Professional Documents
Culture Documents
Panel Threshold Regression Models 2018
Panel Threshold Regression Models 2018
University of Orléans
May 2018
In this simple case, we have two "regimes" for the slope parameters, i.e.
β0 and β0 + β1 .
Remarks
0 h (qt ; θ ) 1
Sometimes, we have
0.9
0.8
)
0.7
transition function h(q,
0.6
0.5
0.4
0.3
=0.5
=1
0.2
=5
0.1
0
-6 -4 -2 0 2 4 6
threshold variable q
Panel data
Objectives
where the dependent variable yit is scalar, αi is a …xed e¤ect, the threshold
variable qit is scalar, the regressor xit is a k vector, and I(.) is the indicator
function and c is a threshold parameter.
Assumptions
The error εit is assumed to be i.i.d. with E (εit ) = 0 and V (εit ) = σ2ε .
or equivalently
0.9
0.8
transition function h(q,c)
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
-6 -4 -2 0 2 4 6
threshold variable q
At a given time t, two cross-section units i and j may have two di¤erent
slope parameters
∂yit ∂yjt
= β1 6 = = β2 if qit c and qjt > c.
∂xit ∂xjt
6 1
0.8
5
0.6
0.4
t/dx t
i
3
0.2
i
Threshold variables
Marginal effect dy
2 0
-0.2
1
-0.4
-0.6
-1
-0.8
-2 -1
0 10 20 30 40 50 0 10 20 30 40 50
Time Time
Estimation by NLS
The PTR model can be rewritten as:
xit I(qit c ) β1
xit (c ) = β =
(2K ,1 )
xit I(qit >c ) (2K ,1 )
β2
Estimation by NLS
In order to eliminate the individual e¤ects, we apply the Within
transformation
yit = β0 xit (c ) + εit
T
yit = yit yi yi = T 1
∑ yit
t =1
T
xit (c ) = xit (c ) x i (c ) x i (c ) = T 1
∑ xit (c )
t =1
T
εit = εit εi εi = T 1
∑ εit
t =1
Let us de…ne
0 1 0 1 0 1
yi ,1 x1,i ,1 (c )0 εi ,1
B y C B x1,i ,2 (c )0 C B ε C
yi = B i ,2 C
@ ... A Xi (c ) = B
@ ...
C
A εi = B i ,2 C
@ ... A
(T ,1 ) (T ,2K ) (T ,1 )
0
yit x1,it (c ) εit
0
1 0 1 0 1
y1 X1 (c ) ε1
B y C B X (c ) C B ε C
Y =B 2 C
@ ... A X (c ) = B 2
@ ...
C
A ε =B 2 C
@ ... A
(Tn,1 ) (Tn,2K ) (Tn,1 )
yn Xn (c ) εn
bε (c ) = Y X (c ) b
β (c )
SSR (c ) = bε (c )0 bε (c )
b
c = arg min SSR (c )
c 2Θ
Given b
c , we can compute the estimates for β as
b β1 (b
c)
β = β (b
c) =
β2 (b
c)
Inference
If one comes to test whether the threshold e¤ect is statically
signi…cant in the model with two regimes, the null hypothesis is
H0 : β 1 = β 2
with β = β1 = β2 .
Inference
The null hypothesis H0 : β1 = β2 can be tested by a standard test.
SSR0 SRR1 (b
c)
F1 = 2
b
σ
Inference
The main problem is that under the null, the threshold parameter c is
not identi…ed (nuisance parameter).
SSR2 (b
c1 , b
c2 ) SSR3 (b
c1 , b
c2 , b
c3 )
F3 = 2
b
σ
where SSR3 (bc1 , b
c2 , b
c3 ) denotes the residual sum of squares of the
model with four regimes and three threshold parameters.
Example
Candelon, Colletaz, Hurlin (2012) investigate the threshold e¤ects in the
productivity of infrastructure investment in developing countries.
(
ai + α1 kit + β1 hit + γ1 xit + εit if qit λ
yit =
ai + α2 kit + β2 hit + γ2 xit + εit if qit > λ.
where yit is the aggregate added value, kit is physical capital, hit is human
capital, xit is infrastructure stock.
Objectives
González, A., Teräsvirta, T., van Dijk, D., 2005. Panel smooth transition
regression model. Working Paper Series in Economics and Finance, vol. 604.
Transition function
0 g (qit ; γ, c ) 1
De…nition
Gonzalez, Teräsvirta and van Dijk (2005) consider a logistic transition
function
1
g (qit ; γ, c ) = , γ>0
1 + exp ( γ (qit c ))
0.9
0.8
)
0.7
transition function h(q,
0.6
0.5
0.4
0.3
=0.5
=1
0.2
=5
0.1
0
-6 -4 -2 0 2 4 6
threshold variable q
Remark 1
If γ ! ∞, the logistic function tends to an indicator function and the
PSTR corresponds to a PTR model
(
1 if qit c
lim g (qit ; γ, c ) = I(qit c ) = ,
γ!∞ 0 if qit < c
Remark 2
If γ ! 0, the logistic function tends to an indicator function and the
PSTR corresponds to a linear panel model
1
lim g (qit ; γ, c ) = 8qit
γ!∞ 2
β0 + β1
yit = αi + β0 xit + εit β=
2
β0 βit β0 + β1
Remark 3
Even there are two "extreme" regimes/values for the slope parameters, in
fact there is an in…nity of regimes (possible values) for the slope
parameters βit given the observed values for qit :
βit = β0 + β1 g (qit ; γ, c )
β0 βit β0 + β1
Marginal e¤ects
If the transition variable does not belong to the set of regressors, the slope
parameters become:
r
δyit
βit = = β0 + ∑ βj gj (qit ; γj , cj )
δxit j =1
Marginal e¤ects
If the transition variable does not belong to the set of regressors, the slope
parameters become:
r r δgj (qit ; γj , cj )
δyit
βit = = β0 + ∑ βj gj (qit ; γj , cj ) + ∑ βj qit
δxit j =1 j =1 δqit
Estimation
The estimation of the parameters of the PSTR model consists of
eliminating the individual e¤ects αi and then in applying NLS to the
transformed model
See Gonzàlez et al. (2005) or Colletaz and Hurlin (2006), for more
details.
González, A., Teräsvirta, T., van Dijk, D., 2005. Panel smooth transition
regression model. Working Paper Series in Economics and Finance, vol. 604.
Colletaz, G., Hurlin, C., 2006. Threshold e¤ects in the public capital
productivity: an international panel smooth transition approach. Document
de Recherche du Laboratoire d’Economie d’Orléans. 2006-1.
Inference
Gonzàlez et al. (2005) propose a testing procedure in order (i ) to test
the linearity against the PSTR model and (ii ) to determine the
number, r , of transition functions, i.e. the number of extreme regimes
which is equal to r + 1.
H0 : γ = 0 or H0 : β0 = β1
But in both cases, the test will be non standard since under H0 the
PSTR model contains unidenti…ed nuisance parameters.
Inference
Thus, testing the linearity of the model against the PSTR simply
consists of testing
H0 : θ 1 = 0
with SSR0 the panel sum of squared residuals under H0 (linear panel
model with individual e¤ects) and SSR1 the panel sum of squared residuals
under H1 (PSTR model with two regimes). Under the null hypothesis, the
F-statistic has an approximate F (1, TN N 1) distribution.
Testing procedure
1 Given a PSTR model with r = r , we test the null H0 : r = r
against H1 : r = r + 1.
2 If H0 is not rejected the procedure ends.
3 Otherwise, the null hypothesis H0 : r = r + 1 is tested against
H1 : r = r + 2.
4 The testing procedure continues until the …rst acceptance of H0 .
5 Given the sequential aspect of this testing procedure, at each step of
the procedure the signi…cance level must be reduced by a constant
factor 0 < ρ < 1 in order to avoid excessively large models. Gonzàlez
et al. (2005) suggest ρ = 0.5.
Example
Hurlin, Rabaud and Fouquau (2008) considers a PSTR model for
determining the relative in‡uence of …ve factors on the Feldstein and
Horioka result for OECD countries.
They consider …ve main factors (as potential threshold variable) generally
considered in this literature: (i) economic growth, (ii) demography, (iii)
degree of openness, (iv) country size and (v) current account balance.
Hurlin C., Rabaud I., and Fouquau J. (2008) The Feldstein-Horioka Puzzle:
a PSTR Approach. Economic Modelling, 25, 284-299.