The Panel Vector Autoregression (PVAR) Model
There are three considerations taken in developing this empirical study. First, identify variables customised in previous studies on the real sector and monetary policy shock. Ref. [
38] examined the impact of a negative Chinese GDP shock as a real sector shock. Ref. [
35] investigated the interaction mechanism of economic and trade shocks between China and ASEAN countries using real output as the real GDP index and foreign trade fluctuation by using import and export growth. Refs. [
47,
48] examined interest rates to formulate monetary policy shocks. Second, the data source for the Chinese shocks to the BRI economic activities for these empirical studies needs to be restricted to certain important variables within this framework to describe all the shocks due to data availability. Lastly, the model should identify potential endogeneity problems for all the variables used.
Following a few articles that investigate the interaction of variables [
14,
38,
39], our econometric methodology is based on the panel VAR framework, and we utilise a panel VAR generalised method of moments (GMM) estimator to estimate the model. The PVAR model is given by:
and to be simplified into:
Zit is an endogenous variables matrix relating to the real sector (cgdpt, copent, gdpi,t, cpii,t, openi,t) and monetary policy (cnirt, dniri,t) shocks. A(L) is a polynomial matrix in the lag operator L, with country i = 1, …50, μi is a vector of unobserved fixed effects, and eit is a vector of random errors. Their parameter will suffer from estimation bias if the traditional mean difference is adopted due to μi being correlated to the lag term. Hence, forward orthogonal deviation (Helmert Transformation) is used to eliminate the μi.
Let
be the means obtained from the future values of
which is a vector variable in the vector
with
is the data sample last period. In addition,
is the mean obtained from the future values of
which is a vector variable in the vector
. Next, the transformed variable will become
−
and
−
respectively. Considering there is no future value for creating the forward mean, we cannot derive the transformation value of the last period data. Consequently, the final transformed model is as follows:
where
and
.
The advantage of the PVAR method stems primarily from its remedy on the considered potential endogenous variable. PVAR is said to be able to gain a degree of freedom by analysing a country’s panel. Additionally, we can identify a spillover effect from one country to another because it captures unobserved country-level heterogeneity [
49]. Furthermore, the PVAR method can provide dynamic interaction between Chinese shocks from the real sector, monetary policy, and other macroeconomic factors from the BRI countries. The technique is parsimonious and straightforward and appears to fit the estimation suitably.
The estimation process in the PVAR model above (5) requires four steps. The initial step is to ensure all variables discussed have suitable temporal properties where panel unit root will be implemented [
50]. In the second step, we need to set the PVAR lag order using the optimal moment and model selection criteria (MMSC). For this, three criteria can be considered as proposed by [
51], namely the Bayesian information criterion (MBIC), the Akaike information criterion (MAIC), and the Hannan-Quinn information criterion (MQIC). The GMM estimator is then applied in the PVAR technique to solve the endogeneity problem in estimating the model. Forward mean-differencing or Helmert transformation is used to trace out the individual fixed effects as suggested by [
49]. The Helmert transformation can conserve the orthogonally transformed variables with the lagged dependent variables used as an instrumental variable in this GMM estimation. The third step is to verify the stability condition for all the variables used in the PVAR estimation. Finally, the dynamic relationship between Chinese shocks and other factors in the BRI countries has been examined through the impulse response function (IRF) and the forecast error variance decomposition (FEVD).
Both shocks from the real sector and monetary policy need to be ordered correctly into structural shocks in the PVAR system using the Cholesky decomposition of the variance-covariance matrix method. Hence, the sequence will start with Chinese shocks from its real sector (cgdp and copen) and its monetary policy shocks (cnir) and transmitted to the macroeconomics variables (gdp, open, nir, and cpi) of BRI countries. This comprises a total of seven variables. The real sector shock is placed first for orderliness since it usually has no immediate effect and responds very slowly. Next is the monetary policy since the government usually has a time lag for transmitting information, thus not appropriate as the initial response. Thus the ordering system adopted is initiated with the least endogenous variable and ends with the most endogenous one.
In each model, Chinese shocks (represented respectively by
gdp, trade openness, and nominal interest rate) will be ordered first because, being a large country, it has the least endogenous variable, as shown in Equation (7). The assumption is that a large country like China is highly exogenous towards a smaller one. Subsequently, the gross domestic product will be in the first order on the small country variable since it usually has no immediate effect and responds very slowly. Monetary policy will be next in line since the government usually has a time lag for the information to respond initially. The consumer price index follows it since it is the least responsive among the remaining two variables, given its slow evolution over time. The consumer price index is assumed to react with lags to innovations in the monetary policy. Finally, trade openness is placed last because of its response according to the exchange rate and the most responsive to all shocks. The matrix below does the Cholesky decomposition as follows:
Based on the Cholesky decomposition method, the value of the matrix in the diagonal form is set to zero, and the free parameters are below the diagonal. In this case, a, c, f, j, and o are diagonals representing their shocks in the system. Other alphabets in the matrix are the free parameters. The error terms of structural shock in Equation (7) are isolated variables in the system (recursive structure), becoming orthogonal as explained by the alphabetically ordered weights linked to the model’s structural shock and retaining all other responses at zero value. This concept confines the variables at first ordering, which will affect other variables contemporaneously and with an exogenous lag. At the same time, delayed factors will affect the front variables with an endogenous lag. Equation (7) specified the immediate shocks from = dlncgdp, dcopen, and dcnir will have a contemporaneous effect on dlngdp, dnir, dlncpi, and dopen. The dlngdp, dnir, dlncpi, and dopen will deliver a lagged effect on the dlncgdp, dcopen, and dcnir, respectively.
Following the panel VAR estimation, the IRF and FEVD analyses will be conducted. The variables will be in an alternative ordering in the Cholesky decomposition to check the robustness of previous findings. The IRF measures the impact of the shock from one variable on another future value of the endogenous variables while holding other shocks constant. The different graphs of the IRF are shown with 95% confidence interval bands implemented by Monte Carlo simulation. FEVD is used to assess the role of each random variable in affecting the variables in the VAR system. In contrast, IRF describes the responses by the variable to the effect of a shock in VAR, which is the percentage of variable variation accumulated over time as explained by the shock from another variable.