSS

Due to regression ∑𝑦 2 ∑𝑦
2
Due to residual ∑𝑦 𝑛 3 ∑𝑦
𝑛 3
Total ∑𝑦 𝑛

For the three-variable case (4.5.9) becomes

.7 77
2 = 2 = 73.8726
𝑛 3 .7 77 6
This is about the same as obtained before, except for the rounding errors. One advantage of the F test
expressed in terms of R2 is its ease of computation: All that one needs to know is the R2 value.

Testing the Equality of Two Regression Coefficients

Suppose in the multiple regression = + 𝛽 + 𝛽 +𝛽 + 𝑢 . .

We want to test the hypotheses

= 𝛽 𝑜𝑟 𝛽 =
------------------4.5.11
𝛽 𝑜𝑟 𝛽
That is, the two slope coefficients 𝑎𝑛𝑑 𝛽 are equal. Such a null hypothesis is of practical
importance. For example, let (4.5.10) represent the demand function for a commodity where Y =
amount of a commodity demanded, X1 = price of the commodity, X2 = income of the consumer, and
X3 = wealth of the consumer. The null hypothesis in this case means that the income and wealth
coefficients are the same. Or, if Yi and the X‘s are expressed in logarithmic form, the null hypothesis
in (4.5.11) implies that the income and wealth elasticities of consumption are the same. (Why?) How
do we test such a null hypothesis? Under the classical assumptions, it can be shown that
(𝛽̂ 𝛽̂ ) 𝛽 𝛽
𝑡= 4.5. 2
𝑠𝑒(𝛽̂ 𝛽̂ )
follows the t distribution with (n− 4) df because (4.5.10) is a four-variable model or, more generally,

27 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

with (n− k) df, where k is the total number of parameters estimated, including the constant term. The
se (𝛽̂ 𝛽̂ ) is obtained from the following well-known formula

𝑠𝑒(𝛽̂ 𝛽̂ ) = √𝑣𝑎𝑟(𝛽̂ + 𝑣𝑎𝑟 𝛽̂ ) 2 𝑐𝑜𝑣(𝛽̂ 𝛽̂ ) 4.5. 3

If we substitute the null hypothesis and the expression for the 𝑠𝑒(𝛽̂ 𝛽̂ ) into (4.5.12), our test
statistic becomes
(𝛽̂ 𝛽̂ )
𝑡= 4.5. 4
√𝑣𝑎𝑟(𝛽̂ + 𝑣𝑎𝑟 𝛽̂ ) 2 𝑐𝑜𝑣(𝛽̂ 𝛽̂ )

Now the testing procedure involves the following steps:

1. Estimate 𝛽̂ 𝑎𝑛𝑑 𝛽̂ . Any standard computer package can do that.
2. Most standard computer packages routinely compute the variances and covariances of the
estimated parameters. From these estimates the standard error in the denominator of (4.5.14) can be
easily obtained.
3. Obtain the t ratio from (4.5.14). Note the null hypothesis in the present case is (β2 – β3) = 0.
4. If the t variable computed from (4.5.14) exceeds the critical t value at the designated level of
significance for given df, then you can reject the null hypothesis; otherwise, you do not reject it.
Alternatively, if the p value of the t statistic from (4.5.14) is reasonably low, one can reject the null
hypothesis. Note that the lower the p value, the greater the evidence against the null hypothesis.
Therefore, when we say that a p value is low or reasonably low, we mean that it is less than the
significance level, such as 10, 5, or 1 percent. Some personal judgment is involved in this decision.
Example: The Cubic Cost Function
Recall the cubic total cost function estimated in which for convenience is reproduced below:
̂ = 4 .7667 + 63.4777𝑋 2.96 5 𝑋 + .9396 𝑋 + 𝑢

𝑠𝑒 = 6.3753 4.7786 .9857 . 59 -------4.5.15

𝑐𝑜𝑣 𝛽̂ 𝛽̂ = . 576 .9983

Where Y is total cost and X is output, and where the figures in parentheses are the estimated
standard errors. Suppose we want to test the hypothesis that the coefficients of the X2and X3 terms in
the cubic cost function are the same, that is, β2 = β3 or (β2 – β3) = 0. In the regression (4.5.15) we
have all the necessary output to conduct the t test of (4.5.14). The actual mechanics are as follows:

28 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

(𝛽̂ 𝛽̂ ) 2.96 5 .9396

𝑡= =
√𝑣𝑎𝑟(𝛽̂ √ .9867 + . 59 2 . 576
+ 𝑣𝑎𝑟 𝛽̂ ) 2 𝑐𝑜𝑣(𝛽̂ 𝛽̂ )

3.9
= = 3.3 3
. 442
The reader can verify that for 6 df (why?) the observed t value exceeds the critical t value even at the
0.002 (or 0.2 percent) level of significance (two-tail test); the p value is extremely small, 0.000006.
Hence we can reject the hypothesis that the coefficients of X2and X3 in the cubic cost function are
identical.

Restricted Least Squares: Testing Linear Equality Restrictions

There are occasions where economic theory may suggest that the coefficients in a regression model
satisfy some linear equality restrictions. For instance, consider the Cobb–Douglas production
function:

𝑌 =𝛽 𝑋 𝑋 𝑒 4.5. 6

Where Y = output, X1 = labor input, and X2 = capital input. Written in log form, the equation
becomes

𝑌𝑖 = 𝛽 + 𝛽 𝑋 + 𝛽 𝑋 + 𝑢 4.5.17

Now if there are constant returns to scale (equiproportional change in output for an equiproportional
change in the inputs), economic theory would suggest that

H0:β1 + β2 = 1----------------------------------------------------4.5.18

This is an example of a linear equality restriction.

How does one find out if there are constant returns to scale, that is, if the restriction (4.5.18) is valid?
There are two approaches.

The t-Test Approach

The simplest procedure is to estimate (4.5.17) in the usual manner without taking into account the
restriction (4.5.18) explicitly. This is called the unrestricted or unconstrained regression. Having

29 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

estimated β1 and β2 (say, by OLS method), a test of the hypothesis or restriction (4.5.18) can be
conducted by the t test of (4.5.12), namely

(𝛽̂ + 𝛽̂ ) 𝛽 +𝛽
𝑡= 4.5. 9
𝑠𝑒(𝛽̂ 𝛽̂ )

(𝛽̂ + 𝛽̂ )
𝑡=
√𝑣𝑎𝑟(𝛽̂ + 𝑣𝑎𝑟 𝛽̂ ) 2 𝑐𝑜𝑣(𝛽̂ 𝛽̂ )

Where (𝛽̂ + 𝛽̂ ) = under the null hypothesis and where the denominator is the standard error of
(𝛽̂ + 𝛽̂ ). Then following if the t value computed from (4.5.19) exceeds the critical t value at the
chosen level of significance, we reject the hypothesis of constant returns to scale; otherwise we do
not reject it.

The F-Test Approach: Restricted Least Squares

The preceding t test is a kind of postmortem examination because we try to find out whether the
linear restriction is satisfied after estimating the ―unrestricted‘‘ regression. A direct approach would
be to incorporate the restriction (4.5.18) into the estimating procedure at the outset. In the present
example, this procedure can be done easily. From (4.5.18) we see that

𝛽 = 𝛽 4.5.2 𝑎

𝑜𝑟 𝛽 = 𝛽 4.5.2 𝑏

Therefore, using either of these equalities, we can eliminate one of the β coefficients in (4.5.17) and
estimate the resulting equation. Thus, if we use (4.5.20a), we can write the Cobb–Douglas
production function as:

𝑌𝑖 = 𝛽 + 𝛽 𝑋 + 𝛽 𝑋 + 𝑢 4.5.21

𝑌𝑖 = 𝛽 + 𝑋 𝛽 𝑋 + 𝛽 𝑋 +𝑢

= 𝛽 + 𝑋 𝛽 𝑋 + 𝑋 +𝑢

30 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

= 𝛽 + 𝑋 + 𝛽 𝑋 𝑋 +𝑢

𝑌𝑖 𝑋 = 𝛽 + 𝛽 𝑋 𝑋 + 𝑢 𝑜𝑟
4.5.22𝑎

𝑋
𝑌𝑖 𝑋 = 𝛽 + 𝛽 𝑙𝑛 ( ) + 𝑢 4.5.22𝑏
𝑋

( 𝑌𝑖 𝑋 ) = output/labor ratio ( 𝑋 𝑋 ) = capital labor ratio, quantities of great

economic importance.
Notice how the original equation (7.5.17) is transformed. Once we estimate β2 from (4.5.22a) or
(4.5.22b), β1 can be easily estimated from the relation (4.5.20a/b). Needless to say, this procedure
will guarantee that the sum of the estimated coefficients of the two inputs will equal 1. The
procedure outlined in (4.5.22a) or (4.5.22b) is known as restricted least squares (RLS). This
procedure can be generalized to models containing any number of explanatory variables and more
than one linear equality restriction. How do we compare the unrestricted and restricted least-squares
regressions? In other words, how do we know that, say, the restriction (4.5.18) is valid? This
question can be answered by applying the F test as follows.
𝑚
= 4.5.23
𝑛 𝑘

Where RSSUR =∑ 𝑢̂ RSS of the unrestricted regression; ∑ 𝑢̂ 𝑜𝑟 RSSR = RSS of the restricted
regression m= number of linear restrictions (1 in the present example) k = number of parameters in
the unrestricted regression n = number of observations. Follows the F distribution with m, (n − k) df.
(Note: UR and R stand for unrestricted and restricted, respectively.) The F test above can also be
expressed in terms of R2 as follows:

𝑚
= 4.5. 24
𝑛 𝑘

where R2 UR and R2R are, respectively, the R2 values obtained from the unrestricted and restricted
regressions, that is, from the regressions (4.5.17) and (4.5.22). It should be noted that

𝑎𝑛𝑑 ∑ 𝑢̂ ∑ 𝑢̂ 4.5.25

31 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

Example: The Cobb–Douglas Production Function for the Ethiopian Economy, 1955–1974

Table 4.6: Real GDP, Employment, and Real Fixed Capital-ethiopia

Year GDP in millions of $ Employment (thousands of people) Fixed capital(millions of $)

1955 114043 8310 182113
1956 120410 8529 193749
1957 129187 8738 205192
1958 134705 8952 215130
1959 139960 9171 225021
1960 150511 9569 237026
1961 157897 9527 248897
1962 165286 9662 260661
1963 178491 10334 275466
1964 199457 10981 295378
1965 212323 11746 315715
1966 226977 11521 337642
1967 241194 11540 363599
1968 260881 12066 391847
1969 277498 12297 422382
1970 296530 12955 455049
1971 306712 13338 484677
1972 329030 13738 520553
1973 354057 15924 561531
1974 374977 14154 609825

By way of illustrating the preceding discussion consider the data given in Table 4.6. Attempting to
fit the Cobb–Douglas production function to these data, yielded the following results:

̂ = .6524 + .3397𝑙𝑛𝑙𝑎𝑏𝑜𝑟 + .846 𝑙𝑛𝑐𝑎𝑝𝑖𝑡𝑎𝑙 + 𝑢

𝑡 = 2.7259 .8295 9. 625

-------4.5.26

𝑝 𝑣𝑎𝑙𝑢𝑒 = . 44 . 849 .
= .995 = . 36
32 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module
 Debre Tabor University, CAES, Department of Agricultural Economics

Where RSSUR is the unrestricted RSS, as we have put no restrictions on estimating (4.5.26). We will
see in nest section how to interpret the coefficients of the Cobb– Douglas production function. As
you can see, the output/labor elasticity is about 0.34 and the output/capital elasticity is about 0.85. If
we add these coefficients, we obtain 1.19, suggesting that perhaps the Ethiopian economy during the
stated time period was experiencing increasing returns to scale. Of course, we do not know if 1.19 is
statistically different from 1.
To see if that is the case, let us impose the restriction of constant returns to scale, which gives the
following regression:
̂ = .4947 + . 53 𝑙𝑛𝑐𝑎𝑝𝑖𝑡𝑎𝑙 𝑙𝑎𝑏𝑜𝑟 + 𝑢

𝑡 = 4. 6 2 28. 56
-------4.5.27
𝑝 𝑣𝑎𝑙𝑢𝑒 = . 7 .
= .9777 = . 66

Where RSSR is the restricted RSS, for we have imposed the restriction that there are constant returns
to scale.
Since the dependent variable in the preceding two regressions is different, we have to use the F test
given in (4.5.24). We have the necessary data to obtain the F value.
. 66 . 36
= == 3.75
. 36 2 3
Note in the present case m = 1, as we have imposed only one restriction and (n − k) is 17, since we
have 20 observations and three parameters in the unrestricted regression. This F value follows the F
distribution with 1 df in the numerator and 17 df in the denominator. The reader can easily check
that this F value is not significant at the 5% level.
The conclusion then is that the Ethiopian economy was probably characterized by constant returns to
scale over the sample period and therefore there may be no harm in using the restricted regression
given in (4.5.27). As this regression shows, if capital/labor ratio increased by 1 percent, on average,
labor productivity went up by about 1 percent.
Testing the Functional Form of Regression: Choosing Between Linear and Log–Linear
Regression Models:
The choice between a linear regression model (the Regressand is a linear function of the regressors)

33 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

or a log–linear regression model (the log of the Regressand is a function of the logs of the
regressors) is a perennial question in empirical analysis. We can use a test proposed by MacKinnon,
White, and Davidson, which for brevity we call the MWD test to choose between the two models.
To illustrate this test, assume the following
H0: Linear Model: Y is a linear function of regressors, the X‘s.
H1: Log–Linear Model: ln Y is a linear function of logs of regressors, the logs of X‘s. Where, as
usual, H0 and H1 denote the null and alternative hypotheses.
The MWD test involves the following steps:
Step I: Estimate the linear model and obtain the estimated Y values. Call them Yf (i.e.,𝑌̂).
Step: II: Estimate the log–linear model and obtain the estimated ln Y values; call them ln f (i.e.,
̂ ).
𝑙𝑛𝑌
Step III: Obtain Z1 = (𝑙𝑛 𝑌 𝑓 𝑙𝑛 𝑓 ).
Step IV: Regress Y on X‘s and Z1 obtained in Step III. Reject H0 if the coefficient of Z1 is
statistically significant by the usual t test.
Step V: Obtain Z2 = (antilog of ln f - Y f ).
Step VI: Regress log of Y on the logs of X‘s and Z2. Reject H1 if the coefficient of Z2 is statistically
significant by the usual t test.
Example:
The demand for roses.* Table 4.7 gives quarterly data on these variables:
Y = quantity of roses sold, dozens
X2 = average wholesale price of roses, $/dozen
X3 = average wholesale price of carnations, $/dozen
X4 = average weekly family disposable income, $/week
X5 = the trend variable taking values of 1, 2, and so on, for the period
1971–III to 1975–II in the Detroit metropolitan area
You are asked to consider the following demand functions:
a). Estimate the parameters of the linear model and interpret the results.
b. Estimate the parameters of the log-linear model and interpret the results.
c. β2, β3, and β4 give, respectively, the own-price, cross-price, and income elasticities of demand.
What are their a priori signs? Do the results concur with the a priori expectations?

34 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

Table 4.7: Demand for rose’s quarterly data for 5 years

Year and quarter Y X2 X3 X4 X5

1971: III 11484 2.26 3.49 158.11 1
IV 9348 2.54 2.85 173.36 2
1972: I 8429 3.07 4.06 165.26 3
II 10079 2.91 3.64 172.92 4
III 9240 2.73 3.21 178.46 5
IV 8862 2.77 3.66 198.62 6
1973: I 6216 3.59 3.76 186.28 7
II 8253 3.23 3.49 188.98 8
III 8036 2.6 3.13 180.49 9
IV 7476 2.89 3.2 183.33 10
1974: I 5911 3.77 3.65 181.87 11
II 7950 3.64 3.6 185.00 12
III 6134 2.82 2.94 184.00 13
IV 5868 2.96 3.12 188.20 14
1975: I 3160 4.24 3.58 175.67 15
II 5872 3.69 3.53 188.00 16

For illustrative purposes, we will consider the demand for roses as a function only of the prices of
roses and carnations, leaving out the income variable for the time being. Now we consider the
following models:
Linear model: 𝑌𝑡 = + 𝑋 + 𝑋 + 𝑢𝑡 4.5.28
– 𝑙𝑛𝑌𝑡 = 𝛽 + 𝛽 𝑙𝑛𝑋 + 𝛽 𝑙𝑛𝑋 + 𝑢𝑡 4.5.29
Where Y is the quantity of roses in dozens, X2 is the average wholesale price of roses ($/dozen), and
X3 is the average wholesale price of carnations ($/dozen). A priori, α2 and β2 are expected to be

negative (why?), and α3 and β3 are expected to be positive (why?). As we know, the slope
coefficients in the log–linear model are elasticity coefficients.
The regression results are as follows:

35 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

𝑌̂ = 9734.2 76 3782. 956𝑋 + 28 5.25 5𝑋 = .77 96

𝑡 = 3.37 5 6.6 69 2.97 2 = 2 .84
̂ = 9.2278
𝑙𝑛𝑌 .76 7𝑙𝑛𝑋 + .3398𝑙𝑛𝑋 = .7292
𝑡= 6.2349 5.9 44 2.54 7 = 7.5
As these results show, both the linear and the log–linear models seem to fit the data reasonably well:
The parameters have the expected signs and the t and R2 values are statistically significant.
To decide between these models on the basis of the MWD test, we first test the hypothesis that the
true model is linear. Then, following Step IV of the test, we obtain the following regression:
𝑌̂ = 9727.5685 3783. 623𝑋 + 28 7.7 57𝑋 + 85.23 9𝑍 = .77 7
𝑡 = 3.2 78 6.3337 2.8366 . 2 7 = 3.44
Since the coefficient of Z1 is not statistically significant (the p value of the estimated t is 0.98), we
do not reject the hypothesis that the true model is linear. Suppose we switch gears and assume that
the true model is log–linear. Following step VI of the MWD test, we obtain the following regression
results:
𝑙𝑛𝑌 = 9. 486 .9699𝑙𝑛𝑋 + .589 𝑙𝑛𝑋 . 3𝑍 = .7798
𝑡= 7. 825 6.4 89 3. 728 .66 2 = 4. 7
The coefficient of Z2 is statistically significant at about the 12 percent level (p value is 0.1225).
Therefore, we can reject the hypothesis that the true model is log–linear at this level of significance.
Of course, if one sticks to the conventional 1 or 5 percent significance levels, then one cannot reject
the hypothesis that the true model is log–linear. As this example shows, it is quite possible that in a
given situation we cannot reject either of the specifications.

4.6. Other functional forms

4.6.1 The Cobb–Douglas Production Function

In this section we demonstrate transformations of nonlinear relationships into linear ones so that we can work
within the framework of the classical linear regression model by taking up the multivariable extension of the
two-variable log–linear model; The specific example we discuss is the celebrated Cobb–Douglas production
function of production theory .The Cobb–Douglas production function, in its stochastic form, may be

expressed as: 𝑌 = 𝛽 𝑋 𝑋 𝑒 4.6.

36 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

The properties of the Cobb–Douglas production function are quite well known:
1. β1 is the (partial) elasticity of output with respect to the labor input, that is, it measures the percentage
change in output for, say, a 1 percent change in the labor input, holding the capital input constant.
2. Likewise, β2 is the (partial) elasticity of output with respect to the capital input, holding the labor input
constant.
3. The sum (β1 + β2) gives information about the returns to scale, that is, the response of output to a
proportionate change in the inputs. If this sum is 1, then there are constant returns to scale, that is, doubling
the inputs will double the output, tripling the inputs will triple the output, and so on. If the sum is less than 1,
there are decreasing returns to scale doubling the inputs will less than double the output. Finally, if the sum is
greater than 1, there are increasing returns to scale doubling the inputs will more than double the output.
A Numerical Example: we obtained the data shown in Table 4.8; these data are for the agricultural
sector of Ethiopia for 1958–1972. Estimate and interpret the coefficient of X1 and X2 if the Cobb–
Douglas production function model, in its stochastic form, is given as follows:
𝑌 =𝛽 𝑋 𝑋 𝑒 4.6.2𝑎
where Y = output X1 = labor input X2 = capital input u = stochastic disturbance term
e = base of natural logarithm or
𝑌𝑖 = 𝛽 + 𝛽 𝑋 + 𝛽 𝑋 + 𝑢 4.6.2𝑏
𝑌𝑖 = + 𝛽 𝑙𝑛 𝑋 + 𝛽 𝑙𝑛 𝑋 + 𝑢 4.6.2𝑐
where = 𝑙𝑛𝛽 . Thus written, the model is linear in the parameters β0, β2, and β3 and is
therefore a linear regression model. Notice, though, it is nonlinear in the variables Y and X
but linear in the logs of these variables. In short, (4.6.2b/c) is a log-log, double-log, or log-
linear model, the multiple regression counterpart of the two-variable log-linear model.

37 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

Table 4.8: Real Gross Product, Labor Days, and Real Capital Input in the Agricultural Sector of Taiwan, 1958–1972

Real GDP (in Labor days (millions of Real capital input

Observation Year millions of $) Y days) X1 (millions of $) X2
1 1958 16607.7 275.5 17803.7
2 1959 17511.3 274.4 18096.8
3 1960 20171.2 269.7 18271.8
4 1961 20932.9 267 19167.3
5 1962 20406 267.8 19647.6
6 1963 20831.6 275 20803.5
7 1964 24806.3 283 22076.6
8 1965 26465.8 300.7 23445.2
9 1966 27403 307.5 24939
10 1967 28628.7 303.7 26713.7
11 1968 29904.5 304.7 29957.8
12 1969 27508.2 298.6 31585.9
13 1970 29035.5 295.5 33474.5
14 1971 29281.5 299 34821.8
15 1972 31535.8 288.1 41794.3
Sum 371030 4310.2 382599.5
Mean 24735.33333 287.3466667 25506.63333

𝑙𝑛𝑌̂ = 3.3384 + .4988𝑙𝑛𝑋 + .4899𝑙𝑛𝑋

𝑠𝑒 = 2.4495 .5398 . 2

𝑡= .3629 2.7765 4.8 5 ---------------------4.6.3

= .889 𝑎𝑛𝑑 ̅ = .87 5 𝑑𝑓 = 2

From Equation (4.6.3) we see that in the Ethiopian agricultural sector for the period 1958–1972 the
output elasticities of labor and capital were 1.4988 and 0.4899, respectively. In other words, over the
period of study, holding the capital input constant, a 1% increase in the labor input led on the

38 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

average to about a 1.5 % increase in the output. Similarly, holding the labor input constant, a 1%
increase in the capital input led on the average to about a 0.5% increase in the output. Adding the
two output elasticities, we obtain 1.9887, which gives the value of the returns to scale parameter. As
is evident, over the period of the study, the Taiwanese agricultural sector was characterized by
increasing returns to scale.
From a purely statistical viewpoint, the estimated regression line fits the data quite well. The R 2
value of 0.8890 means that about 89% of the variation in the (log of) output is explained by the (logs
of) labor and capital. We were seen in previous section how the estimated standard errors can be
used to test hypotheses about the ―true‖ values of the parameters of the Cobb– Douglas production
function for Ethiopian economy.

4.6.2 Polynomial Regression Models

We now consider a class of multiple regression models, the polynomial regression models that
have found extensive use in econometric research relating to cost and production functions. In
introducing these models, we further extend the range of models to which the classical linear
regression model can easily be applied.
Consider the cost function given by = 𝛽 + 𝛽 𝑋 + 𝛽2𝑋 4.6.4𝑎

which is called a quadratic function, or more generally, a second-degree polynomial in the variable
X—the highest power of X represents the degree of the polynomial.
The stochastic version of (4.6.3) may be written as
= 𝛽 + 𝛽 𝑋 + 𝛽2𝑋 + 𝑢̂ 4.6.4𝑏
The general kth degree polynomial regression may be written as
= 𝛽 + 𝛽 𝑋 + 𝛽2𝑋 + +𝛽𝑘𝑋 + 𝑢̂ 4.6.5
Notice that in these types of polynomial regressions there is only one explanatory variable on the
right-hand side but it appears with various powers, thus making them multiple regression models.
Incidentally, note that if Xi is assumed to be fixed or nonstochastic; the powered terms of Xi also
become fixed or nonstochastic.

Do these models present any special estimation problems? Since the second-degree polynomial
(4.6.4b) or the kth degree polynomial (4.6.5) is linear in the parameters, the β‘s, they can be
estimated by the usual OLS or ML methodology. But what about the collinearity problem? Aren‘t
the various X‘s highly correlated since they are all powers of X? Yes, but remember that terms like
39 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module
 Debre Tabor University, CAES, Department of Agricultural Economics

X 2, X 3, X 4, etc., are all nonlinear functions of X and hence, strictly speaking, do not violate the no
Multicollinearity assumption. In short, polynomial regressions models can be estimated by the
techniques presented in this chapter and present no new estimation problems.

Example: Estimating the Total Cost Function: As an example of the polynomial regression, consider the
data on output and total cost of production of a commodity in the short run given in Table 4.9. What type of
regression model will fit these data?
Table 4.9: Total Cost (Y) and Output(X)

Output(X) Total cost(birr)(Y)

1 193
2 226
3 240
4 244
5 257
6 260
7 274
8 297
9 350
10 420

This S shape of the total cost curve can be captured by the following cubic or third-degree polynomial:
𝑌 = 𝛽 + 𝛽 𝑋 + 𝛽2𝑋 + 𝛽3𝑋 + 𝑢̂ 4.6.6
Where Y = total cost and X = output
When the third-degree polynomial regression was fitted to the data of Table 4.3, we obtained the following
results:
𝑌̂ = 4 .7667 + 63.4776𝑋 + 2.96 5𝑋 + .9396𝑋 4.6.7
𝑠𝑒 = (6.3753 4.7786 .9857 . 59 = .9983

40 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

Review Question

1. based on the data given in the Table 4.7 gives quarterly data on each variables and You are asked
to consider the following demand functions:
Linear model: 𝑌𝑡 = + 𝑋 + 𝑋 + 𝑋 + 𝑋 + 𝑢𝑡
– 𝑙𝑛𝑌𝑡 = 𝛽 + 𝛽 𝑙𝑛𝑋 + 𝛽 𝑙𝑛𝑋 + 𝛽 𝑙𝑛𝑋 + 𝛽 𝑙𝑛𝑋 + 𝑢𝑡
Y = quantity of roses sold, dozens
X2 = average wholesale price of roses, $/dozen
X3 = average wholesale price of carnations, $/dozen
X4 = average weekly family disposable income, $/week
X5 = the trend variable taking values of 1, 2, and so on, for the period 1971–III to 1975–II in the
Detroit metropolitan area
a) Estimate the parameters of the linear model and interpret the results.
b) Estimate the parameters of the log-linear model and interpret the results.
c) β2, β3, and β4 give, respectively, the own-price, cross-price, and income elasticities of demand.
What are their a priori signs? Do the results concur with the a priori expectations?
d) How would you compute the own-price, cross-price, and income elasticities for the linear
model?
e) On the basis of your analysis, which model, if either, would you choose and why?
2. Is it possible to obtain the following from a set of data?

a) 𝑟 = .9 𝑟 = .2 𝑟 = .8
b) 𝑟 = .6 𝑟 = .9 𝑟 = .5
c) 𝑟 = . 𝑟 = .66 𝑟 = .7
3. From the following data estimate the partial regression coefficients, their standard errors, and the
adjusted and unadjusted R2 values:

∑ 𝑌 𝑌̅ = 66 42.269 𝑌̅ = 367.693 𝑋̅ = 4 2.76 𝑋̅ = 8. 𝑛 = 5

∑ 𝑋 𝑋̅ = 84855. 96

∑ 𝑋 𝑋̅ = 28 .

41 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

∑ 𝑋 𝑋̅ 𝑋 𝑋̅ = 4796.

∑ 𝑋 𝑋̅ 𝑌 𝑌̅ = 74778.346

∑ 𝑋 𝑋̅ 𝑌 𝑌̅ = 425 .9

4. Consider the data in Table 4.1

Y X1 X2
1 1 2
3 2 1
8 3 -3
Based on these data, estimate the following regressions:

𝑌 = + 𝑋 +𝑢
𝑌 =𝛽 +𝛽 𝑋 +𝑢
𝑌 = + 𝑋 + 𝑋 +𝑢

Note: Estimate only the coefficients and not the standard errors.
a) is = ?? Why or why not?
b) Is = 𝛽 ? ? Why or why not?
c) What important conclusion do you draw from this exercise?
( )
5. Show that 𝑟 . = 𝑎𝑛𝑑 𝑖𝑛𝑡𝑒𝑟𝑝𝑟𝑢𝑡 𝑡𝑕𝑒 𝑒𝑞𝑢𝑎𝑡𝑖𝑜𝑛

6. If the relation 𝑋 + 𝑋 + 𝑋 = for all values of X1, X2, and X3, find the values of the
three partial correlation coefficient?
7. Consider the following regression results 𝑌̂ = 6899 2978.5𝑋 𝑟 = .6 49
𝑡 = 8.5 52 4.728
̂𝑌 = 9734.2 3782.2𝑋 + 28 5𝑋
𝑡 = 3.37 5 6.6 7 2.97 2 = .77 6
8. Based on our discussion of individual and joint tests of hypothesis based, respectively, on the t
and F tests, which of the following situations are likely?
a) Reject the joint null on the basis of the F statistic, but do not reject each separate null on the

42 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

basis of the individual t tests.

b) Reject the joint null on the basis of the F statistic, reject one individual hypothesis on the
basis of the t test, and do not reject the other individual hypotheses on the basis of the t test.
c) Reject the joint null hypothesis on the basis of the F statistic, and reject each separate null
hypothesis on the basis of the individual t tests.
d) .Do not rejects the joint null on the basis of the F statistic, and do not reject each separate null
on the basis of individual t tests.
e) Do not reject the joint null on the basis of the F statistic, reject one individual hypothesis on
the basis of a t test, and do not reject the other individual hypotheses on the basis of the t test.
f) Do not reject the joint null on the basis of the F statistic, but reject each separate null on the
basis of individual t tests.
9. From a sample of 209 firms, Wooldridge obtained the following regression results*:

̂
𝑠𝑎𝑙𝑎𝑟𝑦 = 4.32 + .28 𝑠𝑎𝑙𝑒𝑠 + . 74𝑟𝑜𝑒 + . 24𝑟𝑜𝑠
𝑠𝑒 = .32 . 35 . 4 . 54 = .283
Where salary = salary of CEO
sales = annual firm sales
roe = return on equity in percent
ros = return on firm‘s stock and where figures in the parentheses are the estimated standard
errors.
a. Interpret the preceding regression taking into account any prior expectations that you may have
about the signs of the various coefficients.
b. Which of the coefficients are individually statistically significant at the 5 percent level?
c. What is the overall significance of the regression? Which test do you use? And why?
d. Can you interpret the coefficients of ―roe‖ and ―ros‖ as elasticity coefficients? Why or why not?
10. Return to the child mortality example that we have discussed several times. In regression (4.7.3)
we regressed child mortality (CM) on per capita GNP (PGNP) and female literacy rate (FLR).
Now we extend this model by including total fertility rate (TFR). The data on all these variables
are already given in Table 4.5. We reproduce regression (4.7.3) and give results of the extended
regression model below:
Model I: ̂ = 263.64 6 . 56 2.23 6
43 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module
 Debre Tabor University, CAES, Department of Agricultural Economics

4.7.3𝑎
𝑠𝑒 = .5932 . 9 .2 99 = .7 77
Model II: ̂ = 68.3 67 . . 55 .768 + 2.8686
4.7.3𝑏
𝑠𝑒 = 32.89 6 . 8 .248 = .7474
a. How would you interpret the coefficient of TFR? A priori, would you expect a positive or
negative relationship between CM and TFR? Justify your answer.
b. Have the coefficient values of PGNP and FR changed between the two equations? If so, what
may be the reason(s) for such a change? Is the observed difference statistically significant?
Which test do you use and why?
c. How would you choose between models 1 and 2? Which statistical test would you use to
answer this question? Show the necessary calculations.
d. We have not given the standard error of the coefficient of TFR. Can you find it out? (Hint:
Recall the relationship between the t and F distributions.)
11. Given the table below which gave data on advertising impressions retained and advertising
expenditure for a sample of 21 firms. Decide on an appropriate model about the relationship
between impressions and advertising expenditure. Letting Y represent impressions retained and
X the advertising expenditure, the following regressions were obtained:
Model I: 𝑌̂ = 22. 63 + .363𝑋
𝑠𝑒 = 7. 89 . 97 𝑟 = .424=
Model II: 𝑌̂ = 7. 59 + . 847𝑋 + . 4 𝑋
𝑠𝑒 = 9.986 .3699 . 9 = .53
a. Interpret both models.
b. Which is a better model? Why?
c. Which statistical test(s) would you use to choose between the two models?
d. Are there ―diminishing returns‖ to advertising expenditure, that is, after a certain level of
advertising expenditure (the saturation level) it does not pay to advertise? Can you find out
what that level of expenditure might be? Show the necessary calculations.
Table 4.9: impact of advertising expenditure

Firm Impressions, in millions Expenditure, millions of birr

1 32.1 50.1

44 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

2 99.6 74.1
3 11.7 19.3
4 21.9 22.9
5 60.8 82.4
6 78.6 40.1
7 92.4 185.9
8 50.7 26.9
9 21.4 20.4
10 40.1 166.2
11 40.8 27.0
12 10.4 45.6
13 88.9 154.9
14 12.0 5.0
15 29.2 49.7
16 38.0 26.9
17 10.0 5.7
18 12.3 7.6
19 23.4 9.2
20 71.1 32.4
21 4.4 6.1

12. Suppose the demand for commodity; its price and the price of related goods at different market
outlet recoded in the table 1 as follows:
Table 4.10: The Demand for Commodity Y Depend on its own price (X1) and the Price of Related Goods(X2)

Y X1 X2 Observation
64 57 8 1
71 59 10 2
53 49 6 3
67 62 11 4
55 51 8 5
58 50 7 6
77 55 10 7
57 48 9 8
a) a). Estimate the coefficient of the parameter in the econometric demand model of commodity Y given
as: 𝑌̂ = 𝛽̂ + 𝛽̂ 𝑋 + 𝛽̂ 𝑋 + 𝑢̂
b) Estimate the variance and standard error of the model as well as the parameters?
c) Compute TSS, RSS and ESS?
d) Compute R2 and F- statistic for the model?
e) Test whether the coefficient for price of related goods becomes zero at 5% significance level?
f) Test whether the coefficient for both X1and X2 are becomes zero?
g) Test whether the coefficient for both X1and X2 are equal?

45 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

13. suppose the demand for tea and its own price as well as price of coffee is shown is given as
follows in different coffee house owners
Table 4.11: demand for tea and its price

Y X1 X2
-3.7 3 8
3.5 4 5
2.5 5 7
11.5 6 3
5.7 2 1

a) Estimate the coefficient of the parameter in the econometric demand model of commodity Y given
as: 𝑌̂ = 𝛽̂ + 𝛽̂ 𝑋 + 𝛽̂ 𝑋 + 𝑢̂
b) Estimate the variance and standard error of the model as well as the parameters?
c) Compute TSS, RSS and ESS?
d) Compute R2 and F- statistic for the model?
e) Test whether the coefficient for price of related goods becomes zero at 5% significance level?
f) Test whether the coefficient for both X1and X2 are becomes zero?
g) Test whether the coefficient for both X1and X2 are equal?

46 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

CHAPTER FIVE: VIOLATIONS OF THE CLASSICAL ASSUMPTIONS

5.1 Non-Normality

5.1.1 Nature of the problem

The normality assumption for multiple regressions is one of the most misunderstood in all of
statistics. In multiple regressions, the assumption requiring a normal distribution applies only to the
residuals, not to the independent variables as is often believed. Perhaps the confusion about this
assumption derives from difficulty understanding what the residuals are simply put, the residuals are
the error in the relationship between the independent variables and the dependent variable in a
regression model. Each case in the sample has a residual value that represents the difference in the
observed and predicted values produced by a regression equation. It is the distribution of the
residuals or noise for all cases in the sample that should be normally distributed.
Classical normal linear regression (CNLR) assumes that each is distributed normally
𝑤𝑖𝑡𝑕
Mean = E (Ui) = 0
Variance = E (Ui2) =
Cov (Ui, Uj) = E (Ui, Uj) = 0 (i # j)
Note: For two normally distributed variables, the zero covariance or correlation means independence
of them, so Ui and Uj are not only uncorrelated but also independently distributed. Therefore
NID (0, ) is Normal and independently distributed(Note: ID stands for Normally and
Independently Distributed).

Why the Normality Assumption?

Why do we employ the normality assumption? There are several reasons:
1. 𝑢 Represent the combined influence (on the dependent variable) of a large number of
independent variables that are not explicitly introduced in the regression model. As noted, we hope
that the influence of these omitted or neglected variables is small and at best random. Now by the
celebrated central limit theorem (CLT) of statistics, it can be shown that if there are a large number
of independent and identically distributed random variables, then, with a few exceptions, the
47 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module
 Debre Tabor University, CAES, Department of Agricultural Economics

distribution of their sum tends to a normal distribution as the number of such variables increase
indefinitely.1 It is the CLT that provides a theoretical justification for the assumption of normality of
𝑢.
2. A variant of the CLT states that, even if the number of variables is not very large or if these
variables are not strictly independent, their sum may still be normally distributed.
3. With the normality assumption, the probability distributions of OLS estimators can be easily
derived because, one property of the normal distribution is that any linear function of normally
distributed variables is itself normally distributed. As we discussed earlier, OLS estimator‘s 𝛽̂ and
𝛽̂ are linear functions of ui. Therefore, if ui are normally distributed, so are 𝛽̂ and 𝛽̂ , which makes
our task of hypothesis testing very straightforward.
4. The normal distribution is a comparatively simple distribution involving only two parameters
(mean and variance); it is very well known and its theoretical properties have been extensively
studied in mathematical statistics. Besides, many phenomena seem to follow the normal distribution.
5. Finally, if we are dealing with a small, or finite, sample size, say data of less than 100
observations, the normality assumption assumes a critical role. It not only helps us to derive the
exact probability distributions of OLS estimators but also enables us to use the t, F, and χ2 statistical
tests for regression models.. As we will show subsequently, if the sample size is reasonably large, we
may be able to relax the normality assumption. A cautionary note: Since we are ―imposing‖ the
normality assumption, it behooves us to find out in practical applications involving small sample
size data whether the normality assumption is appropriate. Later, we will develop some tests to do
just that. Also, later we will come across situations where the normality assumption may be
inappropriate. But until then we will continue with the normality assumption for the reasons
discussed previously.
Violations of normality create problems for determining whether model coefficients are significantly
different from zero and for calculating confidence intervals for forecasts. Sometimes the error
distribution is "skewed" by the presence of a few large outliers. Since parameter estimation is based
on the minimization of squared error, a few extreme observations can exert a disproportionate
influence on parameter estimates. Calculation of confidence intervals and various significance tests
for coefficients are all based on the assumptions of normally distributed errors. If the error
distribution is significantly non-normal, confidence intervals may be too wide or too narrow.

48 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

Technically, the normal distribution assumption is not necessary if you are willing to assume the
model equation is correct and your only goal is to estimate its coefficients and generate predictions
in such a way as to minimize mean squared error. The formulas for estimating coefficients require
no more than that, and some references on regression analysis do not list normally distributed errors
among the key assumptions. But generally we are interested in making inferences about the model
and/or estimating the probability that a given forecast error will exceed some threshold in a
particular direction, in which case distributional assumptions are important. Also, a significant
violation of the normal distribution assumption is often a "red flag" indicating that there is some
other problem with the model assumptions and/or that there are a few unusual data points that should
be studied closely and/or that a better model is still waiting out there somewhere.

5.1.2 Source of non-normality

There are four main sources of non –normality problem in our data.
 Extreme values or outliers: examples the income data in developing countries may have
extreme values or have outliers. That is some observations have minimum value but the other
have maximum value. In this case the data have large range.
 Two or more process overlapping: happens when the distribution of data which is the
combination of two or more normal distributions. Therefore we have bimodal or multimodal
distribution.
 Insufficient data: the more data you have the more likely to follow normal distribution. You
should have the data at least from 30 observations. Example, if we increase observation to
50; the data the more likely to follow normal distribution.
 Subset of main sample: if the data is the subset of main sample and you have not taken the
indirect data. This is happen because of sampling error.
 Data follows same other distribution: no reason every data should be follow normal
distributions. In real world the data may follow other distributions like beta distribution,
logistic distribution, exponential distribution, Poisson distribution, gamma distribution etc.

5.1.3 Consequence of non-normality

There are few consequences associated with a violation of the normality assumption, as it does not
contribute to bias or inefficiency in regression models. It is only important for the calculation of p
49 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module
 Debre Tabor University, CAES, Department of Agricultural Economics

values for significance testing, but this is only a consideration when the sample size is very small.
When the sample size is sufficiently large 2 ), the normality assumption is not needed at all as
the Central Limit Theorem ensures that the distribution of residuals will approximate normality.

5.1.4 Detecting non –normality

When dealing with very small samples, it is important to check for a possible violation of the
normality assumption. This can be accomplished through an inspection of the residuals from the
regression model (some programs will perform this automatically while others require that you save
the residuals as a new variable and examine them using summary statistics and histograms). There
are several statistics available to examine the normality of variables, including skewness and
kurtosis, as well as numerous graphical depictions, such as the normal probability plot.
Unfortunately, the statistics to assess it are unstable in small samples, so their results should be
interpreted with caution. When the distribution of the residuals is found to deviate from normality,
Quantile plot: the quantile plot for residuals also used to detect the problem of normality in the
data.
Normality plot p plot: such as histogram probability plot for residuals also shows whether there is
the problem or not.

5.1.5 Remedial measure

What do you if your data are not normal? : Possible solutions include transforming the data,
removing outliers, or conducting an alternative analysis that does not require normality (e.g., a
nonparametric regression).
1. Remove outliers: If the observation is large enough we can remove outlier individual data‘s if
they are few in number.
2. Data transformation: use the uniform log transformation or other mathematical change used
across the board to increase normality. This is ok and legitimate to minimize the gap between
outliers and make the distribution normal. For example log transformations are good when the
data are positively skewed. However keep in mind interpretability of transformation.
3. Fit another distribution:
4. Use non-parametric test: if you do not want to deal with transformations you can use non-
parametric tests that do not require data to be normally distributed.
50 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module
 Debre Tabor University, CAES, Department of Agricultural Economics

5.2 MULTICOLLINEARITY

5.2.1 The nature of the problem

One of the assumptions of the classical linear regression model (CLRM) is that there is no
perfect Multicollinearity among the regressors included in the regression model. Although the
assumption is said to be violated only in the case of exact Multicollinearity (i.e. an exact linear
relationship among some of the regressors), the presence of Multicollinearity (an approximate
linear relationship among some of the regressors) leads to estimation problems.
Multicollinearity does not depend on any theoretical or actual linear relationship among any of the
regressors; it depends on the existence of an approximate linear relationship in the data set at hand.
Unlike most other estimation problems, this problem is caused by the particular sample available.
The existence of Multicollinearity will affect seriously the parameter estimates. Intuitively, when any
two explanatory variables are changing in nearly the same way, it becomes extremely difficult to
establish the influence of each regressor on the dependent variable separately.
Consider the consumption–income model
𝑜𝑛𝑠𝑢𝑚𝑝𝑡𝑖𝑜𝑛 = 𝛽 + 𝛽 + 𝛽 +𝑢
It may happen that when we obtain data on income and wealth, the two variables may be highly, if
not perfectly, correlated: Wealthier people generally tend to have higher incomes. Thus, although in
theory income and wealth are logical candidates to explain the behavior of consumption expenditure,
in practice (i.e., in the sample) it may be difficult to disentangle the separate influences of income
and wealth on consumption expenditure.
Ideally, to assess the individual effects of wealth and income on consumption expenditure we need a
sufficient number of sample observations of wealthy individuals with low income, and high-income
individuals with low wealth. Although this may be possible in cross sectional studies (by increasing
the sample size), it is very difficult to achieve in aggregate time series work.
In general, the problem of Multicollinearity arises when individual effects of explanatory variables
cannot be isolated and the corresponding parameter magnitudes cannot be determined with the
desired degree of precision. Though it is quite frequent in cross section data as well,
Multicollinearity tends to be more common and more serious problem in time series data.
Building a linear regression model is only half of the work. In order to actually be usable in practice,
the model should conform to the assumptions of linear regression.
51 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module
 Debre Tabor University, CAES, Department of Agricultural Economics

5.2.2 Sources of Multicollinearity

The data collection method employed, for example, sampling over a limited range of the values
taken by the regressors in the population.
Constraints on the model or in the population being sampled: For example, in the
Regression of electricity consumption on income and house size there is a physical constraint in the
population in that families with higher incomes generally have larger homes than families with
lower incomes.
Model specification, for example, adding polynomial terms to a regression model, especially when
the range of the X variable is small.
An over determined model. This happens when the model has more explanatory variables than the
number of observations. This could happen in medical research where there may be a small number
of patients about whom information is collected on a large number of variables.
An additional reason for Multicollinearity, especially in time series data, may be that the regressors
included in the model share a common trend, that is, they all increase or decrease over time. Thus,
in the regression of consumption expenditure on income, wealth, and population, the regressors
income, wealth, and population may all be growing over time at more or less the same rate, leading to
collinearity among these variables.

5.2.3 Consequences of Multicollinearity

I. The case of perfect Multicollinearity
Consider a three-variable regression model in deviation form

𝑦 = 𝛽̂ 𝑥 + 𝛽̂ 𝑥 + 𝑢̂

From which it can be obtained

∑𝑥 𝑦 ∑𝑥 ∑𝑥 𝑦 ∑𝑥 𝑥
𝛽̂ =
∑𝑥 ∑𝑥 ∑𝑥 𝑥

52 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

∑𝑥 𝑦 ∑𝑥 ∑𝑥 𝑦 ∑𝑥 𝑥
𝛽̂ =
∑𝑥 ∑𝑥 ∑𝑥 𝑥
Suppose
X 2i = kX1i where k is non-zero constant.
Then
∑𝑥 𝑦 𝑘 ∑𝑥 𝑘∑𝑥 𝑦 𝑘∑𝑥
𝛽̂ = =
∑𝑥 𝑘 ∑𝑥 𝑘 ∑𝑥 𝑥

This is an indeterminate expression. It can also be shown that the expression for 𝛽̂ is
indeterminate.

Recall that 𝛽̂ gives the rate of change in the average value of Y as X1 changes by a unit, holding X2
constant. But if X1 and X2 are perfectly collinear, there is no way X2 can be kept constant: As X1
changes, so does X2 by the factor k. What it means, then, is that there is no way of disentangling the
separate influences of X1 and X2 from the given sample.
Moreover, for a three variable model:
∑𝑥
𝛽̂ =
∑𝑥 ∑𝑥 ∑𝑥 𝑥
∑𝑥
𝛽̂ =
∑𝑥 ∑𝑥 ∑𝑥 𝑥
Substituting 𝑥 = 𝑘𝑥

𝑘 ∑𝑥 ∑𝑥
𝛽̂ = = =
𝑘 ∑𝑥 ∑𝑥 𝑘 ∑𝑥 𝑥

This shows that the variances of the estimates become infinite.

Therefore, in the case of perfect Multicollinearity, the regression coefficients remain
Indeterminate and their standard errors are infinite.

II. High but imperfect Multicollinearity

The perfect Multicollinearity situation is a pathological extreme. Generally, there is no exact linear
relationship among the X variables, especially in data involving economic time series.

53 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

Suppose in the three variable model.

𝑋 2𝑖 = 𝑘𝑋 𝑖 + 𝑣 𝑤𝑕𝑒𝑟𝑒 𝑘 = 𝑎𝑛𝑑 𝑣 is the stochastic error term such that ∑ .
𝑥𝑦 =
In this case, it is possible to estimate the coefficients. However, in the case of near or high
Multicollinearity, one is likely to encounter the following consequences:
Although BLUE, the OLS estimators have large variances and covariances, making precise
estimation difficult. For example, in a three variable linear regression,

𝑣𝑎𝑟 𝛽̂ =
∑𝑥 𝑟
It is apparent from the above formula that as 𝑟 (which is the coefficient of correlation between X1
and X2) tends towards 1, that is, as collinearity increases, the Variance of the estimator increases.
The same holds for 𝑣𝑎𝑟 𝛽̂ and the 𝛽̂ 𝛽̂ )
Because of consequence (1), the confidence intervals tend to be much wider, leading to the
acceptance of the ―Zero null hypothesis‖ (i.e., the true population coefficient is zero). Because of
consequence (1), the t-ratio of one or more coefficients tends to bestatistically insignificant.
Although the t-ratio of one or more coefficients is statistically insignificant, R2, the overall
measure of goodness of fit, can be very high. The OLS estimators and their standard errors can
be sensitive to small changes in the data.

5.2.4 Detecting Multicollinearity

Note that Multicollinearity is a question of degree and not of a kind. It should also be noted that
since Multicollinearity refers to the condition of the explanatory variables that are assumed to be
nonstochastic, it is a feature of the sample and not of the population. Therefore, we do not ―test for
Multicollinearity‖ but can, if we wish, measure its degree in any particular sample. The following
are some rules of thumb and formal rules to detection of Multicollinearity.
a) High R2 but few significant t-ratios. If R2 is high, say in excess of 0.8, the F-test in most cases
will reject the hypothesis that the partial slope coefficients are simultaneously equal to zero, but
the individual t tests will show that none or very few of the partial slope coefficients are
statistically different from zero.
b) High pair-wise correlation among regressors. If the pair-wise correlation coefficient among
two regressors is high, say in excess of 0.8, then Multicollinearity is a serious problem.
c) Auxiliary Regression. Since Multicollinearity arises because one or more of the regressors are

54 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

exact or approximately linear combinations of the other regressors, one way of finding out which
X variable is related to other X variables is to regress each Xi on the remaining X variables and
compute the corresponding R2. As a rule of thumb, Multicollinearity may be a troublesome
problem only if the R2 obtained from an auxiliary regression is greater than the overall R2 (that
obtained from the regression of Y on all the regressors).
d) Tolerance (TOL) and variance inflation factor (VIF): For a linear regression with two
explanatory variables (X1 and X2) =

VIF shows how the variance of an estimator is inflated by the presence of Multicollinearity.
12
As 𝑟 approaches 1, the VIF approaches infinity. That is, as the extent of collinearity increases,
the variance of an estimator increases, and in the limit it can become infinite. If there is no
collinearity between X1 and X2, VIF will be 1.

𝑣𝑎𝑟(𝛽̂ ) = = 𝐼
∑𝑥 𝑟 ∑𝑥

𝑣𝑎𝑟(𝛽̂ ) = = 𝐼
∑𝑥 𝑟 ∑𝑥
Similarly for k-variables model

𝑣𝑎𝑟(𝛽̂ ) = = 𝐼
∑𝑥 ( ) ∑𝑥

Where (𝛽̂ )= estimated partial regression coefficient of regressor 𝑋

= 𝑖𝑛 𝑡𝑕𝑒 𝑟𝑒𝑔𝑟𝑒𝑠𝑠𝑖𝑜𝑛 𝑜𝑓 𝑋 on the remaining regressions
Shows that the variance of 𝛽̂ 𝑎𝑛𝑑 𝛽̂ are directly proportional to the VIF

It is also possible to use TOL as the measure of Multicollinearity in the view of its intimate
connection with VIF

= =
𝐼
The larger the value of the VIFi, the more ―troublesome‖ or collinear the variable Xi. As a rule of
Thumb, if the VIF of a variable exceeds 10, which will happen if R 2 exceeds 0.90, that variable is
i

said to be highly collinear. In other words, the closer is TOLi to zero, the greater the degree of
collinearity of that variable with the other regressors. On the other hand, the closer TOLi is to 1, the

55 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

greater the evidence that Xi is not collinear with the other regressors.

5.2.5 Remedial Measures

The existence of Multicollinearity in a data set does not necessarily mean that the coefficient
estimators in which the researcher is interested have unacceptably high variance. Because
Multicollinearity is essentially a sample problem there are no infallible guides. However one can
try the following rules of thumb, the success of which depends on the severity of the collinearity
problem.
a) Obtain more data: - Because the Multicollinearity is essentially a data problem, additional data
that do not contain the Multicollinearity feature could solve the problem. For example, in the
2
three variable models we saw that now as the sample size increases, x will generally
increases. Thus, for a given r12 the variance of ̂ 1 decreases (thus, the standard error
decreases), which will enable us to estimate ̂ 1 more precisely.
b) Transformation of variables: - In time series analysis, one reason for high Multicollinearity
between two variables is that over time both variables tend to move in the same direction. One
way of minimizing then dependence is to transform the variables.
Suppose𝑌 = 𝛽 + 𝛽 𝑋 + 𝛽𝑋 +
This relation must also hold at time t-1 because the origin of time is arbitrary anyway. Therefore
we have 𝑌 = 𝛽 + 𝛽𝑋 + 𝛽𝑋 +
Subtracting this from the above gives
𝑌 𝑌 = +𝛽 𝑋 𝑋 + 𝛽 𝑋 𝑋 +
This is known as the first difference form because we run the regression, not on the original
variables, but on the difference of successive values of the variables. The first difference regression
model often reduces the severity of Multicollinearity. Although the levels of X1 and X2 may be
highly correlated, there is no a priori reason to believe that their difference will also be highly
correlated.
Another commonly used transformation in practice is the ratio transformation.
Consider the model:
𝑌 = 𝛽 + 𝛽𝑋 + 𝛽𝑋
where Y is consumption expenditure in real dollars, X1 is GDP, and X2 is total population. Since GDP
and population grow over time, they are likely to be correlated. One ―solution‖ to this problem is to
56 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module
 Debre Tabor University, CAES, Department of Agricultural Economics

express the model on a per capita basis, that is,

𝑌 𝛽 𝛽𝑋 𝛽𝑋 𝑢 𝑌 𝑋
= + + + = =𝛽 ( )+𝛽 +𝛽 +𝑢
𝑋 𝑋 𝑋 𝑋 𝑋 𝑋 𝑋 𝑋 𝑋
Such a transformation may reduce collinearity in the original variables.
c) Formalize relationships among regressors: - If it is believed that the Multicollinearity
arises not from an unfortunate data set but from an actual approximate linear relationship
among some of the regressors, this relationship could be formalized and the estimation could then
proceed in the context of a simultaneous equation estimation problem.
d) Combining cross-sectional and time series data: that is, the combination of cross sectional
and time-series data, known as pooling the data.
e) Drop a variable: - when faced with severe Multicollinearity, one of the ―simplest‖ thing to do is
to drop one of the collinear variables. But note that in dropping a variable from the model we
may be committing a specification bias or specification error. Specification bias arises from
incorrect specification of the model used in the analysis. Thus, if economic theory requires some
variables to be included in the model, dropping one of the variables due to Multicollinearity
problem would constitute specification bias.

5.3 HETEROSCEDASTICITY

5.3.1 The Nature of Heteroskedasticity

The assumption of homoscedasticity (or constant variance) about the random variable u is that its
probability distribution remains the same over all observations of X, and in particular that the
variance of each Ui is the same for all values of the explanatory variable. That is, the variation of
each ui around its zero mean does not depend on the value of X.
Symbolically we have Var (ui) = E {(ui – E (ui)}2 = E (u 2) =
If the above condition is not satisfied in any particular case, we say that the ui‟s are heterosckedastic.
That is, Var (ui) =
The problem of heteroskedasticity is more serious in cross section data rather than time series data.
Suppose we have a cross-section sample of family budget from which we want to measure the
savings function. That means Saving = f(income). In this case, the assumption of constant variance
of the ui‟s is not appropriate, because high-income families show a much greater variability in their
saving behavior than do low income families. Families with high income tend to stick to a certain
57 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module
 Debre Tabor University, CAES, Department of Agricultural Economics

standard of living and when their income falls they cut down their savings rather than their
consumption expenditure. But this is not the case in low income families. Hence, the variance of
ui‘s increases as income increases.

5.3.2 Causes of Heteroskedasticity

Heteroskedasticity can arise as a result of several cases.

1. The presence of outliers: An outlying observation, or outlier, is an observation that is much
different (either very small or very large) in relation to the observations in the sample. More
precisely, an outlier is an observation from a different population to that generating the
remaining sample observations. The inclusion or exclusion of such an observation, especially if
the sample size is small, can substantially alter the results of regression analysis.
2. Incorrect specification of the regression model: Very often what looks like heteroskedasticity
may be due to the fact that some important variables are omitted from the model. In such
situation the residuals obtained from the regression may give the distinct impression that the
error variance may not be constant. But if the omitted variables are included in the model, the
impression may disappear.
3. Skewness in the distribution of one or more regressors included in the model: Examples are
economic variables such as income, wealth, and education. It is well known that the distribution
of income and wealth in most societies is uneven, with the bulk of the income and wealth being
owned by a few at the top.
4. Incorrect data transformation (e.g., ratio or first difference transformations).
5. Incorrect functional form (e.g., linear versus log–linear models).
In summary, we may say that on a priori grounds there are reasons to believe that the assumption of
homoscedasticity may often be violated in practice. It is therefore, important to examine the
consequences of heteroskedasticity.

5.3.3 The consequence of Heteroskedasticity

If the assumption of homoscedastic disturbance is not fulfilled we have the following consequences:
Heteroskedasticity does not destroy the unbiasedness and consistency properties of OLS
estimators.
The OLS estimates do not have the minimum variance property in the class of unbiased

58 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

estimators. That is, they are not BLUE.

In the presence of heteroskedasticity, the variances of OLS estimators are not
Provided by the usual OLS formulas. But if we persist in using the usual OLS formulas, the t and F
tests based on them can be highly misleading, resulting in erroneous conclusions.

5.3.4 Detecting the problem of Heteroskedasticity:

As in the case of Multicollinearity, there are no hard-and-fast rules for detecting heteroskedasticity,
only a few rules of thumb.
A) Informal method
Nature of the problem: As a matter of fact, in cross-sectional data involving heterogeneous units,
heteroskedasticity may be the rule rather than the exception. For example, in a cross-sectional
analysis involving the investment expenditure in relation to sales, rate of interest, etc.,
heteroskedasticity is generally expected if small, medium and large-size firms are sampled
together.
Visual Inspection of Residuals / graphical method
This is a postmortem approach when there is no a priori information as the existence of
heteroskedasticity. Hence, this approach examines whether the error term depicts some systematic
pattern or not. To this end, the residuals are plotted against the dependent or independent variable to
which it is suspected the disturbance variance is related. Although are not the same thing as u 2, they
can be used as proxies especially if the sample size is sufficiently large.
B) Formal methods
Park Test: park formalizes the general method by suggesting that is some functions of
explanatory variable 𝑋
The functional form he suggest is 𝑣𝑎𝑟 𝑢 = + 𝑋 𝑒 which could be written as logarithmic
form as 𝑙𝑛 = 𝑙𝑛 + 𝛽𝑙𝑛𝑋 + 𝑣
Where 𝑣 is stochastic disturbance term
Since is generally not known, park suggest using 𝑢̂ as a proxy and running the following
regression
𝑙𝑛𝑢̂ = 𝑙𝑛 + 𝛽𝑙𝑛𝑋 + 𝑣
𝑙𝑛𝑢̂ = + 𝛽𝑙𝑛𝑋 + 𝑣
If 𝛽 turns out to be statistically significant; it would suggest that heteroskedasticity is present in the
59 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module
 Debre Tabor University, CAES, Department of Agricultural Economics

data. If it turns out to be insignificant, we may accept the assumption of homoscedasticity. The park
test is thus a two-stage procedure. In the first stage we run the OLS regression disregarding the
heteroskedasticity question.
We obtain 𝑢̂ from this regression and then in the second stage we run the regression
𝑙𝑛𝑢̂ = 𝑙𝑛 + 𝛽𝑙𝑛𝑋 + 𝑣
Example: Consider a relationship between Compensation (Y) and Productivity (X). To illustrate the
Park approach, the following regression function is used.
𝑌 =𝛽 +𝛽 𝑋 +𝑢
𝑌 = 992.35 + .23𝑋 + 𝑢
𝑠𝑒 = 936.48 . 99 𝑟 = .44
𝑡 = 2. 3 2.33
Suppose that the residuals obtained from the above regression were regressed on Xi giving the
following results.
𝑙𝑛𝑢̂ = 35.82 2.8 𝑙𝑛𝑋
𝑠𝑒 = 38.2 4.22 𝑟 = .46
𝑡= .93 .67
In the above result, the coefficient of lnXi is not significant. That is, there is no statistically
significant relationship between the two variables. Following the Park test, one may conclude that
there is no heteroskedasticity in the error variance.
Although empirically appealing, the Park test has some problems. For instance, the error term,Vi may
not satisfy the OLS assumptions and may itself be heterosckedastic. Nonetheless, as a strictly
exploratory method, one may use the Park test.
Spearman’s Rank Correlation Test
This test requires calculating rank correlation where its coefficient can be used to detect
heteroskedasticity. The rank correlation coefficient is given by
∑𝑑
𝑟 = 6[ ]
𝑛 𝑛
Where 𝑑 = difference in the ranks assigned to two different characteristics of the 𝑖 individual or
phenomenon and n = number of individuals or phenomena ranked. The steps required in this test are
stated as follows.
Assume

60 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

𝑌 =𝛽 +𝛽 𝑋 +𝑢
Step 1: Fit the regression to the data on Y and X and obtain the residuals 𝑢̂
Step 2: Ignoring the sign of 𝑢̂ that is taking their absolute value 𝑢̂ rank both 𝑢̂ and Xi (or 𝑌̂ )
according to an ascending or descending order and compute the Spearman‘s rank correlation
coefficient
Step 3: Assuming that the population rank correlation coefficient ƥs is zero and n > 8, the
significance of the sample rs can be tested by the t test as follows:
𝑟 √𝑛 2
𝑡= 𝑤𝑖𝑡𝑕 𝑑𝑒𝑔𝑟𝑒𝑒 𝑜𝑓 𝑓𝑟𝑒𝑒𝑑𝑜𝑚 𝑛 2
√ 𝑟
If the computed t value exceeds the critical t value, we may accept the hypothesis of
heteroskedasticity; otherwise we may reject it. If the regression model involves more than one X
variables, rs can be computed between | û i | and each of the X variable separately and can be
tested for statistical significance by the t-test given.
Example To illustrate the rank correlation test considers the regression
𝑌 = 𝛽 + 𝛽 𝑋 + 𝑢 . Suppose 10 observations are used to this equation. The following table makes
use of the rank correlation approach to test the hypothesis of heteroskedasticity. Notice that column 6
and 7 put rank of |û i | and Xi in an ascending order.
Table 5.1 Rank Correlation Test of Heteroskedasticity
Obs Rankof d (difference between d2
Y X = Xi the two ranking)
1 12.4 12.1 11.37 1.03 9 4 5 25
2 14.4 21.4 15.64 1.24 10 9 1 1
3 14.6 18.4 14.4 0.20 4 7 -3 9
4 16 21.7 15.78 0.22 5 10 -5 25
5 11.3 12.5 11.56 0.26 6 5 1 1
6 10.0 10.4 10.59 0.59 7 2 5 25
7 16.2 20.8 15.37 0.83 8 8 0 0
8 10.4 10.2 10.50 0.10 3 1 2 4
9 13.1 16.0 13.16 0.06 2 6 -4 16
10 11.3 12.0 11.33 0.03 1 3 -2 4
Total 0 110

Then,

61 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

∑𝑑
𝑟 = 6[ ]= 6[ ] = .33
𝑛 𝑛

and
.33√ 2
𝑡= = .99
√ .
Note that for 8 (=10-2) df, this t-value is not significant even at the 10% level of significance. Thus,
there is no evidence of systematic relationship between the explanatory variable and the absolute
value of the residuals, which might suggest that there is no heteroskedasticity.
The Goldfield – Quandt Test
This popular method is applicable if one assumes that the heterosckedastic variance, σi2 is positively
related to one of the explanatory variables in the regression model. The test is commonly applicable
to large samples. The observation must be at least twice as many as the parameters to be estimated.
The test assumes normality and serially independent disturbance term, Ui‟s. Consider the
following:
𝑌 = 𝛽 + 𝛽 𝑋 + 𝛽 𝑋 … +𝛽 𝑋 + 𝑢

The hypothesis to be tested is

H0: Ui‟s are homoscedastic
H1: Ui‟s are heterosckedastic (with increasing variance)
To test this, Goldfield-Quandt performs the following steps.
Step 1: The observations are ordered according to the magnitude of the independent variable thought
to be related to the variance of the disturbances.
Step 2: A certain number of central observations (represented by ―c‖) are omitted, leaving two
equal-sized groups of observations, one group corresponding to low values of the chosen
independent variable and the other group corresponding to high values. Note that the observations
are omitted to sharpen or accentuate the difference between the small variance and the large
variance group.
Step 3: We fit separate regression to each sub-sample, and we obtain the sum of squared residuals
from each of them and the ratio of their sum of squared residuals is formed. That is,
∑ 𝑢̂ = residuals form the sub-sample of low values of X1 with 𝑛 𝑐 2 – 𝑘 degrees of freedom,

62 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

where k is the total number of parameters in the model.

∑ 𝑢̂ = residual from the sub sample of high values of X, with the sample degree of freedom,
𝑛 𝑐 2 – 𝑘.
If each of these sums is divided by the appropriate degrees of freedom, we obtain estimates of the
variances of the 𝑢̂ 's in the two sub samples.
Step IV: Compute the ratio of the two variances given by
∑ 𝑢̂ 2 2 𝑛 𝑐 2–𝑘 ∑ 𝑢̂ 2 2
= 2 = 2 5.2.
∑ 𝑢̂ 𝑛 𝑐 2–𝑘 ∑ 𝑢̂
has an F distribution (with numerator and denominator each 𝑛 𝑐 2𝑘 2 degree of freedom,
where n = total number of observations, c = central observations omitted, k = number of parameters
estimated from each regression). If the two variances are the same (that is, if the 𝑢̂ ' s are
homoscedastic) the value of F* will tend to one. If the variance differ, F* will have a large value
(given that by the design of the test∑ 𝑢̂ 2 2 ∑ 𝑢̂ 2
. Generally, the observed F* is compared with
the theoretical value of F with 𝑛 𝑐 2𝑘 2 degrees of freedom (at a chosen level of
significance. The theoretical value of F (obtained from the F-tables) is the value of F that defines the
critical region of the test.
If F* > F we accept that there is heteroskedasticity (that is we reject the null hypothesis of no
difference between the variances of U‟s in the two sub samples). If F* < F, we accept that the U‟s are
homoscedastic (in other words we accept the null hypothesis). The higher the observed F* ratio the
stronger the heteroskedasticity of the U‟s.
Example: Suppose that we have data on consumption expenditure in relation to income for a cross
section of 30 families. Suppose we postulate that consumption expenditure is linearly related to
income but that heteroskedasticity is present in the data. Suppose further that the middle 4
observations are dropped after the necessary reordering of the data. Suppose we obtain the following
result after we perform a separate regression based on the two 13 observations.
536.8
=[ ] = 4. 7
377. 7

Note from the F- table in the appendix that the critical F value for 11 numerators and 11
denominator df at the 5% level is 2.82. Since the estimated F* value exceeds the critical value, we

63 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

may conclude that there is heteroskedasticity in the error variance.

Note, however, that the ability of the Goldfeld-Quadent test to perform successfully depends on how
c is chosen. Moreover, its success depends on identifying the correct X (i.e., independent) variable
with which to order the observations.
Breusch-Pagan (BP) test
The Breusch–Pagan test is a Lagrange multiplier test for heteroskedasticity. The main characteristics
of Lagrange multiplier tests are that they do not require that the model is estimated under the
alternative and that they are often simply computed from the R2 of some auxiliary regression.

To illustrate this test, consider the k-variable linear regression model

𝑌 = 𝛽 + 𝛽 𝑋 + 𝛽 𝑋 … +𝛽 𝑋 + 𝑢 5.2.2

The Breusch-Pagan test for heteroskedasticity is carried out as follows:

1. Estimate equation (5.2.2) by OLS, and obtain the residual, ûi .

2. Run the auxiliary regression 𝑢̂ = + 𝑋 + 𝑋 …+ 𝑋 + 𝑣 and obtain the R-

squared from this regression, ̂

3 . Form either the F statistic: F = R2 (k 1) or the LM statistic: LM = n ̂

(1 R2 (n  k)

Under the null hypothesis, LM is distributed asymptotically as

with degrees of freedom equal to number of regressors (excluding the intercept). If the p- value is
sufficiently small, that is, below the chosen significance level, then we reject the null hypothesis of
homoscedasticity.
The White Test
White (1980) proposed a test for heteroskedasticity that adds the squares and cross products of all
of the independent variables. The test is explicitly intended to test for forms of heteroskedasticity
that invalidate the usual OLS standard errors and test statistics.
Suppose a model with three explanatory variables

𝑌 = 𝛽 + 𝛽 𝑋 + 𝛽 𝑋 +𝛽 𝑋 + 𝑢 5.2.3
Steps

64 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

1. Run the regression on the above equation and obtain 𝑢̂

2. Run the auxiliary regression:
𝑢̂ = + 𝑋 + 𝛽 𝑋 +𝛽 𝑋 + 𝑋 +𝛽 𝑋 + 𝑋 +𝛽 𝑋 + 𝑋 +𝛽 𝑋 +𝑢
The rationale of including the independent variables, their squares, and cross products is that the
variance may be systematically correlated with either of the independent variables linearly or non-
linearly.
The null hypothesis is that: H0: α1 =α 2 =α 3 = . . . = α 9
3. Form LM statistic: LM = n ̂
with nine degrees of freedom. If we fail to reject

the null, we conclude that there is homoscedasticity. We can also use an F test of thishypothesis; both
tests have asymptotic justification.

5.3.5 Remedial Measures

Heteroskedasticity does not destroy the unbiasedness and consistency properties of the OLS
estimators, but they are no longer efficient, not even asymptotically (i.e., large sample size). This
lack of efficiency makes the usual hypothesis testing procedure of dubious value. Therefore, remedial
measures are clearly called for. There are two approaches to remediation: when σ2 is known and when
σi 2 is not known.

When σi2 is known: The Method of Weighted Least Squares (WLS)

If σi2 is known, the most straightforward method of correcting heteroskedasticity is by means of
weighted least squares. The OLS method assigns equal weight or importance to each observation.
However, a method of estimation known as generalized least squares (GLS) takes such
information into account explicitly and is therefore capable of producing estimators that are BLUE.
Though WLS is just a special case of the more general estimating technique, GLS, in the context of
heteroskedasticity, one can treat the two terms WLS and GLS interchangeably. To see how this is
accomplished, consider the familiar two-variable model:
𝑌 = 𝛽 + 𝛽𝑋 + 𝑢
𝑌 = 𝛽 𝑋 + 𝛽 𝑋 + 𝑢 𝑤𝑕𝑒𝑟𝑒 𝑋 = 𝑓𝑜𝑟 𝑒𝑎𝑐𝑕 𝑖.
Dividing both sides by σi

= + + Which the exposition, can be written as

𝑌 = 𝛽 + 𝛽 𝑋 +𝑢
65 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module
 Debre Tabor University, CAES, Department of Agricultural Economics

What is the purpose of transforming the original model? To see this, notice the following feature of
the transformed error term ui*
𝑢
𝑣𝑎𝑟 𝑢 = 𝑢 = ( ) = 𝑢 𝑠𝑖𝑛𝑐𝑒 𝑖𝑠 𝑘𝑛𝑜𝑤𝑛

= 𝑠𝑖𝑛𝑐𝑒 𝑢 =

=
Which is a constant. That is, the variance of the transformed disturbance term ui* is now
homoscedastic. Since we are still retaining the other assumptions of the classical model, ui* is
homoscedastic suggests that if we apply OLS to the transformed model it will produce estimators that
are BLUE. In short, the estimated β0* and β1* are now BLUE and not the OLS estimators
𝛽̂ 𝑎𝑛𝑑 𝛽̂
To obtain GLS estimators, we minimize

∑ 𝑢̂ =∑ 𝑌 𝛽 𝑋 𝛽 𝑋

𝑢̂ 𝑌 𝛽𝑋 𝛽𝑋
∑ =∑ 5.3.

Which can also be written as:

∑ 𝑤 𝑢̂ =∑𝑤 𝑌 𝛽 𝑋 𝛽 𝑋

Where 𝑤 =

Thus, in GLS we minimize a weighted sum of residual squares with wi = 1/σ 2 actingi as the weights,
but in OLS we minimize non-weighted or (what amounts to the same thing) equally weighted RSS.
In GLS, the weight assigned to each observation is inversely proportional to its σi, that is,
observations coming from a population with larger σi will get relatively smaller weight and those
from a population with smaller σi will get proportionately larger weight in minimizing the RSS.
When σi2 Is Not Known: White’s Heteroskedasticity-Consistent Standard Errors
If true σi2 are known, we can use the WLS method to obtain BLUE estimators. Since the true σi2 are
rarely known, there is a way of obtaining consistent (in the statistical sense) estimates of the
variances and covariances of OLS estimators even if there is heteroskedasticity. White has shown that
this estimate can be performed so that asymptotically valid (i.e., large-sample) statistical inferences
can be made about the true parameter values. White‘s heteroskedasticity-corrected standard errors

66 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

are also known as robust standard errors.

Note: In addition to the above measures, a log transformation such as lnYi = β0 + β1lnXi + Ui
very often such transformation reduces heteroskedasticity when compared with the regression Yi = β0
+ β1Xi + Ui
This result arises because log transformation compresses the scales in which the variables are
measured. For example, log transformation reduces a ten-fold difference between two values (such
as between 8 and 80) into a two-fold difference (because ln 80 = 4.32 and ln 8 = 2.08).

5.4 AUTOCORRELATION

5.4.1 The Nature of Autocorrelation

The term autocorrelation may be defined as ―correlation between members of series of

observations ordered in time [as in time series data] or space [as in cross-sectional data]. An
important assumption of the classical linear model is that there is no autocorrelation or serial
correlation among the disturbances entering into the population regression function. This assumption
implies that:
Cov (UiUj) = E {[ i –– 𝑖 𝑗– 𝑗
= E(UiUj)
= 0 (for i ≠ j)
If this assumption is violated, the disturbances are said to be auto correlated.
Since auto correlated errors arise most frequently in time series models, the discussion in the rest of
this chapter is couched in terms of time series data.
There are a number of time-series patterns or process that can be used to model correlated errors.
The most common is what is known as ―the first order autoregressive process‖ or AR(1) process.
Consider
Yt = β0 + β1Xt + ut
where, t denotes data or observation at time t (i.e., a time series data) with this one can assumethat
the disturbances are generated as follows.
ut = ρut-1 + ϵt
where ρ is known as the coefficient of auto covariance and ϵt is the stochastic term such that it
satisfies the standard OLS assumptions, namely
E(ϵt) = 0
67 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module
 Debre Tabor University, CAES, Department of Agricultural Economics

Var(ϵt) = σ2
Cov (ϵt, ϵt+s) = 0 where subscript „s‟ represent the exact period of lag.
The above specification is of first order because the regression of ut is on itself lagged one period
(where the coefficient ρ is the first order coefficient of autocorrelation). Note that the above
specification postulates that the movement or shift in ut consists of two parts: a part ρut-1, which
accounts for systematic shift, and the other ϵt which is purely random.
Relationships between ut‟s can be shown as:
Cov (ut, ut-1) = E[(ut – E(ut) (ut-1 – E(ut-1)]
= E[ut ut-1]
by substituting ut = ρut-1 + ϵt we obtain:
= E[(ρut-1 + ϵt) ut-1]
= ρE[u2t-1] + E[ϵt ut-1]
Note that E(ϵt) = 0 thus E(ϵt Ut-1) = 0
Since with the assumption of homoscedasticity (i.e., constant variance) Var(ut) = Var (ut-1) = σ2
The result would be Cov (ut, ut-1) = ρ σ2
Now, correlation of ut, ut-1 is given by
𝑐𝑜𝑣 𝑢 𝑢
𝑐𝑜𝑟𝑟 𝑢 𝑢 = = = = 5.4.
√𝑣𝑎𝑟 𝑢 𝑣𝑎𝑟 𝑢 𝑣𝑎𝑟 𝑢

Where
Hence, ρ(rho) is simple correlation of the successive errors of the original model.
Note that when ρ > 0 successive errors are positively correlated and when ρ < 0 successive errors are
negatively correlated. It can be shown that corr (Ut, Ut-s) = ρs (where s represents the exact period of
lag). It implies that the correlation (be it negative or positive) between any two period diminishes as
time goes by; i.e., as s increases.

5.4.2 Sources of Autocorrelation

Inertia: A salient feature of most economic time series is inertia, or sluggishness. As is well known,
time series such as GNP, price indices, production, employment, and unemployment exhibit
(business) cycles. Starting at the bottom of the recession, when economic recovery starts, most of
these series start moving upward. In this upswing, the value of a series at one point in time is greater
than its previous value. Thus, there is a ―momentum‟‟ built into them, and it continues until something

68 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

happens (e.g., increase in interest rate or taxes or both) to slow them down. Therefore, in regressions
involving time series data, successive observations are likely to interdependent.
Data manipulation: published data often undergo interpolation or smoothing, procedures that
average true disturbances over successive time periods.
Specification bias
Specification Bias: Excluded Variables Case. In empirical analysis the researcher often starts with
a plausible regression model that may not be the most ―perfect‟‟ one. After the regression analysis,
the researcher does the postmortem to find out whether the results accord with a priori expectations.
For example, suppose we have the following demand model:
Yt = β0 + β1X1t + β2X2t + β3X3t + ut
However, for some reason we run the following regression:
Yt = β0 + β1X1t + β2X2t + vt
Now if the first model is the ―correct‟‟ model, running the second is tantamount to letting vt = β3X3t
+ ut. To the extent that X3 affects Yt, the error term v will reflect a systematic pattern, thus creating
(false) autocorrelation. A simple test of this would be to run both models and see whether
autocorrelation, if any, observed in the second model, disappears when the first model is run.
Specification Bias: Incorrect Functional Form. Suppose the ―true‟‟ or correct model in a cost-
output study is as follows:
Marginal costi = β0 + β1 outputi + β2 output 2 + ui
But we fit the following model:
Marginal costi = α0 + α1 outputi + vi
Because the disturbance term vi is, in fact, equal to output2 + ui , it will catch the systematic effect of
the output2 term on marginal cost. In this case, vi will reflect autocorrelation because of the use of
an incorrect functional form.
Lags. For instance, in a time series regression of consumption expenditure on income, it is not
uncommon to find that the consumption expenditure in the current period depends, among other
things, on the consumption expenditure of the previous period. That is, (Consumption)t= β0 + β1(
income)t + β2 (consumption)t−1 + ut
A regression like this is known as auto regression because one of the explanatory variables is the
lagged value of the dependent variable. The rationale is consumers do not change their consumption
habits readily for psychological, technological, or institutional reasons. Now if we neglect the lagged

69 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

consumption in the model, the resulting error term will reflect a systematic pattern due to the
influence of lagged consumption on current consumption.

5.4.3 Consequences of Autocorrelation

As in the case of heteroskedasticity, in the presence of autocorrelation the OLS estimators are still
linear unbiased as well as consistent and asymptotically normally distributed, but they are no longer
efficient (i.e., minimum variance). As a consequence, the usual t, F, and χ2 tests cannot be
legitimately applied.
The prediction based on ordinary least squares estimate will be inefficient with auto correlated
errors. This is because of larger variance as compared with predictions based on estimates obtained
from other econometric techniques.

5.4.4 Testing (Detecting) for Autocorrelation

Autocorrelation is potentially a series problem. Hence, it is essential to find out whether Auto-
correlation exists in a given situation. Since the population disturbances Ut, cannot be observed
directly, we use its proxy, the residual Û t which can be obtained from the usual OLS procedure. The
examination of Û t can provide useful information not only about autocorrelation but also about

heteroskedasticity, model inadequacy, or specification bias.

Graphical Method
Some rough idea about the existence of autocorrelation may be gained by plotting the residuals
either against time or against their own lagged variables.
For instance, suppose plotting the residual against its lagged variable bring about the following
relationship.

70 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

As the above figure reveals, most of the residuals are bunched in the first and the third quadrants
suggesting very strongly that there is positive correlation in the residuals. However, the graphical
method is essentially subjective or qualitative in nature. There are quantitative tests that can be used
to supplement the purely qualitative approach.
Durbin-Watson d Test
The most celebrated test for detecting serial correlation is the one developed by Durbin and Watson.
It is popularly known as the Durbin-Watson d-Statistic and it is defined as
2
∑ ̂ ̂
= ∑̂
5.4.2

Which is simply the ratio of the sum of squared differences in successive residuals to the residual
sum of squares, RSS. Note that in the numerator of the d statistic the number of observations is n-1
because one observation is lost in taking successive differences.
The proof of d-statistic is as follows

∑ (̂ +̂ + 2̂ ̂ ) ∑ ̂ +∑ ̂ +∑ 2̂ ̂
= = 5.4.3
∑ 𝑢̂ ∑ 𝑢̂
However for large samples ∑ ̂ ∑ ̂ and ∑ 𝑢̂ are approximately equal there fore it
can be written as
2 ∑̂ 2 ∑̂ ̂ 2 ∑̂ ̂
= 2[ ]=2 ̂
∑ 𝑢̂ ∑ 𝑢̂ ∑ 𝑢̂

The assumptions underlying the d-statistic are:

a) The regression model includes an intercept term
b) the explanatory variables are non-stochastic or fixed in repeated sampling
c) the disturbances Ut are generated by the first order autoregressive scheme.
Ut = ρUt-1 + ϵt. Therefore, it cannot be used to detect higher-order autoregressiveschemes.
d) The error term ut is assumed to be normally distributed.
e) the regression model does not include lagged value(s) of the dependent variable as one ofthe
explanatory variables
f) There are no missing observations in the data.
In the absence of the autocorrelation we can expect to d take a value close to 2, when negative
autocorrelation is present a value in excess of 2 and may be as high as 4, and when positive

71 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

autocorrelation is present a value lower than 2 and may be close to zero.

The Durbin-Watson statistic tests the hypothesis that H0: ρ = 0 (implying that the error terms are not
auto correlated with a first order scheme) against the alternative. However, the sampling distribution
for the d-statistic depends on the sample size n, the number of explanatory variables and also on the
actual sample values of the explanatory variables. Thus, the critical values at which we might, for
example, reject the null hypothesis at 5 percent level of significance depend very much on the sample
we have chosen. Unfortunately, it is impracticable to tabulate critical values for all possible sets of
sample values. What possible however is, for given values of sample size and number of
explanatory variables, to find upper and lower bounds such that actual critical values for any set of
sample values will fall within these known limits.
The Durbin-Watson test procedure in testing the null hypothesis of ρ = 0 against the alternative
hypothesis of positive autocorrelation is illustrated in the figure below.
The decision criterion for the Durbin-Watson test is therefore, of the following form.
For 𝑑 𝑑 reject the null hypothesis of no autocorrelation in favor of positive
autocorrelation;
For 𝑑 𝑑 do not reject null hypothesis, i.e., insufficient evidence to suggest positive
autocorrelation;
For 𝑑 𝑑 𝑑 : inconclusive.
Because of the symmetry of the distribution it is also possible to use the tables for d *L and d*U to test
the null hypothesis of no autocorrelation against the alternative hypothesis of negative
autocorrelation, i.e. ρ < 0. The decision criterion then takes the form.
For d > 4 - d*L reject the null hypothesis of no autocorrelation in favor of negative
autocorrelation.
For d < 4 - d*U do not reject null hypothesis, i.e., insufficient evidence to suggest negative
autocorrelation
For 4 - d*L > d > 4- d*U: inconclusive.
Note that tables for d*U and d*L are constructed to facilitate the use of one-tail rather than two tail
tests. The following representation explains better the actual test procedure which shows that the
limits of d are 0 and 4.

72 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

Reject H0 Reject H0*

Zone of Do not Reject H0 or Zone of

+Ve -Ve
indecision H0* or both indecision
Autocorrelation Autocorrelation

0 dL dU 2 4-dU 4-dL 4

Note: H0 : no positive autocorrelation and H0*: no negative autocorrelation

Note that from the above presentation we can develop the following rule of thumb. That is, if d is
found to be closer to 2 in an application, one may assume that there is no first order autocorrelation
either positive or negative. If d is closer to 0 it is because ρ is closer to 1 indicating strong positive
autocorrelation in the residuals. Similarly the closer d is to 4, the greater the evidence of negative
serial correlation. This is because ρ is closer to –1.
Example: Suppose in a regression involving 50 observations with 4 regressors, the estimated d
statistic was 1.43. From the Durbin Watson table we find that at the 5% level, the critical d value are
dL = 1.38 and dU = 1.72. On the basis of the d test we cannot say whether there is positive
autocorrelation or not because the estimated d value lies in the indecisive range.
The Breusch–Godfrey (BG) Test
To avoid some of the pitfalls of the Durbin–Watson d test of autocorrelation, statisticians Breusch and
Godfrey have developed a test of autocorrelation that is general in the sense that itallows for:
non-stochastic regressors, such as the lagged values of the Regressand higher-order autoregressive
schemes, such as AR(1), AR(2), etc.; and simple or higher-order moving averages of white noise
error terms.
Consider Yt = β0 + β1Xt + ut -----------------------------------------------------5.3.1
Assume that the error term ut follows the pth order autoregressive, AR (p), scheme as follows:
𝑢 = 𝑢 + 𝑢 + 5.3.2
𝑢 + + 𝑢 +

= 𝑖𝑠 𝑎 𝑤 𝑕𝑖𝑡𝑒 𝑛𝑜𝑖𝑠𝑒 𝑒𝑟𝑟𝑜𝑟 𝑡𝑒𝑟𝑚

The null hypothesis H0 to be tested is that:

= = = = 5.3.3

73 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

The BG test involves the following steps:

Estimate equation (1) by OLS and obtain the residuals
Regress ût on the original Xt (if there are more than one X variables in the original model,

include them also) and ût -1 , ût -2 , . . . , ût -p , where the latter are the lagged values of
the estimated residuals in step 1. Note that to run this regression we will have only (n −
p) observations.
In short, run the following regression:
𝑢̂ = + 𝑋 + ̂ 𝑢̂ + ̂ 𝑢̂ + ̂ 𝑢̂ 5 .3 .4
+ + ̂ 𝑢̂ +
2
and obtain R from this (auxiliary) regression.
If the sample size is large , Breusch and Godfrey have shown that 𝑛 𝑝
If (n − p) R2 exceeds the critical chi-square value at the chosen level of significance, we reject the
null hypothesis, in which case at least one rho in equation (5.3.4) is statistically significantly
different from zero. That is, there is autocorrelation.
A drawback of the BG test is that the value of p, the length of the lag, cannot be specified apriori.

5.4.5 Remedial Measures

With time series data, auto correlated residuals are often indications of some error in the way we have
specified the regression equation than genuine autocorrelation in the disturbances. Mostly, positive
autocorrelations in economic data are caused by omission of relevant variables.
Incorrect functional form may also be the cause for auto correlated residuals.
Therefore, we should find out if the autocorrelation is pure autocorrelation and not the result of mis-
specification of the model. If the source of the problem is suspected to be due to omission of
important variables, the remedy is to include those omitted variables. Besides if the source of the
problem is believed to be the result of misspecification of the model, then the solution is to
determine the appropriate mathematical form.
If it is pure autocorrelation, one can use appropriate transformation of the original model so that in
the transformed model we do not have the problem of (pure) autocorrelation. As in the case of
heteroskedasticity, we will use some type of generalized least-square (GLS) method. In large
samples, the Newey–West method can be applied to obtain standard errors of OLS estimators that
are corrected for autocorrelation.

74 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

The method of generalized least squares (GLS)

Knowing the consequences of autocorrelation, especially the lack of efficiency of OLS estimators,
we may need to remedy the problem. The remedy depends on the knowledge one has about the
nature of interdependence among the disturbances, that is, knowledge about the structure of
autocorrelation.
Consider the two-variable regression model:
Yt = β0 + β1Xt + Ut ------------------------------------------------------------------------------- (i)
And assume that the error term follows the AR (1) scheme, namely
Ut = ρUt-1 + ϵt where -1<ρ <1
There are two cases:
When ρ is known and
when ρ is not known but has to be estimated.
When ρ is known
The serial correlation problem can be satisfactorily resolved if the coefficient of autocorrelation,
ρ, is known.
Recall the two variables model.
Yt = β0 + β1Xt + Ut ------------------------------------------------------------------------------ (ii)
At time t-1 the above model will be
Yt-1 = β0 + β1Xt-1 + Ut-1-------------------------------------------------------------------------- (iii)
Multiplying both sides by ρ, we obtain
ρYt-1 = ρβ0 + ρβ1Xt-1 + ρUt-1 -------------------------------------------------------------------(iv)
Subtracting (iii) from (i) gives
(Yt - ρYt-1) = (β0 - ρβ0) + (β1Xt - ρβ1Xt-1) + (Ut - ρUt-1)
= β0 (1-ρ) + β1(Xt -ρXt-1) + εt -------------------------------------------------------------------(v)
Where εt= Ut - ρUt-1
The transformed model can be expressed as
Y*t = β*0 + β*1X*t + εt------------------------------------------------------------------------- (vi)
Where Y*t = Yt - ρYt-1, β*0 = β0 - ρβ0, X*t = (Xt - ρXt-1) and β*1=β1
Since εt satisfies the OLS assumptions, one can apply OLS to the transformed variables Y* and X*
and obtain estimators with all the optimum properties, namely BLUE. Regression of the transformed
model is equivalent to using generalized least squares (GLS). In short, GLS is nothing but OLS

75 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

applied to the transformed model that satisfies the classical assumptions. Regression of equation (iv)
is known as the generalized, or quasi, difference equation. It involves regressing Y on X, not in the
original form, but in the difference form, which is obtained by subtracting a proportion (= ρ) of the
value of a variable in the previous time period from its value in the current time period. Note that in
this differencing procedure we lose one observation because the first observation has no
antecedent.
When ρ is not known
Although straight forward to apply, the method of generalized difference is difficult to run because,
ρ, population correlation coefficient is rarely known in practice. Therefore, alternative methods need
to be devised.
The First-Difference Method: Since ρ lies between 0 and ±1, one could start from two extreme
positions. At one extreme, one could assume that ρ = 0, that is, no (first-order) serial correlation, and
at the other extreme we could let ρ = ±1, that is, perfect positive or negative correlation.
As a matter of fact, when a regression is run, one generally assumes that there is no autocorrelation
and then lets the Durbin–Watson or other test show whether this assumption is justified. If, however,
ρ = +1, the generalized difference equation in equation (iv) above reduces to the first-difference
equation:
(Yt - Yt-1) = β1(Xt - Xt-1) + (Ut - ρUt-1)
𝑌𝑡𝑡 == β1 ∆Xt +εt
∆𝑌
The first difference transformation may be appropriate if the coefficient of autocorrelation is very
high, say in excess of 0.8, or the Durbin–Watson d is quite low. Strictly speaking, the first-
difference transformation is valid only if ρ = 1. Maddala has proposed this rough rule of thumb: Use
the first difference form whenever d < R2. An interesting feature of the first-difference model is that
there is no intercept in it. Hence, we have to use the regression through the origin.
Computing ρ from Durbin–Watson d Statistic. If we cannot use the first difference transformation
because ρ is not sufficiently close to unity, we have an easy method of estimating it from the
relationship between d and ρ as follows:
𝑑
2
Thus, in reasonably large samples one can obtain rho and use it to transform the data as shown in the
generalized difference equation. However, the relationship between ρ and d may not hold true in
small samples.
76 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module
 Debre Tabor University, CAES, Department of Agricultural Economics

Estimating ρ from the residuals. If the AR(1) scheme ut = ρut−1 + εt is valid, a simple way to
estimate rho is to regress the residuals ût on ût -1 , for û t the are consistent estimators of the true
ut. That is, we run the following regression: ût = ρ ût-1 + vt
where, ût are the residuals obtained from the original (level form) regression and where v t are the
error term of this regression. Note that there is no need to introduce the intercept term, because the
OLS residuals sum to zero.
Iterative Methods of Estimating ρ: All the methods of estimating ρ explained above provide us
with only a single estimate of ρ. But there are the so-called iterative methods that estimate ρ
iteratively, that is, by successive approximation, starting with some initial value of ρ. Among these
methods are:
Cochrane–Orcutt iterative procedure,
Cochrane–Orcutt two-step procedure,
Durbin two–step procedure, and
Hildreth–Lu scanning or search procedure.
Of these, the most popular is the Cochrane–Orcutt iterative method. One advantage of this method is
that it can be used to estimate not only an AR(1) scheme, but also higher-order autoregressive
schemes. Having obtained the two rhos, one can easily extend the generalized difference equation.
The Cochrane-Orcutt interactive Procedure
This procedure helps to estimate ρ from the estimated residuals ̂ so that information about the
unknown ρ will be derived.
To explain the method, consider the two-variable model
Yi = β0 + β1Xi + Ui ------------------------------------------------------------------------------------- (a)
and assume that Ut is generated by the AR(1) scheme namely
Ut = ρUt-1 + εt --------------------------------------------------------------------------------------------- (b)
Cochrane and Orcutt then recommended the following steps to estimate ρ:
Step 1: Estimate the two variables model by the standard OLS routine and obtain the residuals
Û t

Step 2: Using the estimated residuals, run the following regression

̂ = ̂̂ +
Step 3: Using ̂ obtained from step 2 regressions, run the generalized difference equation asfollows

77 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

(Yt - ̂ Yt-1) = β0 (1- ̂ ) + β1(Xt - ̂Xt-1) + (Ut - ̂ Ut-1)

or Y*t = β*0 + β*1X*t + Uˆ * t

Step 4: Since a priori it is not known that the ̂ obtained from the regression in step 2 is the best

estimate of substitute the values of 𝛽̂ * and 𝛽̂ * obtained from the regression in step 3
0 1
into the original regression (a) and obtain the new residuals, say ̂ as
̂ =𝑌 𝛽̂ 𝛽̂ 𝑋
Note that this can be easily computed since Yt , Xt , 𝛽̂ 𝑎𝑛𝑑 𝛽̂ are all known.
Step 5: Now estimate this regression
̂ = ̂̂ ̂ +𝑤
Where ̂̂ is the second round estimate of .
Since we do not know whether this second round estimate ̂̂ is the best estimate of ρ, we can go
into the third estimate, and so on. That is why the Cochrane-Orcutt method is said iterative. But how
long should we go on? The general procedure is to stop carrying out iterations when the successive
estimates of ρ converges. Thus, we select that chosen ρ to transform the model and apply a kind of
GLS estimation that minimizes the problem of autocorrelation.
Note that:
1. Since the OLS estimators are consistent despite autocorrelation, in large samples, it makes little
difference whether we estimate ρ from the Durbin–Watson d, or from the regression of the
residuals in the current period on the residuals in the previous period, or from the Cochrane–
Orcutt iterative procedure because they all provide consistent estimates of the true ρ.
2. The various methods discussed above are basically two-step methods. In step 1 we obtain an
estimate of the unknown ρ and in step 2 we use that estimate to transform the variables to
estimate the generalized difference equation, which is basically GLS. But since we use ̂
instead of the true ρ, all these methods of estimation are known in the literature as feasible
GLS (FGLS) or estimated GLS (EGLS) methods.
3. It is important to note that whenever we use an FGLS or EGLS method to estimate the
parameters of the transformed model, the estimated coefficients will not necessarily have the
usual optimum properties of the classical model, such as BLUE, especially in small samples. In
short, whenever we use an estimator in place of its true value, the estimated OLS coefficients
may have the usual optimum properties asymptotically, that is, in large samples. Also, the

78 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

conventional hypothesis testing procedures are, strictly speaking, valid asymptotically. In small
samples, therefore, one has to be careful in interpreting the estimated results.
4. In using EGLS, if we do not include the first observation (as was originally the case with the
Cochrane Orcutt procedure), not only the numerical values but also the efficiency of the
estimators can be adversely affected, especially if the sample size is small and if the regressors
are not strictly speaking nonstochastic. Therefore, in small samples it is important to keep the
first observation à la Prais Winsten.
The Newey–West method of correcting the OLS standard errors
Instead of using the FGLS methods, we can still use OLS but correct the standard errors for
autocorrelation by a procedure developed by Newey and West. This is an extension of White‘s
heteroskedasticity-consistent standard errors. The corrected standard errors are known as HAC
(heteroskedasticity- and autocorrelation-consistent) standard errors or simply as Newey– West
standard errors.
This method is strictly speaking valid in large samples and may not be appropriate in small samples.
Therefore, if a sample is reasonably large, one should use the Newey–West procedure to correct
OLS standard errors not only in situations of autocorrelation only but also in cases of
heteroskedasticity, for the HAC method can handle both, unlike the White method, which was
designed specifically for heteroskedasticity.

Conclusion
Violating Multicollinearity does not impact prediction, but can impact inference. For example, p-
values typically become larger for highly correlated covariates, which can cause statistically
significant variables to lack significance.
Violating linearity can affect prediction and inference. We saw that prediction and precision in
estimating coefficients were only hindered slightly. However, these things will be exacerbated when
stronger levels of non-linearity are unaccounted for.
The no endogeneity assumption was violated in the model due to an omitted variable. This created
biased coefficient estimates, which lead to misleading conclusions. Prediction was also poor since
the omitted variable explained a good deal of variation in housing prices.
This simulation gives a flavor of what can happen when assumptions are violated. Depending on a
multitude of factors (i.e. variance of residuals, number of observations, etc.), the model‘s ability to
predict and infer will vary. Of course, it‘s also possible for a model to violate multiple assumptions.
79 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module
 Debre Tabor University, CAES, Department of Agricultural Economics

CHAPTER SIX: NON-LINEAR REGRESSION AND TIME SERIES ECONOMETRICS

6.1. Non-linear regression models Overview

Intrinsically Linear and Intrinsically Nonlinear Regression Models: When we started our
discussion of linear regression models in Chapter 3, we stated that our concern in this book is
basically with models that are linear in the parameters; they may or may not be linear in the
variables. On the other hand, if a model is nonlinear in the parameters it is a nonlinear (in-the-
parameter) regression model whether the variables of such a model are linear or not.
However, one has to be careful here, for some models may look nonlinear in the parameters but are
inherently or intrinsically linear because with suitable transformation they can be made linear-in-
the-parameter regression models. But if such models cannot be linearized in the parameters, they are
called intrinsically nonlinear regression models. From now on when we talk about a nonlinear
regression model, we mean that it is intrinsically nonlinear. For brevity, we will call them NLRM.
Example: Consider now the famous Cobb–Douglas (C–D) production function. Letting Y =
output, X1 = labor input, and X2= capital input, we will write this function in three different ways:
𝑌 =𝛽 𝑋 𝑋 𝑒 6. .
𝑜𝑟 𝑙𝑛𝑌 = 𝑙𝑛𝛽 + 𝛽 𝑙𝑛𝑋 + 𝛽 𝑙𝑛𝑋 + 𝑢 6. .2𝑎
Where = 𝑙𝑛𝛽 . Thus in this format the C–D function is intrinsically linear.
𝑜𝑟 𝑙𝑛𝑌 = + 𝛽 𝑙𝑛𝑋 + 𝛽 𝑙𝑛𝑋 + 𝑢 6. .2𝑏
𝑌 =𝛽 𝑋 𝑋 𝑢 6. .3
𝑜𝑟 𝑙𝑛𝑌 = + 𝛽 𝑙𝑛𝑋 + 𝛽 𝑙𝑛𝑋 + 𝑙𝑛𝑢 6. .4
But now consider the following version of the C–D function:

𝑌 =𝛽 𝑋 𝑋 +𝑢 6. .5
As we just noted, C–D versions (6.1.2a) and (6.1.2b) are intrinsically linear (in the parameter)
regression models, but there is no way to transform (6.1.5) so that the transformed model can be
made linear in the parameters. Therefore, (6.1.4) is intrinsically a nonlinear regression model.
Another well-known but intrinsically nonlinear function is the constant elasticity of substitution
(CES) production function of which the Cobb– Douglas production is a special case.

Example: 𝑌 = + 6. .6
Where Y = output, K = capital input, L = labor input, A = scale parameter, δ = distribution parameter

80 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

(0 < δ < 1), and β = substitution parameter (β ≥ −1). No matter in what form you enter the stochastic
error term ui in this production function, there is no way to make it a linear (in parameter) regression
model. It is intrinsically a nonlinear regression model.
Therefore, although we can apply the method of least squares to estimate the parameters of the
nonlinear regression models, we cannot obtain explicit solutions of the unknowns. Incidentally, OLS
applied to a nonlinear regression model is called nonlinear least squares (NLLS). So, what is the
solution?
Estimating Nonlinear Regression Models:
There are several methods of obtaining estimates of NLRMs, such as
The Trial-and-Error Method
Nonlinear least squares (NLLS), and
Linearization through Taylor series expansion.
1. Although linear regression models predominate theory and practice, there are occasions where
nonlinear-in-the-parameter regression models (NLRM) are useful.
2. The mathematics underlying linear regression models is comparatively simple in that one can
obtain explicit, or analytical, solutions of the coefficients of such models. The small-sample and
large-sample theory of inference of such models is well established.
3. In contrast, for intrinsically nonlinear regression models, parameter values cannot be obtained
explicitly. They have to be estimated numerically, that is, by iterative procedures.
4. There are several methods of obtaining estimates of NLRMs, such as (1) trial and error, (2)
nonlinear least squares (NLLS), and (3) linearization through Taylor series expansion.
5. Computer packages now have built-in routines, such as Gauss– Newton, Newton–Raphson, and
Marquard. These are all iterative routines.
6. NLLS estimators do not possess optimal properties in finite samples, but in large samples they do
have such properties. Therefore, the results of NLLS in small samples must be interpreted carefully.
7. Autocorrelation, heteroskedasticity, and model specification problems can plague NLRM, as they
do linear regression models.
8. We illustrated the NLLS with several examples. With the ready availability of user-friendly
software packages, estimation of NLRM should no longer be a mystery. Therefore, the reader should
not shy away from such models whenever theoretical or practical reasons dictate their use.

81 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

6.2. Time series Analysis

6.2.1. Univariate Time Series Analysis

One objective of analyzing economic data is to predict or forecast the future values of economic
variables. One approach to do this is to build a more or less structural econometric model, describing
the relationship between the variable of interest with other economic quantities, to estimate this
model using a sample of data, and to use it as the basis for forecasting and inference. Although this
approach has the advantage of giving economic content to one‘s predictions, it is not always very
useful. For example, it may be possible to adequately model the contemporaneous relationship
between unemployment and the inflation rate, but as long as we cannot predict future inflation rates
we are also unable to forecast future unemployment.
A time series is a sequence of numerical data in which each item is associated with a particular instant
in time. Time series data is, as its name suggests, ordered by time. One can quote numerous examples:
monthly unemployment, weekly measures of money supply, daily closing prices of stock indices, and
so on.
In general we consider a time series of observations on some variable, e.g. the unemployment rate,
denoted as Y1… YT. These observations will be considered realizations of random variables that can
be described by some stochastic process. It is the properties of this stochastic process that we try to
describe by a relatively simple model. It will be of particular importance how observations
corresponding to different time periods are related, so that we can exploit the dynamic properties of
the series to generate predictions for future periods.
Stationary processes
A stochastic process {yt} is said to be weakly stationary or covariance stationary if it satisfies the
following conditions.
1. E (yt) is constant (constant mean)
2. Var (yt) < ∞ (finite variance)
3. Cov (yt, yt-k) is a function of k, but not of t. where k = 1, 2, 3……
In other words, a stochastic process {yt} is said to be weakly stationary (or covariance stationary) if
its mean, variance and autocovariances are unaffected by changes of time. Note that condition (3) is
the equivalent to saying that the covariance between observations in the series is a function only of
how far apart the observations are in time.

82 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

Example: A simple way to model dependence between consecutive observations states that Yt
depends linearly upon its previous value Yt−1.
That is, Yt = δ + θYt−1 + ut |θ| < 1. …………………………………………………6.2.1
where Ut denotes a serially uncorrelated innovation with a mean of zero and a constant variance. The
process in (7.1) is referred to as a first order autoregressive process or AR (1) process. It says that
the current value Yt equals a constant δ plus θ times its previous value plus an unpredictable
component ut. For the moment, we shall assume that |θ| < 1. The process for ut is an important
building block of time series models and is referred to as a white noise process. In this chapter, ut
will always denote such a process that is homoscedastic and exhibits no autocorrelation.
The expected value of Yt can be solved from
E (Yt) = δ + θ E (Yt -1), which, assuming that E (Yt) does not depend upon t, allows us to write

E (Yt) = δ + θ E (Yt),   E (Yt )  ……………………………………………6.2.2
1
Defining yt ≡ Yt - μ, we can write (7.1) as
Yt = θyt-1 + εt. ………………………………………………….……………………………6.2.3
Writing time series models in terms of yt rather than Yt is often Notationally more convenient, and
we shall do so frequently in the rest of this chapter. One can allow for nonzero means by adding an
intercept term to the model. While Yt is observable, yt is only observed if the mean of the series is
known. Note that V (yt) = V (Yt).
The model in (5.321) is a parsimonious way of describing a process for the Yt series with certain
properties. That is, the model in (5.3.2) implies restrictions on the time series properties of the
process that generates Yt. In general, the joint distribution of all values of Yt is characterized by the
so-called autocovariances, the covariances between Yt and one of its lags, Yt-k.
For the AR (1) model, the dynamic properties of the Yt series can easily be determined using (9.1) or
(9.3) if we impose that variances and autocovariances do not depend upon the index t. This is a so-
called stationarity assumption and we return to it below. Writing

 
Var (Yt )  Var (Yt 1  ut )  E Yt 1  ut  2   2 Var (Yt 1 )  Var (ut )

And imposing that Var (Yt) = Var (Yt-1), we obtain

83 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

2
Var (Yt )  …………………………………………………….………………..6.2.4
1 2
It is clear from the resulting expression that we can only impose Var (Yt) = Var (Yt-1) if |θ| < 1, as
was assumed before. Furthermore, we can determine that

2
Cov (Yt Yt 1 )  E (Yt Yt 1 )  E Yt 1  Yt 1    Var (Yt 1 )   .......6.2.5
1 2
And, generally (for k = 1, 2, 3, . . .),

2
Cov (Yt Yt  k )   K
...................................................................................6.2.6
1 2

As long as θ is nonzero, any two observations on Yt have a nonzero correlation, while this
dependence is smaller (and potentially arbitrary close to zero) if the observations are further apart.
Note that the covariance between Yt and Yt−k depends on k only, not on t. This reflects the
stationarity of the process.
Example: Suppose y follows the random walk process:
Yt = yt-1 + ut ………………………………………………………………………………… 6.2.7
Where the stochastic term ut is white noise. One key trait of random walks is that the most recently
observed value of the variable is the best forecaster of future values. By repeated substitution
(assuming y0 = 0), we can write equation (9.7) as:

t
yt   ut
t 1

t
E ( yt )  E (ut )  0 (cons tan t )
t 1

2
Var ( yt )  E ( yt )  E   ut    E (ut 2 )  t u 2
2 t t

 t 1  t 1
The variance depends on t, meaning that the variance of the series is diverging to infinity with t.
thus, the series is not stationary.

84 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

Another simple time series model is the first order moving average process or MA (1) process, given
by
Yt = μ + εt + αεt−1. …………………………………………………………………………..6.2.8

Apart from the mean μ, this says that Y1 is a weighted average of ε1 and ε0, Y2 is a weighted average
of ε2 and ε1, etc. The values of Yt are defined in terms of drawings from the white noise process εt.
The variances and autocovariances in the MA (1) case are given by

 
Var (Yt )  E  t  t 1 2  E ( t 2 ) 2 E ( t 2 )  1 2  2  
Cov (Yt , Yt 1)  E   t  t 1  ( t 1  t  2  E ( t 12 )   2

Cov (Yt , Yt  2 )  E   t  t 1  ( t  2  t 3   0

Or, in general, Cov (Yt , Yt  k )  0 for k  2,3,4......

Consequently, the simple moving average structure implies that observations that are two or more
periods apart are uncorrelated. Clearly, the AR (1) and MA (1) processes imply very different
autocovariances for Yt.
The Box-Jenkins Methodology and Estimation Techniques
The Box-Jenkins approach is one of the most widely used methodologies for the analysis of time-
series data. It is popular because of its generality; it can handle any series, stationary or not, with or
without seasonal elements, and it has well-documented computer programs. It is perhaps the last
factor that contributes most of its popularity.
The basic steps in Box-Jenkins methodology are:
1. Differencing the series so as to achieve stationarity: How do we conclude whether a time series is
stationary or not? We can do this by studying the graph of the correlogram of the series. The
correlogram of a stationary series drops off as k, the number lags, becomes large, but this is not
usually the case for non-stationary series. Thus the common procedure is to plot the correlogram of
the given series yt and successive differences y, 2 y, and so on, and look at the correlogram at each
stage. We keep differencing until the correlogram dampens.
2. Identification of a tentative model: Once we have used the differencing procedure to get a

85 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

stationary time series, we examine the correlogram to decide on the appropriate orders of the AR and
MA components. The correlogram of a MA process is zero after a point. That of an AR process
declines geometrically. The correlogram of ARMA processes show different patterns (but all
dampen after a while). Based on these, one arrives at a tentative ARMA model. This step involves
more of a judgmental procedure than the use of any clear-cut rules.
3. The next step is the estimation of the tentative ARMA model identified in step 2.
4. The next step is diagnostic checking to check the adequacy of the tentative model.
5. The final step is forecasting
Forecasting from Box-Jenkins Models

Suppose that we have estimated the model with n observations. We want to forecast X . This is
t k

called a k-period ahead forecast. It is denoted by X . The first subscript gives the time period
n, k
when the forecast is made, and the second subscript denotes the time periods ahead for which the

forecast is made. Let us start with k = 1 so that we need a forecast of X at time period n. We
n 1
have

X n 1 1 X n  2 X n 1  n 1  1 n   2 n 1
We observe Xn and Xn-1. We can replace  n and  n1 by the predicted residuals.
Predicted residuals: Suppose that we take sample data of n observations and estimate the regression
equation with (n-1) observations at a time by omitting one observation and then use this estimated
equation to predict the y value for the omitted observation. Let us denote the prediction error by
 
   
U *
yt  y  i  . The U * are the predicted residuals. By y  i  we mean a prediction of yt from
 
a regression equation that is estimated from all observations except the ith observation. This is in

contrast to yt , which is the predicted value of yt from a regression equation that is estimated using
all the observations.
The only unknown is  n 1 . Thus we replace by its expected value, zero. Hence

86 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

 
    

X n ,1

1 X n 
2 X n 1  1  n 
2  n 1

Now let us go to k = 2. We have

X n  2 1 X n 1  2 X n  n  2  1 n 1   2 n

We replace  n  2 and  n 1 by zero, their expected value. X n 1 is not known, but we have

the forecast X n ,1
. Thus we get

 
   

X n, 2

1 X n,1 
2 X n  2  n
We continue like this. The procedure is:

1. Write out the expression for X nk .

2. Replace all future values X n  j  j  0, j  k  by their forecasts.

3. Replace all  n  j  j  0 by zero.

4. Replace all  n  j  j  0 by the predicted residuals.

6.2. 2 Multivariate Analysis

In the previous section we considered models for the stochastic process of a single economic time
series. One reason why it may be more interesting to consider several series simultaneously is that it
may improve forecasts. For example, the history of a second variable, Xt say, may help forecasting
future values of Yt. It is also possible that particular values of Xt are associated with particular
movements in the Yt variable. For example, oil price shocks may be helpful in explaining gasoline
consumption. In addition to the forecasting issue, this also allows us to consider ‗what if‘ questions.
For example, what is the expected future development of gasoline consumption if oil prices are
decreasing by 10% over the next couple of years?
Dynamic Models with Stationary Variables

87 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

Considering an economic time series in isolation and applying techniques from the previous
discussion to model it may provide good forecasts in many cases. It does not, however, allow us to
determine what the effects are of, for example, a change in a policy variable. To do so, it is possible
to include additional variables in the model. Let us consider two (stationary) variable Yt and Xt, and
assume that

Yt  Yt 1 0 X t 1 X t 1  t ........................................................................................6.2.9

As an illustration, we can think of Yt as ‗company sales‘ and Xt as ‗advertising‘, both in month t. If

we assume that  t is a white noise process, independent of Xt, Xt-1 … and Yt, Yt-1 …, the above
relation is sometimes referred to as an autoregressive distributed lag model. On the same token if the
model includes one or more lagged values of the dependent variable among its explanatory
variables, then it is called an autoregressive distributed lag model. To estimate it consistently, we
can simply use ordinary least squares.

The interesting element in (7.9) is that it describes the dynamic effects of a change in X t upon
current and future values of Yt.

Taking partial derivatives, we can derive that the immediate response is given by
Yt
 0 ...................................................................................................................................... 6.2.10
X t
Sometimes this is referred to as the impact multiplier. An increase in X with one unit has an
immediate impact on Y of  0 units.

In regression analysis involving time series data, if the regression model includes not only current but
also the lagged (past) values of the explanatory variables, then it is called a distributed lag-model.
Thus,

Yt   0  1 X t 1   2 X t  2 ...  k X t  k  ut .................................... 6.2.11

represents the general form of a distributed lag model.
Both distributed lag model and the autoregressive models are dynamic models because the influence
of the explanatory variable on the dependent variable is distributed over a number of past values of
the explanatory and dependent variables.
88 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module
 Debre Tabor University, CAES, Department of Agricultural Economics

In this sub section, we will discuss the importance of lag in economics. Then we provide the
techniques to estimate a model with lagged values of the explanatory variables. That means it
provides techniques for estimation a distributive lagged model.

The role of “time” or “lag”

The number of lags, s, may be either finite or infinite. Assume that the 𝛽‘s have a finite sum, i.e.

s
 i  
i 0
Lagged values of the variables are important explanatory variables, because most economic
variables are influenced by past patterns of the variable. For example, take the consumption
function. It postulates that the current level of consumption depends on past levels of consumption
and current and past levels of income, i.e.
Ct = f (Ct-1, Yt, Yt-1, X1t, X2t…)
The investment function postulates that it depends on past outputs, on expectation about future
profits, on capital stock and other factors, i.e.
It= f (Qt, Qt-1, Qt-2, …, t, Kt-1,it,…)
Where: Q is the level of output
is profit
K is capital stock
i is interest rate
Very often, dependent variable responds to the explanatory variables with a lapse of time. Such a
lapse of time is called a lag.
Lags are important for decision making especially, for government officials to know how fast, after
how many time periods the economic units will react to changes of various policy variables
(instruments). For example, how fast will producers or consumers react to the imposition of sales tax
and other incentives for investment? How fast will investors react to changes in the interest rate?
Similarly, the lags involved in the demand function following a change in the policy instruments like
price, quantity or advertising of a firm are important for managerial decisions.
Lagged variables are one-way for taking into account the length of time in the adjustment processes
of economic behavior and for handling of expectations about future events. However, economic
89 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module
 Debre Tabor University, CAES, Department of Agricultural Economics

theory never suggests the precise number of lags that should be included in a function, even if it
recognizes the importance of time lags. The researcher will choose among different lags patterns the
one that gives the most satisfactory fit on the basis of statistical criteria.
Estimation of Distributed lag models
Assume that Y depends on the value of X over s periods. This is called a finite (lag) distributed lag
model because the length of lag is specified.

Yt   0  1 X t   2 X t 1 ...  s X t  s  ut .................................... 6.2.12

However, if the model is given as:

Yt   0  1 X t   2 X t 1 ... ut .................................................... 6.2.13

Such a model is called an infinite (lag) model.
To estimate the𝛽‘s in a infinite model, we may adopt two approaches: ad hoc estimation and a prior
restrictions on the𝛽‘s by assuming that𝛽‘s follows some systematic pattern.
Suppose the model includes only lagged values of the exogenous variable (s) in the set of
explanatory variables, make the usual assumptions about the error term U.

U N (0,  2)
U
E (UiUj) = 0 for i j
E (UiXj) =0 for j=1, 2, 3…K
Since the explanatory variable Xt is non-stochastic, it is uncorrelated with the disturbance term Ut.
Thus, Xt, X t-1 and so on are non-stochastic and the ordinary least squares can be applied.
This method suggests that to estimate such a model, one may proceed sequentially., i.e. first regress
Yt on Xt, then regress Yt on Xt and X t-1 and so on. This sequential procedure stops when the
regression coefficients of the lagged variables start becoming statistically insignificant and/or the
coefficient of at least one of the variables changes signs from positive to negative or vice versa.
However, the following problems will arise in attempting to apply this approach.
 If the number of lags is large and the sample is small (in the case of time series data), we may be
unable to estimate the parameters because there will be no adequate degrees of freedom to carry out
the statistical tests of significance.
 There will be a Multicollinearity problem, since there is strong correlation between successive
values of the same variable. With strong collinearity, the values of the estimates will be imprecise
90 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module
 Debre Tabor University, CAES, Department of Agricultural Economics

and their standard errors will be large so that we may be led to mis-specification of the model by
dropping variables.
To avoid these problems, various methods have been suggested to reduce the number of lagged
variables. This is achieved by imposing restrictions on the 𝛽‘s and constructing new variables from a
linear combination of the lagged variables. The methods differ in the weights which are used in
constructing these new variables.
One of the most popular distributed lag models with endogenous lagged variables is Koyck‘s
Geometric lag scheme. This model assumes that the weights/lag coefficients are declining
continuously following the pattern of geometric progression.
Characteristic equation and unit roots
Define the lag operator Z as: ZkXt= Xt-k k = 1, 2, 3…
e.g. Z1Xt = ZXt = Xt-1 Z2Xt = Xt-2 etc.
Suppose that a stochastic process {yt} follows an autoregressive process of order p AR (p):

yt  1 yt 1   2 yt  2   3 yt 3 ...  p yt  p  ut ................................ 6.2.14

where the error term ut is white noise. We can write equation (7.14) using the lag operator as:

yt  1Zyt   2 Z 2 yt .....  p Z p y p  ut

 1 1Z   2 Z 2 ..........  p Z p yt  ut 
 a (Z ) yt  ut

where a (Z )  1 1Z   2 Z 2 .....  p Z p is a polynomial in Z. Setting a (Z) to zero,

we get the characteristic equation:

1 1Z   2 Z 2 .....  p Z p  0......................................................... 6.2.15

Note:
1. The stochastic process {yt} in equation (7.14) is said to be stationary if all the roots of the characteristic
equation (7.15) lie outside the unit circle.
2. The stochastic process {yt}in equation (7.14) is said to be non-stationary if any of the roots of the
characteristic equation (7.14) lie on or inside the unit circle.

91 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

3. In particular, if Z = 1, is a root of the characteristic equation (7.15), then the stochastic process is said to
have a unit root.

Example: Suppose yt follows the AR (1) process

yt  yt 1  ut .............................................................................. 6.2.16

where the error term ut is white noise. Using the lag operator we have:

yt  Zyt  ut
 (1 Z ) yt  ut
 a (Z ) yt  ut

The characteristic equation is given by: 1 Z  0 and the root of the equation is Z  1 .

a) If |  | < 1, then Z = 1/p will be larger than one (outside the unit circle), and hence, the process will be
stationary.
b) If  = 1 (meaning the process is a random walk), then the root of the characteristic equation is Z = 1
(on the unit circle). Hence, the process has a unit root (is integrated of order one) and is a non-stationary
time series.
Integrated processes and differencing
The difference operator  is defined as:

d yt 1 Z d yt d 1,2,3,...
where Z is the lag operator: Zyt  yt 1
e.g. First difference: 1 yt yt  1 Z  yt  yt  Zyt  yt  yt 1

Second difference: 2 yt  1 Z 2 yt  1 2Z  Z 2 yt  

 yt  2Zyt  Z 2 yt  yt  2 yt 1  yt  2
We have seen that the random walk process:
yt = yt-1 + ut where the error term ut is white noise, is non-stationary. This equation can be written as:

92 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

yt - yt-1 = ut  y t  u t

Since the error term ut is white noise (which is stationary by definition), the first difference yt is also
stationary. The series yt is said to be integrated of order one, denoted by I (1), since taking a first
difference produces a stationary process. A series is said to be integrated of order d, denoted by I (d),
if the series becomes stationary after being differenced d times.
Most macroeconomic time series such as output or unemployment rate are I (1). An I (2) series is
growing at an ever-increasing rate. But series that are I (3) or greater are extremely unusual.
Correcting for error autocorrelation of AR (1) scheme
Consider the model:

Yt  1   2 X t  t t 1, 2,......T ....................................... 6.2.17

Where the errors are generated according to the AR (1) scheme:

 t   t 1 ut Where ut fulfills all assumptions of the CLRM. Suppose by applying any one of

the above tests you come to the conclusion that the errors are auto correlated. What to do next?

Lagging equation (17) by one period and multiplying throughout by  , we get:

Yt 1   1   2 X t 1   t 1 .......................................... 6.2.18

Subtracting equation (18) from equation (17), we get:

Yt  Yt 1  1(1  )   2 ( X t  X t 1)  ( t   t 1 )

Y *  1*   2 X t * ut .................................................... 6.2.19

Where ut  t   t 1.
The above transformation is known as the Cochrane-Orcutt transformation. Since ut fulfills all
assumptions of the CLRM, we can apply OLS to equation (***) to get estimates which are BLUE.

93 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module

 Debre Tabor University, CAES, Department of Agricultural Economics

APPENDIX

F Distribution Tables

The F distribution is a right-skewed distribution used most commonly in Analysis of Variance. When referencing the F
distribution, the numerator degrees of freedom are always given first, as switching the order of degrees of freedom
changes the distribution (e.g., F(10,12) does not equal F(12,10) ). For the four F tables below, the rows represent
denominator degrees of freedom and the columns represent numerator degrees of freedom. The right tail area is given in
the name of the table. For example, to determine the .05 critical value for an F distribution with 10 and 12 degrees of
freedom, look in the 10 column (numerator) and 12 row (denominator) of the F Table for alpha=.05. F (.05, 10, 12) =
2.7534

94 Bish200821@[Link] | Prepared by: Bishaw A. (MSc) : Econometrics Module