Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
0% found this document useful (0 votes)
46 views

Econometrics (EM2008) Specification Error in The The K-Variable Model

The document outlines specification error in k-variable linear regression models. It discusses possible problems with the error term, regressors, and parameter assumptions. Tests of parameter constancy include the Chow forecast test and Ramsey RESET test. Tests of structural change analyze whether coefficients change over different time periods using an F-test to compare restricted and unrestricted models.
Copyright
© © All Rights Reserved
0% found this document useful (0 votes)
46 views

Econometrics (EM2008) Specification Error in The The K-Variable Model

The document outlines specification error in k-variable linear regression models. It discusses possible problems with the error term, regressors, and parameter assumptions. Tests of parameter constancy include the Chow forecast test and Ramsey RESET test. Tests of structural change analyze whether coefficients change over different time periods using an F-test to compare restricted and unrestricted models.
Copyright
© © All Rights Reserved
You are on page 1/ 31

Econometrics [EM2008]

Lecture 3
Specification error in the the k-variable model

Irene Mammi

irene.mammi@unive.it

Academic Year 2018/2019

1 / 35
outline

I specification error in k-variable linear regression


I specification error
I model evaluation
I tests of parameter constancy
I tests of structural change
I dummy variables

I References:
I Johnston, J. and J. DiNardo (1997), Econometrics Methods, 4th
Edition, McGraw-Hill, New York, Chapter 4.

2 / 35
specification error

I if any of the underlying assumptions are wrong, there is a


specification error
I the specification of the linear model centers on the error vector u and
the X matrix
I recall the assumptions:
y = Xβ + u
ui are iid (0, σ2 ) i = 1, . . . , n
or
ui are iid N (0, σ2 ) i = 1, . . . , n
E(Xit us ) = 0 for all i = 1, . . . , k and t, s = 1, . . . , n
X is nonstochastic with full column rank k

3 / 35
possible problems with u

1. ui are white noise but not normally distributed: this does not destroy
the BLUE property of OLS, but inference procedures are only
asymptotically valid
2. E(uu 0 ) = diag[σ12 · · · σn2 ]: the assumption of homoskedasticity is
violated The variance-covariance matrix for u is diagonal
3. E(ut ut −s ) 6= 0,(s 6= 0): the errors are pairwise correlated

4 / 35
possible problems with X

1. exclusion of relevant variables


2. inclusion of irrelevant variables
3. incorrect functional form: we may have an appropriate list of variables
but have embedded them in an incorrect functional form. E.g. we use
equation
Y = β 1 + β 2 X2 + β 3 X3 + u
when the correct specification would be

Y = β 1 + β 2 X2 + β 3 X3 + γ2 X22 + γ3 X32 + δ(X2 X3 ) + u

4. the X matrix has less than full column rank: this precludes estimation
of a unique b vector
5. nonzero correlations between the regressors and the errors
6. nonstationary variables

5 / 35
possible problems with β

I the implicit assumption for model y = X β + u is that the β vector is


constant over all actual or possible sample observations
I there may be structural breaks in coefficients
I the effects may be heterogeneous across individuals

6 / 35
tests of parameter constancy

the Chow forecast test

I if the parameter vector is constant, out-of-sample predictions will


have specified probabilities of lying within bounds calculated from the
sample data
I “large” prediction errors cast doubts on the constancy hypothesis
I instead of using all the sample observations for estimation, divide the
sample of n observations into n1 observations to be used for
estimation and n2 = n − n1 observations to be used for testing

7 / 35
tests of parameter constancy (cont.)

I the test of predictive accuracy, or Chow test, is as follows:

1. estimate the OLS vector from the n1 observations, obtaining

b 1 = (X 10 X 1 )−1 X 10 y 1

where X i , y i (i = 1, 2) indicate the partitioning of the data into


n1 , n2 observations
2. use b 1 to obtain a prediction of the y 2 vector, namely.

ŷ 2 = X 2 b 1

3. obtain the vector of prediction errors and analyze its sample


distribution under the null hypothesis of parameter constancy

8 / 35
tests of parameter constancy (cont.)

I the vector of prediction errors is

d = y 2 − ŷ 2 = y 2 − X 2 b 1

I if the equation y = X β + u, with E(uu 0 ) = σ2 I, holds for both sets


of data, the vector of prediction errors may be written as

d = y 2 − X 2 b 1 = u 2 − X 2 (b 1 − β )

I thus E(d ) = 0, and it may be shown that the variance-covariance


matrix for d is

var(d ) = E(d d 0 )
= σ2 I n2 + X 2 · var(b 1 ) · X 20
= σ2 [I n2 + X 2 (X 10 X 1 )−1 X 20 ]

9 / 35
tests of parameter constancy (cont.)

I there is an alternative way to derive the test of predictive accuracy


I suppose we allow for the possibility of a different coefficient vector in
the forecast period so that the complete model could be rewritten as

y 1 = X 1 β + u1

y 2 = X 2 α + u 2 = X 2 β + X 2 (α − β) + u 2 = X 2 β + γ + u 2
where γ = X 2 (α − β). If γ = 0, then α = β, and the coefficient
vector is constant over the estimation and forecast periods

I the resultant test statistic for γ = 0 is

(e 0∗ e ∗ − e 10 e 1 )/n2
F = ∼ F (n2 , n1 − k )
e 10 e 1 /(n1 − k )

11 / 35
tests of parameter constancy (cont.)

I the Chow test may thus be implemented as follows:

1. using n1 observations, regress y 1 on X 1 and obtain the RSS, e 10 e 1


2. fit the same regression using all (n1 + n2 ) observations and obtain
the restricted RSS, e 0∗ e ∗
3. substitute in the F statistic and reject the null of parameter
constancy if F exceeds the relevant critical value

15 / 35
tests of parameter constancy (cont.)
the Ramsey RESET test

I Ramsey argues that various specification errors (omitted variables,


incorrect functional form, correlation between X and u) give rise to a
nonzero u vector
I thus the null and alternative hypotheses are

H0 : u ∼ N (0, σ2 I )
H1 : u ∼ N (µ, σ2 I ) µ 6= 0

I the test of H0 is based on an augmented regression

y = Xβ + Zα + u

I the test for specification error is the α = 0


I Ramsey’s suggestion is that Z should contain powers of the predicted
values of the dependent variable

16 / 35
tests of parameter constancy (cont.)

I using the second, third and fourth power gives

Z = ŷ 2 ŷ 3 ŷ 4
 

0
where ŷ = Xb and ŷ 2 = ŷ 21 ŷ 22 · · · ŷ 2n etc. The first power,


ŷ, is not included since it is an exact linear combination of the


 
columns of X. Its inclusion would make the regressor matrix X Z
have less than full rank.

17 / 35
tests of structural change

test of one structural change

I three alternative (equivalent) ways to carry out the test


I let y i , X i (i = 1, 2) indicate the appropriate partitioning of the data
I the unrestricted model may be written as
    
y1 X1 0 β1
= + u u ∼ N (0, σ2 I )
y2 0 X 2 β2

where β1 and β2 are k-vectors of, say, peacetime and wartime


coefficients respectively
I the null hypothesis of no structural break is

H0 : β 1 = β 2

18 / 35
tests of structural change (cont.)

I first approach: straightforward application of the test for linear


 
restrictions (R β = r ): the null hypothesis defines R = I k −I k
and r = 0

19 / 35
tests of structural change (cont.)

I second approach: a test of linear restrictions may also be formulated


in terms of an unrestricted RSS and a restricted RSS: in this case, the
null hypothesis gives the restricted model as
   
y1 X1
= β+u
y2 X2

I denoting RSS from the restricted model as e 0∗ e ∗ , the test of the null
is given by
(e 0 e ∗ − e 0 e )/k
F = ∗0 ∼ F (k, n − 2k )
e e/(n − 2k )

20 / 35
tests of structural change (cont.)

I third approach: consider an alternative setup of the unrestricted


model,      
y1 X1 0 β1
= + +u
y2 X2 X2 β2 − β1
I now to test H0 simply test the joint significance of the last k
regressors

21 / 35
tests of structural change (cont.)
tests of slope coefficients

I frequently one does not wish to impose restrictions on the intercept


term
I partition the X matrices as

X 1∗ X 2∗
   
X1 = i1 X2 = i2

where i 1 , i 2 are n1 and n2 vectors of ones, and the X i∗ are matrices of


the k − 1 regressor variables
I the conformable partitioning of the β vectors is

β10 = α1 β1∗0 β20 = α2 β2∗0


   

I the null hypothesis is now

β1∗ = β2∗

22 / 35
tests of structural change (cont.)
I the unrestricted model is
 
 α1
X 1∗
  
y1 i 0 0   α2 
= 1

∗  ∗ + u
y2 0 i2 0 X2 β1
β2∗
I the restricted model is
 
 α
X 1∗  1 
  
y1 i 0
= 1 α2 + u
y2 0 i2 X 2∗
β∗
I the test can be based on the RSS from these two regressions
I an alternative formulation of the unrestricted model is
 
α1
0 X 1∗ 0 
   
y1
= 1
i  α2 −∗ α1  + u

∗ ∗ 
y2 i2 i2 X2 X2 β1 
β2 − β1∗

I the test of the last k − 1 regressors is a test of the null hypothesis


23 / 35
tests of structural change (cont.)
test of intercepts

I a meaningful test of differential intercepts assumes common


regression slopes
I the unrestricted model is again
 
 α
X 1∗  1 
  
y1 i 0
= 1 α2 + u
y2 0 i2 X 2∗
β∗

I the restricted model is

X 1∗
    
y1 i α
= 1 +u
y2 i2 X 2∗ β∗

I contrasting RSS between the two models provides a test of equality of


intercepts, given equal regression slopes

24 / 35
tests of structural change (cont.)

I the alternative setup of the unrestricted model is


 
α1
i 1 0 X 1∗ 
   
y1
= α2 − α1  + u
y2 i 2 i 2 X 2∗
β∗

I now a test of the significance of the second regressor tests the


conditional hypothesis that the intercepts are equal

25 / 35
tests of structural change (cont.)
summary
I there is a hierarchy of the three models, namely,

X 1∗
    
y1 i α
I: = 1 +u Common parameters
y2 i2 X 2∗ β∗
 
 α
X 1∗  1 
  
y1 i 0 Differential intercepts
II : = 1 α2 + u
y2 0 i2 X 2∗ common slope vectors
β∗
 
α1
X 1∗ 0 
   
y1 i 0  α2  + u Differential intercepts
= 1

III :
y2 0 i2 0 X ∗  β∗ 
2 1 differential slope vectors
β2∗

I application of OLS to each model will yield a residual sum of squares,


RSS, with associated degrees of freedom, respectively, of n − k,
n − k − 1, and n − 2k

26 / 35
tests of structural change (cont.)
I the test statistics for various hypotheses are as follows

H0 : α1 = α2 Test of differential intercepts

RSS1 − RSS2
F = ∼ F (1, n − k − 1)
RSS2 /(n − k − 1)

H0 : β1∗ = β2∗ Test of differential slope vectors


(RSS2 − RSS3 )/(k − 1)
F = ∼ F (k − 1, n − 2k )
RSS3 /(n − 2k )

H0 : β1 = β2 Test of differential parameters (intercept and slopes)

(RSS1 − RSS3 )/k


F = ∼ F (k, n − 2k )
RSS3 /(n − 2k )

27 / 35
dummy variables

examples

I the last n2 variables in the augmented matrix


      
y1 X1 0 β u
= + 1
y2 X 2 I n2 γ u2

take the form  


0
I n2
where the 0 matrix is of order n1 × n2 , and I n2 is the identity matrix
of order n2 : each n-vector column is a dummy variable. The effect of
the dummies is to exclude the last n2 observations from the
estimation of the β vector; the coefficients of the dummies are the
forecast errors for the last n2 points

28 / 35
dummy variables (cont.)

I a second type of dummy takes the form


 
0
d2 =
i2

as in the model of structural change

29 / 35
dummy variables (cont.)

Figure 1: Regressions with dummy variables

30 / 35
dummy variables (cont.)

seasonal dummies

I may want to allow for seasonal shift in a relation, e.g. specifying a


quarterly dummy such as

Qit = 1 if observation is in quarter i


= 0 otherwise

for i = 1, . . . , 4
I for the four quarters of each year these dummies are

Q1 Q2 Q3 Q4
1 0 0 0
0 1 0 0
0 0 1 0
0 0 0 1

31 / 35
dummy variables (cont.)
I the relationship may then be written as

Yt = α1 Q1t + · · · + α4 Q4t + x t0 β + ut

I an alternative specification is

Yt = α1 + γ2 Q2t + γ3 Q3t + γ4 Q4t + x t0 β + ut

I comparing coefficients of the dummies in the two equations gives

γ2 = α2 − α1 γ3 = α3 − α1 γ4 = α4 − α1

which shows that the γ’s measure differential intercepts, wrt α1


I the hypothesis of interest is usually

H0 : α1 = α2 = α3 = α4

I alternatively the null hypothesis may be expressed as

H0 : γ2 = γ3 = γ4 = 0
32 / 35
dummy variables (cont.)

qualitative dummies

I assume the interest is on an earning function as

Income = f ((sex, race, education, age)

I the first two variables are qualitative and may be represented by


dummy variables as

S1 = 1 if male
= 0 otherwise

and

S2 = 1 if female
= 0 otherwise

33 / 35
dummy variables (cont.)
I typical entries for the S and R dummies would be

S1 S2 R1 R2 R3
0 1 1 0 0
1 0 0 1 0

I as a different example, let E1 be the dummy for dropouts, and E2 , E3 ,


E4 be the dummies for the highest diploma awarded
I modeling income just as a function of educational level, we have

Y = α1 E1 + α2 E2 + α3 E3 + α4 E4 + u

I the expected level of income, conditional on educational level, is

E(Y |Ei ) = α1 i = 1, . . . , 4

I suppressing E1 , an alternative specification is

Y = α1 + γ2 E2 + γ3 E3 + γ4 E4 + u
34 / 35
dummy variables (cont.)
I the γs measure the marginal increment in expected income for a
diploma over the no-diploma level
I to measure the stepwise marginal increments, let the same dummies
to have value of one if a person has the relevant diploma, irrespective
of whether she has one or more diplomas, and E1 to be the pre-high
school diploma dummy
I the equation to be fitted is

Y = α1 + δ2 E2 + δ3 E3 + δ4 E4 + u

and the expected values are

E (Y |pre-HS diploma) = α1
E (Y |HS diploma) = α1 + δ2
E (Y |bachelor diploma) = α1 + δ2 + δ3
E (Y |graduate diploma) = α1 + δ2 + δ3 + δ4
I the δs provide estimates of the marginal increment from one level to
the next higher level
35 / 35

You might also like