nonstationarity, autocorrelation, and collinearity in · pdf file ·...
TRANSCRIPT
Nonstationarity, Autocorrelation, and Collinearity in Modeling Forest Products Marketsode g o est oducts a ets
Nianfu Song and Sun Joseph ChangSchool of Renewable Natural Resources
Louisiana State Universityy
ObjectivesObjectives
• Show the existence of nonstationarityShow the existence of nonstationarity• Consequences of nonstationarity and
problems in current studiesproblems in current studies• Estimation Methods with nonstationarity
d tdata.• Long-run and short-run• Re-explanation of some past research.
Estimation MethodsEstimation Methods• 2SLS, 3SLS
Adams and Haynes 1996;– Adams and Haynes, 1996;– Adams et al. 1986; – Baek and Yin, 2006;
Bernard et al 1997 (NL3SLS);– Bernard et al. 1997 (NL3SLS);– Latta and Adams, 2000 (NL3SLS);– Lewandrowski et al. 1994.
• OLS• OLS– Rockel and Buongiorno, 1982, Adams et al. 1992, Cardellichio,
1990
• ML– Rao et al. 2004; Myneni, 1994
• Kalman Filter Adams et al 1992• Kalman Filter Adams et al. 1992
Nonstationarity, Unit Root
• Stationarity is a basic requirement for Central Limit Theorem (time series used in lumber models are mostly nonstationary)Theorem (time series used in lumber models are mostly nonstationary).
• OLS with nonstationary {zt} is spurious when there is no cointegrationis no cointegration
• Weakly stationary: E[zt] and Var[zt] are constant, and Cov[zt, zt k] is finite and independent of tCov[ t, t-k] s e d depe de o
• If A(L)∆zt = εt, the time series process {zt} has a unit root. With unit root, {zt} is nonstationary. , { t} y
Softwood Lumber Prices are N iNonstationary
600
500
600
BF
200
250
ndex
300
400
rric
esin
$/M
B
150
mbe
rPri
ceI
100
200
Lum
ber
50
100
onth
lyLu
m
0
1990
1990
1991
1992
1993
1993
1994
1995
1996
1996
1997
1998
1999
1999
2000
2001
2002
2002
2003
2004
P1 P2 P3 P4
094
795
095
395
796
096
496
797
197
497
898
198
598
899
299
599
800
200
5
M
P1 P2 P3 P4 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2
Softwood Timber Price IndexSoftwood Timber Price IndexSoftwood timber price index
250
300
200
250
100
150
0
50
2 3 4 6 7 8 0 1 3 4 5 7 8 0 1 2 4 5
1982
1983
1984
1986
1987
1988
1990
1991
1993
1994
1995
1997
1998
2000
2001
2002
2004
2005
Housing Starts, Mortgage Rate and DPIHousing Starts, Mortgage Rate and DPI
9000) 20
7000
8000
9000
on o
f 200
0$)
1618
20
4000
5000
6000
0), D
PI(B
illio
10
1214
age
Rate
2000
3000
4000
g S
tart
s(10
00
4
68
Mor
ga
0
1000
71 74 77 80 83 86 89 92 95 98 01 04
Hou
sin g
0
24
Apr
-7
Apr
-7
Apr
-7
Apr
-8
Apr
-8
Apr
-8
Apr
-8
Apr
-9
Apr
-9
Apr
-9
Apr
-0
Apr
-0
Nonstationary Time Series Studies FoundNonstationary Time Series Studies Found
• The “law of one price” studies withThe law of one price studies with cointegration tests
--Jung and Doroodian (1994), Nanang (2001)g ( ), g ( )--Buongiorno and Uusivuori (1992), Alavalapati et al. (1997) and Hänninen (1998), Yin et al. (2002), Yin et al (2005)et al. (2005).
• ECM models (MLE)--Restricted ECM Toppinen (1998) for Finnish--Restricted ECM, Toppinen (1998) for Finnish sawlog--Baek and Yin (2006)
Cointegration
• When a linear combination of nonstationary variables yis stationary, these variables are cointegrated.
• Cointegration exists as long as market equilibrium iexists.
• OLS with nonstationary {zt} is “super consistent” when there is cointegrationwhen there is cointegration.
• Cointegration has been used in testing “the Law of One Price” by many in forestry literature and in at O e ce by y o es y e u e dleast one structural models for timber, and one for lumber.
Autoregressive RepresentationAutoregressive Representationqp
∑∑• Or
t1j
jtj0i
itit yxy εφβ ++= ∑∑=
−=
−
Or exists)L(wherex)L(By)L( 1
ttt−=+ ΓεΓ
zxyorzx)1(By)1(relationrunlongWith
+=+=−
αΓ tttttt zxyorzx)1(By)1( +=+= αΓ
Estimate the Autoregressive Model i h 2SLSwith 2SLS
• The coefficients of the autoregressiveThe coefficients of the autoregressive representation can be estimated with 2SLS or 3SLS2SLS or 3SLS.
• The invertability and cointegration of the model must be checkedmodel must be checked.
• Long-run relation can be obtained by t f i th ti t d t itransforming the estimated autoregressive representation of the cointegration.
Triangle RepresentationTriangle Representation
• The source of nonstationarity of endogenousThe source of nonstationarity of endogenous variables is the unit roots of the exogenous variables.
ttt zxy α +=
ttx εδ∆ +=
Estimate the Coingretion Relations i h OLS 2SLSwith OLS or 2SLS
• The OLS is supper consistentThe OLS is supper consistent. • 2SLS is consistent.
With t l ti th t i i• With autocorrelation the true variance is large. However the true variance some ti i t bt i d b fttimes is not obtained by software.
Estimate Structural ECM with MLEEstimate Structural ECM with MLE
• With restrictions an ECM can be estimatedWith restrictions an ECM can be estimated by MLE. However the cointegration test with no restriction must be applied first, ppand the restrictions have to be tested by a ki-square test.
• It is a pure statistical method without any knowledge prior to estimating the
i t ticointegration.• An example is Toppinen (1998).
Misspecificationp
• With cointegration DynamicWith cointegration, Dynamic autoregressive model (ECM equivalent) is the data generating function and thethe data generating function, and the models of differences (ARIMA) are misspecified (Engle and Granger 1987)misspecified (Engle and Granger, 1987).
• By this rule, at least one of the lumber structure models published recently isstructure models published recently is misspecified.
Invertible and Stationary C ffi i M iCoefficient Matrix
• The roots of the absolute value of the matrix Г(L)The roots of the absolute value of the matrix Г(L) equal to zero must be outside the unit circle– In the case of one lag, the coefficient of Y,t-1 or the ratio of the
ffi i f d h b l h d hcoefficient of Y,t-1 and Y,t have to be less than 1 and greater than -1.
Thi diti b i l t d b h ki th• This condition can be implemented by checking the eigenvalues of a matrix F (Hamilton, 1994, page 259) The absolute values of the eigenvalues must259). The absolute values of the eigenvalues must be smaller than unit to ensure the stationarity of the data generation system.of the data generation system.
Past StudiesPast Studies
• The cointegration condition was satisfiedThe cointegration condition was satisfied for models estimated with OLS, 2SLS or 3SLS because of the routine reports of the3SLS because of the routine reports of the DW statistics.
• Unknown for other methods• Unknown for other methods.• Some times the invertability condition is
t ti fi dnot satisfied.
Coefficient Matrix of a Recently Published Lumber Model
-0.35
1.91 1.78 1.70
0.25 -0.36
-1.61 -1.45 -1.39
0.55 -.49 -0.41
-0.16
-0 61-0.61
-0.30
-0.14 0.17 0.29
The Eigenvalues of the F MatrixThe Eigenvalues of the F Matrix
• 0.962813
• 0.508607• 0.508607• 0.026072• 0.026072• -0.060270.06027• -0.06027• -0.19521• -0.19521• -0.48036• -0 480360.48036
• -1.3205 (> 1, showing nonstationarity)
Similar problem in another paperSimilar problem in another paper
• With AR(1) when ρ>1 the expected variance• With AR(1), when ρ>1 the expected variance is infinite
• Values of ρ in the paper• Values of ρ in the paper0.80, 0.76, 0.89, 1.021 25 0 76 0 99 1 961.25, 0.76, 0.99, 1.960.44, 1.02, 1.03, 0.98
Implication of Frequencyp q y• Estimates with monthly prices approximately
equals that with annual prices in the long-run, q p g ,• But the short-run effect implied by the ECM is
different
Observations over TimeObservations over Time
Data
54
56
58
60
62
dex
3334353637
lum
e
44
46
48
50
52
54
Pric
e in
d
272829303132
Out
put V
o
445 6 7 8
year
27
A one time increase in price (small dots over time)A one time increase in price (small dots over time)
The observed outputs (large diamonds over time)
With NO LagsWith NO Lags
• Pink line: regressedPink line: regressed curve with OLS
• Diamonds: output Regression
37pagainst price.
33
34
35
36
37
olum
e29
30
31
32
48 50 52 54 56 58 60 62Vo
48 50 52 54 56 58 60 62
Price
Pink line: regressed curve with OLS
Diamonds: output against price with data on the left graph .
Short-Run EffectShort Run Effect
• The coefficients of differences of a ECM h th t
Regression
36
37
shows the current response.
32
33
34
35
Volu
me
29
30
31
48 50 52 54 56 58 60 62
Price
With Lag VariablesWith Lag Variables
RegressionRegression
36
37
33
34
35
lum
e
30
31
32Vol
29
30
48 50 52 54 56 58 60 62
Price
The Long-Run EffectThe Long Run EffectRegression
36
37
33
34
35
olum
e
30
31
32Vo
2948 50 52 54 56 58 60 62
Price
Lumber Demand ElasticityLumber Demand Elasticity
• The reported elasticities in forest literatureThe reported elasticities in forest literature are mostly short-run elasticity in one period of time (lags included). p ( g )
• or ones below their true numbers when the autocorrelation exists in the error terms.
• According to our recent research, the long-run elasticity could be -0.71 when the yoften referred -0.17 is transformed according to the original estimated model.
ConclusionConclusion
• The long-run elasticity may be greaterThe long run elasticity may be greater than what we thought.
• Some past models were misspecified (the• Some past models were misspecified (the impact of the misspecification is unknown)S ti t d t t ti• Some estimated system are nonstationary or unstable.
•
SuggestionsSuggestions
• With level data if the errors are wellWith level data, if the errors are well behaved, the unit roots of the time series can be ignoredcan be ignored.
• The stationarity of a estimated time series model with lags of endogenous variablesmodel with lags of endogenous variables or autoregressive errors should always be checkedchecked.