Wollo University, College of Business and Economics, Department of Economics

Chapter Two
The Simple Regression Model
2.1 Definition of Simple Regression Model
Given the two variables Y & X, the simple linear regression explains Y interims of X which
shows how Y varies with changes in X. The simple equation that linearly relates Y to X
with the inclusion of disturbance or error term  which represents other factors affecting Y
is called simple linear regression model. Such a simple equation is:
Y   0   1 X  ............................(2.1)

This equation is also called the two variable linear regression model or bivariate linear
regression model.

In order to define each of the quantities in the equation, mainly the variables Y&X have
several different names used interchangeably. That is, Y is called dependent variable,
explained variable, response variable, predicted variable or regresand. X is called the
independent variable, explanatory variable, control variable, predictor variable or regressor.

The variable  , is called disturbance or error term which represents factors other than X that
affect y. The equation also addresses the functional relationship between Y and X. If the
other factors in  are fixed, so that change in  is zero, then X has a linear effect Y.

Thus, change in Y is simple  1 multiplied by the change in X, This means that  1 is the

slope parameter. Finally the term  0 is called the intercept parameter.

The simple linear regression can be best explained by a simple wage equation:
Wage =  0   1educ   .................(2.2)

This equation indicates change in years of education affects the wage level. Thus,  1
measures the change in hourly wage given certain years of education, holding all other
factors,  , fixed. Some of those factors may be labor force experience, ability, work ethics
etc. Although it is not realistic for many economic applications, the linearity feature in the
model refers to a one unit change in the independent variable (years of education) has the
same effect on the dependent variable (wage).

Econometrics I Lecture Notes; 2015 By Addisu Molla (PhD) 1

Wollo University, College of Business and Economics, Department of Economics

The most difficult issue to be addressed by simple linear regression model is that whether
the model really allows us to draw ceteris paribus conclusions about how X affects. That is,

we need to show how  1 does measure the effect of X on Y, holding all other factors (in  )

fixed. So as to estimate the effect of X on Y,  1 , with the help of ceteris paribus principle,
we need to explain how unobserved variable X.

Before stating a key assumption about how X and  are related, we can have one
assumption about then average value of  . As long as the intercept  0 is included in the

equation, nothing is last by assuming that the average value of  in the population is zero.
i.e, E( ) = 0 ………………………….(2.3)

Importantly, this assumption doesn’t say nothing about relationship between X &  but
simply explains the distribution of unobservable in the population.

We now turn to the crucial assumption regarding how X &  are related. The natural
measure of the association between the two variables (X &  ) is the correlation
coefficient. However, even if the correlation coefficient shows those X &  are
uncorrelated, it does not be a guaranty that the two variables are not totally related. That is,
it is possible for  to be uncorrelated with X while being correlated with functions of X,
such as X2 , because correlation only measures linear dependency between X &  . The
better approach to solve such problem is using the expected value of  given X. That is, the
conditional distribution of  given any value of X.

The crucial assumption that is prevailed in this instance is that the expected value of 
given X is zero which explains that the expected (or average) value of  does not depend
on the value of X.
Thus, E (  /X ) = 0 ………………… (2.4)

The above equation indicates that the average value of unobservable is the same for any
value of X. this equation is already developed from the crucial assumption called zero
conditional mean assumption and therefore this equation must be equal to the earlier
assumption which says the average value of  on the entire population is zero which is also
termed as the zero mean assumption.

Econometrics I Lecture Notes; 2015 By Addisu Molla (PhD) 2

Wollo University, College of Business and Economics, Department of Economics

i.e, E( /X ) = E ( ) = 0 …………………(2.5)

Since we have come up with the crucial assumption that explains  is not related with X
and hence it allows us to draw a ceteris paribus conclusion, we can use the simple linear
regression model in econometric analysis.

Taking the expected value of the simple linear regression model conditional on X and the
zero conditional mean assumption as given above E (  /X ) = 0 we can obtain a new

equation as: E (Y/X) =  0   1 X ...................................(2.6)

This equation is a population regression function (PRF). If the zero conditional mean
assumption, (i.e, the variables X &  are uncorrelated) is true, then it is useful to break the
simple linear regression model into two components:

The part  0   1 X is called deterministic or systematic component and the part  is called

un deterministic, unsystematic or random component. In other words, out of the total

explained in Y, part of it is explained by X in the deterministic component and the remained
part of X, which is not explained by X, is explained by  in a random component.

To find out the estimate of the population parameters (  0 &  1 ) , we need to derive the

sample regression function (SRF) from the population regression function (PRF). Implying,
we can estimate the parameters from a sampled date say a randomly selected sample of Y
values for the fixed X values. Since there are sample fluctuations, we may not be able to
estimate the PRF accurately from a particular sample date (or SRF). To avoid such problem
we would critically identify the SRF which is a close approximation of the PRF. In this case
the estimate of the population parameters, the sample parameters  0 &  1 are as close as

possible to their respective values of the true population parameters  0 &  1 though we will

never know the true  0 &  2 values.

The sample counter part of the earlier PRF can be written as: Yi   0   1 X i ..............(2.7)

Where: Y is read as “Y – hat” or “Y-Cap”;

Y i = estimator of E(Y/Xi)
 
 0 & 1 are estimators of  0 &  1 .

Econometrics I Lecture Notes; 2015 By Addisu Molla (PhD) 3

Wollo University, College of Business and Economics, Department of Economics

2.2. The Methods of Ordinary Least Squares (OLS)

Basically there are two methods of estimating the PRF on the basis of SRF. These are
ordinary least squats (OLS) and maximum likelihood (ML). Due to its advantage of
mathematically much simpler than the method of ML, the method of OLS will be our focus
of discussion in this section.

To estimate the PRF from its SRF, we need first to explain the least square principles from
the method of OLS. Recall the PRF was given by: Y1   0   1 X i   i its estimate from
SRF is also given by:
  
Y1   0  1 X i   i
    
Yi  Y1   i , because Y1   0  1 X i

This equation can also be expressed as:

 
 i  Yi  Yi
  
 i  Yi   0   1 X i ..........................(2.8)

The earlier equation shows that the residuals (  i ) are simply the differences between the

actual Y values (Yi) and the estimated Y values (Yi ) .

The next task is determining the SRF itself provided that n-pairs of observations on Y&X
are given (i.e, PRF). So as to make the SRF is as close as the PRF the summation of
differences between the actual and estimated values which is the sum of residuals should
be as small as possible.
  
That is, (Yi  Yi )    i should be as small as possible. If this is so the estimated value, Yi ,

is as close as possible to the actual Y value, Y. consequently the estimated parameters

 
 0 & 1 will be the true estimates of the population parameters  0 &  1 .

 
However in the minimum   i criterion, the sum can be small even though the  i are
   
widely spread about the SRF. For instance, if we assume the four residual  i ,  2 , 3 &  4 in
a SRF have the values 10, -2, +2, and -10 respectively, then the algebraic sum of these
residuals is zero although 1 &  4 are scattered more widely around the graphical
representation of the SRF.

Econometrics I Lecture Notes; 2015 By Addisu Molla (PhD) 4

Wollo University, College of Business and Economics, Department of Economics

Y 3 SRF
 
1  4 Yi   o  1 X i
2

X
Fig 3.1 Least-squares Criterion

To avoid this problem we should adopt another criterion called the least- squares criterion,

which states that the summation of squared of residuals (  2 ) should be as small as
possible.
 
 2  (Yi  Yi ) 2
i.e.   
 2  (Yi   0  1 X i ) 2 .....................................(2.9)


as noted previously, under the minimum   i criterion, the sum can be small even though

the  i are widely spread about the SRF. But this is not possible under the least-squares

criterion, because by squaring  i this method gives more weight to residuals 1 &  4 than
   2
 2 &  3 That is, the larger the  i (in absolute value), the larger  i . A further justification
for the least square method lies in the fact that the estimators obtained by it have some
desirable statistical properties, which will be discussed in latter sections.

The method of ordinary least squares (OLS) provides unique estimates of  0 &  1 that gives
2
the smalls possible value of  i .
 
 2  (Yi  Yi ) 2
i.e.  2  
  (Yi   0  1 X i ) 2 .....................................(2.10)

To minimize the equation (2.10), the necessary condition of differentiation, derivatives of

the equation 2.10 with respect to  0 &  1 should be equal to zero, must be fulfilled.
2 2
i.e,    i  0 and  i  0
 
 0  1

Econometrics I Lecture Notes; 2015 By Addisu Molla (PhD) 5

Wollo University, College of Business and Economics, Department of Economics

2 
The partial derivatives of  i with respect to  0 gives
2  
  i   (Yi   0   1 X i ) 2  0
 
 0  0
 
  (Yi   0  1 X i )  0
 
  Yi  n 0  1  X i  0

  Yi  n 0  1 X i ...........................................(2.11)

Dividing equation 2.11 by n, which is the no of observations in the sample, we can obtain =
 
Y   0  1 X ................................................(2.12)

Where Y and X represents the means of Y&X respectively.

2
Through similar procedure the partial derivatives of  i with respect to  1 gives.
2  
   i =   (Yi   0   1 X i ) 2 = 0
 
1 1
 
 (Yi   0   1 X i )( X i )  0
  2
  X i X i   0  X i  1  X i  0
 
  Yi X i   0  X i   1  X 2 ....................................(2.13)

Note that equations 2.11 and 2.13 are known as normal equations and salving these normal
equations simultaneously, we obtain the least squares estimates of  0 &  1 as:

 1 = n  X i Yi   X i  Yi
2
n  X i  ( X i ) 2

= ( X i  X )(Yi  Y )

( X i  X ) 2

=  xi y i
2
 xi .......................................................................(2.14)

Econometrics I Lecture Notes; 2015 By Addisu Molla (PhD) 6

Wollo University, College of Business and Economics, Department of Economics

Where Y and X are sample means of X & Y and hence we can define xi  X i  X and

y i  Yi  Y . Hence forth we adopt the convention of letting the lower case letters denote

deviations from mean values. n represents the sample size.


 0 =  X i  Yi   X i  X i Yi
2
n  X i  ( X i ) 2 .......................................(2.15)

By simple algebraic manipulations of equation (2.11), the parameter  0 can also be solved

as:
 
 Yi  n 0  1  X i
 
 Y   0  1 X
 
  0  Y   0 X ........................................................(2.16)

2.3 Properties of OLS Estimates

2.3.1 Algebraic Properties
There are several useful algebraic properties f OLS estimates and their associated statistics.
However, only the three most important of these are discussed.
1. The sum, and therefore the sample average of the OLS residuals, is zero.

i.e.  i  0.......................................................(2.17)

2. The sample covariance between the regressors and residuals is zero.


i.e.  X i i  0....................................................(2.18)
3. The total variation in the dependent variable is equal to the sum of the variation in
the explanatory variables and the variation in the residuals (or variation in un
explained variables).
 
Yi  Y ( y i )  Yi  Y ( y i )  Yi Y i(ei )......... .......... ....( 2.19)

Total variation = explained variation + unexp. Variable.


Where = Yi = actual value of Y; Y = estimated value of Y and Y = means of Y.

The third algebraic property of OLS estimates provides us the sample covariance between

the estimated Y value ( Yi ) and the residual. ( ei ) is zero.

Econometrics I Lecture Notes; 2015 By Addisu Molla (PhD) 7

Wollo University, College of Business and Economics, Department of Economics

Note: when we consider the square of summation of the three variations discussed above
we will get the three sum of squares as follow = total sum of squares (TSS), explained sum
of squares (ESS) and residual sum of squares (RSS).
2
TSS = (Yi  Y ) 2  y i
 2
ESS = (Yi  Y ) 2   y i
 2
RSS = (Yi  Yi ) 2   ei

 TSS = ESS + TSS …………………………………..(2.20)

2.3.2 Statistical Properties

The statistical properties of OLS estimates refer to the assumption and theorems developed
on the residuals, explanatory variables and parameters.

A. Assumption of Ordinary least Squares (OLS)

Assumption 1: Homoscedasticity of equal variable of  i .

The variance of  i is constant or the same for all observation in each period so that the error
terms are homoscedastically distributed.
i..e. Var ( i ) =  2 …………………………….(2.21)

Assumption 2: No autocorrelation between the disturbances pr error terms.

The random terms of different observations are independent and hence the covariance
between such error terms is zero.

i..e. Cov ( i ,  i ) = 0 …………………………….(2.22)

Assumption 3: Zero covariance between  i and X i

The random term  i is independent of the explanatory variable X i so that the covariance

between  i & X i is zero.

i..e. Cov (  i , X i ) = 0 …………………………….(2.23)

Assumption 4: The explanatory variables X i are measured without error.

If there is some error in measurement of explanatory variables like problem of aggregation,

round off etc. it will be absorbed by the random term  i . So the explanatory variables are

measured without error.

Econometrics I Lecture Notes; 2015 By Addisu Molla (PhD) 8
Wollo University, College of Business and Economics, Department of Economics

Assumption 5: The regression model is correctly specified.

There is no specification bias or error in the model used in the analysis.

Assumption 6: There is no perfect multicollinearity.

The explanatory variables are not perfectly linearly corrected so that the correlation
coefficient between the two explanatory variables (say X 1 & X 2 ) is not unity.

i..e. rx1x2  1 …………………………….(2.24)

B. The Gauss – Markov Theorem

In addition to the assumptions explained the Gauss Markov Theorem also explains the
statistical properties of the least squares estimators. According to the Gauss Markov
Theorem, the least squares estimators satisfy the basic property that is termed as BLUE
(Best Linear Unbiased Estimators) are said to be best linear unbiased estimator (BLUE) if
the following hold.
1. The parameter estimates are a linear function of the dependent variable. For example,
the parameters  0 &  1 are a linear function of the dependant variable Y in the
following simple regression model.
Yi   0   1 X i   i ...............................................(2.25)

2. The estimated parameters (  0 &  1 ) are unbiased estimators of the population

parameters (  0 &  1 ). So the expected or mean values of the estimated parameters are

equal to the true population parameters.

3. The parameter estimates are efficient estimators. This referred to minimum variances
of the estimators in which the variances of the estimators are as small as possible.
i.e, Var (  0 ) and Var (  1 ) are as small as possible.

Econometrics I Lecture Notes; 2015 By Addisu Molla (PhD) 9

Wollo University, College of Business and Economics, Department of Economics

2.4 Inferences and Predications

2.4.1 Hypothesis testing on significance of the Parameter Estimates
 
The values of the parameter estimates  0 & 1 which are obtained from the sample
observation must be tested whether they are statistically reliable to explain the population
parameter. To test the statistical significance of the estimates we need to do the following
step wise procedures.
a. Computing the mean and variance of estimates
b. Testing the statistical significance of the estimates using standard error and test
techniques
c. Construction a confidence interval for the estimates
A. Mean and Variance of the estimates
 
 Mean of  0 : The mean of the ordinary least squares estimates  0 is equal to the true

value of the population parameter  0 .


i.e, E(  0 ) =  ……………………………..(2.27)
 
 Variance of  0 : the variance of the ordinary least squares estimates  0 is the square

of the expected deviation of the sample estimates  0 from the population parameter

 0 [E(  0 )]
  

i.e, Var (  0 ) = E  0  E (  0 ) 
2


=  ( 0   0 ) 2
2 2
= u  X i

n ( X i  X ) 2 .......... .......... .......... .....( 2.28)

 
 Standard error of  0 : it is the square root of the variance of the estimates  0 .

  2
 u  Xi
2

i.e, Se(  0 ) = Var (  0 )  ......................................(2.29)

n ( X i  X ) 2
 
 Mean of  1 : the mean of  1 is equal to the true value of the population parameter.

i.e, E (  1 ) =  1 …………………………………………………….. (2.30)

Econometrics I Lecture Notes; 2015 By Addisu Molla (PhD) 10

Wollo University, College of Business and Economics, Department of Economics

 
 Variance of  1 : The Variance of  1 is the square of the expected deviation of

sample estimates  1 from the population parameter.

  
i.e, Var (  1 ) = E  (  1 - E(  1 )  2

= E(  1 -  1 ) 2
1
= u2 ..................................................(2.31)
( X i  X ) 2
2
Where  u is the variance of the error term .
 
 Standard error of  1 : if is the square root of the variance of  1 .

  2 1
i.e, S(  1 ) = Var (1 )   u ..............................(2.32)
( X i  X ) 2

B. Statistical Significance Test of Estimates.

 The Standard error test of the Estimates
Before we start to test the statistical significance of the parameter estimates we need
to develop the null hypothesis, H 0 , and alternative hypothesis, H 1 .

H 0 =  i  0  the estimate is not statistically significant.

H 1 =  i  0  the estimate is statistically significant.

Having the null & alternative hypothesis, we need to compute the standard errors
of the estimates S (  i )  in order to apply the standard errors test. After

computing the standard errors, the standard errors of the estimates need to be
compared with the numerical values of the estimates and can be concluded as
follow.

i. If the standard error is less than half of the numerical value of a parameter.
 1 
(i.e, S (  i ) <  / i /  , then we reject H 0 (or accept H 1 ) so that we conclude that
2
the estimate  i is statistically significant.

ii. If the standard error of the estimate is greater than half of the numerical value

Econometrics I Lecture Notes; 2015 By Addisu Molla (PhD) 11

Wollo University, College of Business and Economics, Department of Economics

 1 
(i.e, , S (  i ) >  /  i /  , then we accept H 0 (or accept H 1 ) so that we conclude
2
that the estimate  i is not statistically significant.

 The Student’s t-test of the Estimates

The t-test is the second test of significance next to the standard error test. To apply
the t-test we should follow the following step wise procedures:
i. Define the null & alternative hypotheses.
 
H 0 =  i =0 and H 1 =  i ≠ 0
ii. Choose the level of significance, which is the probability of rejecting the null
hypothesis when it is true, in other words, level of significance refers to the
probability of committing type one error by the researcher.
iii. Define the number of degree of freedom (d.f) as N-K where N is the sample size
and K is No of estimated pentameters,  i ‘s.

iv. Find the theoretical or table value of t, which is obtained from the t-table using
the d.f and the significant level chosen. This t-value is termed as t-tabulated, ttab,
or t-critical, tcr.
v. Find the calculate value of t, tcal using the following formula.
 
t =  i /S (  i )
This t-value is also termed as t-calculated, tcal, t-computed, tcom, or t-statistic ts.
vi. Finally, the t-calculated will be compared with the t-tabulated value and can be
concluded as follow.
a. If tcal > ttab, then we reject the bull hypothesis (or accept alternative

hypothesis) so that we concluded that the estimate  i is statistically
significant.
b. If tcal < ttab, then we accept the null hypothesis (or reject the alternative

hypothesis) so that the estimated parameter  i is not statistically significant.

Econometrics I Lecture Notes; 2015 By Addisu Molla (PhD) 12

Wollo University, College of Business and Economics, Department of Economics

C. Confidence intervals for the estimates.

Confidence interval is an alternative way of expressing the significance level with
percentages. For example, if the confidence interval is given as 95 percent (i.e, CI=95%)
then it refers to the researcher is 95% confident enough not to commit type I error by
rejecting the null hypothesis when it is true. In other words, using significance level of 
=5% the researcher has 5% probability of committing the error.

Thus, the CI of 95% can be computed and presented using the following formula
expression.
   

P  i  t / 2 . S(  i )   i   i  t / 2 . S(  i )   95%........(2.33)
Where  is the significance level and the equation shows that the probability of the

estimated value lied between the two extreme values (to the right & left of  ) is 95%.
Alternatively, the CI can be rewritten as:
 
 i   i  t  / 2 SE (  i )..................................................(2.34)

2.4.2 Prediction with Simple Regression Model

Prediction is one of the major goals of econometric analysis. The estimated regression
 
equation (i.e, Y   0  1 Xi) ) is used for predicting the value of Y as dependent variable at
a iven value of X as independent variable.

Suppose that we obtain the following regression result: Y =20.75+0.75X. The predicted
value of Y will be computed as follow at a given value of X, say 100.

i.e, Y =20.57+0.75(100)

Y =95.57

2.5 The R2 as test of Goodness of fit.

The r/R-squared or a coefficient of determination r2 (two variable case) or R2 (multiple
regression) is a measure that tells how well the sample regression line fits the data. Thus,
the coefficient of determination r2 in this simple regression model measures the goodness of
fit of the sample regression line. In other words, it measures how much of the total variation

Econometrics I Lecture Notes; 2015 By Addisu Molla (PhD) 13

Wollo University, College of Business and Economics, Department of Economics

in the dependent variable Y is explained by the explanatory variable X. that is, it measures
then goodness of fit of the model in explaining the variation.

In order to compute r2, we need to recall the relationship between various sum of squares
appearing in equation (2.20) which states that the total sum of squares (TSS) is the sum of
the explained sum of squares (ESS) and the residual sum of squares (RSS).
i.e, TSS = ESS + RSS ………………………………..(2.35)

This equation tells us that the total variation is the sum of the explained variation and
unexplained variation. Therefore, the above eq(2.35) can be rewritten as:
 
 ( Yi  Y ) 2 =  ( Y  Y ) 2 +  ( Yi  Yi ) 2
Total variation (TSS) = Exp. Var. (ESS) + Unexp.Var. (RSS) ……..(2.36)

Dividing both sides of eq(2.36) by TSS (Yi  Y )

2
 we can obtain:
 
(Yi  Y ) 2 (Yi  Yi ) 2
1= 
(Yi  Y ) 2 (Yi  Y ) 2
1 = ESS + RSS
TSS TSS
Now we can define r2 as:

2 (Yi  Y ) 2 ESS
r =  ...............................................(2.37)
(Yi  Y ) 2 TSS
Thus, alternatively r2 can be computed as:
1 = r2 + RSS
TSS
RSS
r2 = 1- .................................................................(2.38)
TSS
Note: A quantity r as coefficient of correlation is closely related but conceptually very
much different from r2. That is, r2 shows the proportion or percentage of variation in Y
explained by X where as r shows a measures the degree of association between two
variables.

The coefficient of correlation, r, can be computed either from:

r   r 2 ...........................................................................(2.39)

Econometrics I Lecture Notes; 2015 By Addisu Molla (PhD) 14

Wollo University, College of Business and Economics, Department of Economics

Or from its definition:

n  X i Yi   X i .  Yi
r=
n  X i
2
 ( X i ) 2 n  Yi 2  ( Yi ) 2 
 xi yi
= .............................................................(2.40)
2 2
( xi )( y i )

Where the lower cases X &Y refers ( X  X ) and ( Y  Y ) respectively which shows the
deviation of variables X and Y from their respective means.

2.6 Further Issues in Simple Regression Model

2.6.1 Units of Measurement and Functional form
In simple regression model two important issues that deal about measurement and
functional forms should be considered:
(1) Understanding how changes in units of measurement of the dependent and/or
independent variables affect OLS estimates
(2) Knowing how to incorporate popular functional forms in to the regression model

2. The effect of changing units of measurement on OLS statistics

Suppose that we are interested to study the relationship between a particular firm’s
performance intermesh of return on equity (roe) and its chief executive officers – CEO
salary (salary), we may develop a regression model as:
Salary =  0   1 roe …………………………………(2.41)
Suppose the salary is measured in thousands of dollar (Salardol) and the estimated model is

given as: salar dol = 963.191 + 18.501roe………………….(2.42)

An important point here is knowing the effect on OLS estimates when the units of
measurement of the dependent and independent variables change. For example, if the
dependent variable is multiplied by 1000 which means each value in the sample is
multiplied by 1000, then the OLS intercept and estimates are also multiplied by 1000
provided that nothing has changed about the independent variable. Therefore, the above
estimated equation will be changed by (This assumes salary measured in dollar):

salar dol = 963.191 + 18,501roe ………………………(2.43)

Econometrics I Lecture Notes; 2015 By Addisu Molla (PhD) 15

Wollo University, College of Business and Economics, Department of Economics

Note that both eq(2.42) & (2.43) gives the same interpretations. That is, if return on equity
(roe) increased by one unit, then the predicted salary increased by $18,501.

Generally, it is easy to figure out what happen to the intercept and slope estimates when the
dependent variable changes its unit of measurement. That is, if the dependent variable is
multiplied by the constant C, then the OLS intercept and slope estimates are also multiplied
by C.

We can also use CEO salary example to see what happens to the intercept and estimate
when we change the units of measurements of the independent variable. For example, if we
divide the independent variable by 100 - which means each value in the sample is divided
by 100 – then the original regression model in which the variable salary is measured in
thousands of dollar (i.e, eq(2.42) will be changed to Y the following estimated model:

salar dol = 963.191 + 1850.1roedec ………………………(2.44)
Where ‘roedec’ refers to return on equity in decimals.

According to the above estimated equation, in order to preserve the interpretation of the two
equations eq(2.42) and eq(2.44) the slope estimate should be multiplied by 100 when the
independent variable is divided by 100. But the intercept will not be changed. This is so
because the coefficient on roedec is 100 times the coefficient on roe in eq(2.42). Implying
that changing roe by one percentage point is equivalent to roedec = 0.01.

Generally, if the independent variable is divided or multiplied by some non zero constant,
C, then the OLS slope coefficient is multiplied or divided by C respectively but does not
affect the intercept.

2. Incorporating Non linearity in Simple Regression

So far we have focused on linear relationship between the dependent and the independent
variables (i.e, Yi   0   1 X i   i ). However, since linear relationships are not nearly

general enough for all economic applications it is rather better to incorporate much
nonlinearity in to the simple regression model.

Generally, we can come up with the two commonly used nonlinear functional forms of
regression models in applied social sciences. These are:

Econometrics I Lecture Notes; 2015 By Addisu Molla (PhD) 16

Wollo University, College of Business and Economics, Department of Economics

i. The Lon-linear regression model

In the applied work of social sciences, we may encounter regression equations
where the dependent variable appears in logarithmic form while the independent
variable is linear form. The wage education example where we regress hourly
wages on years of education (educ), for example, reveals such nonlinear functional
form.
i.e, log (wage) =  0   1 edu + U …………..…………(2.45)

This functional form recognizes the constant percentage increase of wage for more
years of education: a percentage increase in wage is the same given one more years
of education. Therefore, the main reason for using the log of wage is to impose a
constant percentage effect of education on wage.

ii. The log-log Regression Model

The log-log functional model is also another important form of nonlinear model in
which both the dependent and the independent variables appear in logarithmic
form. This functional form of the regression model is used to obtain a constant
elasticity model, for example, a model relates CEO salary to firm sales.
Log (Salary) =  0   1 log (sales) +u ……………….(2.46)

This is a constant elasticity model in which the coefficient of log (sales),  i is the
estimated elasticity of salary with respect to sales. It implies that 1 percent increase
in firm sales increase CEO salary by  i percent- the usual interpretation of
elasticity.

2.6.2. Regression through the Origin

Although it appears in rare cases, there is a situation of which the expected value of the
dependent variable is zero when the independent variable X is equal to zero in a simple
linear regression model. For example, if income (X) is zero, then the incomes tax revenues
(Y) must also be zero.

Econometrics I Lecture Notes; 2015 By Addisu Molla (PhD) 17

Wollo University, College of Business and Economics, Department of Economics

This type of simple linear regression model is a model without an intercept so that it is
known as regression through the origin and can be presented as:

y =  i X ………………..……………………………(2.47)

Obtaining eq(2.47) is called regression through the origin because the line of this equation
pass through the origin of a graph where X = 0, Y = 0.

To obtain the slope estimate  1 we still rely on OLS method, which in this case minimize
the sum of squared residuals:
2 
 ei  ( y i  y i ) 2

= ( y i   1 X ) 2 .......... .......... .......... .......... .......... .........( 2.48)

Using calculus, it can be shown that  1 must solve the first order condition:

  ei
0
 1

2 ( y i  1 X i ) 2
0
21
2 ( y i  1 X i ) .  X  0
( y i   1 X i ). X  0
2
 y i xi   1  x i  0
2
 1  xi   y i xi
 y i xi (Yi  Y )( X i  X )
1  2
 .........................................(2.49)
 xi ( X i  X ) 2

Econometrics I Lecture Notes; 2015 By Addisu Molla (PhD) 18