Christopher Dougherty EC220 - Introduction to econometrics (chapter 11) Slideshow: assumption c.7 Original citation: Dougherty, C. (2012) EC220 - Introduction.
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Christopher Dougherty
EC220 - Introduction to econometrics (chapter 11)Slideshow: assumption c.7
Original citation:
Dougherty, C. (2012) EC220 - Introduction to econometrics (chapter 11). [Teaching Resource]
This version available at: http://learningresources.lse.ac.uk/137/
Available in LSE Learning Resources Online: May 2012
This work is licensed under a Creative Commons Attribution-ShareAlike 3.0 License. This license allows the user to remix, tweak, and build upon the work even for commercial purposes, as long as the user credits the author and licenses their new creations under the identical terms. http://creativecommons.org/licenses/by-sa/3.0/
Assumption C.7, like its counterpart Assumption B.7, is essential for both the unbiasedness and the consistency of OLS estimators.
ASSUMPTION C.7
ASSUMPTIONS FOR MODEL C
C.7 The disturbance term is distributed independently of the regressors
ut is distributed independently of Xjt' for all t' (including t) and j
(1) The disturbance term in any observation is distributed
independently of the values of the regressors in the same
observation, and
(2) The disturbance term in any observation is distributed
independently of the values of the
regressors in the other observations.
2
It is helpful to divide it into two parts, as shown above. Both parts are required for unbiasedness. However only the first part is required for consistency (as a necessary, but not sufficient, condition).
ASSUMPTION C.7
ASSUMPTIONS FOR MODEL C
C.7 The disturbance term is distributed independently of the regressors
ut is distributed independently of Xjt' for all t' (including t) and j
(1) The disturbance term in any observation is distributed
independently of the values of the regressors in the same
observation, and
(2) The disturbance term in any observation is distributed
independently of the values of the
regressors in the other observations.
3
For cross-sectional regressions, Part (2) is rarely an issue. Since the observations are generated randomly there is seldom any reason to suppose that the disturbance term in one observation is not independent of the values of the regressors in the other observations.
ASSUMPTION C.7
ASSUMPTIONS FOR MODEL C
C.7 The disturbance term is distributed independently of the regressors
ut is distributed independently of Xjt' for all t' (including t) and j
(1) The disturbance term in any observation is distributed
independently of the values of the regressors in the same
observation, and
(2) The disturbance term in any observation is distributed
independently of the values of the
regressors in the other observations.
4
Hence unbiasedness really depended on part (1). Of course, this might fail, as we saw with measurement errors in the regressors and with simultaneous equations estimation.
ASSUMPTION C.7
ASSUMPTIONS FOR MODEL C
C.7 The disturbance term is distributed independently of the regressors
ut is distributed independently of Xjt' for all t' (including t) and j
(1) The disturbance term in any observation is distributed
independently of the values of the regressors in the same
observation, and
(2) The disturbance term in any observation is distributed
independently of the values of the
regressors in the other observations.
5
With time series regression, part (2) becomes a major concern. To see why, we will review the proof of the unbiasedness of the OLS estimator of the slope coefficient in a simple regression model.
ASSUMPTION C.7
ASSUMPTIONS FOR MODEL C
C.7 The disturbance term is distributed independently of the regressors
ut is distributed independently of Xjt' for all t' (including t) and j
(1) The disturbance term in any observation is distributed
independently of the values of the regressors in the same
observation, and
(2) The disturbance term in any observation is distributed
independently of the values of the
regressors in the other observations.
6
The slope coefficient may be written as shown above.
22
2
22
22121
2OLS2
XX
uuXX
XX
uuXXXX
XX
uXuXXX
XX
YYXXb
i
ii
i
iii
i
iii
i
ii
ASSUMPTION C.7
7
We substitute for Y from the true model.
22
2
22
22121
2OLS2
XX
uuXX
XX
uuXXXX
XX
uXuXXX
XX
YYXXb
i
ii
i
iii
i
iii
i
ii
ASSUMPTION C.7
8
The 1 terms in the second factor in the numerator cancel each other. Rearranging what is left, we obtain the third line.
22
2
22
22121
2OLS2
XX
uuXX
XX
uuXXXX
XX
uXuXXX
XX
YYXXb
i
ii
i
iii
i
iii
i
ii
ASSUMPTION C.7
9
The first term in the numerator, when divided by the denominator, reduces to 2. Hence as usual we have decomposed the slope coefficient into the true value and an error term.
22
2
22
22121
2OLS2
XX
uuXX
XX
uuXXXX
XX
uXuXXX
XX
YYXXb
i
ii
i
iii
i
iii
i
ii
ASSUMPTION C.7
10
The error term can be decomposed as shown.
22
222
222
22OLS2
XX
uXX
XX
XXu
XX
uXX
XX
uXX
XX
uXX
XX
uuXXb
i
ii
i
i
i
ii
i
i
i
ii
i
ii
ASSUMPTION C.7
11
u is a common factor in the second component of the error term and so can be brought out of it as shown.
22
222
222
22OLS2
XX
uXX
XX
XXu
XX
uXX
XX
uXX
XX
uXX
XX
uuXXb
i
ii
i
i
i
ii
i
i
i
ii
i
ii
–
ASSUMPTION C.7
12
It can then be seen that the numerator of the second component of the error term is zero.
22
222
222
22OLS2
XX
uXX
XX
XXu
XX
uXX
XX
uXX
XX
uXX
XX
uuXXb
i
ii
i
i
i
ii
i
i
i
ii
i
ii
0
XnXn
XnXXX ii
ASSUMPTION C.7
13
We are thus able to show that the OLS slope coefficient can be decomposed into the true value and an error term that is a weighted sum of the values of the disturbance term in the observations, with weights ai defined as shown.
iiuab 2OLS2
22
OLS2
XX
uXXb
i
ii
2XX
XXa
i
ii
ASSUMPTION C.7
14
iiuab 2OLS2
22
OLS2
XX
uXXb
i
ii
2XX
XXa
i
ii
22
2
22OLS2
0
i
ii
iiii
aE
uEaE
uaEuaEbE
Now we will take expectations. The expectation of the right side of the equation is the sum of the expectations of the individual terms.
iinnnnii uaEuaEuaEuauaEuaE ...... 1111
ASSUMPTION C.7
15
iiuab 2OLS2
22
OLS2
XX
uXXb
i
ii
2XX
XXa
i
ii
22
2
22OLS2
0
i
ii
iiii
aE
uEaE
uaEuaEbE
If the ui are distributed independently of the ai, we can decompose the E(aiui) terms as shown.
ASSUMPTION C.7
16
Unbiasedness then follows from the assumption that the expectation of ui is zero.
iiuab 2OLS2
22
OLS2
XX
uXXb
i
ii
2XX
XXa
i
ii
22
2
22OLS2
0
i
ii
iiii
aE
uEaE
uaEuaEbE
ASSUMPTION C.7
17
The crucial step is the previous one, which requires ui to be distributed independently of ai. ai is a function of all of the X values in the sample, not just Xi. So Part (1) of Assumption C.7, that ui is distributed independently of Xi, is not enough.
iiuab 2OLS2
22
OLS2
XX
uXXb
i
ii
2XX
XXa
i
ii
22
2
2OLS2
0
i
ii
ii
aE
uEaE
uaEbE
ASSUMPTION C.7
18
We also need Part (2), that ui is distributed independently of Xj, for all j.
iiuab 2OLS2
22
OLS2
XX
uXXb
i
ii
2XX
XXa
i
ii
22
2
2OLS2
0
i
ii
ii
aE
uEaE
uaEbE
ASSUMPTION C.7
19
iiuab 2OLS2
22
OLS2
XX
uXXb
i
ii
2XX
XXa
i
ii
22
2
2OLS2
0
i
ii
ii
aE
uEaE
uaEbE
In regressions with cross-sectional data this is usually not a problem.
ASSUMPTION C.7
20
If, for example, we are relating the logarithm of earnings to schooling using a sample of individuals, it is reasonable to suppose that the disturbance term affecting individual I will be unrelated to the schooling of any other individual.
iiuab 2OLS2
22
OLS2
XX
uXXb
i
ii
2XX
XXa
i
ii
22
2
2OLS2
0
i
ii
ii
aE
uEaE
uaEbE
CROSS-SECTIONAL DATA:
LGEARNi = 1 + 2Si + ui
LGEARNj = 1 + 2Sj + uj
Reasonable to assume uj and Si independent (i ≠ j).
The main issue is whether ui is independent of Si.
ASSUMPTION C.7
21
iiuab 2OLS2
22
OLS2
XX
uXXb
i
ii
2XX
XXa
i
ii
22
2
2OLS2
0
i
ii
ii
aE
uEaE
uaEbE
Assuming this, the independence of ui and ai then depends only on the independence of ui and Si.
ASSUMPTION C.7
CROSS-SECTIONAL DATA:
LGEARNi = 1 + 2Si + ui
LGEARNj = 1 + 2Sj + uj
Reasonable to assume uj and Si independent (i ≠ j).
The main issue is whether ui is independent of Si.
TIME SERIES DATA:
Yt = 1 + 2Yt–1 + ut
Yt+1 = 1 + 2Yt + ut+1
The disturbance term ut is
automatically correlated with
the explanatory variable Yt in
the next observation.
22
iiuab 2OLS2
22
OLS2
XX
uXXb
i
ii
2XX
XXa
i
ii
22
2
2OLS2
0
i
ii
ii
aE
uEaE
uaEbE
However with time series data it is different. Suppose, for example, that you have a model with a lagged dependent variable as a regressor. Here we have a very simple model where the only regressor is the lagged dependent variable.
ASSUMPTION C.7
TIME SERIES DATA:
Yt = 1 + 2Yt–1 + ut
Yt+1 = 1 + 2Yt + ut+1
The disturbance term ut is
automatically correlated with
the explanatory variable Yt in
the next observation.
23
iiuab 2OLS2
22
OLS2
XX
uXXb
i
ii
2XX
XXa
i
ii
22
2
2OLS2
0
i
ii
ii
aE
uEaE
uaEbE
We will suppose that Part (1) of Assumption C.7 is valid and that ut is distributed independently of Yt–1.
ASSUMPTION C.7
TIME SERIES DATA:
Yt = 1 + 2Yt–1 + ut
Yt+1 = 1 + 2Yt + ut+1
The disturbance term ut is
automatically correlated with
the explanatory variable Yt in
the next observation.
24
iiuab 2OLS2
22
OLS2
XX
uXXb
i
ii
2XX
XXa
i
ii
22
2
2OLS2
0
i
ii
ii
aE
uEaE
uaEbE
Even if Part (1) is valid, Part (2) must be invalid in this model.
ASSUMPTION C.7
25
iiuab 2OLS2
22
OLS2
XX
uXXb
i
ii
2XX
XXa
i
ii
22
2
2OLS2
0
i
ii
ii
aE
uEaE
uaEbE
TIME SERIES DATA:
Yt = 1 + 2Yt–1 + ut
Yt+1 = 1 + 2Yt + ut+1
The disturbance term ut is
automatically correlated with
the explanatory variable Yt in
the next observation.
ut is a determinant of Yt and Yt is the regressor in the next observation. Hence even if ut is uncorrelated with the explanatory variable Yt–1 in the observation for Yt, it will be correlated with the explanatory variable Yt in the observation for Yt+1.
ASSUMPTION C.7
26
iiuab 2OLS2
22
OLS2
XX
uXXb
i
ii
2XX
XXa
i
ii
22
2
2OLS2
0
i
ii
ii
aE
uEaE
uaEbE
As a consequence ui is not independent of ai and so we cannot write E(aiui) = E(ai)E(ui). It follows that the OLS slope coefficient will in general be biased.
X
ASSUMPTION C.7
TIME SERIES DATA:
Yt = 1 + 2Yt–1 + ut
Yt+1 = 1 + 2Yt + ut+1
The disturbance term ut is
automatically correlated with
the explanatory variable Yt in
the next observation.
27
iiuab 2OLS2
22
OLS2
XX
uXXb
i
ii
2XX
XXa
i
ii
22
2
2OLS2
0
i
ii
ii
aE
uEaE
uaEbE
X
ASSUMPTION C.7
We cannot obtain a closed-form analytical expression for the bias. However we can investigate it with Monte Carlo simulation.
TIME SERIES DATA:
Yt = 1 + 2Yt–1 + ut
Yt+1 = 1 + 2Yt + ut+1
The disturbance term ut is
automatically correlated with
the explanatory variable Yt in
the next observation.
28
We will start with the very simple model shown at the top of the slide. Y is determined only by its lagged value, with intercept 10 and slope coefficient 0.8.
Here are the estimates of the coefficients and their standard errors for 10 samples. We will start by looking at the distribution of the estimate of the slope coefficient. 8 of the estimates are below the true value and only 2 above.
This suggests that the estimator is downwards biased. However it is not conclusive proof because an 8–2 split will occur quite frequently even if the estimator is unbiased. (If you are good at binomial distributions, you will be able to show that it will occur 11% of the time.)
However the suspicion of a bias is reinforced by the fact that many of the estimates below the true value are much further from it than those above. The mean value of the estimates is 0.62.
The chart shows the distribution with 1 million samples. This settles the issue. The estimator is biased downwards.
0
0.5
1
1.5
2
2.5
-0.5 0 0.5 1 1.5
Yt = 10 + 0.8Yt–1 + ut
0.8
mean = 0.6233 (n = 20)
ASSUMPTION C.7
35
There is a further puzzle. If the disturbance terms are drawn randomly from a normal distribution, as was the case in this simulation, and the regression model assumptions are valid, the regression coefficients should also have normal distributions.
0
0.5
1
1.5
2
2.5
-0.5 0 0.5 1 1.5
Yt = 10 + 0.8Yt–1 + ut
0.8
mean = 0.6233 (n = 20)
ASSUMPTION C.7
36
However the distribution is not normal. It is negatively skewed.
0
0.5
1
1.5
2
2.5
-0.5 0 0.5 1 1.5
Yt = 10 + 0.8Yt–1 + ut
0.8
mean = 0.6233 (n = 20)
ASSUMPTION C.7
37
Nevertheless the estimator may be consistent, provided that certain conditions are satisfied.
0
0.5
1
1.5
2
2.5
-0.5 0 0.5 1 1.5
Yt = 10 + 0.8Yt–1 + ut
mean = 0.6233 (n = 20)
0.8
ASSUMPTION C.7
38
When we increase the sample size from 20 to 100, the bias is much smaller. (X has been assigned the values 1, …, 100. The distribution here and for all the following diagrams is for 1 million samples.)
Yt = 10 + 0.8Yt–1 + ut
mean = 0.6233 (n = 20)
0
1
2
3
4
5
6
7
-0.5 0 0.5 1 1.5
mean = 0.7650 (n = 100)
0.8
ASSUMPTION C.7
0
5
10
15
20
25
-0.5 0 0.5 1 1.5
39
If we increase the sample size to 1,000, the bias almost vanishes. (X has been assigned the values 1, …, 1,000.)
Yt = 10 + 0.8Yt–1 + ut
mean = 0.6233 (n = 20)
mean = 0.7650 (n = 100)
mean = 0.7966 (n = 1000)
0.8
ASSUMPTION C.7
0
0.5
1
1.5
2
-1 -0.5 0 0.5 1 1.5 2 2.5 3
40
Here is a slightly more realistic model with an explanatory variable Xt as well as the lagged dependent variable.
Yt = 10 + 0.5Xt + 0.8Yt–1 + ut
0.8
mean = 0.4979 (n = 20)
mean = 1.2553 (n = 20)
ASSUMPTION C.7
0
0.5
1
1.5
2
-1 -0.5 0 0.5 1 1.5 2 2.5 3
41
The estimate of the coefficient of Yt–1 is again biased downwards, more severely than before (black curve). The coefficient of Xt is biased upwards (red curve).
Yt = 10 + 0.5Xt + 0.8Yt–1 + ut
0.8
mean = 1.2553 (n = 20)
mean = 0.4979 (n = 20)
ASSUMPTION C.7
0
1
2
3
4
5
6
-1 -0.5 0 0.5 1 1.5 2 2.5 3
42
If we increase the sample size to 100, the coefficients are much less biased.
Yt = 10 + 0.5Xt + 0.8Yt–1 + ut
0.8
mean = 0.7441 (n = 100)
mean = 0.6398 (n = 100)
mean = 1.2553 (n = 20)
mean = 0.4979 (n = 20)
ASSUMPTION C.7
0
5
10
15
20
-1 -0.5 0 0.5 1 1.5 2 2.5 3
43
If we increase the sample size to 1,000, the bias almost disappears, as in the previous example.
Yt = 10 + 0.5Xt + 0.8Yt–1 + ut
0.8
mean = 0.7947 (n = 1000)
mean = 0.5132 (n = 1000)
ASSUMPTION C.7
44
In both of these examples the OLS estimators were consistent, despite being biased for finite samples. We will explain this for the first example. The slope coefficient can be decomposed as shown in the usual way.
ttt uYY 121
2
11
1122
11
11OLS2
tt
ttt
tt
ttt
YY
uuYY
YY
YYYYb
ASSUMPTION C.7
45
We will show that the plim of the error term is 0. As it stands, neither the numerator nor the denominator possess limits, so we cannot invoke the plim quotient rule.
ttt uYY 121
2
11
1122
11
11OLS2
tt
ttt
tt
ttt
YY
uuYY
YY
YYYYb
2
,
211
11
211
11
211
11
1
1
1 plim
1 plim
1
1
plim plim
t
tt
Y
uY
tt
ttt
tt
ttt
tt
ttt
YYn
uuYYn
YYn
uuYYn
YY
uuYY
ASSUMPTION C.7
46
We divide the numerator and the denominator by n.
ttt uYY 121
2
11
1122
11
11OLS2
tt
ttt
tt
ttt
YY
uuYY
YY
YYYYb
2
,
211
11
211
11
211
11
1
1
1 plim
1 plim
1
1
plim plim
t
tt
Y
uY
tt
ttt
tt
ttt
tt
ttt
YYn
uuYYn
YYn
uuYYn
YY
uuYY
ASSUMPTION C.7
47
Now we can invoke the plim quotient rule, because it can be shown that the plim of the numerator is the covariance of Yt–1 and ut and the plim of the denominator is the variance of Yt–1.
ttt uYY 121
2
11
1122
11
11OLS2
tt
ttt
tt
ttt
YY
uuYY
YY
YYYYb
2
,
211
11
211
11
211
11
1
1
1 plim
1 plim
1
1
plim plim
t
tt
Y
uY
tt
ttt
tt
ttt
tt
ttt
YYn
uuYYn
YYn
uuYYn
YY
uuYY
ASSUMPTION C.7
48
ttt uYY 121
2
11
1122
11
11OLS2
tt
ttt
tt
ttt
YY
uuYY
YY
YYYYb
2222
,2
OLS2
11
10
plim
tt
tt
YY
uYb
If Part (1) of Assumption C.7 is valid, the covariance between ut and Yt–1 is zero. In this model it is reasonable to suppose that Part(1) is valid because Yt–1 is determined before ut is generated.
ASSUMPTION C.7
49
ttt uYY 121
2
11
1122
11
11OLS2
tt
ttt
tt
ttt
YY
uuYY
YY
YYYYb
2222
,2
OLS2
11
10
plim
tt
tt
YY
uYb
Thus the plim of the slope coefficient is equal to the true value and the slope coefficient is consistent.
ASSUMPTION C.7
50
ttt uYY 121
2
11
1122
11
11OLS2
tt
ttt
tt
ttt
YY
uuYY
YY
YYYYb
2222
,2
OLS2
11
10
plim
tt
tt
YY
uYb
You will often see models with lagged dependent variables in the applied literature. Usually the problem discussed in this slideshow is ignored. This is acceptable if the sample size is large enough, but if the sample is small, there is a risk of serious bias.
ASSUMPTION C.7
51
ttt uXY 121
It is obvious that Part (2) of Assumption C.7 is invalid in models with lagged dependent variables. However it is often invalid in more general models. Consider the two-equation model shown above.
ASSUMPTION C.7
ttt vYX 121
52
It may be reasonable to suppose that ut is distributed independently of Xt–1 because it is generated randomly at time t, by which time Xt–1 has already been determined. Then Part (1) of Assumption C.7 is valid for the first equation. The same goes for the second equation.
ASSUMPTION C.7
ttt uXY 121
ttt vYX 121
53
However ut is a determinant of Yt, and hence of Xt+1. This means that ut is correlated with the X regressor in the first equation in the observations for Yt+2, Yt+4, ... etc. Again Part (2) of Assumption C.7 is violated and the OLS estimators will be biased.
ASSUMPTION C.7
ttt uXY 121
ttt vYX 121
21212 ttt uXY
1211 ttt vYX
Xt+1 ← Yt ← ut
54
Since interactions and lags are common in economic models using time series data, the problem of biased coefficients should be taken as the working hypothesis, the rule rather than the exception.
ASSUMPTION C.7
ttt uXY 121
ttt vYX 121
21212 ttt uXY
1211 ttt vYX
Xt+1 ← Yt ← ut
55
Fortunately Part (2) of Assumption C.7 is not required for consistency. Part (1) is a necessary condition. If it is violated, the regression coefficients will be inconsistent.
ASSUMPTION C.7
ttt uXY 121
ttt vYX 121
21212 ttt uXY
1211 ttt vYX
Xt+1 ← Yt ← ut
56
However, Part (1) is not a sufficient condition for consistency because it is possible that the regression estimators may not tend to finite limits as the sample size becomes large. This is a relatively technical issue that will be discussed in Chapter 13.
ASSUMPTION C.7
ttt uXY 121
ttt vYX 121
21212 ttt uXY
1211 ttt vYX
Xt+1 ← Yt ← ut
Copyright Christopher Dougherty 2011.
These slideshows may be downloaded by anyone, anywhere for personal use.
Subject to respect for copyright and, where appropriate, attribution, they may be
used as a resource for teaching an econometrics course. There is no need to
refer to the author.
The content of this slideshow comes from Section 11.5 of C. Dougherty,
Introduction to Econometrics, fourth edition 2011, Oxford University Press.
Additional (free) resources for both students and instructors may be
downloaded from the OUP Online Resource Centre
http://www.oup.com/uk/orc/bin/9780199567089/.
Individuals studying econometrics on their own and who feel that they might
benefit from participation in a formal course should consider the London School