數據運作和描述
 

在前面章節中的敘述說明，一個ER決策很大程度上依賴于由政治操縱和結構xing約束所造成的的多維的政治經濟互動。更具體地說，政治邏輯，明示或暗示地確定分布在相互依存的系統，這些政治、經濟因素是如何構建的EA ER政策的各個方面。本章的主要任務是提供內在的和可行的嵌入在政治、經濟制度的因果機制中的實證分析方法。
 

我的實證調查，主要依賴于理論框架和假設，在第2章，旨在應對三個基本任務：一是開展適當的數據運作，定義了一個模糊的概念，使概念衡量的形式變量包括具體的意見。二是要設計出涵蓋整個變量的關系最強大的統計模型，這樣可以讓我提出更準確的統計數據。最后是選擇一個可行的和有效的評估方法，以指定的模型設置納入了實踐和理論動態。

 

Data Operationalization And Description Economics Essay

 

The narratives in the previous chapter illustrate that, an ER policy decision substantially rests on multidimensional political-economic interactions which are mostly induced by political manipulation and structural constraints. More specifically, political logics, explicitly and implicitly, determine how these political-economic factors distributed in the interdependent system construct the various aspects of ER policy in EA. The main task in this chapter is to provide an empirically analytical approach that inherently and feasibly catch the causal mechanisms embedded in the political-economic system.
 

My empirical investigation, which mainly relies on the theoretical framework and hypotheses in Chapter 2, aims to cope with three fundamental tasks: The first is to carry out a proper data operationalization that defines a fuzzy concept so as to make the concepts measurable in form of variables consisting of specific observations. The second is to design a robust statistical model that encompasses most relationships across variables, so that allows me to present more accurate statistics. The last is to select a feasible and efficient estimation method to specify the practical and theoretical dynamics incorporated in the model settings.

 

Note that the three methodological tasks are by no means able to be accomplished independently. That is, the features of collective data would decide the settings of my models and the corresponding specification methods. The essential reason is because occurrences of social phenomena or public policy decisions that usually involve multiple influential factors. With regard to the topic of ER politics, in theory, I intensively contend that the study of ER politics in EA is too complex to consider an interpretation by merely a few variables. Therefore, to conduct an empirical ER study, a large scale model is intrinsic to be applied in order to catch comprehensive dynamics among variables. However, political and institutional variables are mostly compiled in annual basis, because they are not frequently varied for a certain while. In the first two chapters, I clearly point out that this major challenge -- i.e. too many parameters but too few data -- substantially brings the deficiency problem of degree of freedom while proceeding to modeling specification.#p#分頁標題#e#

 

This chapter aims to provide a feasible research design that combines proper model settings and robust specification methods to compromise the difficulty between model scale and sample size. This chapter arranges three sections: Section 3.1 describes my data collection, measurement and basic statistics in order to correctly approximate the practical and conceptual variables, and thus benefit later statistical testing. Subsequently, section 3.2 addresses why the dynamic multivariate VAR and SVAR models should be adopted in my model settings based on the given theoretic system that ER and relevant variables are closely interrelated. Last, according to different beliefs of statistical testing approaches, section 3.3 advances the discussion toward the estimation methods varied between classic and Bayesian SVAR model specification.This chapter starts from overlooking the data and resolving the measurement issues. These include illustrating the statistical description over each analytical variable, as well as generating innovative measuring methods on ER policies, external influences according monetary constraints and coercion, and regional adaptive diffusion apart from the on-hand indicators provided by existing data sets. This conduction allows readers owning broad understandings regarding the general distribution over variables across EA countries at beginning. Furthermore, the debates and gaps between theoretical hypotheses and practical findings usually result from the various understandings regarding defining variables and setting operationalized methods. The description of variable operationalization in this chapter thus demonstrates a basis for any intentional effort of model falsification.

 

3.1 Data Operationalization and Description

 

I propose an integrated theory to comprehend the sophisticated mechanisms underlying ER policy decisions in Chapter 2. The theoretical framework and correlated variables are shown as Figure 2-1 and Table 2-1. This section demonstrates how I create appropriate indicators that precisely represent the variables by my conceptual definitions. In addition, this section illustrates the summary figure for every variable across nine observed EA countries, so that helps me to catch the general picture of variables’ distribution.

 

3.1.1 ER Policy Variables

 

ER Regime

 

The ERR is mainly defined as how a government manages its ER. The conventional measurements, such as IMF annual de jure reports [1] , Levy Yeyati and Sturzenegger’s (2002) [2] and Reinhart and Rogoff’s (2002) [3] de facto classifications, all treat ERRs as categorical differentiation. However, these categorical data are not efficient to statistical testing because of little varied information to provide, in particular when the numbers of studied countries are little.#p#分頁標題#e#

 

Unlike the conventionally categorical classification, this research adopts Bayoumi and Eichegreen’s (1998) idea to measure the numerical flexibilities of ER on behalf of the actual ERR arrangements. While many alternative data sources employ categorical measures of ERR, using a continuum of flexibility scores for a given state, each ranging from zero to one, avoids the pitfall of having to impose arbitrary threshold values for fixed versus floating ERR status. Unlike many alternative categorical indicators of ERR, the ER flexibility scale allows me to consider degrees of governmental intervention on ER, as well as to observe relatively small discrete regime changes. Second, its measures are synthetic rather than reductionist, allowing a continuum that distinguishes dimensions of various ERR settings.

 

Moreover, this ER flexibility index is designed to combine the information both from ER and foreign reserves movements, which fundamentally assumes that the level of ER intervention is the function of conjunction with ER and foreign reserves volatility (Baig 2001; Wang and Yang 2001; Ogawa and Yang 2004). The details to calculate the ER flexibility index are shown as Appendix 3-1. The originally relevant data are collected by IMF’s International Financial Statistics (IFS) [4] and Taiwan’s Central Bank of R.O.C. yearly statistics [5] .

 

The flexibility index indicates that a much flexible ERR is characterized by little intervention in the ER markets together associated with substantial volatility in ER but less changed reserves. Contrarily, a fixed ERR occurs when the ER does not move while foreign reserves are fluctuated because of active exercise on stabilizing ER. In addition, the flexibility index ranges from zero to one. That the index displays zero or near zero indicates a hard-pegged ERR, because least volatile in ER but great volatile in reserves exists. On the other hand, the index for a free-floating ERR is close to one.

 

Figure 3-1: ERRs in EA

 

Note: Figure 3-1 is as the same as Figure 1-2.

 

Figure 3-1 illustrates the distributions of ERRs in nine EA countries. It clearly displays that Indonesia, Japan and Korea exercise relative floating ERRs in most time in comparison with the rest of regional countries from 1980 to 2004. Further scrutinizing, the ERRs in the newly transitional democracies with relative ERR pegs, i.e. the Philippines, Thailand and Taiwan, are more fluctuated than the political steady countries. The ERRs in Philippines and Thailand become relative floating during the AFC period.

 

ER Level

 

The ERL is defined whether a decision maker decides to appreciate or depreciate its national currency. Therefore, the ERL in this research is measured as the valuation changes of current ER against last year’s ER by appreciation, depreciation or unchanged. To create this ERL indicator are two stages. First, I generate an index to approximate annual valuation of ER as equation 3-1.#p#分頁標題#e#

 

(3-1) .

 

In Chapter 1, I point out the difficulty to gauge the actual valuation of ER. However, by calculating the differences between nominal and real ERs, the ER valuation can be approximately indicated. Figure 1-3 and 1-4 show that most EA countries are used to execute undervalued ERs.

 

Second, by employing the approximated ER valuation, I further calculate the annual change of ER valuation to represent ERL as equation 3-2.

 

(3-2) ,

 

where positive (negative) value indicates ER appreciation (depreciation). Moreover, the nominal and real (PPP adjusted) ERs are obtained from IMF’s World Economic Outlook (WEO) database [6] .

 

Figure 3-2 presents that Singapore’s ERL is most fluctuated compared with the other regional counterparts. Japan depicts greater appreciation movements before 1985, and Korea illustrates the greatest depreciation in 1996 right before the 1997 AFC. The other countries only have slight changes on their ERL.

 

Figure 3-2: ERLs in EA

 

3.1.2 Domestic Variables

 

Regime Type

 

The RT is indicated as how a political system operates under the conceptual framework of democracy. I apply the most recognized indicator of measuring polity from autocracy to democracy in Polity IV (Marshall and Jaggers 2009) [7] . The Polity IV unifies POLITY scale ranging from +10 (strongly democratic) to -10(strongly autocratic).

 

Figure 3-3: Regime Types in EA

 

Figure 3-3 illustrates that China and Singapore are the two steady non-democracies, and Japan and Malaysia are the two steady democracies during the period from 1980 to 2004. The other five new democracies have just experienced regime transformation in the period. Although the transitional democracies are under developing, the statistics exhibit that they already reach higher polity scores in the recent years than the steadily democratic Malaysia.

 

Political Competition

 

The PC entails the intension among political parties when they compete for political resources. Here I define the political resources as the seats in executive and legislative branches. The PC intensifies when the gap between the proportions of seats in the executive and in the legislature dramatically exists. This indication in particular follows my theoretical argument that PC is essentially rooted in the incompatible political opportunities toward political competitors in a political system. For example, the incumbent party does not controls major seats in the legislature, however, the occupation of taking all or most seats in the executive would result in high PC in the country. Undoubtedly, the opposition parties tend to defy incumbent governance because of unrepresented seats in the executive until the political offices are reallocated to proportionate to the actual political forces.#p#分頁標題#e#

 

I utilize two indicators – HERFGOV and HERTOT -- in the World Bank’s Database of Political Institutions (DPI) [8] to create the approximated variable of PC. HERFGOV indicates the centralized level of the parties in the central government; likewise, HERTOT reflects the centralized level of the parties in the national legislature [9] . The mathematic derivation from HERFGOV and HERFTOT to PC is as Appendix 3-2. The PC indicator ranges from zero to one, and indicates lowest to highest political competition respectively.

 

Figure 3-4: Political Competition in EA

 

Figure 3-4 illustrates that China and Singapore are the two countries with lower PC which are compatible with their autocratic system. Japan is a democracy with high PC except with the early 1990s. Korea also features high PC in the 1980s and early 1990s when were the years of prevalent social movements. Taiwan soon becomes a political system with high PC after democratization. On the other hand, the other south EA democracies, i.e. Indonesia, the Philippines, Malaysia and Thailand, demonstrate relative lower PC.

 

Tenure Cycle

 

TC is a variable that addresses the cyclic effects on ER policy decisions from the remained years of each central government’s administration. The data is collected from DPI’s YRCURNT which straightforwardly indicates the “years left in current term”. Note that the YRCURNT codes the election year as a “0”.

 

Figure 3-5: Tenure Cycles in EA

 

Figure 3-5 illustrates most EA countries (except the Philippines) have regularly tenured cycles over times. Some transitional democracies have their tenure reforms in certain periods, e.g. Korea in early 1980s and Taiwan after 1996.

 

Inflation

 

PI is used to specify the reciprocal effects between domestic price changes and ER policy settings. I collect this inflation rate indicator that is coded as annual change of average consumer prices in percentage from IMF’s WEO database. Figure 3-6 illustrates that most EA economies control their PI around or even below than ten percents in most time. Only Indonesia during the AFC period and the Philippines in the early post-Marcos era had experienced fluctuated inflation rates that are over forty percents displayed in the figure.

 

Figure 3-6: Inflation Rates in EA

 

Balance of Payment

 

BP is employed to examine whether the central government’s debts affect ER policy decision-making. The data is collected by the “current account balance in percent of GDP” at IMF’s WEO database. The positive (negative) values indicate that national account balances are deficit (surplus).#p#分頁標題#e#

 

Figure 3-7 displays an interesting pattern that most EA countries, with the exception of Singapore, had lower BP since late 1980s and early 1990s until 1997 AFC when the world order under cold war broke down as well as the third wave of democratization prevailed in EA. China and Japan show relative stable BP variations.

 

Figure 3-7: Balance of Payments in EA

 

Sector Difference

 

SEC is defined as the differences of trade sectors between exports and imports. I collect the trade data from various years’ reports of IMF Direction of Trade Statistics (DOTS) [10] . By using the export and import data, the SEC is calculated as the following formula:

 

(3-3) .

 

The index of SEC ranges from -1(complete imports) to 1 (complete exports) in the scale of total trade.

 

Figure 3-8: Sector Differences in EA

 

Figure 3-8 in general seems to contradict our understanding that EA are export-oriented economies. It illustrates that most EA countries have almost balanced trade between imports and exports, unless the significant exporting countries, such as Indonesia, Japan and Malaysia in early 1990s. In fact, this is understandable because most EA countries export manufacturing goods, and need to import compatible raw materials and resources for production. Japan is a typical exporting economy with high additional-valued products; on the other hand, Indonesia is a typical exporting economy with mass exporting raw materials and natural resources.

 

3.1.3 International Variables

 

Power to Delay

 

P2L measures the monetary power of a country to postpone its payment of national debts. Cohen (2006, 41-46) regarded P2L as monetary power of international liquidity which comprises two main components: owned reserves and borrowing capacity. Through IMF’s IFS, I can obtain foreign exchange (line 1D.DZF) and special draw right (SDR) holdings (line 1B.SZF) [11] to approximate own reserves and borrowing capacity respectively.

 

However, the challenge to operationalize P2L is not to collect the reserves and borrowing capacity data, instead, the difficulty is derived from how I combine the two indicators into one and how I transform it to fit a general scale of power indication. My conduction is to treat the US as the major superpower that every monetary authority must deal with in the world monetary pyramid, and subsequently apply American statistics of reserves and SDR holdings as reference to measure the relative P2L of EA economies. By such adjustment, I am able to measure the P2L that each country owns in a consistent scale. Moreover, I equally weight the two components to create an integrated indicator of P2L. The formula is denoted as follows:#p#分頁標題#e#

 

(3-4)

 

Figure 3-9: Power of Delay in EA

 

Due to P2L is relative indicator compared to the US, the positive (negative) value indicates greater (smaller) power than the US owns. Figure 3-9 illustrates that China and Japan equip with greater liquidity power than the US since early 1990s. The other regional counterparts all have lower P2L than the US, with the exception of Korea and Taiwan after 2000. This empirical illustration is consistent to our common understandings that China and Japan own substantial foreign reserves and play dominant roles in world economy. Moreover, in general, Figure 3-9 reflects a trend that the liquidity power in EA is increased, although the increases in some countries are slow.

 

Power to Deflect

 

P2F measures the monetary power of a country to divert the costs of encountering financial shocks. Cohen (2006, 46-49) defined P2F as the monetary power that is composed of openness and adaptability. The openness indicator is represented by the total trade to as a share of GDP from World Bank’s WDI. On the other hand, the operationalization of adaptability is relatively tricky. Cohen defined the concept of adaptability, however, he cannot borrow an appropriate indicator to denote it. I initiatively apply HHI methodology on World Bank’s Trade, Production, and Protection (TPP) database [12] to measure diversity of national trade goods. Economic adaptability thus can be drew through the presentation of HHI.

 

As the same as measurement of P2L, trade openness and adaptability are equally weighted as well as transformed into a general scale that provides comparability across various observations. Again, I treat the US benchmark as the general scale and then assign the relative P2F for each country in EA. The formula is denoted as follows:

 

(3-5)

 

Figure 3-10: Power of Deflect in EA

 

Figure 3-10 clearly illustrates that almost all EA countries during the period from 1980 to 2004 have relative greater P2F than the US. Singapore is the only economy in the region that owns greatest P2F and five to ten times more than the US. According to its open and diverse economy, Singapore is more sustainable to divert and slowdown its loss when economic crisis occurs. In comparison with P2L in Figure 3-9, I can imply that EA countries may not gain great liquidity power by competing capital in the world market; however, they would turn to reform their industrial structures in order to prevent themselves from economic shocks.

 

Monetary Coercion

 

MC is defined as the monetary pressure from the US if it perceives a country’s manufactured goods have dominantly occupied its domestic market. Because of implementing a trade protection policy that would incur many criticisms and costs from the well-known American free market, the US tends to turn to balance its trade deficits (TD) by exerting monetary adjustments. Most policy makers to business executives believe that the US stands on the top of world monetary pyramid, and thus is mostly capable and desired to exercise its monetary power to suppress other countries’ exports.#p#分頁標題#e#

 

From the Japanese and Chinese experiences against American trade imbalance in the two decades, the US on the strength of super monetary power is used to urge Japan and China to appreciate their ERLs in order to reduce trade TD. Therefore, I indicate MC as the shares of total US TD. This indication implies that the more shares of the US TD the more MC a country would suffer to change its ER policy settings involuntarily. Note that the indication of MC by transforming from TD should be able to apply on every country in theory. However, not all countries have equal monetary capabilities to counteract TD. Weak monetary powers usually undertake enormous TD associated with ineffective and incapable policy tools. For example, Indonesia cannot effectively coerce Japan to appreciate yen value although it perceives Japanese imports prevail Indonesian markets, because Indonesian monetary power is substantially weak vis-à-vis Japan. Thus, in order to simplify the exerciser and receiver while designing the indicator of MC, I merely define American power exertion to represent MC in this research.

 

MC is presented in a scale of percentage. Figure 3-11 illustrates that China, Japan and Taiwan in 1980s encountered greater MC from the US. Most EA countries show decline trends of MC after 1990, with the exception of China.

 

Figure 3-11: Monetary Coercion from the US

 

3.1.4 Regional Variables

 

Regional Intra-Trade

 

TR is defined and measured as the shares of intra-regional trade of total trade amounts.

 

TR is used in the theoretical framework in order to test the OCA theory. I collect the TR data from IMF’s DOTS. Figure 3-12 illustrates the distribution of percentage of TR against total trade in EA. It displays that TR occupies substantial portion of total trade (near or over fifty percent) for an EA economy. Meanwhile, TR in EA demonstrates gradually increased trend. In general, the statistics show that TR becomes more and more important toward EA development.

 

Figure 3-12: Regional Intra-Trade in EA

 

Adaption of ERR

 

WR and the later WL are two variables that indicate the rippled effects of ER policy choices from neighboring countries. Furthermore, WR and WL are used to examine how regional dependences interact with ER policy decisions. Recalling my theoretical hypothesis, I argue that EA ER policies are clustered decided by adaptive diffusion forces. WR in this section is measured as the spatial impacts of ERR from neighboring countries.

 

Beyond the theoretical configuration, empirically, political scientists have employed spatial econometrics (Anselin 1988) to study spatial dependences and diffusion processes of policymaking. Essential assumption to spatial analysis is that degrees of contiguity create externality to influence the behaviors of local actors. Moreover, the core of setting up a spatial econometric model is to transform the structure of dependence in a proper and valid “spatial weights matrix”, which is denoted as W comprise of an N by N by T matrix. There are types of this weights matrix indicating neighborhood connection: One, the neighboring contiguity is coded as a discrete threshold, such as binary coding that specifies the adjacent neighbors as 1, otherwise as 0 (LeSage 1998, 7-17; Lin, Wu, and Lee 2006). Two, the neighboring contiguity is measured with the notions of continuous “distance”, which indicate the spatial connection by closeness between two units (Simmons and Elkins 2004; Beck, Gleditsch, and Beardsley 2006; Elkins, Guzman, and Simmons 2006).#p#分頁標題#e#

 

In this research, following Beck, Gleditsch, and Beardsley’s (2006) approach, I calculate non-geographic distances with notions of figurative space to progress specification on ER policy diffusion effects. To employ this approach has two main reasons: One, EA countries in my sampling pool are not all land bordering. Japan, the Philippines and Taiwan are even sea surrounded. Therefore, a discrete threshold in representing the structure of spatial contiguity is not feasible.

 

Two, the conventional indication by using geographic distances in this ER research is not appropriate. Adaptation of ER movements more relies on the proximity of decision making structures and economic development levels between countries. An application of geographic distances may be useful in immigration flows, ideological shifts and cultural assimilation studies, because the geographic location substantially determines the emergences of these social phenomena. On the other hand, capital mobility and financial coordination are not necessarily restrained by geographic distances in the era of open economy. For example, Taiwan is more geographically closed to the Philippines than Korea, however, Taiwan’s financial and monetary arrangements are more prone to Korean settings. Thus, to create non-geographic distances that describe the spatial contiguity on the metaphysical map would properly explain the spatial effects on clustered ER policy decisions.

 

Although an ERR setting measured as the degree of ER flexibility may be varied year by year, policy makers do not frequently change their tendencies toward their ER management. I thus draw a propensity map that locates the EA countries by the average tendencies and changing scales of arranging their ERR. The median point and interquartile range of a country’s ERR series from 1980 to 2004 determine the spatial position of a country in the propensity map. The technical operationalization on the non-geographic notion of distance is presented in Appendix 3-3.

 

Figure 3-13 demonstrates the distribution of WR statistics which are multiplied by neighboring ERRs and spatial weights ranging from zero to one. In general, it shows that EA countries encounter stronger neighboring collective pressures to float their ERR during the AFC period. The neighboring countries of Indonesia, Japan and Korea mostly tend to liberalize their ER management rather than the other countries.

 

Figure 3-13: Adaption of ERR in EA

 

Adaption of ERL

 

WL defines the spatial effect of adaptive ERL policy decisions from neighboring countries. Following the same measurement method as WR as shown in Appendix 3-3, I create a propensity map for WL that allows me to figure out the spatial distances between countries. By the distance specification, the spatial weights matrix can be generated and thus forms the WL variable. The technical discussion and the ERL propensity map are presented in Appendix 3-4.#p#分頁標題#e#

 

Figure 3-14 illustrates that most EA countries’ neighbors are less likely to adjust their ERs frequently. This demonstrates a general tendency that the EA countries

 

Figure 3-14: Adaption of ERL in EA

 

prefer to keep a stable ER system, instead of manipulating ER to satisfy their political demands and economic interests.

 

Economic Competition

 

EC defines the degree of competing for general economic growth which encourages decision makers to adopt necessary adjustments on their ER policies. I create the EC indicator by measuring the relative differences between current national economic growth rate and the average of the other regional counterparts’ rates. In order to prevent opposition from domestic constituents, decision makers are more likely to exert their monetary tools and change their currency policies, especially when their economic outputs have been behind the other regional competitors. The economic growth rate data is collected from IMF’s WOE database.

 

Figure 3-15 illustrates that China leads in the regional EC; on the other hand, general economic performances in Indonesia, Japan and the Philippines are behind with compared to the average in the EA region.

 

Figure 3-15: Economic Competition in EA

 

3.2 Model Settings

 

3.2.1 VAR: Causal Inference from Endogenous and Dynamic Data

 

Opponents to quantitative approach usually criticize that the conventional statistical analysis mistranslates the social observations into a parsimonious model, and misinterprets its statistical results by the common “other things being equal” assumption. They claim that social outcomes are complexly correlated. Social actors and their surroundings are by no means treated as unique under a controllable and static situation. Thus, a statistical analysis that lacks consideration regarding endogenous and dynamic influences apparently misleads the ways to investigate social phenomena. Actually, quantitative political scientists have been aware of these problems. More and more feasible and efficient models are introduced in dealing with the endogeneity and series autocorrelation issues on the specification of data generation processes (DGP).

 

Table 3-1 presents a two by two chart that summarizes current popular statistical models by number of modeling equation and time-variant classification. The number of modeling equation essentially classifies whether an estimated model takes into account the endogenous interactions among variables. As long as the estimated variables are defined as endogenous, in practical observation or theoretic hypothesis, the multiple-equation model as known as system estimation is useful to provide more comprehensive and less biased specification.#p#分頁標題#e#

 

Moreover, unlike single equation model mainly relies on a single theory, a multiple equation model incorporates relevant competing theories for deciding variables altogether while designing model specification. Due to reliance on a leading theory, a single equation model imposes priori restrictions, treats relevant variables as exogenous, and attempts to “normalize” competing theories into one testing equation. It is not too difficult to understand the limitation and invalidity of applying one theory to explain a complex social system. Freeman, Williams, and Lin (1989) argued that these arbitrary treatments in terms of single equation approach [13] are more likely to mislead theoretical interpretation and misspecify causal inference.

 

The time-variant classification distinguishes whether the model incorporates time dependence variables to control series correlation errors. Time dependence mainly regards temporal effects on the dynamic mechanisms. Most of empirical political studies build up their theories based on the contemporaneous effects, i.e. testing the spontaneous statistics toward expected signs. However, this conventional statistical examination lacks to consider whether the impacts from an innovation still keep the same signs and magnitudes for a certain periods. Through time persistent and dynamic testing, researchers are able to catch the causal mechanisms embedded in social phenomena and relevant coordination, and thus provide a more valid estimation.

 

According to the model classification in Table 3-1, a VAR model demonstrates its ability to deal with endogenous and dynamic interactions among variables. Chapter two has presented that my theoretical framework underpinning the ER politics is a system composed of three major theoretical perspectives from domestic to international politics and from rational pursuit for economic interests to non-rational adaption of policy making. Because of the complex interaction in the system, apparently, these analytic variables in the framework are theoretically featured as endogenous and dynamic. Therefore, the adoption of a VAR model setting that appropriately take into account endogeneity across various theoretical equations and dynamic effects in the theoretical system of ER politics would help model specification in my research.

 

Table 3-1: List of Model Classification

 

No Time-Variant Analysis

 

Time-Variant Analysis

 

Single-Equation Model

 

Linear Regression (e.g. conventional LS models)

 

Time Series Analysis (e.g. ARIMA and ADL)

 

Multiple-Equation Model

 

Simultaneous Equation Model

 

VAR (e.g. RF VAR and SVAR)

 #p#分頁標題#e#

3.2.2 VAR: Introduction and Debate

 

Unrestricted RF VAR

 

Sims(1980) innovatively introduced the VAR model to proceed to a macroeconomic analysis without setting priori theoretical identification. Sims argued that the priori restrictions on a traditional structural equation model are incredible. In addition, these theoretical restrictions are easily violated due to dynamic variations and expected uncertainties featured from data. Unlike structural equation modeling, VAR does not impose priori restrictions on the interactions across analytic variables. A basic VAR approach treats every endogenous variable in the system as a function of the lagged values of all of the endogenous variables in the system.

 

A basic VAR model can be formally expressed as the following form:

 

(3-6)

 

where is an vector of endogenous variables; is a matrix of coefficients in lag operator L; is a white noise vector that composes serially uncorrelated disturbances. Note that here is denoted as a nonsingularly identical matrix which normalizes the main diagonal as one and imposes zero restrictions on the contemporaneous correlation across variables. Each endogenous variable can be stacked up in an equation and then forms a system that composes of all these endogenous variables. The VAR is, then, a function of predetermined variable values and the predetermined values of the other variables in the system. This basic VAR model expressed in Equation 3-6 is also as known as RF VAR model.

 

The central principle to unrestricted VAR approach roots in its suspicion against the predetermined construction of model settings by preexisting theories. According to Sims, researchers should let the data speak, without imposing theoretical restrictions, at least as far as the estimation step is concerned. In addition, VAR deals with endogeneity by mainly adding pre-determined variables on the right-hand-side in equation, in order to catch the dynamic mechanisms and gauge specification uncertainty. These features enable VAR modelers to advance a statistical analysis in a relative objective approach. That is, unlike the structural equation approach that treats data analysis as auxiliary procedures in testing theories, VAR analyses build up theories that mainly rely on the findings from data mining.

 

However, because of the unrestricted feature, a RF VAR would encounter the over-parameterization problem. Unrestricted VAR models handle all the endogenous variables across equations in a system. Therefore, even small VAR models have a large number of regression parameters. Too many parameters but insufficient degrees of freedom in a model would produce unreliable and inefficient statistical estimates. This overparameterization problem thus restricts the efficiency and usefulness of VAR analysis for statistical inference (Pagan 1987).#p#分頁標題#e#

 

Likewise, instead of theoretical identification, RF VAR identifies the structural innovation through a moving average representation (MVR) from data per se. MVR is mainly featured to describe the accumulated responses from past shocks in a system. The atheoretical approach by isolating the use of priori restrictions and regressing all variables and their accumulated innovations in an autoregression model has led empirical findings somehow are difficult to test preexisting theories and deliver inconsistent explanation (Cooley and Leroy 1985). In addition, different orders of variables arranged in the RF VAR model results in various estimations on structural parameters. The absence of providing unified modeling estimates in a RF VAR model fails to catch accurate causal mechanisms, which is the main goal for most VAR modelers expect.

 

Contemporaneous Restrictions: SVAR

 

In order to respond the criticisms of atheoretical RF VAR, large part of literature has been dedicated to identify structural shocks and impose restrictions derived from theory or from institutional knowledge, such as Bernanke (1986). The new VAR approach is as known as Structural VAR or SVAR. Central to SVAR is to take into account the contemporaneous relationship among the variables while proceeding to a VAR analysis. Technically, instead of the unrestricted setting in RF VAR, SVAR particularly extends to translate contemporaneous restrictions derived from competing theories into a VAR model. Furthermore, imposition of plausible restrictions on the contemporaneous relationships among the variables benefits to identify unique structural relationships in the SVAR model. Therefore, the causal interactions across variables and equations can be traced out more precisely through this unique identification. A consequence is that the estimation of the VAR coefficients is no longer done on an equation-by-equation basis as in RF version. Instead, SVAR suggests the multivariate time series regression in a full system.

 

Whereas traditional VAR contributes to multivariate time-series specification, lack of considering contemporaneous interdependence between determined variables while studying ER policy would mislead causal inference. In practice, due to the intrinsic feature with respect to price changes, ER is a sensitive economic indicator that simultaneously correlates with many relevant political-economic variables. Likewise, the index of ER frequently changes day-by-day even minute-by-minute. Running a VAR model by my annual ER dataset would provide inaccurate data analysis. Taking into account contemporaneous effects in VAR model helps to reduce inaccurateness of model specification by using annual ER data.

 

Recall the VAR mathematical expression in Equation 3-6, SVAR does not treat as an identical matrix; instead, the off-diagonal components are not restricted to zero according to theoretical identification. Likewise, not all of and are restricted to zero because of the existence of error covariance across variables and equations. Thus, by a succinct but direct definition, each endogenous variable in a SVAR model is determined by the contemporaneous (at time “0”) and past (lagged) values of all the endogenous variables in the system. Following the introduction of Sims and Zha (1998), Waggoner and Zha (2003) and Brandt and Freeman (2006, 2009), a general SVAR model that can be written in matrix form as#p#分頁標題#e#

 

(3-7) ,

 

where and are parameter matrices for the contemporaneous and lagged ( lag and p the maximum of lags) effects of the endogenous variables; D is an parameter matrix for exogenous variables (including an intercept); is the matrix of the endogenous variables; is a vector of exogenous variables (including an intercept) and is a matrix of i.i.d. normal structural shocks such that

 

, and .

 

This structural model in Equation 3-7 can be rewritten in a matrix form:

 

(3-8)

 

where is ,is which defines the contemporaneous correlations of the series, is which contains the lagged Y’s and a column of 1’s corresponding to the constant, is that defines a matrix of the coefficients on the lagged variables and constant, and is . Note that the columns of the coefficient matrices correspond to the equations.

 

Furthermore, the VAR model in Equation 3-8 can be rewritten as a compact form as

 

(3-9)

 

by letting

 

(3-10) , and .

 

Equation 3-9 and 3-10 formally demonstrate that the VAR estimation is a function of linear combination of determined and predetermined factors. Additionally, the innovation accounting on the right-hand-side can be traced out precisely, as long as the linear combination on the left-hand-side is able to be specified unambiguously.

 

Moreover, SVAR can be transformed to the common representation in RF VAR form, in order to compare and thus generalize previous and/or existing empirical findings by using VAR analysis. The RF representation derived from the basic SVAR in Equation 3-7 can be obtained by post-multiplying Equation 3-7 by . The RF is

 

(3-11) ,

 

where , and .

 

Equation 3-11 indicates that plays critical role in model specification. Restrictions on the contemporaneous parameters in substantially affect the computation and interpretation of variances of the RF. This RF error covariance matrix is expressed as

 

(3-12) .

 

Moreover, in a conventional RF VAR analysis, is specified as a just-identified triangular matrix, and contemporaneous causal relationships across equations are recursive through a Cholesky decomposition of (Freeman, Williams, and Lin 1989; Brandt and Williams 2007, 40-41). This is also known as orthogonalization of the residuals. All the shocks entering the system are positive sign. On the other hand, the in a SVAR model is usually non-recursive and over-identified. Therefore, because there is no unique correspondence between shocks in variables and shocks to equations, shocks to a given equation cannot be clearly discerned by positive or negative sign. Waggoner and Zha (2003) and Brandt and Freeman (2009) suggest “sign normalization” to resolve this problem.#p#分頁標題#e#

 

3.3 Specification Method

 

The previous sections have discussed the VAR model is the best setting in describing multivariate dynamic relationships among variables. Likewise, the debates between RF VAR and SVAR in section 3.2 demonstrate the advantages of considering contemporaneous effects in a VAR model. This section centers on the discussion regarding the methods of VAR model specification. Conventionally, VAR models are specified by frequentist approach. That is, model parameters are estimated mainly based on observed data given conditional likelihood function, such as OLS and MLE are belonged to this family. Recently, Bayesian approach which emphasizes the prior information and conditional probability among variables in model specification has been developed and widely applied in dynamic political studies. I summarize these two approaches to VAR application in this section.

 

3.3.1 Frequentist Approach

 

Recall Equation 3-9 and 3-10, the conditional likelihood function to estimate a VAR model is expressed as

 

(3-13)

 

For the frequentist approach to analyzing the unrestricted VAR model, researchers normally adopt seemingly uncorrelated regression (SUR) which assumes the error covariance of residuals to be block-diagonal. Therefore, by normalizing diagonal of in Equation 3-13 as 1 and off-diagonal as 0 (identical matrix), the maximum-likelihood estimator for is

 

(3-14)

 

where denotes the matrix of regression coefficients. As a result, estimators in the VAR model can be estimated by OLS method equation-by-equation. Residuals across various equations are mutually independent. The covariance matrix of the residuals can be estimated from the sample residuals

 

(3-15) ,

 

where is the matrix of residuals from the VAR in Equation 3-6.

 

It is conceptual and technical convenience to obtain statistical results through OLS estimation. Nonetheless, merely a reliance on the data analysis in the basis of the likelihood function renders the frequentist analysis unfit for robust model specification. The first problem as overparameterization as I mention in the previous discussion, when the VAR models are large and degrees of freedom are small, the findings from a typical frequentist method would be unreliable and inefficient as to be nearly useless for inference.

 

The second problem is about the validity of data analysis. The frequentist approach, in particular referred to RF VAR model, has been known that its statistical results “tent to overfit the data, attribute unrealistic portions of the variance in time series to their determined components, and overestimate the magnitude of coefficients of distant lags of variables” (Brandt and Freeman 2006, 5). This problem occurred is because RF VAR models demonstrate their analysis fully relying on data. No theoretical identification is incorporated into the model specification processes. Thus, when data is not randomly collected or data size is not big, the data analysis based on the asymptotical properties of likelihood function is problematic. Statistical results derived from this frequentist approach are biased to interpret theories and invalid to proceed to sampling inference.#p#分頁標題#e#

 

Close dependences on the quality and quantity of data collection somehow become risks to determine the accuracy and efficiency of the statistical estimators derived from frequentist method. SVAR associated with contemporaneous identification can help to overcome the shortcomings of adoption of frequentist analysis at some extents. However, statistical investigation is still risky to be wrong, because the common measurement errors and less randomized collection of data would mislead model specification and thus leads to ill performances of statistical analysis.

 

3.3.2 Bayesian Approach

 

The central to Bayesian approach is to consider both observable data and the probability distribution of prior information while proceeding to statistical analysis. By imposing priors into estimation, Bayesian data analysis would have less chance to conclude a result that significantly deviates from the explanation of existing theories and the realities of previous experiences. The basic function of Bayesian approach can be expressed as

 

(3-16) ,

 

where represents posterior distribution of the parameters given the data, is prior distribution of the parameters, and is likelihood function of the data given the parameters. Equation 3-16 presents that the posterior distribution, which is the main research target in a Bayesian model, is proportioned to the multiplier by prior information and likelihood distribution of observation. Therefore, apart from deciding likelihood function for current observations, a well-performed Bayesian analysis must be able to assign appropriate priors into the model.

 

Minnesota Prior

 

In order to cope with the data problems in running classical likelihood estimation, Robert Litterman (1986) and some other well known scholars associated with the University of Minnesota or Minneapolis Federal Reserve Bank, such as Thomas Doan and Christopher Sims, introduce a widely used Bayesian prior probability distribution for RF VAR models to downsize models with large coefficients on distant lags and explosive dynamics. The Bayesian prior settings by these scholars, which are often labeled the Minnesota prior, have contributed to perform better estimation on VAR models. Based on a belief that most time series are best predicted by their mean or their values in the previous periods, Minnesota prior underpins Bayesian VAR (BVAR) to advance multivariate time series analysis in particular when data is not largely and randomly collected.

 

According to Doan, Litterman, and Sims (1984) and Litterman (1986), the features of Minnesota prior can be summarized as: One, prior means for lagged dependent variables are set to unity in the belief that these are important explanatory variables. Two, prior means of coefficients on all variables other than own-lagged dependent variable are assigned to zero. Three, prior means of zero are more tightly as the lag length increases, based on the belief that more distant lags represent less important variables in the model. The Minnesota prior means and variances suggested take the following form#p#分頁標題#e#

 

(3-17) ,

 

where denotes the coefficients associated with the lagged dependent variable in each equation of the VAR, and represents any other coefficient. The prior variances, , specify uncertainty about the prior means , and indicates uncertainty regarding the means .

 

Doan, Litterman, and Sims (1984) further generated the standard deviation -- composed of three main hyperparameters and a weighting matrix -- to translate our beliefs into the prior setting as the form in Equation 3-17. The specification of the standard deviation of the prior imposed on variable j in equation i at lag k is

 

(3-18) ,

 

where is the estimated standard error from an univariate autoregression involving variable i, so that () is a scaling factor that adjusts for varying magnitudes of the variables across equations i and j. denotes the “overall tightness” which reflects the standard deviation of the prior on the first lag of the dependent variable. The term is a lag-decay function with which reflects the decay rate, shrinkage of the standard deviation with increasing lag length. The function specifies the tightness of the prior for variable j in equation i relative to the tightness of the own-lags of variable i in equation i.

 

Sims-Zha Prior

 

Sims and Zha (1998) argued that the Minnesota prior is not a proper prior setting for the full VAR model. This is because Minnesota prior is only formed for each of equations in the model instead of adopting full structural equations. By the same rationale as SVAR against RF VAR, Sims and Zha (1998) proposed modified priors that take into account the contemporaneous dependences and best describes the dynamic relationships among variables rather than Minnesota prior settings.

 

The Sims-Zha prior for a model given (stacked in ) is assumed as the conditional form

 

(3-19)

 

where denotes the prior mean for lagged variable , represents the prior covariance for , and is a multivariate normal density function with the mean and covariance for the conditional . Apply this prior and Likelihood function in Equation 3-13 to the basic Bayesian function as shown in Equation 3-16, the posterior probability for the coefficients is then formed by combining the likelihood function and the prior:

 

(3-20) .

 

Mostly important, the Equation 3-19 and 3-20 express the key feature of this specification by Sims-Zha prior that is the interdependence of beliefs and the conditioning of the prior on the structural contemporaneous relationships, . Because of the feature of conditional specification, much accurate finite-sample inference and estimation from the model’s posterior displayed in Equation 3-20 shall be obtained via a Gibbs sampler [14] (Waggoner and Zha 2003).#p#分頁標題#e#

 

In addition, the Minnesota prior for a RF model expresses a belief that a random-walk model for each variable in the system is a reasonable “center” for beliefs about the behavior of the variables. By the same token, Sims and Zha proposed that the beliefs about the coefficients should be centered on an identity matrix for top m rows in and zeros for the remaining rows. Recall the conditional distribution of prior

 

(3-21) ,

 

where =, and .

 

Sims and Zha further generated hyperparameters to scale the conditional standard deviations of the dynamic simultaneous equation regression coefficients on lag l of variable j in equation i given by

 

(3-22) ,

 

where the controls the tightness of beliefs on , controls what Litterman (1986) called overall tightness of beliefs around the random walk prior, and controls the rate at which prior variance shrinks with increasing lag length. The vector of parameters ,…, are scale factors allowing overall coefficient covariance to be scaled by each variable on its own lagged values. The constant in the model receives a separate prior variance of . Exogenous variables can be assigned a separate prior deviation parameter, so that the prior variance on exogenous variables is . Moreover, Sims and Zha proposed to add two dummy observations account for unit roots, trends and cointegration. The is used to set prior weights on dummy observations for a sum of coefficient prior that implies beliefs about the presence of unit roots. The is the prior weights on dummy observations for trends and initial observations.

 

Brandt and Freeman (2006, 2009) create a table that summarizes the hyperparameters and their function of Sims-Zha prior. The table is displayed as the follows:

 

Table 3-2: Hyperparameters of Sims-Zha Reference Prior

 

Parameter

 

Range

 

Interpretation

 

[0,1]

 

Overall scale of the error covariance matrix

 

SD about (persistence)

 

Weight of own lag versus other lag

 

Lag decay

 

Scale of SD of intercept

 

Scale of SD of exogenous variable coefficients

 

Sum of autoregressive coefficients component

 

Correlation of coefficients/initial condition component

 

Note: This table is cited from the Table 2 of Barndt and Freeman (2009, 8).#p#分頁標題#e#

 

The eight aforementioned hyperparameters are helpful to translate our prior beliefs into a Bayesian model. Literally, these hyperparameters can cover the following list of general beliefs:

 

“1. The standard deviations around the first lag coefficients are proportionate to all the other lags.

 

2. The weight of each variable’s own lags is the same as those of other variables’ lags.

 

3. The standard deviation of the coefficients of longer lags is proportionately smaller than those on the earlier lags. (Lag coefficients shrink to zero over time and have smaller variance at higher lags.)

 

4. The standard deviation of the intercept is proportionate to the standard deviation of the residuals for the equation.

 

5. The standard deviation of the sums of the autoregressive coefficients should be proportionate to the standard deviation of the residuals for the respective equation (consistent with the possibility of cointegration).

 

6. The variances of the initial conditions should be proportionate to the mean of the series. These are “dummy initial observations” that capture trends or beliefs about stationarity and are correlated across the equations.“ (Brandt and Freeman 2006, 13)

 

The Sims-Zha prior applied in multivariate time-series analysis associated with Gibb sampler is as known as so-called BSVAR. Unlike BVAR applying Minnesota prior into model specification with equation-by-equation basis, BSVAR builds up its estimation method essentially based on the assumption that each variable in the model is conditioned to the others. Therefore, the power of BSVAR is that not only the contemporaneous effects but also the lagged influences from the corresponding variables in the structure are taken into account. Unlike the traditional statistical methods usually impose exgoneity or independence assumption on variables, the BSVAR takes over the difficulty of clarifying edogeneity crossover variables simultaneously and/or temporally. This substantially contributes to provide robust empirical analysis for most social studies featured with that almost every social variable is endogenously correlated with the others.

 

Apart from the conditional settings on SBVAR, the implementation of Gibb sampling technique also offers important finite-properties via MCMC’s largely sampling processes. When the original observations are small, model estimates derived from frequentist analysis are usually invalid and not steady to represent the asymptotic properties of presumed distribution. Through Gibb sampling to set up full conditional distributions and MCMC simulation, the BSVAR estimates are successfully demonstrated with the properties from large-sample and stationary distribution. In addition, these Bayesian estimates are presented as the form of probability distribution with dramatically convincing sample sizes. This PDF importantly provides the uncertainty toward our modeling estimates.#p#分頁標題#e#

 

3.4 research design

 

The causal and mutual dependent mechanisms are complicatedly embedded in my political-economic theory on the ER policy choices in EA. To unravel these complicated relationships surround ER policy decisions is not an easy task, especially when the data size is small in my study. I propose that the VAR model can properly catch the multivariate dynamics among variables.

 

Under the VAR modeling structure, in this research, I plan to conduct both generally and country-specifically empirical testing vis-à-vis my theoretical hypotheses. The general setting aims not only to provide a general perspective toward the ER policies and politics in EA, but also to play a reference benchmark to differentiate the empirical results from various countries. The country-specific testing aims at examining and reflecting variances of estimation across countries. The country-specific testing is important, because EA countries are well-known political and economic diverse so that would benefit to see how far from a country’s estimation to the generally regional findings. Likewise, in methodology, the unit-heterogeneity and uncertainty specification across various countries are being examined from the practical observations while applying my theoretical model.

 

However, in this research, the goal to pursuit for a consistent method that can be used for the general and country-specific settings is hardly to achieve. Recall the previous discussion that VAR model settings usually produce numerous parameters and need large samples to obtain valid estimation. Although the Bayesian analysis is able to cope with the difficulty of small data by imposing prior information at some points, current scholarly works have no successful development with respect to BSVAR analysis toward panel data. This means that I cannot use the BSVAR to conduct the general investigation underlined by total 225 pooling observations. With regard to the country-specific investigation, the frequentist approach cannot proceed to the statistical analysis with only 25 observations but associated with more than 25 parameters to be estimated for a country. Therefore, in terms of general and country-specific analyses, the SVAR investigation in this research must be divided into two sections based on different specification methods:

 

One, the statistical analysis on the general perspective is run by the frequentist approach based on the panel data. According to the basic SVAR settings, contemporaneous relationships are going to be identified and translated into the model specification.

 

Mostly important, in term of the number of lagged value in the dynamic model, I arbitrarily decide on one-lag () in the predetermined term for all endogenous variables. Because the ER is a high-frequent series that usually varies day-by-day or even minute-by-minute, the estimated effect on the long-term persistence according to my annual scale is hardly trustable. In practice, once the ER changes, the consumer price would correspondingly change and then dramatically affect the resources allocation. The relevant stakeholders by no means wait for taking necessary adjustments until one year later. Moreover, with regard to model estimation, more lagged variables would extend the model scale and escalate the difficulty of statistical calculation by my limited degrees of freedom.#p#分頁標題#e#

 

The EViews program is assigned to conduct this frequentist general analysis. Additionally, in order to modify our observations which can be stacked up and processed in a commeasurable scale collected from different development levels, I place a control variable -- the national GDP – into the SVAR model and which is treated as exogenous. The concrete data analysis and findings are arranged to be discussed in Chapter 4.

 

Two, the country-specific analysis is designed to present the unit variances by employing BSVAR method. In addition to my observable data sets, the Sims-Zha prior is assigned into the identified SVAR model. Unlike the general perspective along with 225 observations, this country-specific estimation is highly sensitive to the prior setting because there are only 25 observations (time-points) for each analytical county.

 

The model follows the one in the general setting that designs one-lag difference on the predetermined term and incorporates the national GDP as the exogenously control variable. In term of the country-specific presentation, all nine EA countries are divided into three groups based on the general distinction of their regime types: 1. Transitional democracy: Indonesia, Korea, the Philippines, Thailand and Taiwan; 2. Limited/ No democracy: China and Singapore; 3. Persistent democracy: Japan and Malaysia. This would help to catch a rough picture over the subtle information country-by-country, and then benefits to advance a concrete comparison.

 

The task of running the country-specific BSVAR model is undertaken by Patrick Brandt and Justin Appleby’s “MSBVAR Package” for R. Related technical discussion by running R and statistical findings from the BSVAR are presented on Chapter 5.

 

3.5 Conclusion

 

Overall, this chapter presents a measured operation, model selection and method debate, based on both theoretical settings and practical evidences, about analytical and statistical accounts of ER policy choices. ER is not simply an economic indicator reflecting the exchange values between various national currencies. States and any interest related actor are such sensitive and sophisticated to the management and value change of ER, as well as take necessary actions to intervene these ER decisions. Because of the features of my data type and model setting, to proceed to a statistical analysis aiming at unraveling the complicatedly political-economic relationships centered round ER is not an easy task.

 

This chapter starts from overlooking the data and resolving the measurement issues. These include illustrating the statistical description over each analytical variable, as well as generating innovative measuring methods on ER policies, external influences according monetary constraints and coercion, and regional adaptive diffusion apart from the on-hand indicators provided by existing data sets. This conduction allows readers owning broad understandings regarding the general distribution over variables across EA countries at beginning. Furthermore, the debates and gaps between theoretical hypotheses and practical findings usually result from the various understandings regarding defining variables and setting operationalized methods. The description of variable operationalization in this chapter thus demonstrates a basis for any intentional effort of model falsification.#p#分頁標題#e#

 

This chapter subsequently proceeds to conceptual discussion in reviewing model settings and methodological debates. The SVAR provides a better tool to shed light on the macro mechanisms in the multivariate dynamic system. Specifically, the incorporation of contemporaneous identification would accurately catch the structural relationships among variables rather than the other conventional single-equation and multiple-equation model settings. With regard to method specification, apart from the conventionally frequentist analysis, the Bayesian country-specific data analysis on the SVAR would better specify the context-conditional dependences as well as demonstrate the probability distribution of estimates in the modeling system.

 

Appendix

 

Appendix 3-1: Operationalization of ER Flexibility Index

 

According to Baig, and Wang and Yang, the basic operationalized model of the flexibility index which combines the volatility information from the changes of ERs and foreign reserves is as follows:

 

(A3-1) ,

 

where sdrex is the standard deviation of the monthly difference of exchange rate (IFS line AE: the price of a US dollar in terms of local currency) in a year, and sdrev is the standard deviation of monthly changes () in foreign reserves (IFS line 1d) divided by lagged stock of monetary base (IFS line 14) in a year. More specifically, the measurement of intervention (monthly) in the foreign reserves market as

 

(A3-2) ,

 

where is the log average of the absolute monthly change.

 

Appendix 3-2: Operationalization of PC Indicator

 

First, calculates the HERFGOV and HERFTOT by the HHI as the following formula:

 

(A3-3) ,

 

where is the seat share of party i in the central government and legislature, and N is the number of total elected parties. Therefore, based on the index ranging from zero to one, the higher (lower) index represents higher (lower) level of centralization and lower (higher) competition. Fortunately, the HERFGOV and HERFTOT indexes are provided by DPI.

 

Second, the index of PC is measured as the difference between HERFGOV and HERFTOT as follows:

 

(A3-4)

 

Appendix 3-3: Operationalization of WR

 

To assume all things being equal, the basic spatial autoregressive model that describes the relationship between a dependent variable and a spatial autoregressor can be formally expressed as follows:#p#分頁標題#e#

 

(A3-5) ,

 

where W denotes the standardized spatial weights matrix, y is the vector of all neighboring values, represents average y value of i’s neighbors, and is a scalar coefficient that measures the spatial term .

 

Figure A3-1 is a Box-Plot that graphs the median and quartile distribution of ERR is EA. I use the descriptive information to draw the figurative map that describes the ERR propensity in EA which is shown as Figure A3-2. Figure A3-2 clearly marks the relative distances between decision authorities. Moreover, this propensity map particularly helps to produce the spatial weights matrix while proceeding to define the spatially dependent

 

Figure A3-1: Box-Plot for ERR across EA Countries (1980-2004)

 

Figure A3-2: Figurative Map of ERR Propensity

 

variable, WR. The formal expression of WR with the notion of spatial weights matrix is as the follows:

 

(A3-6) ,

 

where and denote the spatial positions of country i and its eight neighboring countries j on the propensity map as Figure A3-2, respectively. indicates the ERR in the various neighboring countries. Recall the essential assumption in a spatial contiguity analysis that the degree of spatial effect is proportioned to the closeness between two analytic units. Equation A3-6 transforms various decision-distances of propensity into the weighting aggregate that represents the neighborhood effects on ERR.

 

Appendix 3-4: Operationalization of WL

 

The center to the non-geographic spatial method introduced in Appendix 3-3 is to apply median and interquartile range statistics to locate every country into the policy propensity map. Figure A3-3 is the Box-Plot that illustrates the descriptive statistics toward ERL in EA. Figure A3-4 is the propensity map that illustrates each country’s spatial position and relative distances to other counterparts. It shows that Korea mostly tends to appreciate its ER, Japan is more likely to depreciate its currency, and Singapore frequently adjusts its ER in a moderate level.

 

Furthermore, by employing the Equation A3-6 and substituting ERL into , WL (i.e. in A3-5) can be obtained to specify the adaptive effects on ERL policy making.