NONLINEAR ECONOMETRIC MODELS: THE SMOOTH

TRANSITION REGRESSION APPROACH∗

Alenka Kavkler
Faculty of Economics and Business, Maribor, Slovenia,

Peter Mikek
Wabash College, Crawfordsville, IN, USA,

Bernhard Böhm
University of Technology, Vienna, Austria,

Darja Boršič
Faculty of Economics and Business, Maribor, Slovenia,

JEL classification: C30, C32, E24, E31, E41.

Keywords: nonlinear models, smooth transition regression, smooth transition vector

autoregression, panel smooth transition regression, real exchange rate, money demand,
Phillips curve, Okun’s law.

∗
This research was supported by a grant from the CERGE-EI Foundation under a program of the Global
Development Network. All opinions expressed are those of the authors and have not been endorsed by CERGE-
EI or the GDN.

1
 NONLINEAR ECONOMETRIC MODELS: THE SMOOTH
TRANSITION REGRESSION APPROACH∗

Abstract 1

In this paper we study econometric models of smooth transition characterized by switching

regimes through continuous transition functions. We discuss the process of specifying,
estimating and evaluating smooth transition regression (STR) models. Next we present an
overview of the first attempts at extending nonlinear STR techniques to vector autoregressive
(VAR) models and to panels that have emerged in the last few years. Additionally, we review
the applications of the STR modelling techniques to a number of different economic
problems. Finally, we provide an illustration by applying the methodology to a particular
example of nonlinear Okun’s Law for Germany.

∗
This research was supported by a grant from the CERGE-EI Foundation under a program of the Global
Development Network. All opinions expressed are those of the authors and have not been endorsed by CERGE-
EI or the GDN.

2
 NONLINEAR ECONOMETRIC MODELS: THE SMOOTH
TRANSITION REGRESSION APPROACH∗

Abstract 2

In this paper we study econometric models of smooth transition between different possible
regimes. The transition dynamics is based on continuous transition functions that allow for
smooth changes during the transition. Smooth transition models can be seen as a
generalization of threshold models. We discuss the process of specifying the models using
statistical tests for nonlinearities and choice of transition variable. Furthermore, we provide
some details on estimating and evaluating smooth transition regression (STR) models. Next
we present an overview of the first attempts at extending nonlinear STR techniques to vector
autoregressive (VAR) models. Extensions to data panels that have emerged in the last few
years are also discussed. Panels are especially interesting since they can easily be applied to
disaggregated data. Additionally, we review the applications of STR modelling techniques to
a number of different economic issues: dynamics of exchange rates, Okun’s Law, Phillips
curve, structure of wages in different sectors, models based on disaggregated data, and others.
Finally, we provide an illustration by applying the methodology to a particular example of
nonlinear Okun’s Law for Germany. We find that the transition function aligns closely with
substantial increases in unemployment rates, reflecting major shifts in economic structure,
such as German reunification, oil shocks, and a very restrictive monetary policy in the
eighties..

Keywords: nonlinear models, smooth transition regression, smooth transition vector

autoregression, panel smooth transition regression, real exchange rate, money demand,
Phillips curve, Okun’s law.

JEL classification: C30, C32, E24, E31, E41.

∗
This research was supported by a grant from the CERGE-EI Foundation under a program of the Global
Development Network. All opinions expressed are those of the authors and have not been endorsed by CERGE-
EI or the GDN.

3
1 INTRODUCTION

From recent studies of univariate models, we learn that there is much to be gained by allowing
nonlinear specification. Additionally, economic variables are frequently subject to switching
regimes. The notion of the regime switch implies a sudden abrupt change. However, most
economic variables change regimes in a smooth manner, with transition from one regime to
another taking some time. To handle this, Smooth Transition Regression (STR) models have
recently been developed. We present the STR methodology, including specification,
estimation and evaluation of STR models; examine its recent applications; and provide an
illustration of its application to a particular nonlinear model of Okun's Law for Western
Germany covering the reunification period.

In contrast to discrete switching models (e.g. Hansen, 1999), smooth transition regression
(STR) models transition as a continuous process dependent on the transition variable. This
allows for incorporating regime switching behaviour both when the exact time of the regime
change is not known with certainty and when there is a short transition period to a new
regime. Therefore, STR models provide additional information on the dynamics of variables
that show their value even during the transition period.

Capturing nonlinearities and regime switching makes STR models good candidates for
analysis of numerous economic variables. First, STR models naturally lend themselves to
modelling institutional structural breaks. Thus, they may be a useful tool to study transition
economies characterised by many structural breaks in the early part of transition. Second,
several authors provide evidence of asymmetries in the dynamics of economic variables,
depending on the magnitudes of parameters, in established market economies. Examples
include Johansen (2002), who shows asymmetric reactions for wages in various sectors, and
Legrenzi et al. (2004), who study asymmetric adjustment of real exchange rates. Third, the
STR methodology has been extensively used to study exchange rates and has recently been
applied to Okun's Law and the Phillips curve. Finally, the methodology has been extended
recently to VAR and to panel data. This allows for a whole spectrum of new applications
modelling several variables and incorporating heterogeneity in disaggregated data.

4
2 SMOOTH TRANSITION REGRESSION

Economic theory frequently asserts that the economy behaves differently if values of certain
variables lie in one region rather than in another, or, in other words, if they follow different
regimes. The first attempt at modelling such phenomena is represented by discrete switching
models, where a finite number of different regimes is assumed. The central tool of this class
of models is the so-called switching variable, which can be either observable or unobservable.

As smooth transitions between regimes are often more convenient and realistic than abrupt
switches, several scientists proposed a generalisation of discrete switching models of the
following form:

yt = xt′ϕ + ( xt′θ ) ⋅ G (γ , c; st ) + u t , t = 1, 2,…, T , (1)

where ϕ = (ϕ0 , ϕ1 ,…, ϕ p )′ and θ = (θ 0 ,θ1 ,…,θ p )′ are the parameter vectors, xt is the vector of

explanatory variables containing lags of the endogenous variable and the exogenous variables,
(i.e., xt = (1, xt1 , …, xtp )′ = (1, yt −1 ,…, yt − m , zt1 ,…, ztn )′ ), whereas ut denotes a sequence of

independent identically distributed errors. G stands for a continuous transition function

usually bounded between 0 and 1. Because of this property, not only can the two extreme
states be explained by the model, but also a continuum of states that lie between those two
extremes. The slope parameter γ > 0 is an indicator of the speed of transition between 0 and
1, whereas the threshold parameter c points to where the transition takes place. The transition
variable st is usually one of the explanatory variables or the time trend.

The most popular functional forms of the transition function are as follows:

1
• LSTR1 Model: G1 (γ , c; st ) = − γ ( st − c )
1+ e

G1 is a monotonously increasing function of the transition variable st , bounded between 0

and 1. Additionally, G1 (γ , c; c) = 0.5 ; therefore, we can say that the location parameter c

represents the point of transition between the two extreme regimes with lim st →−∞ G1 = 0 and

lim st →∞ G1 = 1 . The restriction γ > 0 is an identifying restriction. As we can see from Figure

1, the slope parameter γ indicates how rapidly the transition of G1 from 0 to 1 takes place.

5
While a moderate value of γ = 1 imposes a slow transition, the function with γ = 10 changes
quite fast.

Figure 1: LSTR1 transition functions with c = 1

If γ → ∞ in the definition of G1 , then model (1) converges to a switching regression model

with the extreme regimes yt = xt′ϕ + u t and yt = xt′ (ϕ + θ ) + u t . For γ = 0, the function G1 is

constant and equal to 0.5. In this case, model (1) simplifies to a linear regression model.

1
• LSTR2 Model: G2 (γ , c1 , c2 ; st ) = − γ ( st − c1 )( st − c2 )
1+ e

Monotonous transition may not always be satisfactory in applications. The quadratic logistic
function in the LSTR2 model is a nonmonotonous transition function that is especially useful
c1 + c2
in the case of reswitching. G2 is symmetric about the point and lim st →±∞ G2 = 1 . G2 is
2
never equal to 0; its minimal value lies between 0 and 0.5. Two examples of the function G2
with different values of the parameters are depicted in Figure 2.

6
Figure 2: LSTR2 transition functions with c1 = −1 and c2 = 1

2
• ESTR Model: G3 (γ , c; st ) = 1 − e−γ ( st −c )

Sometimes it is desirable that small absolute values of the transition variable are related to
small values of the transition function. The ESTR model with an exponential transition
function complies with the above condition for c = 0. The function G3 is nonmonotonous and
symmetric about the point c.

Figure 3: ESTR transition functions with c = 0

Both the LSTR2 model and the ESTR model enable reswitching, but they differ in the
rapidity of reswitching. One can see from Figure 3 that for a large value of γ , the transition of

7
st from 1 to 0 and back to 1 is much faster for the ESTR model as compared to the LSTR2
model, where the reswitching can be slower when the gap between c1 and c2 is large.

2.1 Testing Linearity against STR

Let us start by defining a more convenient notation:

⎧G − 0.5, i = 1, 2
Gi* = ⎨ i (2)
⎩ Gi , i=3

Obviously, Gi* = 0 for γ = 0 . The null hypothesis of linearity for model (2) can be expressed

as H 0 : γ = 0 against H1 : γ > 0 or as H 0′ : θ = 0 against H1′ : θ ≠ 0 . This indicates an

identification problem, since the model is identified under the alternative but not identified
under the null hypothesis. Namely, the parameters c and θ are nuisance parameters that are
not present in the model under H 0 and whose values do not affect the value of the log-
likelihood. Consequently, the likelihood ratio test, the Lagrange multiplier, and the Wald test
do not have their standard asymptotic distributions under the null hypothesis and one cannot
use these tests for a consistent estimation of the parameters c and θ . To overcome this
problem, Luukkonen, Saikkonen and Teräsvirta (1998) replaced the transition function with
its Taylor approximation of a suitable order. Let us write the first order Taylor approximation
around γ = 0 for the logistic transition function G1* as a polynomial in the transition variable

st :

T1 = a0 + a1st + R1 (γ , c; st ) . (3)

After replacing G1* by T1 in equation (2), one obtains

yt = xt′b0 + ( xt′st )b1 + ut* , (4)

where b0 and b1 are (p+1)-dimensional column vectors of parameters. The null hypothesis of

linearity can be tested as H 0′′ : b1 = 0 against H1′′: b1 ≠ 0 with a straightforward Lagrange

multiplier test. The test statistic is asymptotically χ 2 -distributed with p+1 degrees of
freedom. We have to emphasize that auxiliary regression (4) is suitable only if the transition

8
variable st is not an element of the vector xt . Otherwise, the variable st appears twice on the

right-hand side of Equation (4). The problem is solved by substituting xt with

′
t = ( xt1 , …, xtp ) in the second term of (4).
x%

To avoid dealing with low power in some special cases, the third order Taylor polynomial is
applied. This leads to the following auxiliary regression:

yt = xt′b0 + ( xt′st )b1 + ( xt′st2 )b2 + ( xt′st3 )b3 + ut* . (5)

Under the null hypothesis of linearity, the parameter vectors b1 , b2 and b3 are jointly tested to
zero. F-version of the linearity test is usually preferred because of its better small sample
properties. Comprehensive discussion of these issues is given in Teräsvirta (1998) and in
Luukkonen, Saikkonen and Teräsvirta (1998).

2.2 Model Specification

The choice of transition variable is not straightforward, since the underlying economic theory
often gives no clues as to which variable should be taken for the transition variable under the
alternative. Teräsvirta (1998) suggests testing the null hypothesis of linearity for each of the
possible transition variables in turn. The candidates for the transition variable are usually the
explanatory variables and the time trend. If the null is rejected for more than one variable, the
variable with the strongest rejection of linearity (i.e., with the lowest p-value) is chosen for
the transition variable. This intuitive and heuristic procedure can be justified by observing that
the test is most powerful when the alternative hypothesis is correctly specified, and this is
achieved for the "right" transition variable. It has to be emphasized that one cannot control the
overall significance level of the linearity test for this heuristic procedure, since several
individual tests have to be performed.

If the transition variable has already been decided upon, the next step in the modelling process
consists of choosing the transition function. The decision rule is based on a sequence of
nested hypotheses that test for the order of the polynomial in auxiliary regression (5):

H 04 : b3 = 0

9
 H 03 : b2 = 0 b3 = 0 (6)

H 02 : b1 = 0 b2 = b3 = 0 .

The 3 hypotheses are tested with a sequence of F-tests named F4, F3 and F2, respectively. If
the rejection of the hypothesis H 03 is the strongest, Teräsvirta (1998) advises choosing the
LSTR2 or the ESTR model. In practice, one usually chooses the LSTR2 model and
additionally tests the hypothesis c1 = c2 after estimation. If it cannot be rejected, it seems
better to select the LSTR2 model; otherwise ESTR should be selected. In case of the strongest
rejection of the hypotheses H 04 or H 02 , LSTR1 is chosen as the appropriate model. This

heuristic decision rule is based on expressing the parameter vectors b1 , b2 and b3 from

auxiliary regression (5) as functions of the parameters γ , c (or c1 and c2 ) and θ and the first

three partial derivatives of the transition function Gi* at the point γ = 0 .

Teräsvirta (1998) conducted a series of simulation experiments to investigate the properties of

the proposed heuristic specification strategy for choosing the transition variable and the
transition function. The study was carried out for smooth transition autoregressive (STAR)
models in the univariate setting. Different types of STAR models were examined and their
parameters were varied. The "true" transition variable was the lagged endogenous variable
yt − d , where the delay parameter d ran from 1 to 5. For each d , the linearity test was

performed for every possible transition variable in turn (i.e., for yt −1 , yt − 2 ,K , yt −5 ) and the
variable with the lowest p-value was chosen. The empirical size of the overall linearity test
was 3 to 4 % when the nominal size was 5 %. The results of the simulation study justified the
heuristic specification procedure and also showed that the power of the linearity test is better
for higher γ values and for lower values of the delay parameter d . The decision rule for
choosing the type of the transition function was tested for distinguishing between LSTAR1
and ESTAR models. It works best when the number of observations of the transition variable
that lie below c is about the same as the number of observations above c . The performance
of the rule improves with the sample size.

10
2.3 Estimation of STR Models

The specified STR model is usually estimated with nonlinear least squares or with maximum
likelihood estimation under the assumption of normally distributed errors. Both methods are
equivalent in this case. Nonlinear optimisation procedures are used to maximize the log-
likelihood or to minimize the sum of squared residuals. Some of the most often used nonlinear
optimisation algorithms are the Newton-Raphson algorithm, the Broyden-Fletcher-Goldfarb-
Shanno (BFGS) algorithm, the steepest descent algorithm, and the Davidon-Fletcher-Powell
(DFP) algorithm.

An additional remark should be made on the slope parameter γ of the transition function. The
magnitude of the parameter γ depends on the magnitude of the transition variable st and is
therefore not scale-free. The numerical optimisation is more stable if the exponent of the
transition function is standardised prior to optimisation. In other words, it is advisable to
divide γ by the sample standard deviation (in the case of LSTR1 models) or by the sample
variance (for ESTR and LSTR2 models) of the transition variable. In this way the magnitude
of the slope parameter is brought closer to the magnitude of other parameters.

2.4 Misspecification Tests

The misspecification tests were first developed by Eitrheim and Teräsvirta (1996) for
univariate time series, i.e. for smooth transition autoregressive (STAR) models, but the
generalisation to STR models is straightforward. Three tests had to be developed especially
for the STAR models, namely the test of no remaining nonlinearity, the test of no error
autocorrelation, and the parameter constancy test. For a detailed derivation of these tests, see
Eitrheim and Teräsvirta (1996) and Lin and Teräsvirta (1994). Other tests, like the LM test of
no autoregressive conditional heteroscedasticity of Engle (1982) and of McLeod and Li
(1983), and the Lomnicki-Jarque-Berra test of the normal distribution of errors, are performed
in the same way as in the linear setting.

11
3 SYSTEMS OF EQUATIONS

As many issues in economics require the specification of several relationships, techniques to

handle nonlinear features in systems are required. Only in recent years have such methods
appeared in the literature. Most of the work has been done in the nonlinear VAR framework.

Anderson and Vahid (1998) devised a procedure for detecting common nonlinear components
in a multivariate system of variables. The common nonlinearities approach is based on the
canonical correlations technique and can help us interpret the relationships between different
economic variables. The specification and estimation of the system of equations is also
simplified, since the existence of common nonlinearities reduces the dimension of nonlinear
components in the system and enables parsimony. This is particularly important in empirical
investigations involving economic time series of shorter length. Namely, most of the
macroeconomic indicators are published on a quarterly basis.

3.1 Smooth Transition Approach to Vector Autoregressive Models

Weise (1999), van Dijk (2001), and Camacho (2004) extended the STR modelling approach
to vector autoregressive models of smooth transition. Similarly, multivariate Markov–
switching models are treated in Krolzig (1997), and multivariate threshold models in Tsay
(1998). Van Dijk (2001) applies the STVAR modelling approach to study the intraday spot
rates and futures prices of the FTSE100 index, whereas Camacho (2004) examines the
nonlinear forecasting power of the composite index of leading indicators to predict both
output growth and the business cycle phases of the US economy. Since all three studies are
similar, and the most comprehensive description of the methodological approach is given by
Camacho (2004), we shall give a short review of his work. The STR specification is limited to
the case where the transition between different parameter regimes is governed by the same
transition variable and the same type of transition function in every equation of the system.
The authors argue that since the economic practice imposes common nonlinear features, all
equations share the same switching regime.

12
3.1.1 Specification and Estimation

Camacho (2004) considers a 2-dimensional smooth transition vector autoregressive (STVAR)

model

yt = ϕ ′y X t + (θ y′ X t ) G y ( s yt ) + u yt
(7)
xt = ϕ x′ X t + (θ x′ X t ) Gx ( sxt ) + u xt ,

where X t = (1, yt −1 , xt −1 ,K , yt − p , xt − p )′ = (1, X%t )′ , ϕ x , ϕ y , θ x ,θ y are the corresponding parameter

vectors, and U t = (u yt , u xt )′ : N (0, Ω) is a vector series of serially uncorrelated errors. The

difference Dit = sit − ci , i = x, y in the exponent of the transition function Gi is called the

switching expression. The letters yt and xt are used for the two variables in the
autoregressive system, since the smooth transition approach is applied to the rate of growth of
the US GDP and the rate of growth of the US composite index of leading indicators,
respectively. The discussion is restricted to the case of sxt = s yt = st and Gx = G y , where the

same transition variable and the same transition function are used in both equations.

After the linear VAR has been specified, the linearity test is applied. The problems with the
transition function parameters are solved with a suitable Taylor series expansion, as usual.
The auxiliary regression to be performed in case the transition variable st belongs to X t is

3
yt = η ′y 0 X t + ∑η ′yh X%t sth + v yt
h =1
3
(8)
xt = η x′ 0 X t + ∑η xh
′ X%s + vxth
t t
h =1

and the null hypothesis of linearity reads as

H 0 : ηi1 = ηi 2 = ηi 3 = 0, i = x, y. (9)

Consequently, the null hypothesis can be tested with the Lagrange multiplier test.

If the null hypothesis of linearity is rejected in favour of the alternative smooth transition
vector autoregressive model, one has to decide which transition function to use. The decision
is based on the sequence of nested hypotheses tests described in Section 2.2. The parameters

13
of the specified model are estimated with the maximum likelihood estimator under the
assumption of normally distributed errors:

U t = (u yt , u xt )′ : N (0, Ω). (10)

3.1.2 Testing the Model Adequacy

As proposed by Eitrheim and Teräsvirta (1994), three tests are performed in order to check for
the adequacy of the estimated model, namely the test of no error autocorrelation, the test of no
remaining nonlinearity, and the parameter constancy test. A detailed description of the
multivariate generalisations of the three tests can be found in Camacho (2004).

4. PANEL SMOOTH TRANSITION REGRESSION (PSTR)

PSTR models are the latest extension of STR modelling to panel data with heterogeneity
across the panel members and over time. First we present the general form of PSTR
(Gonzalez et al., 2005) and then an extension to multilevel PSTR (Fok et al., 2005a and Fok et
al., 2005b).

4.1. Panel STR Models

While Tsay (1998) discusses multivariate threshold models, Hansen (1999) presents a panel
threshold regression model. Building on this, Gonzalez et al. (2005) generalise the framework
to include STR in panels. The panel threshold model allows for regression coefficients to
change upon some observable variable reaching the threshold. Thus, it assumes that the
threshold for the switching of regimes is clearly defined. Gonzalez et al. (2005) relax this
assumption and incorporate smooth transition into this framework, allowing the regression
coefficients to adjust gradually as the system moves from one regime to another.

The general form of the PSTR model is given in the following equation:

r
yit = µi + ϕi ' xit + ∑ θ i j ' xit G j ( sitj ; γ j , c j ) + uit (11)
j =1

14
µi are means for individual members of the panel, and r is the maximum number of different
transition functions (Gj). The rest of the equation is analogous to (1) above, but with the
additional index i counting the panel members. The transition function is determined as
follows:

1
G j ( sitj ; γ j , c j ) = m (12)
−γ j ∏ ( sitj −c jk )
1+ e k =1

with γ > 0 and c j1 ≤ c j 2 ≤ K ≤ c jm . c j = ( c j1 , c j 2 ,K , c jm )′ is an m-dimensional vector of

threshold parameters.

The transition Gj function incorporates one (m = 1) or more (m > 1) regime switches. m = 2 in

Equation (12) allows for two different switches of regime (with identical outer regimes),
which is usually sufficient for most practical applications. With the slope parameter γj Æ ∞
and m = 1, Gj becomes an indicator function I[sjit > cj], with I[A] = 1 when event A occurs,
and 0 otherwise. m = 1 corresponds to two regimes, and m = 2 to the three regime panel
threshold model developed in Hansen (1999). In general, when sjit = sit for j = 1, ... , r, m = 1
and γj Æ ∞, this results in a panel threshold model with r + 1 regimes. Thus, Equation (11) is
a generalisation of Hansen's approach. Additionally, the model (11) and (12) above is a
convenient alternative for testing for remaining heterogeneity.

4.1.1. Testing Homogeneity of a Panel

If a data generating process is homogenous, the PSTR is not identified. Therefore, the first
step in specifying the model is testing for homogeneity. Consider the model in (11) with r = 1.
For such a panel model, the H0: γ = 0 is tested against the PSTR alternative. The null implies
no heterogeneity. As explained above, the testing is based on the Taylor expansion of the
transition function around γ = 0. If the first-order Taylor approximation is applied, the null
hypothesis (based on auxiliary regression) becomes H0: b1 = b2 = ... = bm = 0. The general
approach is the same as in Section 2.1. Some details can be found in Gonzalez et al. (2005) or
in Luukkonen, Saikkonen and Teräsvirta (1988).

The testing procedure from Section 2.2 can also be used for PSTR models to choose the
transition variable (the one with strongest rejection of linearity) and to determine an

15
appropriate form of transition function (12), thus choosing m. This requires the sequential
testing for m from highest to lowest (as in (6) above).

4.1.2. Estimation

First the fixed effects (µi) are eliminated by subtracting panel member specific means from
the data, and then NLS is applied to the transformed variables. Assuming r = 1 and
subtracting the respective means in equation (11) we obtain:

yit* = β ' xit* (γ , c) + uit* (13)

where β = (φ', θ')', xit(γ,c) = (xit', xit'G(sit;γi,ci))', and asterisks denote deviations from the
individual means. NLS or maximum likelihood can be applied to Equation (13) with a caveat
that transformed x* depends on γ and c through both the levels and the individual means.
Thus, one has to compute x*(γ,c) at each iteration of the nonlinear optimization procedure.

Conditionally on γ and c, the PSTR model is linear in β. As pointed out by Hansen (2000), the
computationally easiest method to obtain the LS estimates in this case is through
concentration. The parameters of the model are estimated by minimising the concentrated sum
of squared errors.

4.1.3. Model Evaluation

Gonzalez et al. (2005) develop the parameter constancy test. This is done using the alternative
that the parameters in (11) change smoothly over time. The model under the alternative is
time varying PSTR. If m = 1 in (12), then the alternative is defined as:

2
yit = µi + ∑ f j (t ; γ 2 , c2 ) ∗ (ϕi j ' xit + θi j ' xit G ( sit ; γ 1 , c1 )) + uit (14)
j =1

where f1(.) = 1 and f2(.) is as given in (12). If the parameter γ2 = 0, then this reduces to (11)
and indicates parameter constancy, implying the following null H0: γ2 = 0. Eitrheim and
Teräsvirta (1996) discuss some numerical problems in computing the test (see also Gonzalez
et al., 2005), but the F test can be applied.

Testing for remaining heterogeneity is based on the alternative r = 2 in (11). The null is again:
H0: γ2 = 0 and can be tested using appropriate F distribution. Gonzalez et al. (2005) also

16
discuss the choice of the number of regimes and provide some simulation results for the
model at hand.

4.2. Multi Level Panel STAR Models

Multi level panel STAR models allow for an additional structure that relates the parameters of
the model across various elements of the panel. Fok et al. (2005a) present a two level STAR
model for a panel of time series. They use it to study disaggregated data and argue that
aggregation loses some information embedded in individual series. Furthermore, their
approach makes STAR models more parsimonious by partially pooling parameters across
panel members.

In addition to estimating the usual STAR model, they add to the model the second level where
the parameters of the transition function (γ and c) are dependent on data characteristics in
individual sectors. In particular they estimate:

yit = µi + ϕi ' xit + θi ' xit G ( st ; γ i , ci ) + uit (15)

⎛ log(γ i ) ⎞
⎜ ⎟ = δ ' wi + ηi (16)
⎝ ci ⎠

The vector of explanatory variables in (15) is given by xit = (1, yit −1 ,K , yit − p )′ . Equation (15)

is the same as (11). However, equation (16) shows the second level regression where the
parameters of the transition function G (γi and ci) depend on observable variables (related to
yit) collected in vector wi. δ is a 2-column matrix of unknown coefficients and ηi is a vector of
well-behaved errors.

The estimation of the complex system above is not straightforward. Fok et al. (2005a) use
simulated maximum likelihood and concentrate the likelihood function with respect to the
parameters of the first level model.

Frequently one may want to know the value of the switching function G(st; γi, ci) for
individual sectors at given times. Using simulation, Fok et al. (2005a) calculate the
conditional expectations of the transition function that depends on the sector-specific
observed data:

17
 1
∑ G(st ; δ ' wi + ηil )ωil
L l
Eηi [G ( st ; δ ' wi + ηi ) yi ] = (17)
1
L l
∑ ωil

Thus the conditional expectation of the transition function G, with parameters dependent on
observables as in (16), given the information on yi, is calculated using appropriate weights ω
(some details can be found in Fok et al., 2005a). These conditional expectations are
interpreted as indicators of the state of the system (for example, the state of the business
cycle).

This two-fold strategy places their model in between two extremes: the completely
heterogeneous panel, which imposes no cross restrictions on the parameters, and the fully
pooled panel, where the regime switching process is equal across the panel members. The
approach is interesting since it allows substantial flexibility across the panel members.
However, the estimation is rather complicated (for iterations on concentrated simulated
maximum likelihood, see Fok et al., 2005a for details.)

Fok et al. (2005b) incorporate some generalisations in the multilevel panel STAR model
presented above, which they use for studying forecasting properties of the model.
Generalisations include the transition function that depends on the vector zt=(zt1, ..., ztk)' of
observable variables. Thus, the transition function is of the following form:

1
G ( zt ; π i , γ i , ci ) = ' (18)
1 + e −γ i (π i zt −ci ) )

where πi is a vector of parameters. Furthermore, they generalise the model by including a

vector νt of additional exogenous regressors into Equation (15) above:

yit = µi + λi 'ν it + ϕi ' xit + βi ' xit G ( zt ; π i , γ i , ci ) + uit (19)

λi is a vector of regression coefficients, and the error term is well-behaved. The second level
estimation for this model is based on the following:

⎛ log(γ i ) ⎞
⎜ ⎟
⎜ ci ⎟ = δ ' wi + ηi (20)
⎜ u ⎟
⎝ i ⎠

18
where variables are defined as in (16). The estimation methodology of concentrated and
simulated maximum likelihood from Fok et al. (2005a) is employed here. As above, the
conditional expectations are simulated to obtain information on the current state with respect
to the regime.

5 APPLICATIONS

5.1 Nonlinearities in Real Exchange Rates

Since the real exchange rate in logarithmic form may be viewed as a measure of the deviation
from purchasing power parity (PPP), the question of mean reversion in the real exchange rate
is closely related to the issue of validity of PPP. In order to circumvent the low power
problem of conventional unit root tests, the validity of PPP is usually investigated through
long-span studies or panel unit root studies. Sarno and Taylor (2002) point out the
disadvantages of both of the mentioned approaches. As far as the long-span studies are
concerned, the long samples required to generate a reasonable level of power with univariate
unit root tests may be unavailable for many currencies. Panel studies, on the other hand,
impose the null hypothesis that all of the series under observation are generated by unit root
processes, implying that the probability of rejection of the null hypothesis may be quite high
when as few as just one of the series is stationary. For this reason, Sarno and Taylor develop a
smooth transition autoregressive (STAR) model to study the behaviour of the real exchange
rate. In their model, the real exchange rate in the logarithmic form is explained by its lagged
values. It is shown that the four major real dollar exchange rates are becoming increasingly
mean reverting with the absolute size of the deviation from equilibrium, which is consistent
with the recent theoretical literature on the nature of the real exchange rate dynamics in the
presence of international arbitrage costs.

Traditional empirical analyses of purchasing power parity validity and its deviations are based
on linear framework and mostly suggest that the long run equilibrium is constant. Moreover,
these analyses suggest that real exchange rate dynamics should be explained by a linear
autoregressive process with continuous and constant speed of adjustment, not taking into
account the size of deviations from purchasing power parity (Sarno and Taylor, 2002). Using
a linear framework for a nonlinear dataset, the rejection of a unit root as a null hypothesis is

19
more likely (Taylor 2006), while the assumption of constant speed of adjustment implies
downward bias of the results.

Potential reasons for nonlinearities in real exchange rates include frictions due to transport
costs, tariffs or non-tariff barriers, interaction of heterogeneous agents in the foreign exchange
market at the micro-structural level, and influence of official intervention in the foreign
exchange market (Taylor, 2006). Sarno and Taylor (2002), Sarno (2003), and Taylor (2006)
provide an overview of nonlinear exchange rate models and assess their contribution to
explaining the behaviour of the exchange rates. Numerous authors reject linearity assumption
in favour of STAR models when studying exchange rate dynamics: Liew, Chong and Lim
(2003) for 11 Asian countries, and Rapach and Wohar (2006) and Ahmad and Glosser (2007)
for the US dollar real exchange rate. Moreover, Paya, Venetis, and Peel (2003) take into
consideration two different approaches in solving the purchasing power parity puzzle:
nonlinear adjustment of real exchange rates induced by transaction costs and non-constant
real exchange rate equilibrium induced by different productivity growth rates. Consequently,
the dynamics of real exchange rates can be described as symmetric and nonlinear.
Additionally, these authors show that the estimated half-lives of the shocks are much shorter
than those obtained by linear models.

A growing number of studies apply nonlinear LSTAR or ESTAR models and find (rapid)
mean reversion in both real and nominal exchange rates: Taylor, Peel, and Sarno (2001);
Guerra (2003) for the Swiss frank–German mark; Liew, Bahrumshah and Lim (2004) for the
Singapore dollar-US dollar; Paya and Peel (2005) for high inflation countries; Leon and
Najarian (2005); and Baum, Barkoulas and Caglayan. (2001) using deviations from PPP
obtained by the Johansen cointegration method. Additionally, several authors reject unit roots
when testing real exchange rates: Sollis (2005) for several US dollar exchange rates with
gradually changing deterministic trends, and Leon and Najarian (2005) for PPP deviations.
Lahtinen (2006) uses the US dollar-euro exchange rate and distinguishes between the sudden
and smooth adjustment to long-run equilibrium. He argues that the adjustment for the data
under observation is sudden.

ESTAR models have also been used to forecast the behaviour of real exchange rates. Kilian
and Taylor (2003) find evidence of exchange rate predictability in 2 to 3 years given ESTAR
real exchange rate dynamics. Asymmetries in adjustment of real exchange rate to equilibrium
was studied in Leon and Najarian (2005), and Legrenzi and Milas (2004).

20
Monte Carlo simulations are frequently used to study the dynamics of exchange rates and test
for possible nonlinearity: Taylor, Peel, and Sarno (2001) show the fastest adjustment process
of real exchange rates when transaction costs and nonlinearities in mean reversions are
present; Paya and Peel (2005) show that nonlinear tests provide support for PPP; and Ahmad
and Glosser (2007) claim that the methodology used to detect nonlinearities in the data exhibit
size bias.

In addition, Peel and Venetis (2005) present theoretical limitations of ESTAR models and
propose a new linear model consistent with rational expectations, while the ESTAR model
assumes adaptive expectations.

One of the relatively rare papers examining purchasing power parity deviations in Central
European countries is Arghyrou, Boinet, and Martin (2005). The authors analyse the data
from Czech Republic, Hungary, Poland, Slovakia and Slovenia. Among other results, it is
shown that the short-run dynamics of the real exchange rates display nonlinear and
asymmetric behaviour, while the speed of adjustment depends on the size and sign of the
deviation.

Legrenzi and Milas (2004) study a VAR that includes exchange rate, unemployment rate and
real wages. They find evidence of nonlinearities and explain it as due to asymmetric
adjustments to exchange rate disequilibria: "prices and wages are more flexible when real
output is high."

5.2 Phillips Curve, Okun's Law, and Money Demand

5.2.1 Nonlinear Money Demand

Since representation of asymmetric reactions, structural changes, and other phenomena of

economic developments can be fruitfully investigated by nonlinear modelling techniques, the
issue of a possible nonlinear money demand specification has been studied by several authors.
Chen and Wu (2005) show that employing the conventional linear cointegration approach in
examining long-run money demand may not be appropriate after taking into account the
existence of transaction costs. They provide evidence that deviations from equilibrium money
demand follow an exponential smooth transition autoregressive process that is mean-reverting

21
outside a given range and has a unit-root inside the range. Similarly, Sarno, Taylor and Peel
(2003) argue that several theoretical models of money demand imply nonlinear functional
forms for the aggregate demand for money, characterized by smooth adjustment toward long-
run equilibrium. Their paper proposes a nonlinear equilibrium correction model of US money
demand that is shown to be stable over the sample period from 1869 to 1997. The use of an
exponential smooth transition regression model, with the lagged long-run equilibrium error
acting as the transition variable, implies faster adjustment toward equilibrium, the greater the
absolute size of the deviation from equilibrium. In a similar study, Sarno (1999) presents a
stable empirical model for the demand for narrow money in Italy using annual data spanning
from Italian unification in 1861 through 1991. A nonlinear functional form of the aggregate
demand for money is characterized by smooth adjustment towards long-run equilibrium,
again achieved by estimating a nonlinear error correction model in the form of an exponential
smooth transition regression

5.2.2 Nonlinear Phillips Curve

Substantial theoretical and empirical evidence can be found in the literature suggesting
nonlinearity in the output-inflation relationship, namely a nonlinear Phillips curve. Dolado,
Ramon and Naveira (2005) investigate the implications of a nonlinear Phillips curve for the
derivation of optimal monetary policy rules. They show that combined with a quadratic loss
function, the optimal policy is also nonlinear, with the policy-maker increasing interest rates
by a larger amount when inflation or output are above target than the amount by which they
are reduced when they are below target. The model of Schaling (2004) features a convex
Phillips curve, in which positive deviations of aggregate demand from potential are more
inflationary than negative deviations are disinflationary. Corrado and Holly (2003) consider
the performance of optimal policy rules when the underlying relationship between inflation
and the output gap may be nonlinear. In particular, if the inflation-output trade-off exhibits
nonlinearities, this will impart a bias to inflation when a linear rule is used. To correct this
bias, they propose a piecewise linear rule, which can be thought of as an approximation to the
nonlinear rule of Schaling (2004).

Mayes and Viren (2002) highlight the implications for a single monetary policy when key
economic relationships are nonlinear or asymmetric at a disaggregate level. Using data for the
EU and OECD countries, they show that there are considerable nonlinearities and

22
asymmetries in the Phillips and Okun curves. High unemployment has a relatively limited
effect in pulling inflation down, while low unemployment can be much more effective in
driving it up. To accommodate the potentially important departure from linearity of the
Phillips curve, Huh (2002) employs a vector autoregressive (VAR) model of output, inflation,
and the terms of trade augmented with logistic smooth transition autoregression
specifications. Empirical results indicate that the model captures the nonlinear features
present in the data well. Based on this nonlinear approximation, the output costs for reducing
inflation are found to vary, depending critically on the state of the economy, the size of
intended inflation change, and whether policymakers seek to disinflate or prevent inflation
from rising. This implies that inferences based on the conventional linear Phillips curve may
provide misleading signals about the cost of lowering inflation and thus the appropriate policy
stance. Böhm (2001) also employs the smooth transition regression modelling approach. In a
formulation of an inflation equation for Austria, which includes demand and supply features,
he explores the capacity of STR models to improve upon specification. The nonlinearities and
asymmetries are found to be relevant ingredients in the Austrian inflation equation, and the
change in the unemployment rate is shown to have a larger impact on inflation during periods
of high volatility of price increases. Kavkler and Böhm (2005) investigate a well-known
model of monetary inflation theory which can be shortly characterized by an equation
describing the monetary system augmented by a Phillips curve and the equation of Okun’s
Law. The basic tool for identification and estimation of the model equations is the smooth
transition regression approach. From the simulation of the estimated nonlinear system,
asymmetric policy reactions can be derived.

5.2.3 Nonlinear Okun's Law

While the linear relationship between output and unemployment rate in the US was
established empirically by Okun, Prachowny (1993) provided a theoretical derivation of the
relation in a special case. Under the assumptions that the aggregate production function is of a
Cobb-Douglas type and that the capital stock and a disembodied technology factor are always
at their long-run levels, Prachowny established a log linear relationship between the output
gap and capacity utilization gap, labour supply gap and hours worked gap. Sögner and
Stiassny (2002) use Baysian methods to test for discrete structural breaks in Okun's Law and
the Kalman filter to check for continuous parameter changes. 15 OECD countries are included
in their study. The first approach does not detect any structural breaks, whereas the results of

23
the second approach imply continuous parameter changes for 10 of the countries. The
relationships between output and labour demand and labour supply, respectively, are also
discussed in the paper. The authors conclude that for most countries the change in Okun's
coefficient results mainly from an increased reaction of employment to GDP change. A
nonlinear relationship between cyclical unemployment and cyclical output is proposed by
Cuaresma (2003). For US data, the linear specification is strongly rejected in favour of a
piecewise linear specification. The estimated Okun's coefficient is significantly higher for
expansions than for recessions, implying that output changes cause asymmetric and regime
dependent changes in the unemployment rate. Additionally, unemployment shocks tend to be
more persistent in times of expansion. The findings of Mayes and Viren (2002) for EU and
OECD countries are similar. Asymmetry is built into Okun's Law with the help of the
threshold model and the error correction mechanism, which enables regime dependent
correction paths. Most of the countries included in the study exhibit an asymmetric
relationship between unemployment rate and change in output. Below we present an
application of STR to the Okun's Law.

5.3. Panel Smooth Transition Regression

Gonzalez et al. (2005) apply the PSTR model to study companies' investment decisions under
capital market imperfections. In particular, asymmetric information between borrowers and
lenders causes investment decisions to depend on other financial variables, for example cash
flow or leverage. In this setting, it is likely that both investment opportunities and information
costs change through time in such a way that firms migrate between the constrained and the
unconstrained regime. Since it is unlikely that the regime would switch abruptly, this merits
the use of smooth transition techniques. The authors apply the model to 565 US firms during
the years 1973-1987 and reject homogeneity for two transition variables: Tobin's q and lagged
debt. Moreover, they determine the order of m in Equation (12) as a logistic function with
m=1. Using Tobin's q as a transition variable, their results reveal that investment decisions
depend on Tobin's q, debt, cash flow, and sales of assets. They report that on average nearly
10% of firms switch regimes during a year, and they conclude that there exists a clear
nonlinear relationship between investment and Tobin's q.

24
Fok et al. (2005a) apply their two level panel STR to 19 3-digit NAICS2 sectors of the US
economy. They model the business cycle in these sectors, taking into account asymmetries
between recessions and expansions. While the former are sharp and short, the latter have
longer durations. Their modelling approach allows them to incorporate differences in the
timing of recession onset in different sectors. To account for heterogeneity in regime-
switching properties across industries, they use capital, worker wages, energy, and material
costs as explanatory variables in the second-level regression. They avoid the usual selection
of transition variable through testing for linearity by choosing the interest rate as their
transition variable. They base their decision on the finding that the "term spread is among the
most powerful (leading) indicators of the US business cycle" (Estrella and Mishkin, 1998).
After handling several estimation problems (such as no convergence in numerical
optimisation for some sectors), Fok et al. report the conditional expectations as an indicator
for the state of the business cycle. These align very well with the official NBER recession
dates. There is some variation in the timing of the onset of recession for different industries,
but that is rather limited. Additionally, they show that for aggregate growth the in-sample
predictions are best for univariate STR; however, their model is superior for out-of-sample
predictions.

Fok et al. (2005b) use the coincidence index, measuring economic activity at the
disaggregated state level in the US, to study the forecast properties of the multi-level panel
STR model. The coincidence index is based on a dynamic factor model for non-agricultural
employment, the unemployment rate, average hours worked, and real wages. Fok et al. use
simulation to compare the forecasts obtained by estimating three different STR models: the
STR for aggregate growth rates, the individual STR model, and the multi-level panel STR
model. They find that the STR model for aggregate growth rates performs the worst in this
simulation. The model is then applied to data and they find that the panel STR model
outperforms the aggregate STR model for both in- and out-of-sample forecasts. Thus, they
conclude that forecasting business cycles based on disaggregated data incorporates nonlinear
dynamics in individual industries and therefore is superior to the forecasts based on the
aggregate time series.

2
North American Industry Classification.

25
Using data for 117 industries, Johansen (2002) applies the panel STR techniques to industry
specific wages. He rejects linearity and uses relative wages as a transition variable. After
eliminating the industry-specific fixed effects by first differencing the data, he applies GMM
on instrumental variables to estimate the parameters. He finds strong support for nonlinearity
in industry wage responses to profitability, outside industry wage, and unemployment. He
claims that nonlinearity reflects higher concern on the part of workers in low wage industries
(first regime) as opposed to those in high wage industries. The long-run insider weight and the
unemployment effect are much stronger in low wage industries.

6 ILLUSTRATION: NONLINEAR OKUN'S LAW

Okun’s law describes the short-run relationship between the GDP gap and the unemployment
rate. This empirical relationship, developed by A.M. Okun in the 1970s, can be stated as
follows (compare Frisch, 1990):

⎛ X − X* ⎞
u = u * − a ⎜⎜ *
⎟⎟
⎝ X ⎠ (21)

where a > 0 is a constant term, u and u * denote the actual and the natural rate of
unemployment, X stands for the actual real output, and X * for potential real output. As a
slightly more general relationship between the rate of unemployment and the rate of growth of
real output, we can write

(
u = u −1 − a x − x * ) (22)

with x * denoting the expected rate of real growth following the long-run trend.

We apply the STR approach to modelling Okun’s Law based on seasonally adjusted quarterly
data from West Germany between 1969 and 2000.3 West Germany was chosen to investigate
the impact of the German reunification, and the end of the time span under observation was
determined by the availability of official data published by the Federal Statistical Office of
Germany for West Germany. Following Grant (2002), four approaches for modelling x * via
the business cycle were applied: simple average, linear trend, the Hodrick-Prescott

3
Some further details can be found in Kavkler and Böhm (2005).

26
decomposition, and the Beveridge-Nelson decomposition. As the simple average did not
perform worse than the other three methods, it was chosen as the appropriate method. After
several attempts to specify a model that is linear in the output gap, the following estimation
result including the squared gap was obtained. Additionally, it has proved useful to apply a
difference transformation ∆u t = u t − u t −1 when searching for the appropriate dynamics. In

Equation (23) below, the variable gap stands for the difference x − x * , with the expected rate
of real growth x * set constant and equal to the arithmetic mean of x. The dummy variables
were introduced to reduce the ARCH effects caused by the outliers in the years 1991 and
1992, when German reunification took place. The obtained model proved satisfactory after
being tested for normality, autocorrelations, ARCH effects, and constancy of coefficients. The
estimation results are given in Equation (23) and the results of the tests in Table 1.

∆u t = -0.0194 - 0.0464 gap + 0.0086 gap 2 - 0.1960 u t-1 + 0.4247 ∆u t-1 + (23)
(0.0225) (0.0093) (0.0023) (0.0312) (0.0606)
+ 0.1788 ∆u t-3 - 0.1886 ∆u t-4 - 0.6667 dummy1 + 0.5415 dummy2
(0.0744) (0.0708) (0.1992) (0.1197)
R 2 = 0.6276, S.E. = 0.1946, AIC = -0.3675, T = 128

Table 1: Specification and diagnostic tests (p – values)

Test Jarque-Bera Breusch-Godfrey (4 lags) Ljung-Box (4 lags) ARCH (4 lags)
p-
0.1843 0.1229 0.568 0.4768
value

Figure 4 depicts recursive coefficient estimates obtained by least squares estimation over
gradually increasing time intervals. Sudden changes in the course of the recursive estimates
imply structural change, whereas smooth changes hint at misspecification. In our case, the
coefficients C(1), C(3) and C(5) of the constant and the variables gap 2 and ∆u t-1 display the
most variation. Equation (23) is thus a potential candidate for nonlinear STR specification,
since several coefficients do not seem to be constant over time.

27
Figure 4: Recursive coefficients (following the coefficients in Equation (23) row-wise)

-.02 .014 .06

-.03 .012 .04

.02
-.04 .010
.00
-.05 .008
-.02
-.06 .006
-.04

-.07 .004 -.06

-.08 .002 -.08

1995 1996 1997 1998 1999 2000 1995 1996 1997 1998 1999 2000 1995 1996 1997 1998 1999 2000

Recursive C(1) Estimates ± 2 S.E. Recursive C(2) Estimates ± 2 S.E. Recursive C(3) Estimates ± 2 S.E.

-.12 .6 .36

.32
-.16 .5 .28

.24
-.20 .4
.20
.16
-.24 .3
.12

-.28 .2 .08

.04

-.32 .1 .00
1995 1996 1997 1998 1999 2000 1995 1996 1997 1998 1999 2000 1995 1996 1997 1998 1999 2000

Recursive C(4) Estimates ± 2 S.E. Recursive C(5) Estimates ± 2 S.E. Recursive C(6) Estimates ± 2 S.E.

.00 -0.2 .8

-.04 -0.3
.7
-.08 -0.4

-.12 -0.5 .6

-.16 -0.6
.5
-.20 -0.7

-.24 -0.8 .4

-.28 -0.9
.3
-.32 -1.0

-.36 -1.1 .2
1995 1996 1997 1998 1999 2000 1995 1996 1997 1998 1999 2000 1995 1996 1997 1998 1999 2000

Recursive C(7) Estimates ± 2 S.E. Recursive C(8) Estimates ± 2 S.E. Recursive C(9) Estimates ± 2 S.E.

The first step in modelling the nonlinear relationship is to find an appropriate transition
variable and transition function. The results of the F, F4, F3 and F2 tests are given in Table 2.
The variable ∆u t-1 with the strongest rejection of linearity (i.e., with the lowest p-value of the
F-test) is chosen for the transition variable. The comparison of the p-values of the F4, F3 and
F2 tests for the variable ∆u t-1 indicates the ESTR model as the best choice (see Section 2.2 for
details).

Table 2: Linearity tests (p-values)

Variable F F4 F3 F2
time trend 0.1553 0.0869 0.5435 0.2648
gap 2 0.0055 0.5306 0.0257 0.0056
v t-1 0.0261 0.5446 0.0240 0.0551
∆v t-1 0.0000 0.0057 0.0003 0.0079
∆v t-3 0.0036 0.0201 0.0194 0.2663
∆v t-4 0.0330 0.0131 0.0840 0.8711

28
After eliminating insignificant variables from the model, one obtains the estimated
coefficients as shown in Equation (24):

∆u t = -0.0422 gap - 0.1099 u t-1 + 0.4727 ∆u t-1 + 0.1212 ∆u t-3 - 0.7308 dummy1 + (24)
(0.0084) (0.0341) (0.0599) (0.0657) (0.1683)
+ 0.6099 dummy2 + [-0.1739 + 0.0400 gap 2 - 0.2378 u t-1 - 0.5881 ∆u t-4 ] *
(0.1039) (0.0722) (0.0113) (0.0808) (0.2795)
* [1 - Exp(-1.2706( ∆u t-1 - 0.1142)2)]
(0.5963) (0.0383)
R 2 = 0.7275, S.E. = 0.1686, AIC = -3.4703, T = 128, σˆ nl / σˆ lin = 0.6540

The estimate of the coefficient c makes sense because it lies in the range of the transition
variable. The variable gap² is significant in the nonlinear part, but is dropped as insignificant
from the linear part. The fact that the gap variable may even have increasing effects on
unemployment may seem odd but can be explained by the additional demand for high skilled
labour in periods of excessive growth, say by more than two to three percent deviation from
normal.

Finally, specification and diagnostic tests are performed to evaluate the obtained model. The
p-values of the Jarque-Bera test and the test of no remaining error autocorrelation show that
the null hypotheses of the normally distributed errors and of no error autocorrelation,
respectively, cannot be rejected (Table 3). Table 3 also reveals that there are no ARCH effects
present in our model. The test of parameter constancy detects only problems concerning the
constant term in the nonlinear part of the model.

Table 3: Specification and diagnostic tests (p – values)

Test Jarque-Bera AR LM (4 lags) Ljung-Box (4 lags) ARCH LM (4 lags)
p-value 0.6495 0.4343 0.8926 0.4107

Since the threshold parameter c in Equation (24) is close to zero, the two extreme regimes
with G = 0 and G = 1 are related to small and large changes in the unemployment rate,
respectively. The short-run relationship between the variables ∆u t and gap is linear when the
transition variable is close to threshold,
∆u t = -0.0422 gap (for G = 0)

29
and nonlinear otherwise:
∆u t = -0.0422 gap + 0.0400 gap 2 (for G = 1)
Output changes thus cause asymmetric and regime dependent changes in the unemployment
rate.

A comparison of the linear and nonlinear models reveals an increase in explanatory power
( R 2 increases from 0.63 to 0.73) and a decrease in the standard error of regression from 0.19
to 0.17. The null hypothesis of linearity tested against the alternative of a smooth transition
regression model has to be rejected for every possible transition variable with the exception of
the time trend. Both of these facts confirm our intuition that the linear relationship of Okun’s
Law can be improved by consideration of regime changes.

By plotting the transition function G and the unemployment rate u in the same graph in Figure
5 one can observe that most of the major changes in the transition function occur when the
unemployment rate has risen to new heights. This can be associated with major structural
changes in the German economy in those periods. In particular, we can clearly observe
changes in regimes during three distinct periods. The first covers the oil shocks of the
seventies, especially the first one. The second corresponds to strongly restrictive monetary
policy in both Germany and the US during the eighties. The third and most clearly discernible
covers the period following the reunification of Germany from 1990 to 1995. Each of the
periods is characterized by a sharp rise in unemployment. The German labour market was
characterized by a number of rigidities, ranging from centralized wage bargaining, a rigid
institutional and legal framework for the labour market, and low mobility of the labour force;
to high legal protection against firing (Berthold and Fehn, 2003; Hunt, 1999; Solow, 2000;
Blanchard and Wolfers, 1999; Siebert, 1997).

During each of these periods, large changes in the economic environment affected both the
unemployment rate and GDP growth. However, due to a number of rigidities in the labour
markets mentioned above, the interaction of shocks with the institutions of the labour market
(Berthold and Fehn, 2003) changed the relationship between unemployment and output
growth asymmetrically. Therefore, the major structural breaks of these periods can be seen as
distinct regimes. Figure 5 thus reveals that the nonlinear part of the model captures these
uneven developments in the economy rather well.

30
Figure 5: Unemployment rate and transition function

1.0 10

0.8 8

0.6 6

0.4 4

0.2 2

0.0 0
1970 1975 1980 1985 1990 1995 2000

transition function unemployment rate

7 CONCLUSION

The recently developed methodology of smooth transition regression allows for continuous
smooth changes in regimes. Therefore, it lends itself very well to modelling structural breaks,
asymmetries in dynamics of variables, and many other applications. Additionally, the
methodology easily incorporates the possibility of regime reversals. The recent extensions of
the methodology include VAR with smooth transition and panel smooth transition
regressions.

Numerous fruitful applications have been found for smooth transition regression models in
economics, ranging from modelling exchange rate dynamics and asymmetries in sectoral
wage structure, to nonlinear Phillips Curve, Okun’s Law, and nonlinear disaggregated models
of business cycles. In particular, many papers apply the modelling strategy to exchange rates.

31
We illustrate the methodology on the example of Okun’s Law for Germany. We find that
substantial increases in unemployment during the studied period, including the reunification
of Germany, indicate substantial structural changes in the economy. Changes in the transition
function closely follow major increases in unemployment, reflecting structural breaks such as
the reunification of Germany, oil shocks, and the restrictive monetary policy of the eighties.

References:

Ahmad, Y., S. Glosser (2007), Searching for Nonlinearities in Real Exchange Rates.
Manuscript, University of Wisconsin: Whitewater, WI.
Anderson, H. M., F. Vahid (1998), Testing Multiple Equation Systems for Common
Nonlinear Components. Journal of Econometrics, Vol.84, pp. 1-36.
Arghyrou, M. G., V. Boinet, C. Martin (2005), Beyond purchasing power parity: Nominal
exchange rates, output shocks and non linear/asymmetric adjustment in Central Europe.
Money Macro and Finance Research Group Conference, paper No. 35.
Baum, C. F., J. T. Barkoulas, M. Caglayan (2001), Nonlinear adjustment to purchasing power
parity in the post Bretton-Woods era. Journal of International Money and Finance, Vol. 20(3),
pp. 379-399.
Berthold, N., F. Rainer (2002), Unemployment in Germany: Reasons and Remedies. CESifo
Conference center, Munich.
Blanchard, O. und J. Wolfers (1999), The Role of Shocks and Institutions in the Rise of
European Unemployment: The Aggregate Evidence. NBER Working Paper 7282, Cambridge,
MA.
Böhm, B. (2001), Structural Change in the Austrian Inflation Process. In G. Chaloupek, A.
Guger, E. Nowotny (eds.). Ökonomie in Theorie und Praxis, Berlin: Springer Verlag, pp. 31-
46.
Camacho, M. (2004), Vector Smooth Transition Regression Models for US GDP and the
Composite Index of Leading Indicators. Journal of Forecasting, Vol. 23, pp. 173-196.
Chen, S.L., J.L. Wu (2005), Long-run Money Demand Revisited: Evidence from a Non-linear
Approach. Journal of International Money and Finance, Vol. 24, pp. 19-37.
Corrado, L., S. Holly (2003), Nonlinear Phillips Curves, Mixing Feedback Rules and the
Distribution of Inflation and Output. Journal of Economic Dynamics and Control, Vol. 28, pp.
467-492.
Cuaresma, J.C. (2003), Okun’s Law Revisited. Oxford Bulletin of Economics and Statistics,
Vol. 65, pp. 439-451.
Dolado, J. J., M. D. Ramon, M. Naveira (2005), Are Monetary-Policy Reaction Functions
Asymmetric? The Role of Nonlinearity in the Phillips Curve. European Economic Review,
Vol. 49, pp. 485-503.
Eitrheim, O., T. Teräsvirta (1996), Testing the adequacy of smooth transition autoregressive
models. Journal of Econometrics, Vol. 74, pp. 59-76.

32
Engel, R.F. (1982), Autoregressive Conditional Heteroskedasticity with Estimates of the
Variance of U.K. Inflation. Econometrica, Vol. 50, pp. 987-1008.
Estrella, A., F. S. Mishkin (1998), Predicting U.S. recessions: financial variables as leading
indicators. Review of Economics and Statistics, Vol. 80, pp. 45-61.
Fok D., D. van Dijk, P. H. Franses (2005a), A multi-level panel STAR model for US
manufacturing sectors. Journal of Applied Econometrics, Vol. 20(6), pp. 811-827.
Fok D., D. van Dijk, P. H. Franses (2005b), "Forecasting aggregates using panels of nonlinear
time series. International Journal of Forecasting, Vol. 21(4), pp. 785-794.
Frisch, H. (1990), Theories of Inflation. Cambridge University Press: Cambridge.
Gonzalez, A., T. Teräsvirta, D. van Dijk (2005), "Panel Smooth Transition Regression
Models," Research Paper Series 165, Quantitative Finance Research Centre, University of
Technology, Sydney.
Granger, C.W., T. Teräsvirta (1993), Modelling nonlinear economic relationships. Oxford
University Press: Oxford.
Grant, A.P. (2002), Time-varying estimates of the natural rate of unemployment: A
revisitation of Okun’s law. The Quarterly Review of Economics and Finance, Vol. 42, pp. 95-
113.
Guerra, R. (2003), Nonlinear adjustment towards purchasing power parity: the Swiss Franc-
German Mark case. Swiss Journal of Economics and Statistics, Vol. 139(1), pp. 83-100.
Hansen, B.E. (2000), Sample Splitting and Threshold Estimation. Econometrica, Vol. 68(3),
pp. 575-603.
Huh, H.S. (2002), Estimating Asymmetric Output Cost of Lowering Inflation for Australia.
Southern Economic Journal, Vol. 68, pp. 600-616.
Hunt, J. (1999), Determinants of Non-employment and unemployment durations in East
Germany. NBER Working Paper No. W7128.
Kavkler, A., B. Böhm (2005), A nonlinear monetary model of inflation. In: L. Zadnik Stirn
and S. Drobne (eds). The 8th International Symposium on Operational Research Proceedings.
Ljubljana: Slovenian Society Informatika (SDI), Section for Operational Research (SOR), pp.
291-296.
Kavkler, A., B. Böhm (2005), Nonlinear Okun's law: a smooth transition regression approach.
In: Neichev, S. (ed.). Economic development and growth: 4th International Symposium
Economy & Business Proceedings. Bourgas: Bulgarian Academy of Science, pp. 12-21.
Kilian, L., M. P. Taylor (2003), Why is it so difficult to beat the random walk of exchange
rates? Journal of International Economics, Vol. 60(1), pp. 85-107.
Lahtinen, M. (2006), The purchasing power parity puzzle: a sudden nonlinear perspective.
Applied Financial Economics, Vol. 16(1/2), pp. 119-125.
Leon, H., S. Najarian (2005), Asymmetric adjustment and nonlinear dynamics of real
exchange rates. International Journal of Finance & Economics, Vol. 10(1), pp. 15-39.
Legrenzi, G., C. Milas (2004), Non-linear real exchange rate effects in the UK labour Market,
manuscript, Keele University, Great Britain.

33
Liew, V. K. S., T. T. L. Chong, K. P. Lim (2003), The inadequacy of linear autoregressive
model for real exchange rates: empirical evidence from Asian economies. Applied
Economics, Vol. 35(12), pp. 1387-1392.
Liew, V. K. S., A. Z. Baharumshah, K. P. Lim (2004), On Singapore dollar-US dollar and
purchasing power parity. Singapore Economic Review, Vol. 49(1), pp. 71-84.
Lin, C.F., T. Teräsvirta (1994), Testing the constancy of regression parameters against
continuous structural change. Journal of Econometrics, Vol. 6, pp. 211-228.
Luukkonen, R., P. Saikkonen, T. Teräsvirta (1998), Testing linearity against smooth transition
autoregressive models. Biometrika, Vol. 75, pp. 491-499.
Mayes, D.G., M. Viren (2002), Asymmetry and the Problem of Aggregation in the Euro Area.
Empirica, Vol. 29, pp. 47-73.
McLeod, A.J., W.K. Li (1983), Diagnostic checking ARMA time series models using squared
residual correlations. Journal of Time Series Analysis, Vol. 4, pp. 269-273.
Paya, I., I. A. Venetis, D. A. Peel (2003), Further evidence on PPP adjustment speeds: the
case of effective real exchange rates and the EMS. Oxford Bulletin of Economics and
Statistics, Vol. 65(4), pp. 421-437.
Paya, I., D. A. Peel (2005), The process followed by PPP data. On the properties of linearity
tests. Applied Economics, Vol. 37(20), pp. 2515-2522.
Peel, D. A., I. A. Venetis (2005), Smooth transition models and arbitrage consistency.
Economica, Vol. 72 (287), pp. 423-430.
Prachowny, M.F.J. (1993), Okun’s Law: Theoretical Foundations and Revised Estimates.
Review of Economics and Statistics, Vol. 75, pp. 331-336.
Rapach, D. E., M. E. Wohar (2006), The out-of-sample forecasting performance of nonlinear
models of real exchange rate behaviour. International Journal of Forecasting, Vol. 22 (2), pp.
341-361.
Sarno, L. (1999), Adjustment Costs and Nonlinear Dynamics in the Demand for Money: Italy,
1861-1991. International Journal of Finance and Economics, Vol. 4, pp. 155-177.
Sarno, L. (2003), Nonlinear exchange rate models. A selective overview. IMF Working
Papers 111.
Sarno, L., M. P. Taylor (2002), Purchasing power parity and the real exchange rate. IMF Staff
Papers 49, pp. 65-105.
Sarno, L., M.P. Taylor, D.A. Peel (2003), Nonlinear Equilibrium Correction in US Real
Money Balances, 1869-1997. Journal of Money, Credit, and Banking, Vol. 35, pp. 787-799.
Schaling, E. (2004), The Nonlinear Phillips Curve and Inflation Forecast Targeting:
Symmetric versus Asymmetric Monetary Policy Rules. Journal of Money, Credit, and
Banking, Vol. 36, pp. 361-386.
Siebert, H. (1997), Labor Market Rigidities – At the Root of Unemployment in Europe.
Journal of Economic Perspectives, Vol. 11 (3), pp. 37-54.
Sögner, L. A. Stiassny (2002), An Analysis on the Structural Stability of the Okun’s Law - a
Cross-Country Study. Applied Economics, Vol. 14, pp. 1775-1787.
Solow, R. (2000), Unemployment in the United States and in Europe: A Contrast and the
Reasons. CESifo Working Paper 231.

34
Sollis, R. (2005), Evidence on purchasing power parity from univariate models: the case of
smooth transition trend stationarity. Journal of Applied Econometrics, Vol. 20(1), pp. 79-98.
Taylor, M. P. (2006), Real exchange rates and purchasing power parity: mean-reversion in
economic thought. Applied Financial Economics, Vol. 16(1/2), pp. 1-17.
Taylor, M. P., D. A. Peel, L. Sarno (2001), Nonlinear mean-reversion in real exchange rates:
toward a solution to the purchasing power parity puzzles. International Economic Review,
Vol. 42(4), pp. 1015-1042.
Teräsvirta, T. (1998), Modelling economic relationships with smooth transition regression. In:
Handbook of Applied Economic Statistics. Marcel Dekker: New York. pp. 507-522.
Tsay, R.S. (1998), Testing and Modeling Multivariate Threshold Models. Journal of the
American Statistical Association, Vol. 93, pp. 1188-1202.
van Dijk, D. (2001), Smooth Transition Models: Extensions and Outlier Robust Inference.
Tinberg Institute: Amsterdam.
Weise, C.L. (1999), The Asymmetric Effects of Monetary Policy: A Nonlinear Vector
Autoregression Approach. Money, Credit and Banking, Vol. 31, pp. 85-108.

35
Authors:

Alenka Kavkler
Faculty of Economics and Business
Razlagova 14
2000 Maribor
Slovenia
tel: +386 2 2290-314
fax: +386 2 2510-461
e-mail:[email protected]

Peter Mikek
Wabash College
301 W. Wabash Avenue
Crawfordsville, IN 47933
USA
tel.: +1 317 765-361-6120
e-mail: [email protected]

Bernhard Böhm
Institute for Mathematical Methods in Economics
University of Technology-Vienna
Argentinierstr. 8
1040 Vienna
Austria
tel: +43 1 58801-11945
fax: +43 1 58801-11999
e-mail: [email protected]

Darja Borsic
Faculty of Economics and Business
Razlagova 14
2000 Maribor
Slovenia
tel: +386 2 2290-345
fax: +386 2 2516-141
e-mail: [email protected]

36