国贸专业英文文献翻译

case study using a VAR model

SUMMARY

In this study， a vector auto regression model (VAR) and a vector error correction model (VECM) were estimated to examine the impact of oil price fluctuations on seven key macroeconomic variables for the Kuwaiti economy. Quarterly data for the period 1984 - 1998 were utilized. Theoretically and empirically speaking， VECM is superior to the VAR approach. Also， the results corresponding to the VECM model are closer to common sense.

However， the estimated models indicate a high degree of interrelation between major macroeconomic variables. The empirical results highlight the causality running from the oil prices and oil revenues， to government development and current expenditure and then towards other variables. For the most part， the empirical evidence indicates that oil price shocks and hence oil revenues have a notable impact on government expenditure， both development and current. However， government development expenditure has been influenced relatively more.

The results also point out the significance of the CPI in explaining a notable part of the variations of both types of government expenditure. On the other hand， the variations in value of imports are mostly

1

accounted for by oil revenue fluctuations and then by the fluctuation in government development expenditures. Also， the results from the VECM approach indicate that a significant part of LM2 variance is explained by the variance in oil revenue. It reaches about 46 percent in the 10th quarter， even more than its own variations.

KEY WORDS： vector auto regression (VAR); oil fluctuation; Kuwait

1. INTRODUCTION

The post-1973 effects of the oil boom on the economies of Arab oil producing countries have been diverse， though on balance， many of those governments might look back on the period 1973 – 1986 as a mixed blessing. Income on the oil account certainly rose rapidly， but so did price inflation， wage rates and reliance on foreign labor. Above all， the growth of the oil sector as a contributor to national income tended to reduce the role of nonoil sectors to insignificance in most Arab states of the Gulf. This phenomenon has been termed in the literature ‘the Dutch Disease ’ . Dramatic rises in per capita income were the fruits of rising oil revenues alone， even in the case of the larger more diversified economies of the Gulf such as Iran (Al-Abbasi， 1991).

There is a great deal of theoretical and empirical literature scrutinizing various aspects of the Dutch Disease economies such as Cordon and Neary (1982)， Hamilton (1983)， Neary and van Wijnbergen (1986)， Fardmanesh (1991)， Van Wijnbergen (1984)，

2

Gelb and Associates (1988) and Taylor et al. (1986) to name a few. Recently， several empirical studies have been published on Arab oil producing countries. For instance， Taher (1987) studied the impacts of changes in the world oil prices on the different sectors of the Saudi economy. Furthermore， Al-Mutawa (1991) and Al-Mutawa and Cuddington (1994) analysed the effects of oil shocks and macroeconomic policy changes for the UAE. The results showed that， in the case of UAE， an oil-quantity boom led to higher welfare gains than an oil-price boom. Moreover， an oil-price or quantity bust always led to lower economic growth and created a welfare loss. Also， Al-Mutairi (1993) attempted to identify the sources of output fluctuations and the dynamic response of the economy to changes in key macroeconomic variables for Kuwait. His empirical results suggested that for short horizons of one and two years， shocks to the oil price account for more than 50 per cent of the variance of GDP. However， at longer horizons of three years and more， these shocks are seen to be unimportant in inducing GDP fluctuations， accounting only for less than 10 per cent of the variance. Shocks of real-government expenditure were also found to have a significant role in causing GDP fluctuations.

Kuwait is a typical example of an oil-based economy. The oil sector contributes over two-thirds of GDP and over 90 per cent of exports. Although Kuwait tries hard to lessen its dependence on oil through the development of a non-oil sector， its success has so far been， at the best， very modest. The real problem is that oil prices and hence oil

3

revenues are exogenously determined. As a member of OPEC， Kuwait has no control over the price of its crude oil and at least theoretically speaking cannot exceed its assigned production quota. The objectives of this study are to investigate the impacts of oil price fluctuations on key macroeconomic variables of the Kuwaiti economy， to examine the direction of causality and to determine the significance of such impacts. This will certainly enhance our understanding of how international oil price fluctuations impact key macroeconomic variables and the dynamic response of these economic variables， including policy variables such as government expenditure and money supply.

In this study， the analysis is carried out using two different models， namely， the vector autoregres-sion model (VAR) and the vector error correction model (VECM). The VAR technique is appropriate in this case because of its ability to characterize the dynamic structure of the model as well as its ability to avoid imposing excessive identifying restrictions associated with different economic theories. The use of VAR in macroeconomics has generated much empirical evidence， giving fundamental support to many economic theories (see Blanchard and Watson， 1984， Bernanke， 1986 among others).

In the next section， a brief review of the literature is presented followed by the VAR model along with the data utilized. The empirical results and their interpretation are given in section four， followed by the conclusions.

4

2. THE MODEL

2.1. The background of VAR methodology

The VAR system is based on empirical regularities embedded in the data. The VAR model may be viewed as a system of reduced form equations in which each of the endogenous variables is regressed on its own lagged values and the lagged values of all other variables in the system. An n variable VAR system can be written as

A(l)Yt=A+Ut (1) A(l)=l-A1l-A2l2-Amlm (2)

where Yt is an n × 1 vector of macroeconomic variables， A is an n × 1 vector of constraints， and Ut is an n × 1 vector of random variables， each of which is serially uncorrelated with constant variance and zero mean. Equation(2) is an n × n matrix of normalized polynomials in the lag operator l with the first entry of each polynomial on A ' s being unity.

Since the error terms (Ut) in the above model are serially uncorrelated， an ordinary-least-squares (OLS) technique would be appropriate to estimate this model. However， before estimating the parameters of the model A(l) meaningfully， one must limit the length of the lag in the polynomials. If l is the lag length， the number of coefficients to be estimated is n(nl +c)， where c is the number of constants.

In the VAR model above， the current innovations (Ut) are unanticipated but become part of the information set in the next period.

5

This implies that the anticipated impact of a variable is captured in the coefficients of lagged polynomials while the residuals capture unforeseen contem-poraneous events. Therefore， an important feature of the VAR methodology is the use of the estimated residuals， called VAR innovations， in dynamic analysis. Unlike in conventional economic modelling， these VAR innovations are treated as an inherent part of the system.

In order to analyse the impact of unanticipated policy shocks on the macroeconomic variables in a more convenient and comprehensive way， Sims (1980) proposed the use of impulse response functions (IRFs) and forecast error variance decompositions (FEVDs). IRFs and FEVDs are obtained from a moving average representation of the VAR model [Equations (1) and (2)] as shown below

Yt=Constant+Ht(l)U (3)

And

H(l)=I+Htl+H2l (4)

Where H is the coefficient matrix of the moving average representation which can be obtained by successive substitution in Equations (1) and (2). The elements of the H matrix trace the response over time of a variable i due to a unit shock given to variable j. In fact， these impulse response functions will provide the means to analyse the dynamic behaviour of the macroeconomic variables due to unanticipated shocks in the exogenous variables.

Having derived the variance-covariance matrix from the moving-average representation， the FEVDs can be constructed.

6

FEVDs represent the decomposition of forecast error variances and therefore give estimates of the contributions of distinct innovations to the variances. Thus， they can be interpreted as showing the portion of variance in the prediction for each variable in the system that is attributable to its own innovations and to shocks to other variables in the system.

2.2. Vector error correction methodology

Dickey and Fuller (1979) have emphasized the necessity of analysing the time-series properties of the variables before their relationship can be established. This is necessary because if the variables in question are nonstationary， then the estimated equations will yield spurious and misleading regression results. If the variables in a relationship are stationary then it is generally true that any linear combination of these variables is said to be cointegrated. Johansen’s test (1991， 1995) is commonly used to test for cointegration between more than two time series. It also provides estimates of the possible long-term relationships， i.e. the parameters of the relationships that ensure cointegration. In this study， a vector error correction model (VECM) was also estimated.The VECM is basically a VAR system that builds on Johansen ' s test for cointegration and is usually referred to in the literatures as the restricted VAR. 2.3. The estimated model and data

The first step in estimating a VAR model is to make a choice of the macroeconomic variables that are essential for the analysis. The

7

variables consist of one external shock measured by innovations in the price of Kuwaiti blend crude oil， three key macroeconomic variables， oil revenues， the consumer price index， (CPI) and the value of imports and three policy variables， Money Supply M2， government current expenditure and government development expenditure. The notations of these variables are as follow： OILP = Oil Price of Kuwaiti Blend Crude OILR= Oil Revenue

EXDEV = Government Development Expenditure EXCON =Government Current Expenditure CPI = Consumer Price Index M2 =Money Demand (M2 Definition)

IMPORTS = Value of Imports of Goods & Services

Quarterly data for the period 1984：1-1998：4 were utilized in this study. The data for the period of the Iraqi occupation and the liberation of Kuwait were removed from the time series for obvious reasons (1990-1991). All data are from the Quarterly Monetary Statistics of the Central Bank of Kuwait and OPEC’s Monthly Bulletin. Similar to the previous studies， all the variables are expressed in logarithmic form. This can be partially justified by the fact that logarithmic forms tend to reduce the scale of the variables， which is a desirable quality when analysing the time-series properties of the variables before their relationship can be established. It is also a useful tool in providing estimates of the possible long-term relationships， i.e. the parameters of the relationships that ensure co-integration.

8

A very important point that should be mentioned here is that the major shortcoming of the VAR approach is its lack of theoretical substance (Cooly and LeRoy， 1985; Leamer， 1985). In response to this criticism， Blanchard and Watson (1984) and Bernanke (1986) developed procedures， called the structural vector autoregression (SVAR) approach， which combines the features of the traditional structural modelling with those of the VAR methodology. The major advantage of using SVAR comes from the fact that standard VAR disturbances are generally characterized by contemporaneous correlations. In the presence of such correlations， the response of the system， indicated by IRFs， to an innovation in one of the variables is in fact the response to innovations in all those variables that are contemporaneously correlated with it. Similarly， the ability of FEVDs to quantify the relative contributions of specific sources of variation is confounded in the presence of this correlation.

However， in standard VAR methodology this contemporaneous correlation is purged by the Cholesky orthogonalization procedure. For the most part， the Cholesky procedure implicitly assumes recursivity in the VAR model as it is estimated. Although theoretical considerations may help in determining this ordering and ex-post sensitivity analysis may further help provide insights regarding appropriate ordering， it remains largely at the discretion of the modeller.

The following ordering of equations was adopted in this study; LOILP， LOILR， LEXEDEV， LEXCON， LCPI， LM2 and

9

LIMPORTS. Generally speaking， this ordering reflects the fact that oil prices have an influence on oil revenues and then on all the other variables in the model. However， the behavior of oil prices and to some extent oil revenue are the least determined by other variables included in the model. This is quite a plausible assumption because the oil prices and hence oil revenues which consist of oil export revenues and net factor income from abroad are largely determined by the world market conditions rather than within the Kuwaiti Economy.

Similarly， this ordering assumes that the government expenditure is largely determined by the level of oil revenues which again is quite a plausible assumption considering the dominant role of the public sector in driving the Kuwaiti economy. It is also sensible to assume that the value of imports is largely dependent on the level of government expenditure.

Since the only variables included are those suggested by economic theory， and since theoretical considerations are important in selecting the ordering used here， the SVAR is not followed in this study. Nevertheless， the approach utilized here can be considered to be in the spirit of the SVAR approach.

3. THE EMPIRICAL RESULTS

First， the VAR technique requires stationary data， thus each series should be examined for stationarity. Table I gives the nonstationary test for all the time-series， using the conventional Dicky-Fuller test (DF)， its augmented version (ADF) and Phillips-Parron t-tests. These

10

tests include a constant but no time trend， as recommended by Dickey and Fuller (1979).

First， the reported t-statistics， when compared with the critical values obtained by Engle and Yoo (1987)， indicate that almost all the series， except CPI， M2 and IMPORTS， are stationary in the levels as shown by the DF， ADF and Phillips-Perron t-tests. These tests are reapplied after differencing all terms. The t-statistics on the lagged first-difference terms indicate that， for all series the null hypothesis is rejected， that is to say， all series are first differences stationary. However， in transforming a variable， a usual question arises as to whether one should use the variables in the system in levels or in differences. The overall guideline is that if there are k number of cointegrating vector among the variables used in the system， then VAR could be modelled with k stationary and n-k differences of original variables. But if all the variables in the system are nonstationary， using a VAR in levels is appropriate. On the other hand， estimating a VAR in the levels in the case of cointegration may lead to the omission of important constraints.

In this context， Doan et al. (1984) noted that differencing a variable is ‘important ' in the case of Box-Jenkins ARIMA Modelling. Doan et al. also observed that it is not desirable to do so in VAR models. Fuller (1976) has also shown that differencing the data may not produce any gain so far as the ‘asymptotic efficiency’ of the VAR is concerned ‘even if it is appropriate’. Moreover，Fuller has argued that differencing a variable ‘ throws information away ' while producing no

11

significant gain. Thus， following Doan and Fuller’s argument， the level rather than the difference was utilized here.

Second， the estimation of a VAR model requires the explicit choice of lag length in the equations of the model. Following Judge et al. (1988) and McMillin (1988)， Akaike’s AIC criterion is used to determine the lag length of the VAR model. The chosen lag length is one that minimizes the following：

AIC(n)=Indet∑+(2d2n)/T

where d is the number of variables in the model，T the sample size and∑n an estimate of the residuals’ variance-covariance matrix ∑n obtained with a VAR (n). The maximum lag length is set at five quarters， considering the sample size and number of variables in the model. A maximum lag of greater than five quarters would reduce the degrees of freedom for estimation unacceptably. The result of employing this technique is summarized in Table II， which shows the corresponding AIC values. It can be seen that the AIC criterion is minimized for order 4. This suggests that， for this study， the VAR model should be of order 4.

The next step is to estimate the VAR. The estimates along with their t-values are reported in Table III. Although the estimates of individual coefficients in VAR do not have a straightforward interpretation， a glance at the table generally shows that most of the t-values are significant (except for the CPI equation) and almost all of the equations have high R-squares. It also confirms the assertion that oil prices are exogenously determined than other variables included in the

12

model. However， oil revenue equation has a larger number of significant t-values than current expenditure and CPI.

3.1. Variance decomposition

Table IV presents the variance decomposition for the 10-quarters forecasts. Table IV shows that initially the variations in all of the variables are typically explained by the variables’ own trends. That is to say， at the beginning， the historical trend of each variable explains a large part of its own variations. For the most part， after ten quarters， about 60 per cent of the variance in oil prices is explained by the variable itself which is indicative of its exogenous nature. On the other hand， oil revenue explained about 93 per cent of its own variations at the first quarter and about 20 percent at the 10th quarter. Moreover， variation in oil prices account for about 45 per cent of the variation in oil revenues starting at the second quarter and through to the tenth. This shows that the causality is running from oil prices to oil revenues. Similar results are also evident for government development and current expenditures. Over a time period， about 15-17 per cent of the variance in government expenditures (development and current) is accounted for by the variations in the oil revenues.

Furthermore， looking at the variance decomposition in the government expenditures (development and current) it is observed that following their own variations and oil revenues， the LCPI account for a notable part of the expenditures’ variance. The CPI accounts for one

13

fifth of the current expenditure variations and about 15 per cent of the development expenditure. This is quite a plausible result and very apparent in the case of current expenditure.

On the other hand， it is worth noting that the other variable that also picks up a significant part of the variations in government development expenditure is the value of imports. It accounts for about 16 per cent.

Looking at the variance decomposition of M2， it is apparent that a noticeable part of its variance is explained by the variance in the CPI (about 33 per cent)， even more than its own variations. Also， oil prices and oil revenue， respectively， account for about 23 and 13 per cent of its variations. These results suggest an important role for money supply.

Moreover， over the time period， the results show that 25-45 per cent of the variance in the value of imports is accounted for by the variation in oil revenues alone. Other variables included in the model that exert significant influence on the behaviour of imports are the two types of government expenditure but in particular， the development expenditure.

3.2. Impulse responses

Figure 1 displays the Impulse Response Functions， which are essentially the dynamic multipliers .Since the primary interest is to see the response of major macroeconomic variables to the shocks given to the oil revenues and then to the government expenditure， only ten

14

time periods are reported. An inspection of Figure 1 reveals that innovation in the oil prices and hence oil revenue has a similar effect on most of the variables in the model. Generally， most of these variables show an increase in the first quarter. This increase continues in the second and third quarter and then it gradually tapers off over the successive quarters. The only exceptions are the CPI and the value of imports and M2. This may be attributable to the shortcomings in the data set used to estimate the VAR. Recall that LCPI， LM2 and LIMPORTS were found to be non-stationary in the level.

3.3. Estimation of the vector error correction model

Since most of the variables included in the model pertain to stationary time series data except LCPI， LM2 and LIMPORTS， Johansen’s test (1991， 1995) was applied to check for cointegrating vectors. The test indicated that there are four cointegrating vectors. Therefore， a vector error correction model is warranted. A vector error correction model is a VAR that build-in cointegration. Each co-integrating equation adds the parameters associated with the term involving levels of the series which needs to be added to each equation in the VAR. There is a sequence of nested models in this framework. The Johansen test procedure computes the likelihood ratio for each added co-integrating equation.

On the basis of Johansen’s test， a Vector error correction model (VECM) was estimated. Four co-integrating equations were estimated using the same seven variables that were used in the VAR. However，

15

since the results of estimating the VECM do not have a direct interpretation， they are not reported here.

3.4. Variance decomposition

The variance decomposition results corresponding to the estimated VECM are presented in Table V. They are based on the same ordering as was used in the VAR. Comparing these results with the VAR shows that while the qualitative nature of macroeconomic linkages remains almost the same， the intensity of interaction between the variables is significantly higher when co-integration has been accounted for. For example， looking at the variance decomposition of the oil revenues， it shows that variables like government development and current expenditures have a substantially larger share when compared with the VAR results which increased from 14 per cent to about 40 percent. Similarly， the oil revenue has picked up a relatively larger proportion of the variation in the government current expenditure as well as the value of imports， especially during the first 4-6 quarters. However， the results indicate that variations in LCPI， LM2 and LIMPORTS， after 7-10 quarters， are mainly explained by changes in oil revenue alone.

In particular， the variance decomposition of LM2 indicates that a significant part of its variance is explained by the variations in oil revenues (about 46 per cent after 10 quarters)， even more than its own variations. On the other hand， contrary to the result from VAR， LCPI accounts for only 7 per cent.

16

Empirically speaking， the VECM model shows a relatively higher degree of statistical significance. Theoretically speaking， this is because it yields a closer interaction between macroeconomic variables than what the VAR indicated.

3.5. Impulse response functions

Figure 2 displays the impulse response functions corresponding to the VECM model. Figure 2 indicates that innovations in the oil prices and oil revenue have a similar impact on the variables included in the model. However， similar to the VAR， most of the variables show an increase for the first few quarters then it gradually tapers off over the successive quarters with the exception of CPI， value of imports and LM2.

Comparing these IRFs with those corresponding to the VAR version reveals that it takes a little longer for the multipliers in the VECM version to reach the level of the VAR version. While they generally reached their peak in the VAR version in about 6-7 quarters， it took them 8-9 quarters to reach almost the same level in the VECM version.

4. CONCLUSIONS AND SOME POLICY

IMPLICATIONS

The primary goal of this paper was to investigate how macroeconomic variables react to fluctuations in the world oil prices. Therefore， two different versions have been estimated， namely，the VAR and the VECM. While the qualitative nature of macroeconomic

17

linkages remains almost the same in the two models， the intensity of interaction between the variables is significantly higher when cointegration has been accounted for. Thus， quantitatively the two models give results that are significantly different from each other. However， empirically， the VECM gives better results because it yields a closer interaction between macroeconomic variables than by the VAR estimation. The results corresponding to the VECM are also closer to common sense.

Nevertheless， the two versions estimated indicated a notable degree of interrelation between the major macroeconomic variables. The results have highlighted the causality running from oil prices towards oil revenues and then towards government expenditure and other variables. However， further assessment of the relationship between these variables， based on orthogonal innovations， lead us to believe that oil price shocks do impact macroeconomic variables in Kuwait and in particular， via government development and current expenditures.

The evaluation of the decomposition of the variance of government expenditures suggests that oil revenue fluctuations account for a notable part especially in the case of development expenditure. This result is not surprising and is actually consistent with what is expected in a country in which the government is the sole owner of the main national income source， the oil and gas industry. Thus， government expenditure becomes the major determinant of the level of economic activity and the mechanism by which the government can

18

effect the circular flow of income within the economy.

What is surprising， however， is that one would expect the impact of oil shocks to be much stronger and in particular， in the case of current government expenditure. However， this may be explained by the fact that over the last three decades， the government has accumulated capital reserve (surplus) which is regularly used to finance current government commitments， especially in times of low oil revenue.

Moreover， the results also point out the significance of the CPI in explaining a notable part of the variations of both types of government expenditure. On the other hand， the variations in value of imports are mostly accounted for by oil revenue fluctuations and then by the fluctuation in government development expenditures. These results suggest that fiscal policy appears to be effective in Kuwait as the oil shocks impact government expenditure and then government expenditure accounts for a relatively considerable part of the CPI and the value of import variations.

However， the results from VECM approach indicate that a significant part of LM2 variance is explained by the variance in oil revenue. It reaches about 46 per cent in the 10th quarter even more than its own variations.

This exercise has shown a high degree of sensitivity of the results to specification of the variables， i.e. the theoretical structure underlying the VAR. Part of this sensitivity may be attributable to shortcomings in the data set. However， for the most part， it reflects the limitations of

19

the VAR approach. For example， there are no standard statistical tests available to choose from among various versions and then these are not the only versions. One can obtain an infinitely large number of estimates. It remains at the discretion of the modeler to make a choice. Finally， the VAR and VECM results reported here demonstrate that the VAR approach， which is gaining popularity among the modelers， has serious problems when applied to a small economy， which is highly exposed to the world events and therefore has problems in conforming to the standard macroeconomic theoretical constructs.

--- M. Nagy Eltony and Mohammad

Al-Awadi．INTERNATIONAL JOURNAL OF ENERGY

RESEARCH， 2001， 25： 939-959.

20

石油价格波动及其对科威特的宏观经济

变量的影响：使用VAR模型的案例研究

摘要

在这项研究中，我们用向量自回归模型（VAR）和向量误差修正模型（VECM）来估计检查石油价格波动对科威特经济中的七个关键宏观经济变量的影响。并利用1984–1998期间内的季度数据进行分析。从理论和实证上讲，VECM模型优于VAR方法。因此，对应的向量误差修正模型的结果更接近于常识。

然而，估计的模型表明，主要宏观经济变量之间存在高度的相互关系。实证结果表明石油价格和石油收入的运行，的发展与经常性支出对其他变量有因果关系。在大多数情况下，实证证据表明，石油价格冲击以及石油收入对的支出都有显著影响。然而，发展支出一直受到相对更大的影响。

研究结果还指出了CPI在解释开支的两种类型中一个显著变化的重要组成部分的意义。另一方面，进口值的变化主要受石油收入波动和发展支出的波动的影响。同时，从误差修正方法的结果表明，很大一部分LM2方差是由石油收入的差异来解释。它在第十季度达到46%左右，比其自身的变化更大。 关键词：向量自回归（VAR）；油价波动；科威特；

1.引言

在1973后，石油繁荣对阿拉伯石油生产国经济影响已经多样化，但总的来说，许多可能回顾1973 - 1986年期间是喜忧

21

参半。当然石油帐户收入增长迅速，但价格通胀也一样，工资率和依赖外国劳动力。最重要的是， 在大多数海湾的阿拉伯国家石油行业的发展作为国家收入因素倾向于减少无意义的非石油部门的角色。这种现象被称为“荷兰病”。甚至在更大的更多样化的海湾经济体如伊朗，人均收入大幅增加是石油收入上升的成果， (Al-Abbasi，1991)。

有大量的理论和实证文献审查荷兰病经济体的各个方面如科登和尼瑞(1982)，汉密尔顿(1983)，尼瑞和范温伯根(1986)，Fardmanesh(1991)，范温伯根(1984)，盖尔布、Associates (1988)和泰勒 (1986)等等。

最近，一些实证研究在阿拉伯等石油生产国已经发表。例如，塔希尔(1987)研究了世界石油价格变化在沙特经济的不同领域的影响。此外，Al-Mutawa(1991)和Al-Mutawa Cuddington(1994)分析了石油波动和阿联酋宏观经济的变化的影响。结果表明，在阿拉伯联合酋长国、油量繁荣导致福利收益高于油价的繁荣。此外，油价或产量降低总是导致较低的经济增长和福利损失。

同时，Al-Mutairi(1993)试图识别输出波动的来源和经济的动态响应对科威特关键宏观经济变量变化的反应。他的实证结果表明，在短期内如一到两年时间里，油价冲击占GDP变化的50%以上。然而，在长期内如三年或以上，这些冲击在诱导GDP波动方面是不重要的，只占不到10%的份额。并发现在导致GDP波动上真正的支出的冲击占一个重要的角色。

科威特石油经济是一个典型的例子。石油行业贡献超过GDP的三分之二，超过90%的出口。尽管科威特努力通过非石油行业的发展减少对石油的依赖，其成果迄今为止是最好的，非常温和。

22

然而，真正的问题是油价，因此石油收入是外生确定。作为欧佩克成员国，科威特没有控制其原油的价格，至少理论上说不能超过其指定的生产指标。

本研究的目标是研究石油价格波动对科威特经济的关键宏观经济变量的影响，检验因果关系的方向和确定这些影响的重要性。这肯定会增强我们的了解国际石油价格波动对主要宏观经济变量和这些经济变量的动态响应的影响，包括变量如支出和货币供应。

在这项研究中，使用两个不同的模型进行分析，即向量自回归模型(VAR)和向量误差修正模型(结果)。VAR方法在这种情况下是适当的，因为它能够描述模型的动态结构，以及它能够避免与不同的经济理论施加过度识别。使用VAR在宏观经济学产生了太多的经验证据，提供了许多经济理论的基本支持 (参考布兰查德和华生，1984年、伯南克，1986年)。

在下一节中，先是简要回顾文献，紧随其后的是利用VAR模型和数据进行实证分析。实证结果和他们的解释有四节，最后是结论。

2.模型

2.1 VAR方法的背景

VAR系统是基于经验规律中嵌入数据。VAR模型可能被视为一个简化型方程组，每个内生变量的回归的滞后值和所有其他变量的滞后值系统。一个n变量VAR系统可以写成

A(l)Yt=A+Ut (1) A(l)=l-A1l-A2l2-Amlm (2)

23

Yt是一个n×1矢量的宏观经济变量，是一个n×1的向量的约束， Ut是一个n×1随机变量的向量，其中每个连续与不断变化和零的意思。方程(2)是一个n×n矩阵归一化多项式的滞后算子l与每个多项式的第一个条目的统一。

因为错误条件(Ut)在上面的串行模型是不相关的， ordinary-least-squares(OLS)技术将适当的估计该模型。然而，在估计模型的参数(l)时，一个人必须在多项式滞后的长度。如果滞后长度为l，系数的数量估计是n(n l + c)，c是常量。

在上面的VAR模型，当前的变量(Ut)在接下来的时期但成为信息集合的一部分。这意味着一个变量的预期影响捕获滞后多项式的系数而残差捕捉不可预见的同时发生的事件。因此，VAR方法的一个重要特性是使用估计的残差，称为VAR创新动态分析。与传统的经济模型相比，这些VAR创新被视为系统的一个固有部分。

为了以一个更方便和全面的方式分析未预料到的冲击对宏观经济变量的影响，西姆斯(1980)提出了(IRF)使用脉冲响应函数和预测误差方差分解(FEVDs)。IRF和FEVDs用获得的移动平均值表示VAR模型(方程(1)和(2))，如下所示

Yt=Constant+Ht(l)U (3) H(l)=I+Htl+H2l (4)

H是移动平均线表示的系数矩阵，可以得到连续的替换方程(1)和(2)。由于单位冲击变量j，H矩阵的元素跟踪随变量时间i响应。事实上，这些脉冲响应函数将提供方法分析由于意外冲击的外生变量导致的宏观经济变量的动态行为。

24

FEVDs可以构造派生的移动平均值的方差协方差矩阵表示法，。FEVDs代表预测误差方差的分解，因此给出不同的变量的估计方差的贡献。因此，他们可以被解读为显示每个变量的方差的预测系统部分，可归因于自己的变化和系统中其他变量和冲击。

2.2 向量误差修正方法

Dickey和富勒(1979)强调分析时间序列属性变量可以建立他们之间关系的必要性。这是必要的，因为如果问题中的变量不稳定，估计方程会产生虚假和误导性的回归结果。如果变量的关系是固定的，那么结果一般是这些变量的线性组合是共合体。约翰森的测试(1991、1995)通常用于测试超过两个时间序列之间的协整。它还提供了估计可能的长期关系，即确保协整关系的参数。在这项研究中， 也估计了一个向量误差修正模型(VECM)。VECM基本上是一个基于Johansen协整检验的VAR系统，通常成为的VAR系统。

2.3 估计模型和数据

一个VAR模型的第一步是选择分析必要的宏观经济变量。由科威特混合原油的价格衡量的一个外部冲击变量对三个关键的宏观经济变量（石油收入，消费者价格指数(CPI)和进口）和三个变量（当前货币供应量M2、支出和开发支出）的值的影响。这些变量的符号如下：

OILP =科威特混合原油的石油价格 OILR =石油收入您=开发支出 EXCON =当前的支出 CPI =消费者价格指数(CPI)

25

M2 =货币需求(M2定义)

IMPORTS=进口商品和服务的价值

这项研究所用的数据为1984：1-1998：4的季度数据。由于显而易见的原因删除了伊拉克占领时期和科威特时期的数据 (1990 - 1991)。所有数据都来自银行的季度货币统计以及科威特和欧佩克的月报。类似于之前的研究，所有的变量都是对数形式表示。采用对数形式可以合理的减少变量的规模，这是在可以建立他们之间的关系之前分析一个变量的时间序列属性的理想的方式。它也是一个估计可能的长期关系的有用的工具，即确保协整关系的参数。

应该提到的一个非常重要的一点是，VAR方法的主要缺点是缺乏理论的支持(苦力和勒罗伊，1985;利默尔，1985)。为了回应这些批评，布兰查德和华生(1984)和伯南克 (1986)的开发过程，称为结构向量自回归(SVAR)方法，结合传统的特征结构造型与VAR方法。使用SVAR的主要优势来自这样一个事实：标准VAR干扰通常特点是同时发生的相关性。这种相关性的存在，系统的响应，由IRFs表示，一个创新变量实际上是对所有的这些创新变量同时相关。同样，FEVDs能够量化具体来源的相对贡献差的相关性变化。

然而，在标准的VAR方法中，同期相关性是净化了的柯列斯基正交化过程。在大多数情况下，隐式地假定柯列斯基过程估计VAR模型的循环性。虽然理论上确定这种排序和事后灵敏度分析可能有助于进一步帮助提供关于适当的排序的见解，但它在很大程度上仍然是分析员的自由裁量权。

这项研究采用了以下方程的顺序，LOILP， LOILR，

26

LEXEDEV， LEXCON， LCPI ，LM2 LIMPORTS。一般来说，这种排序反映了这样一个事实： 在模型中油价会影响石油收入和其他所有变量。然而，石油价格的波动和至少某种程度上的石油收入是最不可能由模型中其他变化所决定的。这是一个相当合理的假设，因为石油价格和石油收入，包括石油出口收入和来自国外的净要素收入，在很大程度上取决于世界经济市场环境而不是取决于科威特的经济。

同样，这个命令假设支出在很大程度上取决于石油收入的水平，假设公共部门在推动科威特经济方面的主导作用也是相当合理的。也是合理的假定进口的价值在很大程度上是依赖于支出的水平。

经济理论提出唯一的变量以来，考虑选择这里的使用顺序很重要，所以SVAR不是在这项研究中。然而，这里使用的方法可以被认为是本着SVAR精神的方法。

3．实证结果

首先，VAR技术需要静止的数据，因此每个系列应该为平稳性检验。表I给所有的时间序列非平稳的测试，使用常规Dicky-Fuller测试(DF)，其增强版本(ADF)和Phillips-Parron t检验。这些测试包括一个常数但没有时间趋势，如同Dickey和富勒(1979)推荐的。

首先，t-statistics报道， 通过与恩格尔和Yoo(1987)得到的关键值相比，表明几乎所有的系列，除了CPI，M2和IMPORTS，静止在水平外，如图DF、ADF和Phillips-Perron t检验图所示，。这些测试在差分后重新应用所有条款。滞后一阶差分的t-statistics

27

条款表明，拒绝所有系列的零假设，也就是说，所有的系列都是固定第一差异。

然而，改变一个变量，一个通常的问题是是否应该在水平或差异上使用系统中的变量。总的指导原则是，如果系统中有k个使用的变量之间的协整向量，然后用常量k和VAR可以模仿n-k原始变量的差异。但如果系统中所有的变量不稳定，使用VAR水平是适当的。另一方面， 在协整的情况下估计一个VAR水平可能导致遗漏重要的约束条件。

在这种背景下，Doan等(1984)指出，差分变量对于Box-Jenkins ARIMA模型是重要的。Doan等人也注意到， 在VAR模型中这样做是不可取的。富勒(1976)也表明，差分数据可能不会产生任何VAR的渐近效率，即使它是适当的。此外，富勒认为， 当生产无显著增加时差分变量依旧 “抛出信息”。因此，跟随Doan和富勒的论证，在这里使用的是标准而不是区别。

第二，VAR模型的估计需要显式的滞后长度选择模型的方程。跟随Judge等(1988)和McMillin(1988)，Akaike AIC准则是用来确定VAR模型的滞后长度。所选的滞后长度由下式最小化得出：

AIC(n)=Indet∑+(2d2n)/T

用d表示模型中变量的数量，T表示样本大小，∑n表示估计残差的方差协方差矩阵，由∑u获得VAR(n)。考虑到样本的大小和在模型中变量的数量，最大延迟大于五个季度会减少估计的自由度让人难以接受，因此最大滞后长度被设置在五个季度。采用这种技术的结果总结在表II，显示相应的AIC值。可以看出AIC准则最小秩序是4。这表明，在这项研究中， VAR模型的秩序应该是4。

28

下一步是估计VAR。相应的估计值及其t值在表III中。虽然在VAR对个别系数的估计没有一个简单的解释，通过表III可以看出大多数t值显著(CPI方程除外)，几乎所有的方程都具有较高的广场。这也证实了一个观点，即石油价格比这个模型中其他变量更具有外生决定性。然而，石油收入方程比当前的支出和CPI具有更大的t值。

3.1 方差分解

表4给出了10个季度预测方差分解。表4表明， 所有变量最初的变化通常解释为变量的趋势。也就是说，在一开始，每个变量解释了很大一部分的历史趋势变化。在大多数情况下，十季后，大约60%的石油价格的方差解释为变量本身即表明其外源性性质。另一方面， 在第一季度石油收入能解释93%的变异、在第十季能解释约20%的差异。

此外， 从第二季度开始到第十季度石油价格的变化大约占石油收入的变化45%的。这表明从石油价格运行的石油收入的因果关系。发展和当前的支出之间类似的结果也很明显。每隔一段时间，大约15 - 17日的方差在支出(发展和当前)占石油收入的变化。在一段时间内，石油收入的变化约占支出（发展和当前）方差的15-17％。

此外，看支出的方差分解(发展和当前)后可以观察到，按自己的变化和石油收入，所述LCPI占支出方差的一个显著部分。CPI占经常性支出变化的五分之一和开发支出的15％左右。这是一个相当合理的结果，并在经常性支出的情况非常明显。

另一方面，值得一提的是，在发展费用变化显著部分的

29

另一变量是进口的值。它约占16％。

观察M2的方差分解，很明显，它的方差一部分在于CPI的变化中（约33％），甚至超过了自己的变化。此外，石油价格和石油收入，分别约占23%及其变化的13％。这些结果表明货币供应量的重要角色。

此外，在一段时间周期，结果表明：进口石油收入的变化占方差值的25%-45%。模型中的其他变量对进口的行为有显著影响得是两个类型的支出，尤其是的发展支出。

3.2 脉冲响应

图1显示了脉冲响应函数，它本质上是动态乘数。由于乘数的主要兴趣是看主要宏观经济变量对石油收入冲击的反应，然后是支出，所以只有10季度的报告。图1的检查表明，油价的创新，因此在模型中大部分的石油收入也有类似的影响变量。一般来说，大部分的这些变量显示增加在第一季度。这种增长在第二和第三季度仍将保持，然后是连续不断下降的季度。唯一的例外是CPI、进口和M2的值。这可能是由于使用估计VAR数据集的缺陷。回想一下，LCPI，LM2和LIMPORTS被认为是不固定的水平。

3.3 向量误差修正模型的估计

除了LCPI、 LM2、 LIMPORTS外由于大多数的变量涉及平稳时间序列数据，因此约翰森的测试(1991、1995)应用于检查协整向量。测试表明，有四个协整向量。因此，向量误差修正模型是十分必要的。向量误差修正模型是一个内置的VAR协整。每个协整方程涉及的参数和术语水平相关的系列都需要添加到在VAR

30

中的每个方程。在这个框架中有一系列的嵌套模型。所述约翰森测试程序为每个添加的协整方程计算似然比。

在约翰森的测试的基础上，一个向量误差修正模型估计(VECM)。四协整方程估计，在 VAR中使用相同的七个变量。然而，由于VECM的评估没有直接的解释，所以他们在这里并没有报道。

3.4 方差分解

对应于所述估计VECM方差分解结果列于表Ⅴ给出在VAR使用中它们基于相同的顺序。与VAR比较这些结果表明，虽然宏观经济联系的定性性质几乎保持相同，变量之间的相互作用的强度是明显高于协整占的比例的。例如，观察石油收入的方差分解， VAR结果从14%上升到40%左右，这表明发展和经常性支出等变量占有一个相当大的份额。

同样，石油收入在经常新支出以及进口值的变化上已经占了一个相对较大比例，特别是在前4 - 6季度。然而，结果表明，LCPI，LM2 LIMPORTS在第7 - 10个季度后主要是靠石油收入的变化来解释。

特别是LM2的方差分解表明其方差变化的重要组成部分来解释10季度后的石油收入 (约46%)，这甚至超过了其本身的变化。另一方面，从VAR的结果看，LCPI只占7%。

凭经验来说，VECM模型显示了一个相对更高程度的统计学意义。从理论上讲，这是因为它比VAR表示的指标更能展现宏观经济变量之间的密切互动。

3.5 脉冲响应函数

31

图2显示对应于VECM模型的脉冲响应函数。图2表明，石油价格和石油收入对包含在模型中的变量会产生类似的影响。然而，类似于VAR， 除了CPI，进口和LM2的值外，在刚开始连续几个季度中大多数的变量显示增加然后它们在后面几个季度里逐渐减少。

比较这些IRF与对应VAR的版本显示，它在VECM版的乘数达到VAR版的水平需要一段时间。他们通常在6 - 7季度达到他们在VAR版本的峰值， 而他们花了8 - 9季度才达到VECM版本相同水平。

4 结论和含义

本文的主要目的是研究国际油价波动对宏观经济变量的影响。因此，我们用了两个不同版本来估计其影响，即VAR和VECM。尽管宏观经济联系的定性性质在两个模型中几乎相同，但是变量之间的相互作用的强度明显高于协整方程。因此， 从定量角度分析，两个模型给出的结果明显不同。不过，根据实证分析， VECM给出更好的结果，因为它在宏观经济变量间能够产生比VAR估计模型更密切的互动关系。VECM对应的结果也更接近常识。

然而，两个估计版本表明主要宏观经济变量之间的相互关系的显著程度。结果凸显了石油价格对石油收入的因果关系，然后对支出和其他变量。然而， 基于正交创新进一步评估这些变量之间的关系，引导我们相信在科威特石油价格冲击的确影响宏观经济变量，尤其是通过发展和经常性支出。

支出的方差分解的评估表明，石油收入的波动在发展支出中占有很大的部分。

32

这一结果并不令人意外，实际上是符合在一个国家的预期的，是国家收入主要的来源--石油和天然气行业的唯一拥有者。因此，支出成为经济活动水平的主要决定因素和可以影响经济内循环流动收入的机制。

然而，令人惊讶的是，人们所预计的石油危机的影响更强，特别是在当前的支出的情况下。然而，这可能是由于这样一个事实：在过去的三十年里，已经积累的资本公积金(盈余)，经常用于资助当前承诺，尤其是在较低的石油收入的时期内。

此外，研究结果还指出CPI在解释支出的这两种类型显著变化部分的意义。另一方面，进口的变化值大多来自石油收入波动，然后是发展支出的波动。这些结果表明在科威特当石油危机影响支出时，这些财政似乎是有效的，然后支出占CPI和进口的价值变化的相对可观的部分。

然而，VECM方法的结果表明，石油收入的变化可以解释相当一部分的LM2的变化。它在第十季达到约46%，甚至超过自己的变化。

这个工作对于变量的说明已经显示出高度敏感的结果，即基于VAR的理论结构。这部分敏感性可能归因于数据集内的缺点。然而，在大多数情况下，它反映了VAR方法的局限性。例如， 在不同版本中没有标准的统计检验可供选择，然后这些并不是唯一的版本。人们可以得到一个无限大的数目的估计。它仍然在建模的自由裁量权中做出选择。

最后，VAR和VECM研究结果表明， 建模者中流行的VAR方法当应用于小型经济体有严重的问题，它是一个解决高度暴露的世界事件，符合标准的宏观经济问题的理论结构。

33