Markov: Erratum: The Likelihood Ratio Nonstandard Conditions: Testing Switching Model of

JOURNAL OF APPLIED ECONOMETRICS, VOL.
11,195-198 (1996)
ERRATUM: THE LIKELIHOOD RATIO TEST UNDER

NONSTANDARD CONDITIONS: TESTING THE MARKOV
SWITCHING MODEL OF GNP
BRUCE E . HANSEN
Department of Economics, Boston College, Chestnut Hill, MA 02167-3806,USA
There was an error in Hansen (1992). I am very grateful to James Hamilton for pointing out the
error.
Equations (2) and (3) in the original read
where Q ( a )is a mean zero Gaussian process with covariance function

K(ai,a21 = E ( q ; ( a i ) ~ i ( a d ) .
While equation (2) is correct, (3) is not. Instead, the correct expression is
The reason is that the likelihood components qi(a)will be serially correlated for some values
of a. This will be the case even when the original data are iid, since the likelihood qi(a)is a
function of all data up to time i. It should be noted that this problem does not apply to the
testing methods of Hansen (1994), which involve application of empirical process theory to
specific likelihood scores which are serially uncorrelated.
This error implies that the method of calculating the asymptotic distribution in Section 3.2 is
incorrect. Instead, set d i ( a )= q j ( a ,8 ( a ) )and
i= 1 P-I LlsiGn-k I+t<i<n J

where wkM= 1 - 1 k 1 /(M+ 1) is the Bartlett kernel and M is a bandwidth number (selected to
grow to infinity slowly with sample size). Then a consistent estimate of
CCC 0883-7252/96/020195-4 Received June 1994

0 1996 by John Wiley & Sons, Ltd. Revised May I995
196 B.E.HANSEN
is given by
Sample draws from this process can be obtained by constructing
where [ ui]S:f"is a sample of random N(0,l) variables. The reader may verify that conditional
on the data, G * ( a ) is a mean zero Gaussian process with exact covariance function
c ( a l ,a2),and the latter is an asymptotic approximation to /?(a,, a2).
The theory does not give any particular guidance for choice of M. It therefore seems prudent
to calculate the tests for several choices to assess sensitivity.
The original paper reported estimates of c ( a l az) , and a* ( a ) effectively with M = 0.
Thus all test statistics and Monte Carlo evidence were presented with M = 0. All the numerical
work was recalculated for M = 1, ...,4. Other than the change discussed above, the methods
were essentially identical to those outlined in Hansen (1992). The corrected results are presented
in the following tables. It it interesting to note that the results are not very sensitive to M. None
of the conclusions drawn are affected. The category 'CPU hours' referred to the time required
for the programs to run on a 486/66 computer.
A GAUSS program which produces the empirical results reported here is available on request
from the author.
Table IU. Standardized LR statistics for Hamilton model
p-value
LR: M=O M= 1 M=2 M=3 M=4 CPU hours

Grid 1 1 -24 0.77 0.75 0.75 0.76 0.73 1.2
Grid 2 1.56 0.73 0.68 0.67 0-66 0.62 2.5
Grid 3 1.55 0.74 0.70 0-69 0.69 0.65 4.6
Table IV. Monte Carlo size and power, no autoregressivecomponent (percentages)

~~ -
Null Alternative
Nominal size: M 20 10 5 20 10 5
0 12 4 0 86 80 74
1 14 8 0 86 80 74
2 16 8 2 86 76 74
3 14 8 2 86 76 74
4 14 6 2 86 76 74
ERRATUM: MARKOV SWITCHING MODEL 197
Table VI. Standardized LR statistics for Markov switching intercept model
p-value
L R ~ M=O M=l M=2 M=3 M=4 CPU hours

Grid 1 2.20 0.29 0.26 0.30 0.27 0.28 0.3
Grid 2 1.95 0.49 0.5 1 0.50 0.46 0.48 0.6
Grid 3 2.09 0-44 0.42 0.43 0.41 0.43 1.1
Grid 4 2.2 1 0.4 1 0.38 0.38 0-37 0.38 2.6
Grid 5 2.23 0.39 0.38 0.36 0.34 0.34 5.2
Table VII. Monte Carlo size and power with AR components (percentages)
~ ~~ ~ ~~~
Null Alternative
Nominal size: M 20 10 5 20 10 5
14 10 52 40 30
14 10 52 40 30
14 10 52 38 28
16 10 52 38 28
16 10 52 36 26
26 18 14 44 36 24
30 20 16 44 40 24
28 20 16 44 40 22
28 22 16 44 38 24
30 22 16 44 33 20
Table X. Standardized LR statistics for switching parameters model
p-value
LR: M=O M=l M=2 M=3 M=4 CPU hours

Model A: ( p , &) vary between states, q = 1 - p
Grid 1 3.50 0.01 0.0 1 0.0 1 0.02 0.0 1 0.4

Grid 2 3.53 0.01 0.01 0.01 0.02 0.02 0.6
Grid 3 3.61 0.0 1 0.02 0.01 0.0 1 0.0 1 0.8
Grid 4 3.59 0.01 0.0 1 0.01 0.01 0.0 1 1.2
Model B: ( p , @J vary between states, q unconstrained
Grid 1 3.50 0.02 0.02 0.02 0.02 0.03 1.8

Grid 2 3.55 0.03 0.03 0.03 0.03 0.04 4.0
Grid 3 3.61 0.02 0.03 0.03 0.04 0.03 7.1
Grid 4 3.61 0.04 0.04 0.02 0.03 0.04 15.9
198 B. E. HANSEN
ACKNOWLEDGEMENTS
I thank two referees for helpful comments. Financial support from the National Science
Foundation and the Sloan Foundation is gratefully acknowledged.
REFERENCES
Hansen, B. E. (1992), ‘The likelihood ratio test under nonstandard conditions: Testing the Markov
switching model of GNP’, Journal of Applied Econornetrics, 7 , S61-82.
Hansen, B. E. (1994), ‘Inference when a nuisance parameter is not identified under the null hypothesis’,
Ecoriornetrica, forthcoming.

Markov: Erratum: The Likelihood Ratio Nonstandard Conditions: Testing Switching Model of

Uploaded by

Copyright:

Available Formats

Markov: Erratum: The Likelihood Ratio Nonstandard Conditions: Testing Switching Model of

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Markov: Erratum: The Likelihood Ratio Nonstandard Conditions: Testing Switching Model of

Uploaded by

Copyright:

Available Formats

JOURNAL OF APPLIED ECONOMETRICS, VOL.

ERRATUM: THE LIKELIHOOD RATIO TEST UNDER

where Q ( a )is a mean zero Gaussian process with covariance function

i= 1 P-I LlsiGn-k I+t<i<n J

CCC 0883-7252/96/020195-4 Received June 1994

Sample draws from this process can be obtained by constructing

Table IU. Standardized LR statistics for Hamilton model

LR: M=O M= 1 M=2 M=3 M=4 CPU hours

Table IV. Monte Carlo size and power, no autoregressivecomponent (percentages)

Table VI. Standardized LR statistics for Markov switching intercept model

L R ~ M=O M=l M=2 M=3 M=4 CPU hours

Table X. Standardized LR statistics for switching parameters model

LR: M=O M=l M=2 M=3 M=4 CPU hours

Grid 1 3.50 0.01 0.0 1 0.0 1 0.02 0.0 1 0.4

Model B: ( p , @J vary between states, q unconstrained

Grid 1 3.50 0.02 0.02 0.02 0.02 0.03 1.8

You might also like