3. Econometric methodology
We use an unrestricted reduced-form, nonlinear vector autoregression
following a smooth transition autoregressive form. The
nonlinearity is based on the fact that the dynamic behavior of time
series depends on states or regimes of the variables. Terasvirta and
Anderson (1992) used such a nonlinear model in a single equation
framework, but here we follow Weise (1999) and extend it to a multiequation
setting.
By ignoring moving average terms, our logistic smooth transition
vector autoregressive model is as follows
Δxt = a1 + Σ
p
i=1
Γ1;iΔxt−i
+ a2 + Σ
p
i=1
Γ2;iΔxt−i
Gðzt ; c; γÞ + et ;
ð1Þ
where xt is a (k×1) vector of time series and G(zt ; c,γ) is a function
which lies between zero and one, with these two extreme valuescorresponding to two regimes. We assume that G(zt ; c,γ) is a logistic
function expressed by
Gðzt ; c; γÞ =
1
ð1 + expf−γðzt−cÞgÞ ð2Þ
where c is the ‘threshold’ parameter around which the dynamics of the
model change and γ is the ‘smoothness’ parameter the value of which
indicates whether the transition between the two states is instantaneous
or smooth. If γ→0, the model becomes linear and if γ→∞, the
logistic function reduces to the indicator function, Iðzt N cÞ. Themodel's
behavior changes if zt is greater than or less than the threshold value. If
(zt−c)→−∞, then G(zt ; c,γ)→0 and if (zt−c)→∞, then G(zt ;
c,γ)→1.
Fig. 4 clearly indicates that the United States economy experienced
periods of both low and high oil price volatility during our sample
period. So we perform a test of linearity in the baseline VAR, using the
oil price volatility series as the switching variable. In doing so, we test
the null hypothesis H0 :γ=0 against the alternative H1 :γN0 in
Eq. (1). As in Weise (1999), our tests of linearity are F versions of the
Lagrange multiplier tests described in Granger and Terasvirta (1993)
in the context of a single-equation framework. For the system as a
whole, we test the null hypothesis H0 :γ=0 using a log-likelihood
test. As can be seen in Table 2, linearity is rejected for any lag of the oil
price volatility series in the system as a whole (see the last column)
and also individually rejected for the Δln oilt and volt equations (see
the second and fourth columns). This is strong evidence against
linearity and in favor of our logistic smooth transition VAR. In what
follows, we use the volt−3 series as the switching variable, as our
model does not converge with other lags of the oil price volatility
series. Other variables may also be plausible candidates for the
switching variable but we are interested in oil price volatility series in
order to analyze the asymmetric response of output to oil price shock
in the presence of volatility in oil prices.