Hetroskedasticity, SFU Notes
Hetroskedasticity, SFU Notes
Pure Heteroskedasticity
Tends to be seen in cross-sectional data more than time series data. Tends to be seen
when there is a lot of variation in the dependent variable.
, 1,2, … ,
If this assumption holds, the error term observations are all being drawn from the same
distribution (with mean zero and variance ).
Page 1 of 18
CHAPTER 10: HETEROSKEDASTICITY
If this assumption is not satisfied we have heteroskedasticity:
, 1,2, … ,
Most common form of heteroskedasticity is where the variance of the error term is related
to an exogenous variable :
_______________________________________________________________ Page 2 of 18
Short & Fast Understanding (SFU) Notes
Dr. Muhammad Atiq ur Rehman, Assistant Professor of Economics, Higher Education Department Punjab
& Adjunct Faculty of Economics and Management Sciences, COMSATS University Islamabad/ Lahore campus.
CHAPTER 10: HETEROSKEDASTICITY
Impure Heteroskedasticity
Page 3 of 18
CHAPTER 10: HETEROSKEDASTICITY
The Consequences of Heteroskedasticity
1. Pure heteroskedasticity does not cause bias in the regression coefficient estimates.
Page 4 of 18
CHAPTER 10: HETEROSKEDASTICITY
Testing for Heteroskedasticity
ln ln
Page 5 of 18
CHAPTER 10: HETEROSKEDASTICITY
This test does not assume a particular form for the heteroskedasticity.
Page 6 of 18
CHAPTER 10: HETEROSKEDASTICITY
Step 2: Estimate the auxiliary regression
Include all the explanatory variables, their squares and their cross-products.
Step 3: Test the overall significance of this equation using the test statistic which
follows a chi-square distribution with degrees of freedom equal to the number of
explanatory variables in the auxiliary regression. The is the sample size and the
is the from the auxiliary regression. Table B-8 gives critical values for the
chi-square distribution. If the value of your test statistic is greater than the critical
value, you reject the null hypothesis.
Page 7 of 18
CHAPTER 10: HETEROSKEDASTICITY
Remedies for Heteroskedasticity
where .
Page 8 of 18
CHAPTER 10: HETEROSKEDASTICITY
If we transform the equation by dividing both sides by we obtain a new
regression equation that is homoskedastic:
Adjust the standard errors of the estimated regression coefficients but not the
estimates themselves since they are still unbiased. These standard errors are called
White Heteroskedasticity-Consistent Standard Errors.
Page 9 of 18
CHAPTER 10: HETEROSKEDASTICITY
Example
Imagine a rock band hires us to evaluate their revenues from going on tour. Let’s suppose
we collect data for the band’s most recent tour in 50 US states.
The ticket price is always the same so it is not included in the model.
Page 10 of 18
CHAPTER 10: HETEROSKEDASTICITY
OLS regression:
Dependent Variable: REVENUES
Method: Least Squares
Sample: 1 50
Included observations: 50
Variable Coefficient Std. Error t-Statistic Prob.
ADVERTISING 3.147334 1.328637 2.368843 0.0223
STADIUM 34.66051 7.888484 4.393811 0.0001
CD 8.299202 6.049464 1.371891 0.1771
RADIO 300425.7 70633.17 4.253323 0.0001
WEEKEND 356003.5 84215.38 4.227298 0.0001
C 73215.34 70909.63 1.032516 0.3075
R-squared 0.958248 Mean dependent var 1753187.
Adjusted R-squared 0.953504 S.D. dependent var 1018119.
S.E. of regression 219536.3 Akaike info criterion 27.54859
Sum squared resid 2.12E+12 Schwarz criterion 27.77803
Log likelihood -682.7147 F-statistic 201.9707
Durbin-Watson stat 1.930626 Prob(F-statistic) 0.000000
Page 11 of 18
CHAPTER 10: HETEROSKEDASTICITY
Plot of residuals against the order the observations were recorded:
6000000
5000000
4000000
3000000
2000000
800000
1000000
400000 0
-400000
-800000
5 10 15 20 25 30 35 40 45 50
Using this graph, heteroskedasticity does not appear to be a problem. But thinking about
this problem more carefully, you realized that the 50 concert states vary significantly in
terms of size and that this may cause the error term variance to be proportional to each
state’s population (i.e. this is the Z).
So....you go ahead and attach each state’s population to the data set.
Page 12 of 18
CHAPTER 10: HETEROSKEDASTICITY
Let’s now order the data from low population states to high population states and plot the
same residuals again
6000000
5000000
4000000
3000000
800000
2000000
400000 1000000
0
0
-400000
-800000
5 10 15 20 25 30 35 40 45 50
Page 13 of 18
CHAPTER 10: HETEROSKEDASTICITY
Park Test
Page 14 of 18
CHAPTER 10: HETEROSKEDASTICITY
Weighted Least Squares
Page 15 of 18
CHAPTER 10: HETEROSKEDASTICITY
The White Test
White Heteroskedasticity Test:
F-statistic 6.036562 Probability 0.000007
Obs*R-squared 39.63334 Probability 0.003655
Test Equation:
Dependent Variable: RESID^2
Method: Least Squares
Date: 08/11/09 Time: 21:13
Sample: 1 50
Included observations: 50
Variable Coefficient Std. Error t-Statistic Prob.
C 1.89E+11 4.95E+10 3.820635 0.0006
ADVERTISING 5149460. 2482203. 2.074552 0.0467
ADVERTISING^2 35.36982 27.99974 1.263219 0.2162
ADVERTISING*STADIUM -599.8550 276.1516 -2.172195 0.0379
ADVERTISING*CD -178.4319 177.3528 -1.006084 0.3224
ADVERTISING*RADIO 2284099. 1453695. 1.571237 0.1266
ADVERTISING*WEEKEND 2436290. 1576309. 1.545566 0.1327
STADIUM -15030148 9079681. -1.655361 0.1083
STADIUM^2 1580.498 680.0162 2.324206 0.0271
STADIUM*CD 1731.925 904.7144 1.914333 0.0652
STADIUM*RADIO -12192611 8764398. -1.391152 0.1744
STADIUM*WEEKEND -21643059 9259410. -2.337412 0.0263
CD 10749212 6681980. 1.608687 0.1182
CD^2 274.1774 293.7963 0.933223 0.3582
CD*RADIO -19563691 7208794. -2.713865 0.0109
CD*WEEKEND -1179236. 5923193. -0.199088 0.8435
RADIO -2.95E+11 9.51E+10 -3.105109 0.0041
RADIO^2 1.42E+11 5.27E+10 2.690739 0.0115
RADIO*WEEKEND 1.06E+11 4.94E+10 2.135779 0.0410
WEEKEND 5.14E+10 6.91E+10 0.743333 0.4631
Adjusted R-squared 0.661356 S.D. dependent var 7.12E+10
S.E. of regression 4.14E+10 Akaike info criterion 52.02069
Log likelihood -1280.517 F-statistic 6.036562
Durbin-Watson stat 2.122886 Prob(F-statistic) 0.000007
Page 16 of 18
CHAPTER 10: HETEROSKEDASTICITY
White Heteroskedasticity-Consistent Standard Errors
Dependent Variable: REVENUES
Method: Least Squares
Sample: 1 50
Included observations: 50
White Heteroskedasticity-Consistent Standard Errors & Covariance
Variable Coefficient Std. Error t-Statistic Prob.
ADVERTISING 3.147334 1.948887 1.614939 0.1135
STADIUM 34.66051 10.02266 3.458216 0.0012
CD 8.299202 8.284117 1.001821 0.3219
RADIO 300425.7 80135.47 3.748973 0.0005
WEEKEND 356003.5 78030.18 4.562382 0.0000
C 73215.34 95545.02 0.766291 0.4476
R-squared 0.958248 Mean dependent var 1753187.
Adjusted R-squared 0.953504 S.D. dependent var 1018119.
S.E. of regression 219536.3 Akaike info criterion 27.54859
Sum squared resid 2.12E+12 Schwarz criterion 27.77803
Log likelihood -682.7147 F-statistic 201.9707
Durbin-Watson stat 2.233807 Prob(F-statistic) 0.000000
Page 17 of 18
CHAPTER 10: HETEROSKEDASTICITY
Redefining Variables
Page 18 of 18