Sunteți pe pagina 1din 57

Managerial Economics in a

Global Economy, 5th Edition


by
Dominick Salvatore

Chapter 4
Demand Estimation

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 1
The Identification Problem

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 2
The Identification Problem
• Observed p-q data points result from the
intersection of unobserved D & S curves.
• Therefore, the dashed line connecting
these points is not the demand curve.
• To derive D2 for example we allow the S
curve to shift or be different and then
correct through regression analysis for the
forces that cause the D curve to shift or to
be different.

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 3
Demand Estimation:
Marketing Research Approaches
• Consumer Surveys
• Observational Research
• Consumer Clinics
• Market Experiments
• Virtual Shopping
• Virtual Management
Prepared by Robert F. Brooker, Ph.D.
Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 4
Regression Analysis
Year X Y
1 10 44
Scatter Diagram
2 9 40
3 11 42
4 12 46
5 11 48
6 12 52
7 13 54
8 13 58
9 14 56
10 15 60
Prepared by Robert F. Brooker, Ph.D.
Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 5
Regression Analysis
• Regression Line: Line of Best Fit

• Regression Line: Minimizes the sum of the


squared vertical deviations (et) of each point
from the regression line.

• Ordinary Least Squares (OLS) Method

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 6
Assumptions of Regression Analysis
• The error term (e) (or vertical deviation) is:
– normally distributed.
– Has zero expected value or mean, and
– Has constant variance in each time period and for all
values of X, and that
– Its value in one time period is unrelated to its value in
any other period.
• These assumptions are required to
• obtain unbiased estimates of the slope coefficient
and
• be able to utilize probability theory to test for the
reliability of the estimates.
Prepared by Robert F. Brooker, Ph.D.
Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 7
Regression Analysis

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 8
Ordinary Least Squares (OLS)

Model: Yt  a  bX t  et

ˆ
Yˆt  aˆ  bX t

et  Yt  Yˆt

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 9
The Error Term: et
• It is the vertical deviation of the actual or
observed value from the estimated value of the
dependent variable from the regression line in
period t.
• Reasons for such an error:
– Numerous explanatory variables with only slight or
irregular effect on Y are not included in the model
equation.
– Possible errors in the measurement of Y
– Random human behavior leads to different results
(for example different purchases) under identical
conditions.
Prepared by Robert F. Brooker, Ph.D.
Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 10
Regression Line
• It is the line that best fits the data points.
• It is so because the sum of the e2t is minimum.
• Remember to square the errors before adding
them up in order to avoid the cancellation of
errors of equal size but opposite signs.
• Squaring the errors also penalizes larger errors
relatively more than smaller ones.

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 11
Ordinary Least Squares (OLS)

Objective: Determine the slope and


intercept that minimize the sum of
the squared errors.

n n n

t  t t  t
e 2

t 1
 (Y  Yˆ ) 2
 (Y 
t 1
ˆ
a  ˆ )2
bX
t 1
t

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 12
Ordinary Least Squares (OLS)
Estimation Procedure
n

(X t  X )(Yt  Y )
bˆ  t 1 ˆ
â  Y  bX
n

 t
( X
t 1
 X ) 2

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 13
Ordinary Least Squares (OLS)
Estimation Example
Time Xt Yt Xt  X Yt  Y ( X t  X )(Yt  Y ) ( X t  X )2
1 10 44 -2 -6 12 4
2 9 40 -3 -10 30 9
3 11 42 -1 -8 8 1
4 12 46 0 -4 0 0
5 11 48 -1 -2 2 1
6 12 52 0 2 0 0
7 13 54 1 4 4 1
8 13 58 1 8 8 1
9 14 56 2 6 12 4
10 15 60 3 10 30 9
120 500 106 30
n n n
n  10  X t  120
t 1
 Yt  500
t 1
(X
t 1
t  X ) 2  30 bˆ 
106
30
 3.533

n n n
X t 120 Yt 500
X    12 Y    50 (X t  X )(Yt  Y )  106 aˆ  50  (3.533)(12)  7.60
t 1 n 10 t 1 n 10 t 1
Prepared by Robert F. Brooker, Ph.D.
Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 14
Ordinary Least Squares (OLS)
Estimation Example
n
X t 120
n  10 X    12
t 1 n 10
n
Yt 500
Y 
n n

X
t 1
t  120 Y
t 1
t  500
t 1 n

10
 50

n
106
(X
t 1
t  X )  302 ˆ
b
30
 3.533

(X
t by
t  X )(Yt  Y )  106
1 Robert F. Brooker, Ph.D.
aˆ  50  (3.533)(12)  7.60
Prepared
Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 15
Interpretation of the estimated regression line
ˆ a
Y ˆ
ˆ  bX
t t

• Ŷ t = 7.6 + 3.53Xt
• With zero advertising expenditure, the expected
sales revenue of the firm is $7.6 million.
• With X1 = $10 million, Ŷ 1 = 7.6 + 3.53 (10) = $42.9
million.
• With X10 = $15 million, Ŷ 10 = 7.6 + 3.53 (15) =
$60.55 million.
• Plotting these two points and joining them by a
straight line we obtain the regression line.

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 16
Interpretation of the estimated regression line

• The regression line can be used to estimate the


firm’s sales revenue with advertising expenditure
of $8 million or $16 million.
• But caution should be exercised in using the
regression line to estimate the sales revenue of
the firm from advertising expenditures very
different from those used in the estimation of the
regression line itself.
• Not much confidence can be usually attached to
the value of the estimated constant coefficient.
Therefore, we concentrate on the marginal effect
on Y from each unit change in X.
Prepared by Robert F. Brooker, Ph.D.
Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 17
Tests of Significance
• Using different samples could have yielded
different values for the estimated b.
• The greater is the dispersion of the estimated
values of b, the smaller is the confidence that we
have in our single estimated value of b.
• Therefore, we have to test the hypothesis that b
is statistically significant.
• To do this, we have first to calculate the standard
error of the estimated b.

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 18
Tests of Significance
Standard Error of the Slope Estimate
n is the number of observations, and k is the number of
estimated coefficients. (n-k) = df

sbˆ 
 (Yt  Y )
ˆ 2


e 2
t

( n  k ) ( X t  X ) 2
(n  k ) ( X t  X )2

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 19
Tests of Significance
Example Calculation
Time Xt Yt Yˆt et  Yt  Yˆt et2  (Yt  Yˆt )2 ( X t  X )2
1 10 44 42.90 1.10 1.2100 4
2 9 40 39.37 0.63 0.3969 9
3 11 42 46.43 -4.43 19.6249 1
4 12 46 49.96 -3.96 15.6816 0
5 11 48 46.43 1.57 2.4649 1
6 12 52 49.96 2.04 4.1616 0
7 13 54 53.49 0.51 0.2601 1
8 13 58 53.49 4.51 20.3401 1
9 14 56 57.02 -1.02 1.0404 4
10 15 60 60.55 -0.55 0.3025 9
65.4830 30

n n n  (Y  Yˆ ) 2
65.4830
 e   (Yt  Yˆt )2  65.4830 (X sbˆ    0.52
t
2
 X )  30
2

t 1
t
t 1 t 1
t ( n  k ) ( X  X )
t
2
(10  2)(30)
Prepared by Robert F. Brooker, Ph.D.
Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 20
Tests of Significance
Example Calculation
n n

 t  t t  65.4830
e 2

t 1
 (Y  Yˆ ) 2

t 1
n

 t
( X
t 1
 X ) 2
 30

sbˆ 
 (Yt  Y )

ˆ
65.4830
2

 0.52
( n  k ) ( X t  X ) 2
(10  2)(30)
Prepared by Robert F. Brooker, Ph.D.
Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 21
Tests of Significance
Calculation of the t Statistic
bˆ 3.53
t   6.79
sbˆ 0.52
The higher this ratio, the more confident we are that
the true but unknown value of b that we are seeking
is not equal to zero. (i.e. there is a significant
relationship between sales revenue and Ad).
Degrees of Freedom = (n-k) = (10-2) = 8

Critical Value at 5% level =2.306


Prepared by Robert F. Brooker, Ph.D.
Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 22
Confidence Intervals
• Find a confidence interval such that 95% level
of confidence in the following manner:
• Estimated b ± 2.306(Sb^), i.e.
• 3.53 ± 2.306(0.52) = 3.53 ± 1.2
• Meaning that we are 95% confident that the
true value of b lies between 2.33 and 4.73.

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 23
Tests of Significance
Decomposition of Sum of Squares

Total Variation = Explained Variation + Unexplained Variation

 (Yt  Y )   (Y  Y )   (Yt  Yt )
ˆ 2 2 ˆ 2

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 24
Tests of Significance
Decomposition of Sum of Squares

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 25
Tests of Significance

Coefficient of Determination

R2 
Explained Variation

 (Yˆ  Y ) 2

TotalVariation  t
(Y  Y ) 2

373.84
R 
2
 0.85
440.00

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 26
Coefficient of Correlation

• It measures the degree of association or


covariation that exists between variables X and Y.
• It does not imply any causality or dependence.
• This means that variables X and Y vary together
(for example 92%) of the time.
• It usually ranges between -1 and 1.
• The sign of r is always the same as the sign of the
estimated slope coefficient b^.

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 27
Tests of Significance

Coefficient of Correlation

r  R2 withthe sign of bˆ

1  r  1

r  0.85  0.92

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 28
Multiple Regression Analysis

Model: Y  a  b1 X1  b2 X 2   bk ' X k '

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 29
Assumptions of Multiple Regression Model

• In addition to the assumptions made for


simple regressions analysis we have another
two assumptions:
• (1) the number of independent variables in
the regression be smaller than the number of
observations, and
• (2) there is no perfect linear correlation
among the independent variables
Prepared by Robert F. Brooker, Ph.D.
Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 30
Multiple Regression Analysis
Adjusted Coefficient of DeterminationR2: It is
used when we have more than one
explanatory variable. Remember that
adding more variables reduces the df
and that R2 would be higher.
(n  1)
R  1  (1  R )
2 2

(n  k )

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 31
Multiple Regression Analysis
• Notice that bhat is different now which means
the omission of an important explanatory
variable from the simple regression gives
biased results for the estimated slope
coefficient.
• Therefore, include all important independent
variables in your model.

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 32
Multiple Regression Analysis
Analysis of Variance and F Statistic: It tests the overall
explanatory power of the entire regression, i.e. it tests
the hypothesis that the variation in the Xs explain a
significant proportion of the variation in Y.

Explained Variation /(k  1)


F
Unexplained Variation /(n  k )

R 2 /(k  1)
F
(1  R 2 ) /(n  k )

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 33
Point and Interval Estimates
• Standard Error (SE) of the regression: it is nothing
else than the SE (but not deviation) of the
dependent variable Y from the regression line. In
our example, SE = 2.095
• The smaller SE the better fit of the line.
• If X1 = 12, and X2 = 5, we can find the point
estimate and the interval estimate.
• Point estimate Y^ = 17.944 + 1.915 (12) + 1.873
(5) = 50.289
• At the 95% confidence interval estimate of Y is
given by 50 ± 2(SE), i.e. the true value of Y will lie
between 46.1 and 54.48.
Prepared by Robert F. Brooker, Ph.D.
Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 34
Steps in Demand Estimation
• Model Specification: Identify Variables
• Collect Data
• Specify Functional Form
• Estimate Function
• Test the Results

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 35
Functional Form Specifications

Linear Function:

QX  a0  a1PX  a2 I  a3 N  a4 PY  e

Power Function: Estimation Format:

QX  a( PXb1 )( PYb2 ) ln QX  ln a  b1 ln PX  b2 ln PY

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 36
Prediction Using Regression Equations

• We can predict the range within which the


value of Y can be at a specific X value because
the value predicted by the regression
equation will be subject to error. That is why
that actual values might not be on the
estimated regression line.
• We have to calculate the standard error of the
estimate which is a measure of the probable
error in the predicted values.
Prepared by Robert F. Brooker, Ph.D.
Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 37
Prediction Using Regression Equations

• Se = √ ([(Yi - Y)2 - b^  (Xi – X) (Yi - Y)] /


n- k – 1).
• The predicted value of Y, (i.e. Ŷ) is called a
point estimate to distinguish it from the
confidence interval estimate.
• The 95% confidence interval is given by Ŷ ± tn-
k-1Se.
• This will give the range within which the
actual value of Y would fall with 95%
confidence interval.
Prepared by Robert F. Brooker, Ph.D.
Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 38
Prediction Using Regression Equations

• Assume that X = 22 units.


• From the regression equation we find Ŷ = $355.7.
• Se = 27.14
• Thus a 95% confidence interval estimate for the cost
of producing 22 units is Ŷ ± tn-k-1Se = 355.7 ± 2.571
(27.14) = $285.92 to $425.48.
• It is only 5% of the time will the actual cost be
higher than $ 425.48 or lower than $285.92.

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 39
Multiple Regression
• Multiple regression is the estimation of the parameters
of an equation with more than one independent
variable.
• It is the same as the linear regression. However, it is
more complicated.
• Multiple regressions therefore use computers and
software programs to handle them.
• They are widely used in economics. If we tend to omit
certain variables the estimation of the relationship
between Y and X might be incorrect.
• For interpretation purposes, we must be careful: each
estimated parameter measures the impact of one
variable on the DV, holding constant the influence of
other variables.

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 40
Multiple Regression
• In this kind of regression, the regression
coefficient measures the net or partial effect
of each IV, while holding the effect of the
other IVs constant.
• That is another reason why multiple
regressions are preferred to simple
regressions.

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 41
Demand Estimation
• Four steps in regression analysis:
• (1) to formulate a model based on economic
theory,
• (2) to collect data,
• (3) to choose a functional form, and
• (4) to estimate and interpret the results.
• Economic theory can be used to evaluate the
signs and magnitudes of estimated
coefficients.
Prepared by Robert F. Brooker, Ph.D.
Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 42
Demand Estimation
• Prior knowledge based on economic theory can be
used to assess the empirical results of regression
analysis.
• If the signs of the estimated coefficients are
inconsistent with the predictions of economic
theory, then the process should be reexamined: the
demand equation might be (i) incorrectly
formulated, or (ii) there were errors in the data
collection and entry, or (iii) it might be that relevant
variables were omitted, or (iv) that there is
something unique about the product.

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 43
Data Collection
• Data must be obtained from surveys, market
experiments, or existing historical records of
the firm or government publications.
• It can be time series or cross-section data.
• Time series (or period-by-period
observations) require at least 15 observations
on each variable.
• Cross-section data are based on a number of
markets at a single point in time.

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 44
Choice of Functional Form
• Estimation using regression analysis requires the
choice of a specific functional form for the
equation.
• ① Linear regression equation such as
Qd = B + apP + aII +aOPO + aTT
• It has several advantages: 1- no
transformations of the data are necessary, 2-
the coefficients have a simple interpretation,
3- the estimated changes are constant for
each IV and unaffected by values of the other
variables.
• It is possible to calculate elasticities based on
the estimated coefficients.
Prepared by Robert F. Brooker, Ph.D.
Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 45
Choice of Functional Form
• For example, since Ep = (dQd/dP) (P/Qd),
• And since (dQd/dP) = aP, thus EP can be calculated by
multiplying aP by P/Qd.
• ② Multiplicative functional form, such that Qd =
BPaPIaIPaOoTaT
• In this form the equation cannot be estimated using
OLS because it is not linear. In this case we have to
transform the equation into a logarithm of the both
sides and then we sum the logarithms because the
logarithm of a product is just the sum of the
logarithms.
Prepared by Robert F. Brooker, Ph.D.
Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 46
Choice of Functional Form
• log(Qd) = log(BPaPIaIPaOoTaT (
• log(Qd) = log(B) +log(PaP) + log(aI) + log(PaOo) + log(TaT
(which can be further simplified into:
• logQd = logB +aP log P +aI log I + aO log PO + aT log T
• The coefficients in such estimations are nothing but
the elasticities where they are constant and unaffected
by changes in the IVs.
• Another feature of the multiplicative form is that the
change in Qd is determined not only by the associated
variable, but also by the values of other IVs. That is
why such estimations may be more realistic.
• The choice of the functional form depends on the
underlying theory.

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 47
Estimation & Interpretation of Results
Variable
Constant P I Po
Estimated 50.7836 -4.9892 0.0034 -1.2801
coefficient
Standard 10.2189 1.3458 0.0045 0.5890
error
t-statistic (4.97) (-3.71) (0.76) (-2.17)
Number of observations = 182, R2 = 0.6837
This for Qd = B + aPP + aII + aOPO
Prepared by Robert F. Brooker, Ph.D.
Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 48
Estimation & Interpretation of Results
• Qd = 50.7836 if all other variables are zero. However, it has
little economic meaning in most demand functions.
• The other coefficients estimate the change in Qd per $
change in the associated IV when other IVs are held
constant. This good is a normal good. It is complement to
the other good.
• The coefficients are consistent with the economic theory
and they also are significant (with exception of the income
coefficient) because the t-value is > 1.96 with 95%
confidence interval.
• Finally the R2 indicates the overall explanatory power of the
model. About two-thirds of the total variation in Qd is
explained by price, income, and the price of the other good.

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 49
Problems with Regression Analysis

• Omitted Variables: When regression results


are inconsistent with economic theory, the
omission of important variables could be the
cause.
• Identification Problem: Coefficients are
biased. For example: There is simultaneity
between the supply and demand equations.
Symptom: Independent variables are known
to be part of a system of equations. Solution:
Use as many independent variables as
possible
Prepared by Robert F. Brooker, Ph.D.
Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 50
Problems with Regression Analysis

• Multicollinearity: Sometimes independent variables aren’t


independent. They might be highly correlated.
• EXAMPLE: Q =Eggs
Q = a + b Pd + c Pg
where Pd is for a dozen
and Pg is for a gross.
• Symptoms: Coefficients are UNBIASED, large standard errors
and therefore, t-values are small, R2 is high.
• Solutions:
– Drop a variable.
– Do nothing if forecasting

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 51
Problems in Regression Analysis
• Multicollinearity: Two or more explanatory variables
are highly correlated. Example expenditure on quality
control and advertising expenditure can be so if the
firm keeps the level of the first as a % of the second.
• Result in exaggerated SE and low t values for both
variables.
• Can be overcome by (1) extending the sample size,
• (2) using a priori information like b2 = 2 b1 from
previous studies,
• (3) transforming the functional relationship, and or
• (4) dropping one of the highly collinear variables.
Prepared by Robert F. Brooker, Ph.D.
Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 52
Autocorrelation or serial correlation
• Consecutive error terms are correlated.
• Result in SE biased downward, t values are
exaggerated. Significance tests will be
unreliable.
• Remedy: include time as an additional
explanatory variable, or include an
important missing variable, or re-estimate
the regression in nonlinear form.

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 53
Serial Correlation
• Problem:
– Coefficients are unbiased
– but t-values are unreliable
• Symptoms:
– look at a scatter of the error terms to see if there
is a pattern, or
– see if Durbin Watson statistic is far from 2.
• Solution:
– Find more data
– Take first differences of data: Q = a + b•P
Prepared by Robert F. Brooker, Ph.D.
Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 54
Durbin-Watson Statistic: Test for
Autocorrelation
n

 (e t  et 1 ) 2
d  t 2
n

e
t 1
2
t

The value of d ranges between 0 & 4.


If d=2, autocorrelation is absent.
If the value is greater than dU in the table, the autocorrelation is
absent.
If the value is lower than dL in the table, the autocorrelation is
present.
If the value falls between the critical value of dL and dU in the table,
the test is inconclusive.
Prepared by Robert F. Brooker, Ph.D.
Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 55
Heteroscedasticity
• Problem:
– Coefficients are unbiased
– t-values are unreliable
• Symptoms:
– different variances for different sub-samples
– scatter of error terms shows increasing or
decreasing dispersion
• Solution:
– Transform data, e.g., logs
– Take averages of each sub-sample: weighted least
squares
Prepared by Robert F. Brooker, Ph.D.
Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 56
Heteroskedasticity
• It is that the variance of error term is not
independent of the X variable, i.e. error term may
rise or fall with the size of an independent variable.
• This result in biased SE and incorrect statistical
tests.
• Can be overcome by using the log of the
explanatory variable that leads to the problem or by
running a weighted least-square regression by
dividing the dependent and all independent
variables by the responsible variable, then run.

Prepared by Robert F. Brooker, Ph.D.


Copyright ©2004 by South-Western, a
division of Thomson Learning. All
rights reserved. Slide 57

S-ar putea să vă placă și