Essays on China’s Macroeconomic Fluctuations By Yueqing ...
Transcript of Essays on China’s Macroeconomic Fluctuations By Yueqing ...
Essays on China’s Macroeconomic Fluctuations
By Yueqing Jia
B.A in Economics, July 1995, Shandong University, China
M.A in Economics, January 1998, Shanghai University of Finance and Economics, China
M.A in Economics, January 2008, the George Washington University
A Dissertation submitted to
The Faculty of
The Columbian College of Arts and Sciences in partial fulfillment of the requirements for the degree of Doctor of Philosophy
January 31, 2012
Dissertation directed by
Tara M. Sinclair
Associate Professor of Economics and International Affairs
ii
The Columbian College of Arts and Sciences of The George Washington University certifies that
Yueqing Jia has passed the Final Examination for the degree of Doctor of Philosophy as of
November 14th, 2011 . This is the final and approved form of the dissertation.
Yueqing Jia
Dissertation Research Committee:
Tara M. Sinclair, Associate Professor of Economics and International Affairs, Dissertation Director
Frederick L. Joutz, Professor of Economics, Committee member
Maggie X. Chen, Associate Professor of Economics and International Affairs, Committee member
Essays on China’s Macroeconomic Fluctuations
iii
© Copyright 2011 by Yueqing Jia
All rights reserved
iv
Acknowledgements
I am grateful to my advisor Professor Tara Sinclair for her invaluable support and guidance
throughout my dissertation writing process, and want to express my great appreciation to the
other members of my dissertation committee: Professor Fred Joutz, Professor Maggie Chen,
and dissertation reader Professor Neil Ericsson and Dr. Mark DeWeaver for providing very
helpful comments and suggestions. My appreciation also extends to the participants and
discussants in the Southern Economist Association annual meeting and Georgetown Center
for Economic Research 2011 conference at Washington DC, for their insightful comments
and suggestions. I thank the Institute for International Economic Policy of GWU and GW-
CIBER for the support for my dissertation research.
v
Abstract of Dissertation
Essays on China’s macroeconomic fluctuations
The dissertation contributes to the understanding of the properties of China’s
macroeconomic fluctuations through applying advanced econometric methods to China’s
quarterly real GDP, which have been investigated little due to the shortage of data. The first
essay provides quarterly real GDP estimates for China from 1978q1-1991q4 using an
unobserved components approach. The approach imposes fewer prior restrictions on related
series and is more flexible than other disaggregation methods. The multivariate unobserved
components model with total trade and domestic credit as related series is selected as the best
fit model for temporal disaggregation of China’s real GDP. The estimated quarterly real GDP
data are then evaluated with univariate and multivariate time series analysis techniques. The
constructed quarterly data are shown to be of good quality and to provide valuable
information for the analyses of China’s macroeconomic fluctuations. I extend the study to the
macroeconomic linkages between China and other economies in the second and third essays.
In the second essay a two-country correlated UC model is applied to explore the relationships
between the real output fluctuations for the US and China over the period 1978q1-2009q4.
The two countries are found to share approximately half of their permanent and transitory
shocks. The third chapter investigates the relationships between the real output fluctuations
of China with those of developed countries over the period 1978q1-2009q4. The results
show that the correlations between the real output fluctuations of China and the developed
world are insignificant both in terms of permanent and transitory shocks. The analysis of this
dissertation with China’s quarterly data suggests that domestic factors may be the major
drivers of China’s macro-economic fluctuations during the sample period.
vi
Table of Contents
Acknowledgments…………………………………………………………………………..iv
Abstract of Dissertation…………………………………………………………..……...…..v
List of Figures……………………………………………………………………………...vii
List of Tables…………………………………………………………………….….……...x
Introduction………………………………………………………………………..………..1
Chapter 1 A New Look at China’s Output Fluctuations:
Quarterly GDP Estimation with an Unobserved Components approach………..6
Chapter 2 Permanent and Transitory Macroeconomic Relationships
between the US and China…………………………………..…………………70
Chapter 3 Permanent and Transitory Macroeconomic Relationships
between China and the Developed World……………………….....………….100
References…………………………………………………………………………..……..128
Appendices…………………………………………………………………………..…….143
Appendix 1-1: Literature review on studies of China’s macro data quality
Appendix1-2. the unobserved components decomposion model
Appendix1-3. Standard bivariate Blanchard-Quah model and decomposition
Appendix1-4: More results from the GVAR model estimation
with MUC and DdPS data.
vii
List of Figures
Figure1-1. China’s most recent revised official annual and quarterly real GDP
year on year growth rates (the shaded areas are “slowdown eras”)………………60
Figure 1-2a. The log quarterly real GDP (2000 as base year)
and the potential related series………...………………………………………….60
Figure1- 2b. Log annual data………………………………………………………………...61
Figure 1-2c: quarterly year on year growth rates of real GDP with
monetary related series (the shaded areas are “slowdown eras”)…………….61
Figure 1-2d: quarterly year on year growth rates of real GDP with
international trade related series (the shaded areas are “slowdown eras”) …....62
Figure1- 3: Disaggregation model selection:
Year on year quarterly growth rates (%) 1992-2008………..………….…..……62
Figure1- 4: Year on Year quarterly growth rate
(comparing with A&R from 1979-1991)……………………………..………….63
Figure1-5: Seasonal factors or China’s quarterly real GDP MUC
temporal disaggregation model and X12 seasonal adjustment method………....63
Figure 1-6: HP cycles with different value of λ……………………………………………64
Figure 1-7: Unobserved components decomposition: filtering and smoothing…………….64
Figure 1-8: HP and Christiano-Fitzgerald cycles of MUC temporal disaggregation
and A&R estimation of China’s quarterly GDP 1978-1992………………………65
viii
Figure1- 9: Linear in time residual, HP and BP cycles…………………………………..…65
Figure 1-10: HP, Christiano-Fitzgeral and UC cycles………………………………….…..66
Figure 1-11 a: Seasonal adjusted inflation (CPI) and real GDP level 1986-2010………….66
Figure1-11. b: First difference of log seasonal adjusted inflation
and real GDP 1986-2010………...…………………………………………………………..67
Figure1-12: Impulse responds functions on real output and inflation……………………...67
Figure1-13: Blanchard-Quah output gap with HP and Christiano-Fitzgerald cycles……….68
Figure 1-14: Blanchard and Quah cycles based on MUC data and A&R data ……………..68
Figure 1-15: The DdPS quarterly real GDP data and the MUC
estimated quarterly real GDP data ……………………………………………..69
Figure 2-1: Estimated permanent and transitory components………………………………94
Figure 2-1-1 : The US………………………………………………………...……………..94
Figure 2-1-2: China…………………………………………………………………..…..….95
Figure2- 2: Transitory Components Comparison………….…………………………..…...96
Figure 2-2-1: US Transitory Component: Comparing with HP Cycle……………………..96
Figure 2-2-2: China Transitory Components: Comparing with HP Cycle…………………96
Figure 2-3: US Transitory Component Comparing Different Information Sets…………..97
Figure 2-3-1: US Transitory Components Comparing: Univariate,
with China, with Inflation and with Oil price…………………………………..97
ix
Figure 2-3-2: US Transitory Component Comparing: with China vs. with Canada ……...98
Figure2- 4: China Transitory Components with Different Information Sets……………..99
Figure 3-1: Estimated permanent and transitory components……………………………123
Figure3-1.a: China Based on Bivariate Model with G7……………...…………………….123
Figure3-1.b: China Based on Bivariate Model with OECD……………………………….123
Figure3-1.c: G7 Based on Bivariate Model with China…………..………………………..124
Figure3-1.d: OECD Based on Bivariate Model with China………………………………..124
Figure 3-2: Comparing the Different Filtered Cycle Estimates:
Univariate and Bivariate Models………………………………….…………..125
Figure 3-3: Comparing the Different Cycle Estimates:
Univariate, Bivariate, and Trivariate Models………………………………………………125
Figure 3-4: Comparing the Cycle Estimates: DW aggregate,
Exports and Trade Balance………………………………….…………………126
.Figure3-5: 2007 – 2009 Chinese Real GDP and Permanent Component Estimates………126
Figure3- 6: 2007 – 2009 G7 and OECD Real GDP and
Permanent Component Estimates……………………………………………127
Figure 4A-1 Impulse response function: MUC data and DdPS data………………………155
x
List of Tables
Table 1-1: Unit root test results
(Augmented Dickey-Full Test on annual data 1978-2009)………….……..…….51
Table1-2: Johansen Co-integration test results of annual data……………………….….….52
Table1- 3: Disaggregation model selection………………………………………………….53
Table 1-4: China quarterly real GDP data: MUC model estimation,
A&R estimation and the official data (1978q1-2011q2)…………… …………….……..….54
Table 1-5: Correlations of year on year growth rates of quarterly real GDP
and potential related series……………………………………………………….57
Table 1-6: Temporal disaggregation parameter estimates---
MUC model with domestic credit and total trade (Log real GDP equation only);
Variance/correlation of cross series components for Log GDP
(final model: 1978q1-2009q4 with 1978q1-1991q4 real GDP missing) ……….....…57
Table 1-7: Unobserved component model parameter estimates (maximum likelihood)…....58
Table 1-8: Cointegration test of DdPS data and MUC data (1979q2-2003q4)……………....59
Table 1-9: Cointegrating analysis of GVAR modeling for China with MUC data
and DdPS original data (1979-2003, replicating of DdPS 2007)…… … ………59
Table2-1: Correlations of cycles of the US and China real GDP with HP,
BP decomposition and the growth rates Quarterly Data, 1978.1 – 2008.4………………..…92
Table 2-2: Estimation Results………………………………………………………………..92
Table 2-3: Standard Deviations of Shocks…………………………………………………..93
Table2- 4: Correlations of Permanent and Transitory Shocks………………………..……..93
Table 3-1: Correlations of Cycles for China and the Developed Country Aggregates……120
Table 3-2: Estimation Results……………………………………………………………..120
xi
Table3- 3: Standard Deviations of Shocks…………………………………………………121
Table 3-4: Within Series Correlations of Shocks…………………………………………..122
Table 3-5: Cross Series Correlations of Shocks……………………………………………122
Table 4A-1. Likelihood ration tests on the equality of cointegrating coefficients
estimated by GVAR modeling for China
with MUC data and DdPS original data………………………….…………...153
Table 4A-2: Short run error correction equation coefficients of GVAR
estimated based on MUC data and DdPS data …………………………………..………....154
1
Introduction
In the past three decades, China has emerged as one of the most important and
influential economies, with its remarkably rapid growth and integration into the world
economy. As the world’s second largest economy1, the largest exporter (since 2009) and the
world’s largest foreign exchange reserves holder, China’s economy, in the short and long
run, is more than ever the focus not only of academic research but also of policy makers and
stakeholders from inside and outside China. Although increasingly important, the properties
of China’s output fluctuations are not well understood. Very limited econometric analysis has
been conducted on China’s macroeconomic fluctuations. This is mainly due to the shortage
of high-frequency data. The complexity of China’s transitional economic and political
structures adds difficulties to the analysis. Lack of proper characterization of China’s output
trend and cycles may mislead the theoretical economic studies on Chinese economy as well
as the policy analysis. While the existing research relies mostly on the available annual data,
the study of low-frequency annual data obviously cannot fulfill the needs of economic
decision making in this fast changing world.
This dissertation contributes to the existing literature on empirical studies on China’s
macro economy in the following aspects: First, it provides a new temporal disaggregation of
quarterly real GDP data through the more flexible and general unobserved components (UC
hereafter) model framework. Second, the application of correlated UC model to China’s
quarterly real GDP data since 1978 adds to the limited existing empirical research on China’s
business cycle with advanced time series econometrics analysis. This dissertation provides
new evidence and policy implications for the understanding of China’s growth fluctuations in
1China passed Japan as the second largest economy in terms of nominal GDP in 2010.
2
terms of permanent and transitory shocks. Third, this study is the first application of
multivariate correlated UC model analysis of China’s quarterly real GDP with other
economies and external variables such as oil prices. With information from other economies
and external variables, I find that China’s macro-economic fluctuations are mainly driven by
domestic factors and economic policies.
The first essay of the dissertation provides temporal disaggregate estimates of China’s
quarterly real GDP data for the years 1978 through 1991, using multivariate unobserved
components models with state-space representation. China’s official quarterly real GDP data
are only available since 1992, which provides only 78 quarterly observations up to 2011 q2.
Small samples can limit the applicable methods and the quality of empirical analysis.
Extending the quarterly real GDP data from 1991 back to 1978 for China would provide a
complete sample of growth fluctuations for the economy since the beginning of the reforms
and thus a better understanding of the evolving properties of China’s macroeconomic
fluctuations along with the implementation of the reforms. Abeysinghe and Rajaguru (2004,
A&R hereafter), the only published study on the temporal disaggregation of China’s GDP
data, applies the original Chow-Lin (CL) related series technique to disaggregate China’s
annual real GDP data into quarterly data and provides quarterly real GDP growth rate
estimates from 1978Q1 to 1994 Q4. However, as I discuss in the first essay, the original CL
model based on univariate regression assumes a linear relationship among the related series
and does not consider the unit root properties of the series. Both assumptions may not be
proper in practice when choices of available related series are very limited.
In the first essay, I generalize the univariate and multivariate unobserved components
modeling for temporal disaggregation, and provides temporal disaggregate estimates of
3
China’s quarterly real GDP data for the years 1978 through 1991, using the selected
multivariate unobserved components model. The unobserved components approach is more
general than the Chow-Lin approach in dealing with unit root, seasonality and irregularity
properties. The method allows simultaneous disaggregation and seasonal adjustment of the
data, and imposes minimal prior restrictions on the data. It provides more flexibility for the
data selection, which is especially important for emerging countries whose high frequency
data are very limited. The multivariate unobserved components (MUC) model with domestic
credit and total trade as related series is selected as the best fit model based on the root mean
squared standard errors of the estimated data and the official published data over the
overlapping period. The estimated data from 1978-1991 with selected MUC model are more
efficient than the estimation from other temporal disaggregation methods.
The MUC estimated quarterly real GDP for China provides a better alternative of
China’s quarterly real output data for different univariate and multivariate time series
analyses. To evaluate the data quality, I apply different univariate trend cycle analyses, such
as Hodrick-Prescott filter (1997, HP filter hereafter), Band-Pass Baxer-King (1999) and
Christiano-Fizgerald (2003) filter and unobserved components decomposition method, and
structural multivariate analysis, such as Blanchard-Quah (1989) decomposition and Global
Vector Autoregression models (Dees et. al 2007, DdPS model hereafter) to the estimated
data. The analyses show that the extension of quarterly real GDP with the MUC model
provides valuable information for the understanding of the output fluctuations during the
sample period. Domestic factors and supply shocks are found to be the main driver of
China’s output fluctuations.
4
I extend the study to the macroeconomic fluctuation linkages using quarterly data
between China and other economies in chapter two2 and chapter three3 . Chapter two
contributes to the literature that applies the multivariate correlated unobserved components
model, a more general model with less restrictions and priors than the simple correlation and
VAR approaches, to investigate economic relationships of two economies at different
development levels and with more divergent economic structures. The relationship between
the macroeconomies of the US and China is explained through the permanent and transitory
components. The chapters also contribute to the limited literature on empirical studies of
properties of China’s macroeconomic fluctuations with a reasonably long sample of quarterly
data.
The model employed in the second chapter is a two-country correlated unobserved
components model based on the correlated unobserved component model proposed by MNZ
(2003) and extended by Sinclair (2009) and Mitra and Sinclair (2009). The model
specifically allows for the distinguishing of cross-country correlations driven by the
relationships between permanent innovations, caused by real shocks such as changes in
technology and economic and social institutions, from transitory or cyclical movements,
caused by changes in aggregate demand or monetary shocks in the two countries. The model
is also capable of exploring the role of information from the dynamics of each country in
identifying fluctuations in the other country.
The economic fluctuations of the US and China are found to be significantly
positively correlated for both permanent and transitory shocks. The two countries share about
2 The second chapter is based on joint work with Tara M. Sinclair that is currently under revision and resubmission to
the Journal of International Money and Finance. 3The third chapter is based on the joint work with Tara Sinclair prepared for the CESifo Venice Summer Institute
workshop on The Evolving Role of China in the Global Economy and to be published in a conference volume by MIT press.
5
half of the shocks both in the long run trend and short run movements. Estimates of China’s
permanent and transitory components do not change too much with information from the US
and alternative external information sets as well, which suggests that domestic factors may be
the major drivers of China’s real GDP fluctuations. The US transitory components estimated
with China data are very different from that estimated with other information sets such as
inflation, GDP of other developed countries and the oil price.
The third chapter investigates the relationships between the real output fluctuations of
China with those of developed countries over the period 1978q1-2009q4. The chapter
focuses on two measures of aggregate developed-country output: real GDP for the G-7
countries and real GDP for 25 OECD developed countries. The results of chapter three show
that the correlations between the real output fluctuations of China and the developed world
are insignificant both in terms of permanent and transitory shocks.
The results of analyses in all chapters, with bivariate unobserved components models,
Blanchard-Quah decomposition, and the GVAR models, are in agreement in that supply side
shocks and domestic factors play an important role in China’s real output movements.
Although China’s economy has been widely open to the world economy, outside shocks,
which may mainly be on the demand side, may have either not been as strong as that from
the domestic economic reforms and productivity changes, or have been effectively offset by
China’s macro-economic policies.
6
Chapter 1: A New Look at China’s Output Fluctuations: Quarterly GDP Estimation
with an Unobserved Components Approach
I. Introduction
“China moves to centre stage”
----Cover story, Economist, Oct. 30, 2008
In the past three decades, China has emerged as one of the most important and
influential economies, with its remarkably rapid growth and integration into the world
economy. China is the world’s second largest economy in 20104, the largest exporter (since
2009) and the world’s largest foreign exchange reserves holder. The recent financial crisis
pushed China to the frontier of world economic development. China’s economic
performance, in the short and long run, is more than ever the focus not only of academic
research but also of policy makers and stakeholders from inside and outside China.
Although increasingly important, the properties of China’s output fluctuations are not
well understood. Very limited econometric analysis has been conducted on China’s
macroeconomic fluctuations. This is mainly due to the shortage of high-frequency data. The
complexity of China’s transitional economic and political structures adds difficulties to the
analysis. Lack of proper characterization of China’s output trend and cycles may mislead the
theoretical economic studies on Chinese economy as well as policy analysis. While existing
research relies mostly on the available annual data, the study from low-frequency annual data
obviously cannot fulfill the needs of economic decision making in a fast-changing world.
4 China passed Japan as the second largest economy in terms of nominal GDP in 2010.
7
China’s official quarterly real GDP data are only available since 1992, which
provides only 78 quarterly observations up to 2011 q2. Small samples can limit the
applicable methods and the quality of empirical analysis. Just shortening the sample period
and ignoring the available annual observations before the quarterly data are available results
in losing important information for the properties of data generating process. Extending the
quarterly real GDP data from 1992 back to 1978 for China would provide a complete sample
of growth fluctuations for the economy since the beginning of the reforms and thus a better
understanding of the evolving properties of China’s macroeconomic fluctuations along with
the implementation of the reforms.
As to flow data such as real GDP, one way to solve the above problem is to
temporally disaggregate or interpolate the low frequency data into higher frequency data.
Using a proxy observed at higher frequency and estimating the real GDP with the production
function would be alternatives. However, for China, quarterly macro-economic data before
1992 are extremely limited. The only available series are from monetary and international
trade statistics. They are not sufficient to estimate the proxy and production function data
construction alternatives. Temporal disaggregation of annual real GDP to quarterly real GDP
with available related information thus becomes the only practical approach for the quarterly
GDP construction for the period. Temporal disaggregation is also a commonly used method
for resolving similar problems to other countries5 . Univariate methods, related series
univariate method or Chow-Lin method (Chow and Lin 1971, CL model hereafter) and
multivariate unobserved components (UC approach hereafter) methods are the three groups
5 For example the Eurostat (1999) documents that the temporal disaggregation method is used in the official statistics
agencies of the European countries.
8
of approaches that have been applied to temporal disaggregate macro-economic data in
literature.
Abeysinghe and Rajaguru (2004, A&R hereafter), the only published study on the
temporal disaggregation of China’s GDP data, applies the Chow-Lin related series technique
to disaggregate China’s annual real GDP data into quarterly data and provides quarterly real
GDP growth rate estimates from 1978Q1 to 1994 Q46. However, as I will discuss below, the
CL model based on univariate regression assumes a linear relationship among the related
series and does not consider the unit root properties of the series. Both assumptions may not
be proper in practice when choices of available related series are very limited.
My study generalizes the univariate and multivariate unobserved components
modeling for temporal disaggregation, and provides temporal disaggregate estimates of
China’s quarterly real GDP data for the years 1978 through 1991 using the selected
multivariate unobserved components model (MUC model hereafter). The unobserved
components approach is more general than the Chow-Lin approach in dealing with unit root,
seasonality and irregularity properties. The method allows simultaneous disaggregation and
seasonal adjustment of the data, and imposes minimum prior restrictions on the data. It
provides more flexibility for the data selection, which is especially important for emerging
countries whose high frequency data are very limited.
I temporally disaggregate China’s real GDP series using the unobserved components
models with different specifications of components and different combinations of related
series for model selection. The multivariate unobserved components (MUC) model with
6On their website (http://courses.nus.edu.sg/course/ecstabey/gdpdata.xls), the authors extended the series through
2007Q1, using quarterly year-on-year real GDP growth rates from the country data of Economist Intelligence Unit. The data resources of EIU country data for China are CEIC and National Bureau of Statistics of China (NBS).
9
domestic credit and total trades as related series is selected as the best fit model based on the
root mean squared standard errors of the estimated data and the official published data over
the overlapping period. The estimated data from 1978-1991 with selected MUC model are
more efficient than the estimation from other temporal disaggregation methods7.
The MUC estimated quarterly real GDP for China provides a better alternative of
China’s quarterly real output data for different univariate and multivariate time series
analyses. To evaluate the data quality, I apply different univariate trend cycle analyses, such
as Hodrick-Prescott filter (1997, HP filter hereafter), Band-Pass Baxer-King (1999) and
Christiano-Fizgerald (2003) filter and unobserved components decomposition method, and
structural multivariate analysis, such as Blanchard-Quah (1989) decomposition and Global
Vector Autoregression models (Dees et. al 2007, DdPS model hereafter) to the estimated
data. The analyses show that the extension of quarterly real GDP with the MUC model
provides valuable information for the understanding of the output fluctuations during the
sample period. Domestic factors and supply shocks are found to be the main driver of
China’s output fluctuations.
The purpose of this paper is to provide quarterly real GDP for China in consistency
with the official real output data. Thus, the availability and reliability of China’s official data,
the big concerns for China’s official GDP statistics, are carefully discussed before the data
construction.
There are six sections of this paper: Section II reviews the relevant literature on the
methodology of temporal disaggregation and the contribution of my paper in high frequency 7The comparison of the estimation of different approaches based on the root mean square errors is presented in section
V. The selected MUC estimation, including the observed annual levels in the estimation, fits the observed annual real growth rates better than the A&R estimation.
10
data construction; Section III presents the unobserved components or structural time series
models for the temporal disaggregation of China’s real GDP. Section IV addresses China’s
macroeconomic data problem. In Section V, the results of the data construction from
different model specifications are presented and evaluated. Section VI analyzes China’s real
GDP fluctuations with the constructed quarterly real GDP data using different univariate and
multivariate approaches and compares the results of permanent and transitory
macroeconomic fluctuations based on the constructed quarterly real GDP data. Section VII
concludes.
II. The literature
This study is related to the following strands of literature: first, the estimation of
missing high-frequency macro economic data from available low-frequency data, i.e.
interpolation and temporal disaggregation methods; second, the research on identifying
China’s output fluctuations and the commonly used univariate and multivariate aggregate
output trend and cycle decomposition methods. To evaluate the quality of the quarterly real
GDP data estimated in this paper, both univariate and multivariate methods are applied to the
estimated quarterly real GDP data for China. The unobserved components model is the key
modeling framework applied in this paper.
2.1 Literature on missing high frequency data and temporal disaggregation methods of
time series
Temporal disaggregation or interpolation has been extensively used by researchers
when high frequency data required by econometric analysis are not available. It is also a
routine practice for official statistical agencies to apply temporal disaggregation methods to
11
generate high frequency data, especially when direct estimation methods are unavailable or
information collection is costly (Proietti 2006, Eurostat, 1999).
Problems of time series disaggregation include interpolation of stock variables and
temporal disaggregation of flow variables. The estimation of China’s quarterly Gross
Domestic Product from the annual official data falls into the second category. There are
broadly three groups of methods that have been developed to temporally disaggregate lower
frequency data into higher frequency data:
1) univariate methods rely on the time series properties of the targeted series only.
For example, Stram and Wei (1986, 1990) derive smoothed estimation of unavailable high
frequency data based on the ARIMA structure of the series. Stram and Wei (1990) method
has been applied to the estimation of macroeconomic indicators by many official statistic
agencies (Eurostat 1999, part 6.45-6.46).
When the missing high frequency data period is short compared to the whole sample
period, a simple univariate interpolation is often convenient. However, when missing high
frequency data are for a relatively long period, the simple univariate interpolation, which
uses the time series properties of the target series itself (usually properties of the series
during more recent period when the high frequency data are available), is likely to distort the
results of high frequency analysis. Especially for the sample period when high frequency data
are missing. The problem may be more serious for data from emerging or transitional
economies. Their underlying economic structure and environment often change substantially.
More sophisticated methods that use more information to disaggregate data should be
considered.
12
2) Related series univariate models that were first proposed by Chow and Lin 1971
and extended by Fernandez (1981) and Litterman (1983) with a random walk and I (1) error
term respectively. Application of temporal disaggregation by the CL approach begins with
running OLS or GLS on a linear model of target series to the related series with low
frequency data. Assuming that the linear relationship of target series and related series is
consistent with low frequency and high frequency data, the estimated coefficients are then
used to predict the target series based on the high frequency related series data with
adjustment to match the annual aggregates. An AR (1) process for the error term is assumed
in the original CL model. To account for the non-stationary residuals, a random walk process
is assumed in the Fernandez model and I (1) in the Litterman model. Santos and Cardoso
(2001) and recently Proietti (2006) add lagged values of the dependent variable
(autoregressive distributed lag or ADL models) into the CL model. Harvey and Pierse (1984)
and Proietti (2006) present these groups of models in state space form and apply the Kalman
filter to estimate the missing observations.
The CL approach and its extension models use more information from observed
related high frequency data. Chen (1993) demonstrates with Monte Carlo simulated data that
the CL procedure is usually more efficient than the univariate only alternatives. The
convenience in application has made this method more popular in practice than the first
group of models. A&R (2004), the only published temporal disaggregation of China’s real
GDP, applies CL approach to the growth rates of GDP with the growth rates of M1 and total
trade as related series.
13
The major problem with this group of models is that the assumption of a linear
relationship between the target series and the related series is often difficult to verify with the
data (Proietti 2006, Moauro and Savio 2005).
As Harvey (1989 section 8.7) and Proietti (2006) mentioned, the univariate related
series models, although widely applied in practice, impose a strong assumption of
cointegrating relationship between targeted series and the related series and/or exogeneity of
the related series regressors. Moauro and Savio (2005) further proves that even with the
existence of cointegrating relationship, univariate with related series models, or CL method
will be efficient only when the constant and the autoregressive parameters are equal for all
included related series, or in terms of unobserved component model specification, the related
series are trend homogeneous. The assumptions are not likely to be verified in the true data,
especially when the choices of related series are very limited8.
The restriction of CL methods limits the choice of related series. The target series and
the sets of available “related” series may not have a linear relationship, but are usually
affected by the same economic environment and thus could provide valuable information and
improve the efficiency of the disaggregation. A&R, who use nominal M1 and the total trade
as related series, apply CL approach to the first-differences and the growth rates instead of
levels of the real GDP and related series to avoid the non-stationary and cointegration
problem. Even with the growth-rate approach, A&R did not find a significant linear
relationship between the growth rate of real GDP and M1, but found including M1 leads to
8 When choices of related series are limited, if no linear relationship is found among the series, there will be no
alternatives available.
14
better disaggregation9. Although the relationship of M1 and real GDP is not significant with
annual data, it provides information for quarterly GDP estimation that improves the
efficiency of the disaggregation.
3) Multivariate unobserved components (MUC) or structural time series models. The
application of UC models to temporal disaggregation was originated by Harvey and Pierse
(1984) and Harvey (1989)10. Harvey and Chung (2000), Moauro and Savio (2005) are
examples of the contributions of this group of models. MUC models set up a system of
unobserved components equations of the targeted and related series and estimate the models
in the system simultaneously. The approach allows cross series correlations for the
components and thus includes the quarterly information provided by the available related
series to the disaggregation of target series.
The MUC approach overcomes the limitation of CL methods. The MUC models are
capable of taking advantage of information from the available high frequency “related” time
series for the disaggregation, without putting prior restrictions on the specific relationship
between the series. Common trends, common cycles and common seasonalities among the
related series can be tested through MUC models (Moauro and Savio 2005). The MUC
approach is also flexible in handling both seasonally adjusted and non-seasonally adjusted
data and allows disaggregation and seasonal adjustment simultaneously. Proietti and Moauro
(2005) further show that the model is capable of handing seasonality very well. In addition,
with the Kalman smoothing estimation, the sample period can be extended to include the
information of the available high frequency observations of the targeted series in the later
9
A&R evaluates the disaggregation by comparing the quarterly data estimated from the model with the real data during
the overlapping period. 10 Harvey (1989) names his model “seemingly unrelated structural time series equation (SUTSE)”.
15
years. For example, China’s official quarterly real GDP data that are available since 1992 can
be included in the estimation of China’s real quarterly GDP during 1978-1991 in MUC
models. Since the time series property of a time series itself is usually stable over time, the
property of the observed high frequency data of the target series can help improving the
temporal disaggregation.
Since Harvey and Pierse (1984) first cast the univariate disaggregation methods into
state-space form and applied Kalman filter technique to estimate the missing high frequency
data, the state space approach has been regularly applied in the temporal disaggregation of
time series. Harvey (1989, section 6.3) proposed the method of using cumulator series to set
up the state space form over series of different time frequencies, where the missing high
frequency data is treated as missing observations, estimated with the Kalman smoothing
algorithm. The method provides more flexibility in modeling components of the series and
can be applied to both univariate and multivariate disaggregation models on both flow and
stock time series. Durbin and Quenneville (1997), Proietti (2006) generalized the state space
method applications to the CL model and its extensions. The related series are modeled as
exogenous regressors that enter into the measurement equation and/or transitory equation.
Cuche and Hess (1999) and Tasdemir (2008) disaggregate European and Turkish data using
the state space methods.
This paper further generalizes the models for the temporal disaggregation of flow
series with unobserved components stemming from Harvey (1989) and Moauro and Savio
(2005). I present the univariate and multivariate unobserved components models, with or
without cyclical components, in state space forms and estimate China’s quarterly real GDP
16
during 1978-1991 with Kalman smoothing algorithm. This paper is the first application of
temporal disaggregation of China’s real GDP with the unobserved components approach.
2.2 Literature on identifying and characterizing China’s output fluctuati ons
To evaluate the information provided by the quarterly real GDP levels of China from
1978-1991 estimated by the selected unobserved components model in this paper, I apply
different univariate and multivariate time series analytic methods to the data. Literature on
the measurement of China’s potential output or output gap is then closely related to this part
of study. The existing studies generally follow three approaches: the production function
approach, univariate trend-cycle decomposition approaches and multivariate time series
approaches or Vector autoregression approaches (VAR).
The trend of aggregate output is generally assumed to correspond to potential output
and the cycle is assumed to correspond to the output gap. Most of the studies on China’s
potential output estimation apply production function approach to annual data, which
estimates the Cobb-Douglas production function with potential capital and labor inputs11.
Decomposition of the aggregate output series into trend and cycle components has
been a common practice for aggregate output fluctuation analysis. Competing approaches
have been developed to decompose macroeconomic series such as the aggregate output into
“trend” and “cycle”, or permanent and transitory components. A number of studies (Morley
2008, Canova 1998, Zarnowitz and Ozyildrim 2006, Park 1996 and King and Rebelo 1993,
11
In this group of studies, Chow (1993 and 2002) intends to find the importance of capital formation and contributions of sectors, Heytens and Zebregs (2003) try to find the growth of Total Factor Productivity, Young (2003) focuses on alternative price levels, and Scheibe and Vine (2005) study the Phillips curve. Scheibe (2003) explores the production function with sector based estimation. This approach is also used for alternative GDP data constructions.
17
Gerlach and Yiu 2003) have shown that the revealed trend cycle properties are sensitive to
the detrending methods12.
For China’s real GDP, the Hodrick-Prescott (HP) filter is the widely used univariate
detrending method in the literature that models China’s business cycles (for example: Ha,
Fan and Shu 2003)13. Gerlach and Peng (2006) estimate annual Chinese output gap from
1982-2003 using the unobserved components (UC) model following Watson (1986) and
Clark (1987). They find that HP filter and UC approaches generate similar cyclical patterns.
Their estimation suffers from limitations due to the small sample size (21 observations) and
very broad confidence bands. Laurenceson and Rodgers (2010) use different frequencies of
cycles to identify the relative importance of demand and supply volatilities occurring at
China’s business cycle.
In recent years, several Chinese scholars applied nonlinear univariate methods such as
Markov-switching process to indentify the phases of business cycles in China (mostly
published in Chinese), examples of the studies include Chen and Liu (2007) Liu (2003, Liu
12
Canova (1998) examines the business cycle properties of US real macroeconomic time series data with seven different decomposition methods and finds that for aggregate output of the US, different decomposition methods generate cycles that differ in time duration and turning points. As for developing countries, Gerlach and Yiu (2003) compare output gaps with annual data for eight Asian economies (not including China) derived from four decomposition methods, and find gaps (cycle component) from HP, UC and BP decompositions are similar for these countries but the gaps derived by BN decompositions are different.
13 Although the production function approach uses more information for potential GDP estimation than the alternative
univariate approaches, it also introduces more potential problems. First, several assumptions are frequently made to set up the production function. Assumptions such as constant returns to scale in production, competitive markets for inputs and outputs may not be appropriate for China. There are no generally accepted potential levels of labor, capital inputs and total factor productivity growth for China. Secondly, the estimation needs capital and labor data, which faces similar data availability and reliability problems as well. Labor or employment data are even less available and reliable for China than GDP data. Missing capital and employment data have to be estimated before they are used in the potential GDP estimation. Third, the estimation must select proper price levels for different inputs and sectors (Young 2003), which are unavailable and have to be estimated. Finally, most research using this approach relies only on annual data due to the data availability problem. Because of the issues above, the results of production function estimations are not consistent among different studies.
18
and Zheng (2008), Zheng, Teng and Song (2010). Most of these researches suffer from small
sample problem.
In this paper, I apply the most widely used univariate methods include the Hodrick
and Prescott (1997, HP) filter, the Band-Pass filter (Baxter and King 1987, Christiano and
Fizgerald 2003, BP) filter, and the unobserved components models (Harvey 1985, Watson
1986, Clark 1987) to the estimated quarterly real GDP data to present the information
provided by the data.
Structural multivariate methods, on the basis of economic theories, introduce other
macroeconomic variables to identify the properties or the origins of output fluctuations.
These methods include the structural VAR, such as Blanchard-Quah (1989) approach and its
extensions (e.g. Clarida and Gali 1994) and multivariate system models such as the global
VAR approaches (Dees, Di Mauro, Pesaran and Smith 2007, DdPS hereafter). The current
applications of these approaches to China’s macroeconomic data suffer from shortage of long
time period quarterly GDP data. Among the limited published applications of Blanchard and
Quah to Chinese data, Zhang and Wan (2005) use real industrial output as a proxy for real
GDP for 1985-2000. Siklos and Zhang (2010), analyzing China’s inflation fluctuation with
the standard Blanchard and Quah framework and a tri-variate extension, have to limit their
analysis to the short sample period of 1990-2003. The original DdPS model estimation with
Chinese data use quarterly real output data that are derived from annual real GDP level by
evenly allocating the annual output to the four quarters of the year.
This paper applies the above univariate and multivariate time series analyses to the
estimated China’s quarterly real GDP and shows that the constructed data provide a better
19
high frequency and long period real GDP data alternative for the analyses of China’s output
fluctuations14.
III. Temporal disaggregation of China’s quarterly real GDP with unobserved
components model
3.1 the Model
The unobserved components models are set up in terms of components that have
direct interpretation of stylized features of the series. The models are capable of including
trend (or long run, permanent component), cycle (short run, transitory component), seasonal
components (if the data are not seasonal adjusted) and irregularities that represent the non-
systematic outlier observations or measurement errors.
Following Harvey (1989), the unobserved component models for temporal
disaggregation of China’s real GDP can be presented in multivariate state space form as
follows:
The measurement equation is:
ittititititit +εX)++γ+s+c=(τy β (1)
Or in a more general form:
ttttt +X+ZY Ε= βα (1’)
Where: Ttniyyy=y ntttit ,...,2,1 and ,...,2,1),,...,,( ,21 == ty1 is the target series (ty1 :
China’s real GDP) and ,...,,2 ntt yy are the related series for multivariate models. tZ is a
14
To focus on the evaluation of quarterly real GDP construction, in this paper I do not include the production function approach, which may involve data problems from capital and labor statistics.
20
mn× matrix, where m is the number of unobserved components, and N is the number of
dependent variables in multivariate models. tX is the matrix of assumed exogenous variables
or the “related variables” which will only be present in the univariate with related series
model15. β is the parameter vector of the explanatory variables. tα is 1×m the state vector
that contains the unobserved components that may include trend tτ , cycle tc , seasonality ts
, and irregularity tγ . tΕ is an 1×n vector of serially uncorrelated disturbances assumed with
mean zero and the covariance matrix tG , or 0)( =tE ε and tt GVar =)(ε
The transition equation is,
ttttt HT ηαα += −1 (2)
Where )' ( ttttt sc γτα = , tT is a mm× matrix, tH is mg × matrix ( mg = when all
components are defined as stochastic, g will not equal to m when some components are
defined as fixed or determinate). tη is a 1×g vector of serially uncorrelated disturbances with
mean zero and the covariance matrix, tQ , or 0)( =tE η and tt QVar =)(η
I use seasonally unadjusted data in the disaggregation models to avoid losing
information from such adjustments. The seasonal components generated through the
disaggregation will be compared with the seasonal factor generated from commonly used
seasonal adjustment procedure in the evaluation section of the paper.
15 I do not add explanatory variables to the unobserved components such as trend, cycles and/or seasonality. Adding
explanatory variables means adding assumptions on the relationship of the related series, which will reduce the generality of the model of data construction.
21
The disaggregation models are specified with choices for stochastic drift and
seasonality to capture the possible time variance of drift and seasonality16.
Disturbances of components within series are assumed uncorrelated from each other
in the disaggregate models following Harvey and Watson (1986), Harvey (1989) and Clark
(1987).
Putting observations with different timing intervals into the state space form is the
key step of using unobserved component model for temporal disaggregation. Harvey (1989,
section 6.3) introduced the technique of using a cumulator variable for mixed frequency data
in state space model. The cumulator variable for quarterly data is defined as following:
tct yy = , where 4),/,...(1,1)1( ==+−= ssnst ττ
11 ++ += ttct yyy
322 +++ ++= tttct yyyy
3213 ++++ +++= ttttct yyyyy ,
44 ++ = tct yy
545 +++ += ttct yyy
…
Where cty is the year up to date cumulated value of the quarterly level of the series,
and ty is the quarterly level of the series. For the years with only annual data available, only
cty 3+ , which equals the annual level of the series, is observable. For China’s real GDP, the
16 Fixed drift and seasonality are also tried for comparison.
22
cumulator variable cty is only observed once every 4 periods during 1978-1991. The series of
cumulator variable is considered as the target series (ty1 ) in all models. The unobserved
values of cty are then treated as missing observations to be estimated with smoothing
algorithms from the Kalman filter.
3.2 The unobserved components specification
The general UC model specification is capable of nesting the three categories of
interpolation models: Univariate models without related series, univariate models with
related series and multivariate unobserved components models. Different specifications of
the components are tried and the constructed quarterly data are evaluated.
The unobserved trend component can be specified as:
),0(~ , t1 τττ ηηµττ Qiidtttt ++= − (3)
There are two widely applied specifications of the slope of the trend: one is based on
Harvey (1985) and Watson (1986), which assumes that the trend is random walk with
constant drift, i. e., µµ =t ; the other is the Clark (1987) model, in which the trend is
assumed to follow random walk with a random walk drift, i.e. ttt ζµµ += −1 . Clark’s
assumption of a random walk with drift is capable of accounting for possible structure
breaks. To avoid losing information in the temporal disaggregation, the choice of time
varying or stochastic slope (the drift of trend) following Clark model is applied in the
disaggregation models with related series17.
17 Fixed slope of trend is also tried for comparison.
23
The estimation starts with general local linear trend (LLT) models, which do not
include cyclical component. The cyclical component is then introduced into the model and
assumed following 2nd order auto-regressive process or AR (2), as for most real output series
in the literature (Harvey 1985, Clark 1987 and Morley et. al. 2003). Harvey (1989) and
Durbin and Koopman (2001) suggest trigonometric expressions for cyclical components
which are not applied in this paper, because it may introduce arbitrary waves in the cyclical
components.
Thus when the unobserved cyclical component is included, the transition equation for
cyclical component is:
),0( ~ , ct2211 cctttt QNiidccc ηηϕϕ ++= −− (4)
The correlations between trend and cycle disturbance are assumed to be zero in the
temporal disaggregation models following Harvey and Watson (1986) and Clark (1987).
Since I will only take the Kalman smoothing estimation results of the missing observations
and not investigate the trend cycle decomposition from the temporal disaggregation, the
assumption of zero cross correlations will not affect the final result of the data construction18.
The unobserved components models are capable of handling seasonal adjustment
simultaneously with the temporal disaggregation, with the underlying assumption that the
seasonality relationships of the disaggregated series are consistent or homogenous with the
related series along the sample period. There are two options of seasonality formulations in
the Structural Time Series Analyser, Modeller and Predictor package (STAMP 8, Koopman,
18 Assumption on the correlation of permanent and transitory shocks is critical to the estimation of permanent and
transitory compositions or the unobserved components. While the data construction only take the result of the estimation of the level of the series and not doing decomposition for the series, thus will not be affected by the assumption. The zero cross correlation assumption reduces the number of coefficients to be estimated and increases the degree of freedom, thus increases the chance of convergence especially for the multivariate models.
24
et al. 2007): trigonometric stochastic seasonality (Harvey 1989, 6.2), which allows for
changes of seasonality pattern along the sample time, and the fixed dummy seasonality. I
chose stochastic seasonality in the models, because it can track the possibility of changing of
seasonality during the time period and avoid losing information for the seasonality in the
constructed data19.
Measurement errors of the data can be a big concern for China’s macro-economic
data. Therefore I include irregularity term in the models. However, I do not find any
significant irregularities in any models.
3.3 Univariate and Multivariate models
Univariate models without related series
When 1=n , the model is a univariate UC model and contains the cumulated quarterly
real GDP level with missing observations only. The modeling starts with a Local linear
Trend (LLT) univariate model without cyclical components. For LLT model, the
measurement equation is simplified as:
tttt sy γτ ++= (5)
Univariate models with related series.
The univariate model with related series as exogenous explanatory variables without
cyclical component is comparable with Chow-Lin approach and its modifications (Harvey
1986). The explanatory variables enter into the measurement equation as:
��� � ��� � ��� � ��� � �� (6)
19 Fixed or determinant seasonality is also tried for comparison.
25
If the components are set as deterministic and tε as AR (1), the model is comparable
with the original CL model (Moauro and Savio 2005). When there is no cyclical component
included, the univariate with related series model with random walk drift is comparable with
modified CL method, also known as Fernadez (1981) model. To find the best model for
temporal disaggregation of China’s real GDP, I also try the models with AR (2) cyclical
component.
Multivariate unobserved components models
Temporal disaggregation using multivariate unobserved components models, as
reviewed in section II, uses the information from related macro-economic series and at the
same time avoids liner relationship assumption on the related series or the weakly exogenous
as in the Chow-Lin model and its extensions. As discussed in the literature section of this
paper, the multivariate UC models may be more appropriate if the cointegrating relationship
is hard to find between the available related series. The problem can be more likely for
emerging countries, where high frequency macro-economic data are very limited.
Another advantage of this framework is that it allows for simultaneous
disaggregation, seasonal adjustment and trend cycle decomposition. While not the focus of
data construction section of this paper, the common trends, cycles and seasonality among the
related series can be easily imposed and tested in the multivariate UC framework.
3.4 Logarithmic transform of the data
All series values are in natural logarithms to ensure positive estimation for real GDP
in the disaggregation models. First, due to the relatively small sample, a few large seasonal
troughs may cause the estimated quarterly real GDP to become negative when using levels of
26
the series20 . Logarithmic transformation of the data guarantees positive values of the
estimated quarterly data. Secondly, as Proietti 2006 shows, since logarithmic transformation
reduces the heteroscedasticity of the series and the underlying assumptions of the
multivariate disaggregation models include homoscedasticity on seasonality and variances
among series, it is more appropriate to use for those models.
The cumulated series of the logarithmic transformed target series can be expressed as:
)log( tct yly = , where 4),/,...(1,1)1( ==+−= ssnst ττ
)log( 11 ++ += ttct yyly
)log( 322 +++ ++= tttct yyyly
)log( 3213 ++++ +++= ttttct yyyyly ,
)log( 44 ++ = tct yly
)log( 545 +++ += ttct yyly
…
==+−=
=+= − otherwise ,1
4),/,....(1,1)1(,0),log( 1
ssnstyyy tt
ctt
ct
ττφφ (7)
Where ctly is the logged cumulated level of the real GDP, and c
ty is the cumulated
level of the real GDP.
20 Estimation of the models using levels of the series does generate a few negative results at certain seasonal troughs.
27
One concern about the logarithmic transformation on China’s data is the relatively
high growth of the series. As Banerjee et al. (1993) discussed about logarithmic
transformation of time series, changes in the logarithm approximately equal to the percentage
change of original levels of the series. When the changes are relatively large, logarithmic
transformation may dampen the growth patterns. However, comparing the results from
disaggregation with the real official data (the overlapping period 1991q1-2008q4) does not
show any significant effect on the magnitude of fluctuations.
IV. China macroeconomic data
My study aims to estimate China’s quarterly real GDP data in consistency with the
official quarterly published data since 1992. I focus on the real output fluctuations since
1978, when China embarked on the market-oriented and openness economic reform. The
annual and quarterly data used in this paper are from the National Bureau of Statistics of
China (NBS), the nation’s statistical authority21 , and official monetary authority and
international trade statistic agency.
4.1 The official GDP data
China’s National Accounts followed the Material Product System (MPS) of the
former Soviet Union from 1949 until 1985. China’s GDP estimation transitioned to follow
the guidance of United Nations System of National Account (SNA) in 1985 and formally
completed the process in 1992. The quarterly GDP data are officially published since 1992.
Empirical studies of the Chinese economy have been plagued by the problems of
availability and reliability of official Chinese macroeconomic data. Despite the data
21 The official data are published as cumulated year on year growth rate at comparable price. Data from 1992-2005 are
from the publication of National Bureau of Statistics of China (NBS 2008).
28
challenges, it is still worthwhile to study the features of China’s economy, the world’s 2nd
largest economy and the most populous country.
China’s official GDP have been criticized for having been overstated during 1980s
and 1990s (Rawski 2001, World Bank 2005, etc), understated in the middle of 2000s
(Economist May 1st 2008), and then overstated again during the most recent financial crisis.
Most recently, Huang (2011) argues that the true China’s GDP is likely to be much higher
than reported due to the understatement of consumption estimation. As a transitional
economy, China has undergone continuous changes, and has complex political, social and
economic structure. Despite the efforts made by NBS to improve and explain the GDP
estimates over the years, confidence in the accuracy of official data quality remains the
primary problem that must be addressed for empirical research on Chinese macroeconomic
issues22.
After carefully reviewing the literature on Chinese data quality and the national statistical
accounting system (Appendix 1-1), and comparing different data resources and data
construction methods23 , I agree with many researchers (Holz 2006, Chow 2006, Klein and
Ozmucur 2003) and most international organizations (OECD, IMF24). Although there are
weaknesses or short-comings in the statistical system that derives Chinese national accounts
22 A partial list of the recent media news and reports on China’s data problem includes (from the latest): “Reflating the dragon, can the world’s fastest-growing economy avoid a sharp downturn?” Economist Nov. 15th 2008,
which claims that China’s official growth fluctuations are smoothed” “An aberrant abacus—coming to terms with China’s untrustworthy economic numbers” Economist, May 1st 2008,
which ranks the reliability of Chinese statistics. 23 Besides the estimation with the data presented here, I apply the same methods to data covering shorter periods and
from other informal resources (the IMF World Economic Outlook dataset and Fudan University dataset). I compare the results and check if the data from different resources and the subsample data have significant different features.
24 The World Bank criticized the Chinese national account statistics and revised their GDP estimation for China upward for 34% from the officially reported number in 1993. In 1996, the World Bank accepted China’s reformed statistic system and the official GDP number again. But the World Bank revision and method of estimation was also questioned by many researchers.
29
estimation, Chinese official macroeconomic data after 1978 do not appear to be politically
manipulated or systematically biased. The official data can serve as “a reliable guide”
(OECD 2006) to the level and growth pattern of GDP, even though the margins of error are
“certainly larger than that of the most developed countries” (OECD 2006). Any other
alternative data series constructed or corrected by researchers has not been proven to be more
precise or reliable (Holz 2006). When I run the dissaggregation models in STAMP, I cannot
find the existence of any significant measurement errors or irregularity. Thus “Official
Chinese data should be the first port of call” (Scheibe 2003) for my study. The data resources
of my study, CEIC and IFS, both use the official data from China NBS as their final data
source.
The official quarterly real GDP year-on-year growths are shown in Figure 125. The
year-on-year real growth rates suggest that instead of “recessions” or negative growth rates at
the troughs of the cycles, China’s economy experienced “slowdowns” or “growth recessions”
at times but always had positive growth rates during the sample period. There are five major
slowdowns in year-on-year growth rates, which happened in 1980, 1984, 1989, 1996-1998
and most recently in 2008 (Chinese Academy Of Social Science, 2008-2010, Liu 2009). The
slowdowns in output growth in 1984 and 1996 were accompanied by hyperinflations. The
“Tiananmen Square” political chaos in 1989 significantly halted lots of the economic
activities across the country. The Asian financial crisis occurred in 1998 had an adverse
effect on the economic growth. In 2008, China’s economic growth dropped to 6.5% in the
last quarter, adversely shocked by the global financial crisis.
4.2 Related series
25
The data construction of this paper is based on log level data. The discussion in this section documents the information provided by the raw official real GDP data, which are published as year on year growth rates only.
30
The only available high frequency quarterly macro-economic data for the sample
period is monetary statistics and international trade statistics. To estimate the missing
quarterly real GDP data for China, I consider different combinations of monetary and trade
variables as related series in the multivariate UC modeling. The monetary series, including
domestic credit, international reserves, M1 and M2, are available quarterly since 1978;
International trade series include total exports and imports, total trade volume, which are
available quarterly since 1981. Domestic credit, international reserves, M1 and M2 are
nominal outstanding balance. Each series may carry different information associated with the
economic development and outputs fluctuations.
The monetary and trade data used in the temporal disaggregation of GDP are not only
available quarterly for the sample period, but also of good quality. According to the
Economist (2008) the quality of the related series’ data is among the top two most reliable
official macroeconomic data of China.
Figures 2.a and 2.b shows the log quarterly and annual real GDP and the potential
related series data used in the models. All data are not seasonally adjusted. The series appear
to follow similar upward trend in the long run. Figure 2.c and Figure 2.d present the year on
year growth rates26 of the available quarterly real GDP with the monetary and trade related
series respectively. Table1-5 documents the correlations of the year on year growth rates of
real GDP and the potential related series for the whole sample, data construction sample and
the fully available sample period.
26 The real GDP data construction of this paper is based on log level data rather than growth rates. However, the
relationships of series based on properties of growth rates, although they may be different from that based on the level data, provide useful information.
31
The correlations of the fluctuations of real GDP with most of the potential related
series appear to be high and stable except exports. The close to zero correlation of real GDP
and exports during 1978-1991 shows that the openness of China economy was very limited
during the period. With China’s expediting integration into the global economy in later years,
the correlation of economic growth with exports increased substantially. All joint tests of
cointegrating relationships between combinations of related series that include exports show
no evidence of cointegration (Table2).
V. Temporal disaggregation model selection and estimation results
This section presents the procedure of model selections and temporal disaggregating
estimation using the unobserved components model specified in section III.
5.1 Unit root and cointegration tests
The procedure starts with unit root test and cointegration test for the real GDP and
related series. The tests are important in finding whether Chow-Lin related series models are
valid or not. As I have discussed above, the estimation of univariate models with related
series (Chow-Lin method) will only be valid if there is a linear relationship between the
included related series and the target series. With nonstationary series, this assumption is
only valid when there is a cointegrating relationship among the series.
The stationarity of the annual logarithms level of the real GDP and related series is
tested using the Augmented Dickey-Fuller (ADF)27 . Table 1 reports the results of the ADF
tests. All series appear to have a unit root in the level and are stationary in first differences.
27
Other unit root tests methods lead to the same conclusion. The results are available upon request.
32
Thus, any temporal disaggregation methods, such as the original Chow-Lin model,
that not consider the nonstationarity of the series are invalid for Chinese real GDP level
disaggregation. The data must be first differenced before applying those methods. Or data
other than the real GDP level but stationary, such as the real growth rates used by A&R,
should be used. However, important information, especially on the level and trend, may be
lost during the first difference or using growth rates data. Plus the growth rates series may
have difference properties than the level data.
As I discussed in section III, the unobserved component approach is not only capable
of nesting the disaggregation of stationary series, but also capable of modeling the
nonstationarity with different specifications of the permanent component to capture the
property of the series.
I then use the Johansen cointegration test to check for a cointegrating relationship
among the different combination of the annual real GDP and related series. Table 2 presents
the results of the tests. The tests provide evidence in favor of cointegration among the annual
real GDP, total trade, imports and monetary indicators (domestic credit, M1 and M2). There
is no evidence of a cointegrating relationship when including exports in the system. Including
international reserves may introduce more than one cointegrating relation among the series.
The existence of a cointegrating relationship among the related series ensures that the Chow-
Lin method is applicable to the disaggregation of China real GDP with selected related
series. Thus the univariate with related series models should be included in the model
selection of temporal disaggregation.
5.2 Temporal disaggregation model selection and results
33
The unobserved components models for temporal disaggregation are estimated using
the STAMP8 program. The program applies the Kalman filter to obtain the components of
the series and uses maximum likelihood methods to estimate the parameters. Missing
quarterly data are generated with smoothing algorithm of the Kalman filter.
To select the model for disaggregation, I estimate the models with different
specifications of components and different combinations of related series. All models are
estimated with the full sample period from 1978q1-2009q4, but missing quarterly real GDP
from 1978q1-2008q4. Official quarterly real GDP for 2009 q1-q4 are used to initiate the
estimation, which is required by the STAMP program28. Once the estimations for 1978 q1 to
2008q4 are obtained, I calculate the root mean squared standard errors (RMSE) of the
estimated data and the official published quarterly data over overlapping period 1991q1-
2008q4. The best fit model for data disaggregation is then determined by the minimum
RMSE29. Based on the selected model, I estimated China’s real GDP series over the period
1978q1-1991q4. The RMSE criterion suggests the multivariate UC model including domestic
credit and total trade as related series with stochastic trend and AR (2) cyclical component
(Table 3). To further check the stability of the model, I replicate the model selection
procedure using subsample period from 1992-2009, when the quarterly real GDP are fully
observed. MUC model with domestic credit and total trade as related series is still the best fit
MUC model among all MUC combinations I have tried.
28
The initial value can be changed but the one year official quarterly data help in finding the convergence and reduce the length of iteration procedures.
29 Other statistic criteria are also shown in the table. Since the purpose of the modeling is not finding the best explanatory of GDP, the best fit of data disaggregation, or the RMSE is used to determine the selection of disaggregation model.
34
Table 6 presents the parameter coefficients and variances/correlations of components
estimates of the selected MUC domestic credit and total trade model, with full sample from
1978-2009 all quarterly real GDP observations missing, subsample from 1992-2009 with all
quarterly real GDP observations missing and the real temporal disaggregation model on full
sample with 1978-1991 quarterly real GDP to be estimated. The estimates of slope and AR
coefficients are very stable cross sample periods.
Figure 3 compares the year on year growth rate of official data, comparable Chow-
Lin methods and the MUC domestic credit and total trade model.
Using the selected model the final data construction estimation includes all available
official quarterly real GDP observations, leaving only 1978q1-1990q4 missing.
Figure 4 shows the year on year quarterly growth rates of the final results of
disaggregated real GDP, compared with series constructed by A&R. The MUC estimates has
similar but a little larger fluctuations than A&R estimation, except that the growth
accelerating started from 1981 peaked in 1984q4 in my estimation, while the A&R
estimation peaked 1 year later in 1985q4. Based on the annual official real growth data and
the official analysis from China NBS (Xu 2010), the MUC estimation is more reliable.
Bounded by the available annual level directly in the model estimation, the MUC estimates
follow the observed annual level better than A&R estimates. The multivariate structural
model analysis in the next section shows that this different affects the identification of the
property of shocks to China’s economy during this period. Both MUC estimation and A&R
estimation show a big drop in the last quarter of 1989. MUC estimate drops below zero.
VI. Univariate and multivariate time series analysis of China’s real GDP
fluctuations using the disaggregated quarterly data
35
To further evaluate the quality and information provided by the MUC temporal
disaggregated China quarterly real GDP data, I apply different univariate and multivariate
time series analysis techniques to the quarterly real GDP data from 1978q1-2010q4, with
data from 1978q1-1991q4 disaggregated from annual data by MUC model. The univariate
time series analytic methods include Linear in time functions, the Hodrick-Prescott (HP)
filter, the Band-pass or BP filter proposed by Baxter-King (1999) and Christiano and
Fitzgerald (2003), the unobserved components (UC) techniques. The multivariate structural
time series approaches are Blanchard-Quah decomposition (Blanchard and Quah 1989) and
the global vector autoregressive approach building on the work of Dees et al. (2007) (DdPS
approach hereafter). The MUC temporal disaggregated quarterly data provides a better
alternative of high frequency real output data that covers the whole period after China’s
economic reform and openness, and adds valuable information to the empirical investigation
on the properties of China’s output fluctuations.
6.1 Seasonal adjustment
Before conducting analysis, China’s quarterly real output data are seasonally adjusted
using the X-12 ARIMA method.
As discussed above, temporal disaggregation with the unobserved component
approach is capable of conduct seasonal adjustment simultaneously with the temporal
disaggregation. Figure 5 presents the seasonal components of the logged real GDP generated
from the MUC temporal disaggregation estimation and the seasonal factor based on X-12
ARIMA method. The two seasonal series are exactly the same, except slightly different at the
beginning of the sample period (5 observations). The seasonal adjustment through MUC
model appears to be convenient and reliable.
36
The X-12 ARIMA (2, 1, 2) and Tramo/seat (Time series Regression with ARIMA
noise, Missing Values and Outliers/Signal Extraction in ARIMA Time series) methods also
give similar results. The finding is consistent with Blades (2007), who performed similar
tests on current price quarterly GDP of China. The seasonal pattern of China’s quarterly real
GDP is regular and predictable.
The seasonally adjusted real GDP levels are then transformed to natural logarithms.
For calculation and explanation convenience, the natured log seasonally adjusted real GDP
level is annualized (multiplied by 4) and multiplied by 100.
6.2 Univariate statistical filters
Economic theories on economic fluctuations and growth, including real business
cycle theory, Keynesian theory and monetarism, all suggest that economies react differently
to permanent shocks with long-run effects than to transitory shocks whose effects dissipate in
the short run. Permanent or trend component of the real GDP is also considered as potential
output of a economy, while the transitory or cyclical component is used as measures of the
output gap. Understanding the relative role of permanent versus transitory movements in the
macroeconomic fluctuations of the countries is important for economists, forecasters, and
policy makers.
In contrast to the “classical business cycles” first defined by the Burns and Mitchell
(1946) as recurrent expansion, downturn, contraction and upturn in economic activity, the
“cycles” studying here follow the definition of “growth cycle”, which are “recurrent
fluctuations in the series of deviations from trend” (OECD Glossary). The later definition of
cycle is more appropriate for China’s real GDP fluctuations because the economy
experienced slowdowns in growth rates but the growth rates have always remained positive.
37
The contractions by definition should include slowdowns using a “growth cycle” definition
instead of only include absolute declines or recessions in economic activity. Morley and
Piger (2009) denote a more general “transitory-deviation definition” of the business cycle,
which are the short-run or transitory fluctuations in economic activity around the permanent
or “trend” level. The unobserved components (UC) techniques applied in this paper fit in the
“transitory-deviation” definition, while the cycles isolated from the series with a statistic
filter such as Hodrick-Prescott (HP) or Band-Pass (BP) filter fit in the traditional business
cycle or growth cycle definition.
Linear in time (LIT) in time
As benchmark for comparison, I begin the decompositions from the most naive linear
in time and polynomial in time models. The models assume a deterministic linear (or
polynomial) in time trend30. With the LIT estimation, I check for structural breaks by
applying Quandt-Andrews unknown breakpoint test to the constant and time coefficient with
trimmed 15% data and find that 1992q4 is a significant breakpoint during the sample
period31. The breakpoint is confirmed by Chow known breakpoint test. The Chow known
breakpoint test on linear in time model with official annual real GDP level confirms 1993 as
a significant (at 10% significant level) breakpoint during 1978-2009. There are evidences of
30 The linear and polynomial in time model specification is:
When n=1, the model is linear in time model. When n=2, it is quadratic in time model. The models are estimated with least squares method. The trend is the predicted value of yt and the cyclical component
is the residual from the estimation. The residuals are significantly auto-correlated. Although it is well known that LIT or polynomial in time models often fail to reveal the changes in slope or intercept of the trend over time, the estimations are simple and straightforward, thus still can be used as benchmarks. Since larger power polynomial in time model fit the data no better than the linear trend model, I use LIT result for comparison with results from other methods.
31 The test is conduct in Eviews. Same exercise using Pcgive package shows similar result, the recursive estimation breakpoint Chow test with the MUC quarterly data show evidence of breakpoint around 1992q3-1993q1.
38
breakpoint around the end of 1992 in both official annual data and the MUC quarterly data.
This breakpoint is mostly based on true underlying economic activity rather than the data
construction procedure.
In China’s economic development history, 1992q4 was the start of an era of stable
and high growth following the former leader Mr. Deng Xiaoping’s speech on his “Tour the
South of China” in early 1992. The speech re-affirmed the national policy of market oriented
economic reform and openness that was halted by the 1989 “Tiananmen” square political
chaos. This breakpoint thus is considered in the other trend cycle decomposition approaches
and the possible different properties of the economic fluctuations before and after 1992q4 is
investigated.
Hodrick-Prescott (HP) filter
HP filter is the most widely used approach of decomposition to obtain the smoothed
long term trend of China’s output (or potential output) and output gaps in the literatures so
far. The HP trend is determined by minimizing the weighted sum of the squared cycle and
changes in the growth rate of the trend32.
Although very popular and convenient, HP filter may generate artificial cycles when
applied to first-order integrated series. As shown by Cogley and Nason (1995) and Park
(1996), the HP filter is convenient but subject to several limitations: HP filter implicitly
assumes that the business cycle of the economy is symmetric during expansion and recession
32 I generate HP trend and cycle in E-views, where the HP filter chooses s to minimize:
λ is the smoothness control parameter that penalizes the fluctuations of trend. The HP filter get smoothed and stochastic
trend with is uncorrelated with the cyclical components. For quarterly data, the standard selection is λ=1600.
39
time, and the assumption of periodicity is sensitive to the end of the sample period. In
addition, the HP filter may generate artificial cycles when applied to first-order integrated
series. King and Rebelo (1993) pointed out the arbitrarily picked smooth parameter is based
on the observations of the US business cycles and may not be the optimal choice for other
economies. The HP filter also is criticized for lacking fundamental economic justification
and arbitrarily picking smooth parameter33. The cycle components are significantly sensitive
to the arbitrarily set smoothness control parameter λ. The standard choice of λ=1600, which
is based on the observations of US business cycles, leaves the duration of the cycles average
at 4-6 years. For China real GDP data, the average durations are above 8 years with λ=1600.
I tried different values for λs (8, 40) to check the sensibility of cycles (Figure 6). The
magnitudes of cycle appears much bigger with λ=1600, while similar when λ is set at 8 or 40.
However smaller λs make the cycle cross more frequently from the zero line. Since the λ
choices are arbitrary and there is no generally accepted criteria for choosing λ for China’s
quarterly real GDP, I use the standard λ=1600 result to compare with decompositions from
other approaches.
Band-Pass (BP) Baxer-King and Christiano-Fizgerald filter
Band-Pass (BP) filter, also called frequency filter (Sims 1974), isolates the cyclical
component of a time series by specifying a frequency band or a range of duration for cycles.
The filter takes a two-sided weighted moving average (Baxer and King 1999) of data where
cycles pass through the band that is arbitrarily selected. The BP procedure does not make
deterministic or stochastic assumption about the trend. The frequency of the cycle is the only
criteria used to identify trend and cycle. The selection of band is critical to the decomposition
33. See King and Rebelo(1993) for optimal conditions of HP filter.
40
results. The typical choice for quarterly data is usually set at between 6 to 32 quarters (fixed
length symmetric, Baxer and King 1999), which isolates all cycles that completed greater
than 6 quarters and less than 32 quarters into the cyclical component. Christiano and
Fitzgerald (2003) proposed a filter that considers the nonstationary and asymmetry of the
underlying data, thus is more proper for time series that have unit roots such as China’s real
GDP.
Unobserved component decomposition
As discussed above, the frequency filters impose assumptions that may generate
artificial cycles, thus may overstate the importance of the cyclical component. The statistical
filters provide very little information on the evolution of permanent or trend of the series,
which is important for a fast growing and transitional economy such as China. The fast
changes of China economic structure and gradually but continuously implemented economic
reforms may have impacts on the economy permanently or transitorily. To understand
China’s output fluctuations beyond the spurious statistical filters, I apply the structural time
series modeling or unobserved components modeling to further investigate the property of
permanent and transitory changes of China’s output.
Recent development of univariate time series econometric approaches favors “let the
data speak for itself”. The Unobserved Components (UC) models that I used for the temporal
disaggregation are also widely applied decomposition methods (Harvey 1985, Watson 1986
and Clark 1987). These methods explicitly account for the unit roots property of the series
without imposing any prior assumptions. By assuming a stochastic trend, the method capture
the property of the trend for aggregate output, which for most economies, grows or changes
over time and thus is not stationary in levels.
41
The unobserved component (UC) model, as discussed in the temporal disaggregation
part of the paper, assumes a stochastic trend and stationary cycle. Although theoretically the
temporal disaggregation and trend cycle decomposition can be done simultaneously with UC
models, only Kalman smoothing algorithm, which uses the information from the whole
sample, can be applied with missing quarterly observation of the real GDP series. Here I use
both Kalman smoothing and filtering algorithms on the full sample. The Kalman smoothing
calculations of China’s real GDP permanent and transitory components include information
from future data. The model specification is documented in appendix1- 2.
Figure 7 shows that the smoothing cycle is slightly larger in amplitude than the
filtering cycle. The turning points are also slightly different when using the future
information other than the historical information till the estimated point. The smoothing
estimates, including more information from the whole sample, usually fit the data better. The
filtering estimates is still very useful in forecasting when only historical information is
available.
The drift and parameters of AR terms can be estimated along with the decomposition
using maximum likelihood method. The parameter estimates of the models are reported in
Table 5. The estimated drift term, which can be interpreted as the average growth of the trend
or permanent component of China real output, is 2.46% quarterly or about 9.8% annually.
The estimated autoregressive coefficients, which represent the dynamics of the cyclical
components, are summed at 0.977, which implies that the fluctuations of the transitory
components are highly persistent.
42
Figure 8 shows the most commonly used HP and BP filtered cycles of China’s real
GDP with MUC and A&R disaggregated data. There are slight differences in the turning
pattern and magnitudes of cycles.
6.3 Comparison of output fluctuation results from different univariate time series
analytic methods.
HP and BP decompositions are similar in that both isolate low-frequency fluctuations
to the trend and keep certain high frequency fluctuations in the cycle. The methods impose
smoothness prior assumptions on the components and thus subject to restrictions of the
condition of properly using the priors. Both methods may distort the characteristics of trend
and cyclical components of integrated or I (1) series (Baxer-King BP filter). The LIT, HP, BP
Baxer-King and Christiano-Fitzgerald asymmetric BP filter cycles of China’s quarterly real
GDP generated from E-views are shown in Figure 9. The HP and BP Baxer-King cycles
appear to have similar cyclical patterns in peaks and troughs, while BP Baxer-King cycle is
smoother than HP cycle. Similar results have been found for output data of other countries.
(See Gerlach and Yiu 2003 for Asian economies, Park 1996 for the US). The identified
turning points of Christiano-Fitzgerald cycle are different from others for sample period after
1992q3. Christiano-Fitzgerald cycles turn earlier than other cycles. Christinao-Fitzgeral filter
considers the I(1) property and asymmetric length of cycles during the sample period. Based
on the shaded areas, which are peaks to troughs of real growth rates, the Christiano-
Fitzgerald cycle appears more reasonable. I use Christiano-Fitzgerald filter result for
comparison.
Figure 10 compares the estimated permanent and transitory components from HP, BP
Christiano and Fitzgerald filters and the UC model.
43
Different decomposition methods generate similar cyclical components before 1992
but different ones after 1992. The cycle periods appear shorter before 1992 (averaged at
about 4-6 years), while longer after (about 10 years for the cycle before the most recent
global financial crisis)34. Similar to the US economy, China experienced “moderation” in
economic fluctuation during the period. Explanation of this difference would be, during the
earlier period, the economy had to adapt to some fundamental economic reforms and transit
from full planned economy to a market oriented economic structure, and the lack of
adjustment mechanism result in stronger reactions to any shocks.
All approaches identify big transitory drops in 1982, when stimulating effects from
the first round of economic reforms faded, and a less negative gap in 1987, following the first
peak of inflation since the reform. The late 1989 drop was a combination of hyper-inflation
and political chaos. The most recent financial crisis period is first identified as spike in the
transitory components based on all methods, which shows the effects of stimulation package.
6.4 Structural multivariate analysis
The estimated China’s quarterly real GDP with selected multivariate UC model also
provide a better high frequency and long period real GDP data alternative for structural
multivariate analyses on China’s macroeconomic fluctuations. The following excises on
Blanchard-Quah approach and the global VAR method (Dees, Di Mauro, Pesaran and Smith
2007, DdPS hereafter) are two examples of the structural multivariate analyses using the
estimated China’s quarterly real GDP data35.
The structural VAR approach: Blanchard-Quah decomposition
34 Note that length of HP and BP cycles may just due to the choice of smoothness parameters. 35 Due to data limitation on quarterly capital investment and employment as I mentioned in earlier sections, production
function analysis cannot be conducted for the period before 2000 without data construction on those series.
44
Due to the shortage of long time period quarterly GDP data, the applications of
Blanchard and Quah method to Chinese quarterly data either use a very short sample period
of data (Siklos and Zhang 2010) or an alternative series as proxy for real output (Zhang and
Wan 2005). The MUC estimation of China’s real GDP fills this gap. Using the MUC
estimation of China’s quarterly real GDP data and inflation rate from IMF-IFS database, I
derive the demand component of China’s real output fluctuations using the standard
Blanchard and Quah bivariate structural VAR approach for 1986-201036. Compared with the
literature that uses other alternatives for aggregate output of China, the new real GDP data
provides more information for China’s macroeconomic fluctuation for a longer time period
and better coverage of the economy.
Blanchard and Quah (1989) identify structural supply shocks and demand shocks
with a structural vector autoregression method by assuming that the supply shocks, which are
usually driven by changes in productivity, affect the real output permanently whereas the
demand shocks only have temporary impacts on output. The original Blanchard and Quah
model uses the unemployment rate as the additional macroeconomic variable. Due to the
shortage of reliable employment data, application of the approach to China and other
emerging countries often use inflation as demand side related series (Bayoumi and
Eichengreen 1992, Bersch and Sinclair 2011). Following these literature, I use inflation rate
as additional macroeconomic series to identify supply and demand shock.
The unit root tests result indicates that the logged seasonal adjusted real GDP and
logged inflation rate are integrated of order 1. The Blanchard-Quah decomposition can be
applied to the bivariate VAR of first difference of logged seasonally adjusted real GDP (∆y)
36 The quarterly CPI data are available only since 1986.
45
and inflation rate (∆π). The data are presented in Figure11.a and 11.b. Appendix1-3 presents
the model specification. The lag length selection criteria suggest including six lags in the
VAR. The structural VAR is then set by imposing the Blanchard and Quah long-run
restriction, assuming the aggregate demand shocks do not have long-run effects on real
output.
Figure 11 shows the impulse response functions that trace out the impact on the levels
of real GDP and inflation by the identified supply and demand shocks. It shows that one unit
of positive demand shock increases the output by about 0.8 percent and the effect diminishes
in about 4 years. One unit of positive supply shock pushes the real output up by about 1.6
percent in the long run. A demand shock results in a sharp increase in inflation and a supply
shock leads to a slight drop in inflation first and then the effect quickly reverses upward. The
economy would face strong same direction price level changes from both demand and supply
shocks. It will be difficult for the central bank to control the inflation solely through
monetary policy which mainly affects the demand side. The impulse response function
results are in agreement with the results of previous studies (Zhang and Wan 2005).
The estimated forecast error variance decomposition of real output based on the MUC
data provides different information about China’s real output fluctuations than the
decomposition based on industrial production, an alternative proxy of real output used by
Zhang and Wan (2005). Based on MUC data, the output fluctuations are largely explained by
aggregate supply shock, while aggregate demand shocks are the main driver of inflation
changes. The share of supply shock to the variance of forecast error on real GDP is much
more stable (above 60% of the variance of output fluctuations) than the result of Zhang and
Wan (2005) indicated (increasing from 55% to 92% over 2 years forecasting horizon). Use
46
industrial production as proxy for the aggregate output can be problematic for China. The
share of industrial production in China’s aggregate output changes dramatically during the
sample period. As discussed in the data section, share of production from service sector, once
trivial during early 1980s, has greatly increased to over 40% of the total GDP in recent years.
Industrial production, which does not cover the service sector, could not reflect the overall
properties of the macroeconomic fluctuations.
The Blanchard-Quah output gap, defined as the accumulation of demand shocks on
output, can be derived using the estimated structural demand shocks and the noncumulative
impulse response function. By definition, the cyclical component of the real output should be
mean zero in the long run. Thus the gap is set to be closed at the mean of the Blanchard-Quah
demand components. Figure13 compares the Blanchard-Quah output with HP and Christiano
Fitzgerald cycles. The differences shall reflect the fluctuations caused by supply shocks
within the frequency band of the statistical filters.
Figure 14 compares the Blanchard-Quah output gaps based on MUC estimation and
A&R estimation. The Blanchard-Quah output gap of A&R data appears very similar to that
of MUC estimated data.
Global VAR
DdPS (2007) developed a multivariate system—a global VAR (GVAR) model—to
explore the international linkages among economies in an increasingly globalized world.
DdPS’s GVAR model for China includes the following country-specific variables as
dependent variables: real outputs, inflation, interest rates, and real exchange rates. Trade
weighted output, price level, equity price, interest rates and long run interest rates of the rest
of the world and the world oil price enter into the GVAR as exogenous variables. The model
47
includes two lags for all variables. The authors assume that all variables are integrated of
order one. To compare the estimation results with DdPS 2007, I replicate the model using
quarterly data of same sample period from 1979q2 to 2003q4.
A problem of the original DdPS model estimation with Chinese data is that the
quarterly real output data are derived from annual real GDP by evenly allocating the annual
output to the four quarters of the year. This simple disaggregation smoothes the quarter by
quarter changes of the series, thus loses the information on the quarterly macroeconomic
dynamics. I replicate the China DdPS GVAR model with the quarterly real GDP data
estimated by MUC model. The result of the replication shows that the new data provide
important information to the model. Simple disaggregaion of the data distorts the results37.
Figure 15 shows the level and the first difference of logged seasonal adjusted real
GDP estimated through the MUC model and the original DdPS data of logged real GDP38.
The MUC quarterly real GDP data introduce quarterly dynamics to the GVAR
system. The two real GDP series move within the boundary of the same annual real GDP
movement. Thus, they provide similar information on signs of the long run relationship of the
variables in the cointegrating vector. Table 8 shows the evidence of cointegration of the two
series (Erisson, Herdry and Tran 1994)39. However, the estimated effects of domestic
37 Note that the quarterly CPI data are available only since 1986. DpDS appear to construct the quarterly inflation data
based on annual inflation statistics. Same situation may exist for the interest rate data as well. In this analysis I focus on the discussion of real GDP data quality. To check the difference of the estimation based on the different real GDP data only, I keep the other series the same as the original DpDS dataset. However, there may be data construction problems to other series in DpDS estimation as well.
38 The level differences are due to the difference of base year setting. It will not affect the result of VAR estimation, which is based on the first difference of the series.
39 Ericsson, Hendry and Tran (1994) theoretically explain why seasonal adjusted and non seasonal adjusted data are cointegrated. The MUC real GDP and DdPS real GDP can be considered as seasonally adjusted through different procedures, thus should be cointegrated.
48
inflation and interest rates on real output based on the MUC data are much stronger than the
original DdPS data reveals40.
Table 9 presents the estimated cointegrating vector coefficients based on MUC
temporal disaggregated quarterly real GDP data and on the original DdPS data. The
estimation based on MUC data and original DdPS data both find that only domestic inflation
and short term interest rate are statistically significant. The high significance of domestic
inflation suggests the existence of a strong Phillips curve type relationship in China’s
macroeconomic fluctuations.
All foreign variables are insignificant for China’s macroeconomic fluctuations in the
long run based on both datasets. Although rapidly integrated into the world economy,
China’s economic fluctuations are still mainly driven by domestic factors rather than the
foreign and global factors during the sample period41.
Appendix 1-4 presents the results of short run coefficients and the graphs of impulse
response functions for each domestic variable in the GVAR system. The real GDP and
inflation appear to be exogenous based on the MUC data estimation. While based on the
original DdPS data, only inflation is exogenous. The plots of impulse response functions for
China’s real GDP based on the MUC data show that the economy recovers much quicker
than the original DdPS data estimated from shocks on other domestic variables. Seasonal
40 Table A4-1 in appendix 4 shows the result of likelihood ratio tests of equality of the coefficients estimated with MUC
data with the estimated coefficient original DdPS data. MUC coefficients are larger in absolute value but not significantly different from the DdPS estimated coefficients, and vise versa, which can be explained again based on Ericsson, Hendry and Tran (1994), that the two datasets can be considered as results of different seasonal adjustments and should show similar long run property of the underlying series.
41 This result is in agreement with the result of the other two papers I coauthored with Tara Sinclair using multivariate unobserved components model to investigate the relationship of China’s real output fluctuations with the US and the developed world economies using quarterly data from 1978-2009
49
dynamics introduced by the MUC data may cause these difference in short run analysis
(Ericsson et al. 1994)42.
VII. Conclusion
This paper provides quarterly real GDP estimates from 1978q1-1991q4 using
multivariate unobserved components models. The selected disaggregation model estimates
the quarterly real GDP levels of China from annual data with Kalman smoothing technique,
using information from the available quarterly domestic credit and total international trade
data without prior assumption of cointegration among the series. Although the traditional
Chow-Lin method of temporal disaggregation is valid for China data because of the evidence
of cointegration among the related series, the MUC model is found to be more efficient.
To evaluate the MUC model estimated China’s real quarterly GDP data, I apply the
temporal disaggregated quarterly real GDP series, lengthen by the temporal disaggregation
from 1978-2010, to different univariate and multivariate methods. The constructed quarterly
data are shown to be a better alternative than other proxies and estimations. The data provide
valuable information to the empirical study on China’s macroeconomic fluctuations.
The multivariate unobserved component temporal disaggregation approach could be
easily applied to the missing data problem of other macroeconomic indicators, and to the data
of other developing and transitional economies, where lack of high frequency data has been a
big obstacle of macroeconomic analysis.
Through evaluate the MUC model estimated China’s real quarterly GDP that covers
the 32 years since China started the economic reform and openness, the properties of China’s
42 Ericsson, Hendy and Tran (1994) provide analysis on the possibility of the difference in short run or error correction
modeling due to the difference in seasonal adjustment.
50
output fluctuations can be better understood. The results of unobserved components
decomposition, Blanchard-Quah decomposition, and the GVAR model suggest that supply
side shocks and domestic factors play an important role in China’s real output movements.
Although China’s economy has been widely open to the world economy, outside shocks,
which may mainly be on the demand side, may have either not been as strong as that from
the domestic economic reforms and productivity changes, or have been effectively offset by
China’s macro-economic policies.
Where is China’s economy today? Is it below or above potential output or trend? The
different trend cycle decompositions give different answers to this question: HP and
Blanchard-Quah decomposition find that China’s aggregate output since 2010 is slightly
below trend, while the Christiano Fitzgerald filter and UC model shows it still slightly above
permanent level. All methods of analyses show that China’s economy is now very close to
the potential level. The answer to whether the growth will speed up or slow down looks
ambiguous43. Given the importance of China in the global economy, this suggests further
research on China’s economy is clearly warranted.
43 Further research is desirable to better understand the features of China’s macroeconomic fluctuations. In Chapter two
and three I extend the study with multivariable unobserved components model on the relationship of output fluctuations with other macro-economic series. Other possible extension includes applying the multivariate approaches with inflation (the Philips curve), monetary policy indices and/or consumption.
51
Figures and Tables
Table 1-1: Unit root test results (Augmented Dickey-Full Test on annual data 1978-
2009)
Series test statistics a Lag-length
b Deterministic
c
Log GDP 3.608 2 Constant
log Export 7.391 0 none
Log Import 0.439 2 constant
Log Total trade 7.332 0 none
Log M1 4.480 1 none
Log M2 2.827 1 none
Log Domestic credit 3.489 1 none
Log international reserves 0.917 4 none
dLog GDP -4.002 *** 5 Constant
dlog Export -3.470 ** 5 constant
dLog Import -4.918 *** 1 constant
dLog Total trade -3.502 ** 0 constant
dLog M1 -5.744 *** 0 Constant
dLog M2 -3.547 ** 0 constant
dLog Domestic credit -4.288 *** 0 constant
dLog international reserves -4.344 *** 3 constant
Note: a. * , ** and ***denote rejection of the null hypothesis of a unit root at the10%, 5% and 1% significant levels critical values respectively; Critical values for the level series without constant are -1.954 and -2.653 for 5% and 1% significance levels respectively. Critical values with constant are -2.986 and -3.724 respectively. b. Optimal Lag length is determined by Akaike information criterion (AIC).
c. Deterministic components in the test are determined by AIC.
52
Table 1-2. Johansen Co-integration test results of annual data
Log GDP with selected combination of related series
Selection of
related series
in the system
Deterministic
components in
the cointegrating
equations a
Hypothesized
No. of CE(s)
Eigen-
value
Trace
Statistic Prob.
Max-
Eigen
Statistic Prob. c
M1,
Total trade
Constant
+trend
None * b 0.652 54.141 0.003 31.635 0.008
At most 1 0.463 22.506 0.124 18.634 0.064
At most 2 0.121 3.871 0.761 3.871 0.761
M2,
Total trade
Constant
+trend
None * 0.730 55.839 0.002 39.254 0.001
At most 1 0.387 16.585 0.447 14.663 0.213
At most 2 0.062 1.922 0.973 1.922 0.973
Domestic
credit,
Total traded
Constant
+trend
None * 0.630 49.016 0.011 29.792 0.014
At most 1 0.411 19.224 0.268 15.900 0.150
At most 2 0.105 3.324 0.836 3.324 0.836
Domestic
credit,
Exports
Constant
+trend
None 0.532 36.590 0.186 22.754 0.121
At most 1 0.297 13.836 0.671 10.583 0.557
At most 2 0.103 3.254 0.845 3.254 0.845
Domestic
credit,
Imports
Constant
+trend
None * 0.719 56.398 0.001 38.133 0.001
At most 1 0.408 18.265 0.326 15.708 0.158
At most 2 0.082 2.557 0.925 2.557 0.925
Intl.
Reserves,
Total trade
constant
+trend
None * 0.770 74.395 0.000 44.151 0.000
At most 1 * 0.515 30.244 0.013 21.678 0.023
At most 2 0.248 8.566 0.209 8.566 0.209
Note: a. The VAR systems all include a single lag and a linear trend (a constant and trend) on each variables, selected by the deterministic components in the cointegrating equation, chosen by Akaike Information Criteron (AIC) and Schwarz criteron (SC);
b. * denotes rejection of the null hypothesis at 5% significant level. The critical values for the 5% significant level on the Trace statistic are 42.915, 25.827 and 12.518 for the null hypothesis of no cointegrating equation, at most 1 cointegrating equation and at most 2 cointegrating equation respectively; the critical values for Maximun Eigen statistics for the 5% significant level on the Trace statistic are 25.823, 19.387 and 12.518 for the null hypothesis of no cointegrating equation, at most 1 cointegrating equation and at most 2 cointegrating equation respectively.
c. The p-values by MacKinnon-Haug-Michelis (1999) p-values; d. The combination of related series in the selected model of quarterly real GDP data estimation with
multivariate unobserved component approach.
53
Table1- 3: Disaggregation model selection
Note: a. The sample period covers 1978q1-2009q4, with only 2009 q1-q4 quarterly data observed as initiate
value; b. Stochastic slope: the slope is specified as random walk with drift c. RMSE: Root mean square errors of the estimated quarterly data with the published official quarterly data
over period 1991-2008 d. The cyclical component choice may change to find convergence
Explanatory
variable
Model specification Model comparison criterion
Slope Seasonality Cyclical
component
Log
Likelihood
Akaike
Inform
ation
Criteri
on
(AIC)
Bayesian
Schwartz
Criterion
(BIC)
DW
test RMSE
c
Univariate
models
M1-1 Stochastic b Stochastic No cycle 107.85 -8.21 -8.08 1.950 0.0143
M1-2 Stochastic Stochastic AR(2) 111.03 -9.54 -9.40 1.998 0.0144
M1-3 fixed fixed AR(2) 112.09 -9.52 -9.38 1.998 0.0144
Univariate
models with
Explanatory
variables a
M1, TR
[Chow-Lin
comparable ]
fixed Stochastic
AR(1)
94.95 -7.15 -7.01 1.885 0.0721
M1,
TR[Fernandez
1981 A&R
comparable ]
Stochastic Stochastic
No cycle
99.26 -8.60 -8.43 1.422 0.0134
DC, TR Stochastic Stochastic AR(2) 97.56 -9.12 -8.74 1.713 0.0139
M2 TR Stochastic Stochastic AR(2) 98.04 -8.63 -8.46 1.687 0.0130
M2 IM EX Stochastic Stochastic AR(2) 96.08 -8.72 -8.52 1.623 0.0133
DC IM Stochastic Stochastic AR(2) 97.99 -8.85 -8.67 1.581 0.0142
Multivariate
models d
M1, TR Stochastic Stochastic No cycle 768.98 -8.45 -8.32 1.806 0.0163
M1, TR Stochastic Stochastic AR(2) 773.31 -8.44 -8.30 1.815 0.0162
DC TR Stochastic Stochastic AR(2) 806.56 -9.74 -9.61 1.986 0.0128
M2 TR Stochastic Stochastic no cycle 796.86 -8.40 -8.27 1.981 0.0148
DC EX Stochastic Stochastic AR(2) 810.41 -9.17 -9.04 1.995 0.0147
DC IM Stochastic Stochastic AR(2) 760.95 -9.33 -9.20 1.906 0.0207
DC TR M1 Stochastic Stochastic AR(2) 1234.03 -8.90 -8.77 1.683 0.0249
DC TR IR Stochastic Stochastic no cycle 1051.56 -8.59 -8.46 1.978 0.0168
DC IM EX Stochastic Stochastic no cycle 1067.73 -9.06 -8.92 1.950 0.0196
54
Table1- 4: China quarterly real GDP data: MUC model estimation, A&R estimation
and the official data (1978q1-2011q2)
Continue:
Quarter
MUC estimated
quarterly real
GDP level (2000
as base year)
Standard
Errors a
MUC
estimated
year on
year real
growth
rates
A&R
estimates
year on
year
growth
rates
Official year
on year real
growth rates
(updated
2011 Q2) b
Official published
cumulated year on
year real growth
rates (updated 2011
Q2) c
Official
annual
growth rates
Annual
real GDP
level(2000
as base
year)
1978-1 267.4 0.03216
1978-2 310.0 0.02482
1978-3 317.5 0.03308
1978-4 412.0 11.7 1306.8
1979-1 286.2 0.04353 7.0 6.4
1979-2 332.7 0.03545 7.3 7.3
1979-3 340.9 0.03704 7.4 7.9
1979-4 446.3 8.3 9.1 7.6 1406.1
1980-1 306.5 0.01653 7.1 7.5
1980-2 359.4 0.04006 8.0 8.4
1980-3 367.5 0.05357 7.8 8.2
1980-4 482.4 8.1 7.2 7.8 1515.8
1981-1 327.4 0.03612 6.8 4.8
1981-2 380.1 0.04336 5.8 4.1
1981-3 384.6 0.05673 4.7 3.9
1981-4 502.5 4.2 4.9 5.2 1594.6
1982-1 345.6 0.03061 5.6 6.9
1982-2 407.7 0.03837 7.3 7.8
1982-3 422.4 0.04840 9.8 9.3
1982-4 564.0 12.3 9.0 9.1 1739.7
1983-1 377.0 0.02947 9.1 7.8
1983-2 449.1 0.03604 10.2 9.0
1983-3 471.0 0.04526 11.5 12.1
1983-4 632.2 12.1 13.7 10.9 1929.4
1984-1 421.4 0.02753 11.8 14.9
1984-2 509.4 0.03365 13.4 14.2
1984-3 542.8 0.04210 15.2 14.0
1984-4 749.0 18.5 15.3 15.2 2222.6
1985-1 485.4 0.02550 15.2 16.3
1985-2 586.4 0.03114 15.1 16.3
1985-3 610.6 0.03886 12.5 15.8
1985-4 840.2 12.2 16.8 13.5 2522.7
1986-1 542.9 0.02325 11.8 7.3
1986-2 650.4 0.02841 10.9 10.6
1986-3 663.9 0.03538 8.7 8.9
1986-4 887.4 5.6 8.6 8.8 2744.7
1987-1 593.9 0.02077 9.4 11.0
1987-2 717.8 0.02539 10.4 10.7
1987-3 749.6 0.03158 12.9 11.9
1987-4 1001.6 12.9 13.4 11.6 3063.0
1988-1 666.0 0.01794 12.1 11.4
1988-2 812.3 0.02196 13.2 12.5
1988-3 836.9 0.02728 11.6 11.8
1988-4 1094.0 9.2 9.5 11.3 3409.2
55
Continue:
1989-1 729.6 0.01460 9.6 6.2
1989-2 866.2 0.01790 6.6 5.4
1989-3 862.8 0.02219 3.1 3.2
1989-4 1090.4 -0.3 0.2 4.1 3548.9
1990-1 757.6 0.01025 3.8 2.1
1990-2 891.0 0.01259 2.9 2.3
1990-3 900.8 0.01553 4.4 4.4
1990-4 1134.4 4.0 7.3 3.8 3683.8
1991-1 797.1 5.2 8.6
1991-2 957.1 7.4 8.2
1991-3 1017.9 13.0 9.7
1991-4 1250.6 10.2 10.3 9.2 4022.7
1992-1 905.5 13.6 13.6
1992-2 1082.0 13.1 13.3
1992-3 1153.2 13.3 13.3
1992-4 1453.2 16.2 14.2 14.2 4593.9
1993-1 1042.3 15.1 15.1
1993-2 1239.4 14.5 14.8
1993-3 1308.2 13.4 14.3
1993-4 1642.6 13.3 14.0 13.9 5232.5
1994-1 1176.7 12.9 12.9
1994-2 1387.9 12.0 12.4
1994-3 1470.4 12.4 12.4
1994-4 1882.9 14.6 13.1 13.1 5917.9
1995-1 1317.9 12.0 12.0
1995-2 1528.8 10.2 11.0
1995-3 1616.0 9.9 10.6
1995-4 2100.2 11.5 10.9 10.9 6563.0
1996-1 1461.6 10.9 10.9
1996-2 1678.4 9.8 10.3
1996-3 1769.1 9.5 10.0
1996-4 2310.3 10.0 10.0 10.0 7219.3
1997-1 1613.6 10.4 10.4
1997-2 1846.7 10.0 10.2
1997-3 1920.1 8.5 9.6
1997-4 2510.4 8.7 9.3 9.3 7890.7
1998-1 1736.2 7.6 7.6
1998-2 1973.2 6.9 7.2
1998-3 2074.5 8.0 7.5
1998-4 2722.4 8.4 7.8 7.8 8506.2
1999-1 1894.2 9.1 9.1
1999-2 2123.0 7.6 8.3
1999-3 2235.1 7.7 8.1
1999-4 2900.3 6.5 7.6 7.6 9152.6
2000-1 2064.7 9.0 9.0
2000-2 2310.1 8.8 8.9
2000-3 2434.0 8.9 8.9
2000-4 3112.7 7.3 8.4 8.4 9921.5
2001-1 2240.2 8.5 8.5
2001-2 2489.0 7.7 8.1
2001-3 2624.3 7.8 8.0
2001-4 3391.5 9.0 8.3 8.3 10744.9
56
Notes: a. The standard errors are for the estimated log cumulated year up to date levels estimated by the MUC model.
b. Calculated by the author based on the official published cumulated year on year quarterly growth rates. c. The official year on year real growth rates from 1992-2004 are from Historical data on china Quarterly
GDP estimator 1992-2005 (National Bureau of Statistics of china, 2008). Data from 2005-2011 are from the website of the NBS (http://www.stats.gov.cn/tjsj/jidusj/). All data include official revisions up to date.
.
2002-1 2439.6 8.9 8.9
2002-2 2710.5 8.9 8.9
2002-3 2880.0 9.7 9.2
2002-4 3692.8 8.9 9.1 9.1 11722.7
2003-1 2703.1 10.8 10.8
2003-2 2946.6 8.7 9.7
2003-3 3191.4 10.8 10.1
2003-4 4054.0 9.8 10.0 10.0 12895.0
2004-1 2984.2 10.4 10.4
2004-2 3281.2 11.4 10.9
2004-3 3504.0 9.8 10.5
2004-4 4428.1 9.2 10.1 10.1 14197.4
2005-1 3318.4 11.2 11.2
2005-2 3636.2 10.8 11
2005-3 3899.2 11.3 11.1
2005-4 4948.0 11.7 11.3 11.3 15801.7
2006-1 3729.9 12.4 12.4
2006-2 4135.8 13.7 13.1
2006-3 4377.4 12.3 12.8
2006-4 5565.5 12.5 12.7 12.7 17808.5
2007-1 4252.1 14.0 14
2007-2 4754.1 15.0 14.5
2007-3 4999.9 14.2 14.4
2007-4 6331.4 13.8 14.2 14.2 20337.3
2008-1 4732.5 11.3 11.3
2008-2 5264.3 10.7 11
2008-3 5493.8 9.9 10.6
2008-4 6799.1 7.4 9.6 9.6 22289.7
2009-1 5040.2 6.5 6.5
2009-2 5696.5 8.2 7.4
2009-3 6008.8 9.4 8.1
2009-4 7572.7 11.4 9.1 9.1 24318.1
2010-1 5639.9 11.9 11.9
2010-2 6288.5 10.4 11.1
2010-3 6592.1 9.7 10.6
2010-4 8302.5 9.6 10.3 10.3 26823.0
57
Table1-5. Correlations of year on year growth rates of quarterly real GDP and
potential related series
Sample period
Domestic
credit M1 M2 Exports Imports
Total
trade
1978-2009 0.35 0.43 0.38 0.19 0.49 0.43
1978-1991* 0.43 0.42 0.32 -0.02 0.56 0.51
1992-2009 0.32 0.47 0.56 0.35 0.38 0.40
Note: * official quarterly GDP real growth rates are not available during this period. Annual growth rates are used to get the correlations.
Table1-6. Temporal disaggregation parameter estimates---MUC model with domestic
credit and total trade (Log real GDP equation only)
1978q1-2009q4
(all quarterly real
GDP missing)
1992q1-2009q4
(all quarterly real GDP
missing
1978q1-2009q4
(1978q1-1991q4
real GDP missing)
Log
likelihood 805.566 479.22 1063.98
coefficients
Slope 0.024 0.020 0.025
(0.000) (0.000) (0.000)
AR(1)+AR(2) 0.999 0.999 0.973
Variances of components
Level 0.0000000 0.0000000 0.0000000
Slope 0.0000000 0.0000042 0.0000000
Seasonal 0.0000001 0.0000000 0.0000000
ARs 0.0000161 0.0000053 0.0000000
Irregular 0.0000000 0.0000000 0.0000000
Note: standard errors in parentheses.
58
Variance /correlation of cross series components for Log GDP (final model: 1978q1-
2009q4
With 1978q1-1991q4 real GDP missing)
log GDP
(var)
LDCq
(correlations) LTRq(correlations)
level
0.000000 0.022340 0.096930
Slope
0.000013 0.870300 0.998500
Seasonal
0.000003 0.426000 0.280800
AR(1)
0.000000 0.000980 0.002656
AR(2)
0.000000 0.005417 0.014180
Irregular
0.000000 0.000000 0.000000
Table1-7. Unobserved component model parameter estimates (maximum likelihood)
Drift(μ) Phi1(φ1) Phi2 (φ2)
S.E of
permanent
shocks
S.E of
transitory
shocks
Log
likelihood
2.460 1.876 -0.899 0.937 0.227 -188.684
(0.087) (0.147) (0.151) (0.094) (0.210)
59
Table 1-8: Cointegration test of DdPS data and MUC data (1979q2-2003q4)
hypotheses
Trace
test [ Prob]
Max
test [ Prob]
r=0 21.4 [0.005]** 21.39 [0.002]**
r≤1 0.01 [0.921] 0.01 [0.921]
Note: The tests are Johansen trace eigenvalue test and maximul eigenvalue test. ** denote the rejection of hypotheses at 1% critical value. Rejection of r=0 is evidence in favor of the existence of at least one cointegrating vector.
Table1- 9. Cointegrating analysis of GVAR modeling for China with MUC data and
DdPS original data (1979-2003, replicating of DdPS 2007)
Note: Note: *. **,*** indicate the rejection (at the 10%, 5% and 1% critical values) of the null hypothesis that a particular coefficient is zero. The tests are based on the likelihood ratio statistic that are asymptotically distributed as Chi^2 (1).
MUC data
estimates DdPS estimates
Variables β Standard
Errors(SE) Chi^2(1) Prob β
Standard
Errors(SE) Chi^2(1) Prob
Endogeous variables
China GDP 1 1
China inflation 13.267 (2.180) 18.313 [0.0000]*** 6.791 (1.058) 22.720 [0.0000]***
real exchange rates 0.303 (0.229) 0.916 [0.3386] 0.083 (0.129) 0.258 [0.6118]
ST interest rate of China -43.895 (8.714) 15.464 [0.0001]*** -23.011 (4.286) 20.406 [0.0000]***
Exogenous variables
foreign aggregate GDP 2.887 (2.749) 0.930 [0.3350] 1.747 (1.406) 1.561 [0.2116]
foreign inflation 4.998 (8.738) 0.242 [0.6230] 1.464 (4.327) 0.091 [0.7626]
foreign real equity price -0.114 (0.258) 0.157 [0.691] -0.096 (0.142) 0.462 [0.4968]
foreign ST interest rates 21.501 (12.191) 1.580 [0.2088] 9.835 (5.986) 1.442 [0.2298]
foreign LT interest rates -24.928 (23.187) 1.103 [0.3140] -7.452 (11.557) 0.368 [0.5442]
oil price 0.211 (0.170) 1.466 [0.2260] 0.076 (0.085) 0.730 [0.3929]
TREND -0.049 (0.024) 2.647 [0.1037] -0.038 (0.012) 5.849 [0.0156]**
Figure 1-1. China’s most recent revised official
on year growth rates (the shaded areas are “slowdown eras”
Figure1-2a. The log quarterly real GDP (2000 as base year) and the potential related
series.
0
2
4
6
8
10
12
14
16
181
97
8.1
19
79
.2
19
80
.3
19
81
.4
19
83
.1
19
84
.2
19
85
.3
60
. China’s most recent revised official annual and quarterly real GDP year
the shaded areas are “slowdown eras”)
. The log quarterly real GDP (2000 as base year) and the potential related
19
86
.4
19
88
.1
19
89
.2
19
90
.3
19
91
.4
19
93
.1
19
94
.2
19
95
.3
19
96
.4
19
98
.1
19
99
.2
20
00
.3
20
01
.4
20
03
.1
20
04
.2
20
05
.3
20
06
.4
Annual growth Quarterly growth
quarterly real GDP year
. The log quarterly real GDP (2000 as base year) and the potential related
20
06
.4
20
08
.1
20
09
.2
20
10
.3
61
Figure 1-2b. Log annual data
Figure1-2c: quarterly year on year growth rates of real GDP with monetary
related series (the shaded areas are “slowdown eras”)
-10
0
10
20
30
40
50
1980 1985 1990 1995 2000 2005 2010
Real GDP (show annual growth rate when quarterly not available)domestic creditM1M2
62
Figure 1-2d: quarterly year on year growth rates of real GDP with international
trade related series (the shaded areas are “slowdown eras”)
Figure 1-3. Disaggregation model selection: Year on year quarterly growth rates
(%) 1992-2008
Note: Year on Year quarterly growth rates are calculated as g=log(Yt)-Log(Yt-4)
-40
-20
0
20
40
60
1980 1985 1990 1995 2000 2005 2010
Real GDP (show annual growth rate when quarterly not available)ExportsImportsTotal trades
4
6
8
10
12
14
16
18
92 94 96 98 00 02 04 06 08
official real GDP MUC Chow-Lin Fernandez1981
63
Figure1-4. Year on Year quarterly growth rate (comparing with A&R from 1979-1991)
Figure1-5 Seasonal factors or China’s quarterly real GDP MUC temporal
disaggregation model and X12 seasonal adjustment method
-4
0
4
8
12
16
20
79 80 81 82 83 84 85 86 87 88 89 90 91
MUC estimatesA&RAnnual growth
-1.0
-0.8
-0.6
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
1980 1985 1990 1995 2000 2005
X12 seasonal factorMUC seasonality
64
Figure 1-6. HP cycles with different value of λ
Figure 1-7. Unobserved components decomposition: filtering and smoothing
-8
-6
-4
-2
0
2
4
6
1980 1985 1990 1995 2000 2005 2010
HPCYCLE HP_40 HP_8
-6
-4
-2
0
2
4
6
700
800
900
1,000
1,100
1980 1985 1990 1995 2000 2005 2010
UC trend filteringUC trend smoothingUC cycle filteringUC cycle_smoothing
65
Figure 1-8. HP and Christiano-Fitzgerald cycles of MUC temporal disaggregation and
A&R estimation of China’s quarterly GDP 1978-1992
Figure 1-9. Linear in time residual, HP and BP cycles
-8
-6
-4
-2
0
2
4
6
1978 1980 1982 1984 1986 1988 1990 1992
HPCYCLE_MUC HPCYCLE_AR
-6
-4
-2
0
2
4
6
1978 1980 1982 1984 1986 1988 1990 1992
BPCYCLE_MUC BPCYCLE_AR
-10
-8
-6
-4
-2
0
2
4
6
8
1980 1985 1990 1995 2000 2005 2010
LIT residualBP Baxer-King cycleChristiano-Fitzgerald cycleHP cycle
66
Figure 1-10. HP, Christiano-Fitzgeral and UC cycles
Figure 1-11 a. Seasonal adjusted inflation (CPI) and real GDP level 1986-2010
1-11. b. First difference of log seasonal adjusted inflation and real GDP 1986-2010
-8
-6
-4
-2
0
2
4
6
1980 1985 1990 1995 2000 2005 2010
UC cycleHP cycleChristiano-Fitzgerald cycle
20
40
60
80
100
120
1990 1995 2000 2005 2010
seasonal adjusted CPI level
600
650
700
750
800
850
900
1990 1995 2000 2005 2010
log seasonal adjusted real GDP*100
67
Figure1-12 Impulse responds functions on real output and inflation
Figure1-13. Blanchard-Quah output gap with HP and Christiano-Fitzgerald cycles
-2
0
2
4
6
8
10
86 88 90 92 94 96 98 00 02 04 06 08 10
first difference of log real GDPfirst difference of log CPI
-0.2
0.0
0.2
0.4
0.6
0.8
1.0
1.2
5 10 15 20 25 30 35 40
Output response to demand shockInflation response to demand shock
-1.0
-0.5
0.0
0.5
1.0
1.5
2.0
5 10 15 20 25 30 35 40
Output response to supply shockInflation response to supply shock
68
Figure 1-14 Blanchard and Quah cycles based on MUC data and A&R data
Figure 1-15. The DdPS quarterly real GDP data and the MUC estimated quarterly real
GDP data
-8
-6
-4
-2
0
2
4
6
1980 1985 1990 1995 2000 2005 2010
HP cycleBlanchard-Quah cycleChristiano-Fitzgerald cycle
-8
-6
-4
-2
0
2
4
6
88 90 92 94 96 98 00 02 04 06 08 10
Blanchard-Quah cycle A&R dataBlanchard-Quah cycle MUC data
69
Note: The difference of levels is due to the difference of the base year setting of the two datasets. It will not
affect the VAR analysis, which uses first difference of the series.
2
3
4
5
6
7
8
9
78 80 82 84 86 88 90 92 94 96 98 00 02
Log real GDP_DdPSLog real GDP_MUC
-.01
.00
.01
.02
.03
.04
.05
.06
78 80 82 84 86 88 90 92 94 96 98 00 02
First difference of quarterly real GDP_MUCFirst difference of quarterly real GDP_DdPS
70
Chapter2. Permanent and Transitory Macroeconomic Relationships between the US
and China44
I. Introduction
In the midst of the recent global financial crisis, economic linkages between the US
and China, the largest developed country and the largest developing country respectively,
have become an especially hot topic in the media and among policy makers from both
countries. The nominal GDP of the US and China together accounted for 30% of total world
output in 2008 according to the World Bank Global Economic Monitor estimation. Terms
such as “Chimerica” (Ferguson and Schularick 2007) and “G2” were introduced recently to
describe the ties between the US and Chinese economies and the importance of their
relationship not only to each other, but also to the world economy.
Although bilateral trade and the macroeconomic imbalances experienced by both
countries have been more discussed in the relationship of the US and China, linkages
between these two economies are now substantial in many respects. The two countries have
mutually benefitted from cross-country trade and investment. Concerns, however, have
arisen for both countries due to their close economic linkages. Questions from the US
include: Is China a threat to the US economy? Will the growth of China hurt the
competitiveness of the US? (US Congress research report 2007).45 Questions for China
might be: How is its economic performance affected by the US business cycle and economic
44 The second chapter is based on joint work with Tara M. Sinclair that is currently under revision and resubmission to
the Journal of International Money and Finance. 45 Although the US is still near the top of the list according to the Global Competitiveness Report (World Economic
Forum 2009), China has quickly climbed into the top 30. The US lost its top competitiveness ranking in the World Economic Forum’s Global Competitiveness Report 2009-2010 to Switzerland. The US dropped to second due to the impact of the financial crisis on its financial markets and macroeconomic stability. China inched up from 30 to 29 in the 2009 report.
71
policy? Are the high growth rates China experienced since the economic reform sustainable?
Maintaining a relatively high and stable growth rate is considered to be the top priority for
successful economic reforms and political stability in China. A better understanding of how
the two economies react and interact with respect to macroeconomic shocks is important to
answer the above questions for stake holders from both nations.
Economic theories on economic fluctuations and growth, including real business
cycle theory, Keynesian theory and monetarism, all suggest that economies react differently
to permanent shocks with long run effects than to transitory shocks whose effects dissipate in
the short run. Understanding the relative role of permanent versus transitory movements in
the macroeconomic fluctuations of these two countries and the connections between them is
thus important for economists, forecasters, and policy makers. This paper investigates the
relationships between the macroeconomic fluctuations of the US and China. We do this by
estimating the permanent and transitory components for each country’s real GDP while
allowing for within and cross-country correlations between the permanent and transitory
shocks.
Different economies may experience different types of shocks as well as react
differently to those shocks. Shocks can be shared or transmitted across countries through
trade and financial linkages, through similar economic experiences, or through “contagion”46,
where shocks appear to be transmitted across countries even though there is no fundamental
reason for the transmission. Proper identification and better understanding of the relationship
of the permanent and transitory components of the economic dynamics between the
economies is thus important for proper long term and short term strategy and policy making
46 http://www1.worldbank.org/economicpolicy/managing%20volatility/contagion/definitions.html
72
on the economic relationships between the economies. The issue is of particular importance
for the study of macroeconomic relationships between the US and China. An improved
understanding of the patterns of long term competitiveness and productivity and short term
fluctuations may lead to different domestic and foreign economic and political policies which
influence not only the economic development and future relationships of the two giants but
also the rest of the world.
The model employed in this paper is a two-country correlated unobserved
components model based on the correlated unobserved component model proposed by
Morley, Nelson and Zivot (2003, hereafter MNZ) and extended by Sinclair (2009) and Mitra
and Sinclair (2009). It is estimated with quarterly real GDP data of the two countries from
1978 through 2008. The model specifically allows us to distinguish cross-country
correlations driven by the relationships between permanent innovations, caused by real
shocks such as changes in technology and economic and social institutions, from those
between transitory or cyclical movements, caused by changes in aggregate demand or
monetary shocks in the two countries. The model also allows us to explore the role of
information from the dynamics of each country in identifying fluctuations in the other
country. Bivariate models with alternative information sets are estimated for comparison
purposes.
The structure of the rest of the paper is as following: Section II reviews the related
literatures. Section III presents the econometric models and methods applied. Section IV
discusses the data used in this paper. Section V presents the results of the model estimation.
Section VI concludes.
II. Literature Review
73
2.1 Literature on the Method
Empirical studies examining the macroeconomic relationships across economies
generally apply one of three major approaches. The first method estimates correlations of the
time series of macroeconomic variables or correlations of their filtered cyclical and/or trend
components. The second widely used approach applies vector auto regression (VAR) models
to investigate the co-movement of economic fluctuations among the economies. The third
approach is to use a factor model to capture the correlation among economies in a common
factor or factors.
The first method is the simple correlation method, based either on classical
correlation, which estimates a static correlation between time series, or dynamic correlation
(Croux et al 2001), which takes into consideration the frequency of the business cycles. This
method is very limited and depends heavily on the decision on how to handle the
nonstationarity which is regularly found in macroeconomic time series data. Competing
econometric tools have been developed to decompose macroeconomic series such as the
aggregate output into “trend” and “cycle”, or permanent and transitory components. Among
them, the most widely used univariate methods include the Hodrick and Prescott (1997, HP)
filter, the Baxter and King (1987, BP) filter, the Beveridge and Nelson (1981)
decomposition, and the unobserved components models (Harvey 1985, Clark 1987, and
MNZ 2003). These methods, however, tend to produce very different estimates of trend and
cycle, thus we may find very different correlations depending upon the detrending approach
used. Researchers often report the correlation only for the detrended series, which ignores
the possibility of correlation across permanent shocks. Furthermore, the most commonly
used HP and BP filters are known to be problematic when applied to non-stationary series
74
such as the level of GDP for most countries (Cogley and Nason, 1995; Murray, 2003). In
addition, for this method trends and cycles are first estimated and then the correlation
between these estimated components is estimated in a second stage, which is inferior to
directly estimating the correlation at the same time as estimating the components. As an
alternative to filtering the data, first differenced data can be used, but then again information
is lost and the correlation may reflect a combination of the permanent and transitory
relationships.
The VAR approach on the other hand can be used to identify the effects of underlying
structural shocks, such as monetary and technology shocks, across economies, which can be
much more informative than simply identifying permanent and transitory correlations.
However, structural identification of shocks is sensitive to the identification assumptions of
the structural model. Furthermore, this approach depends on cointegration for finding long
run co-movements in series with unit roots (Granger 1983, Engle and Granger 1987, Vahid
and Engle 1993, Stock and Watson 1988). Highly correlated time series are not necessarily
restricted as cointegrated or having common trend and common cycle. Everaert (2007) finds
that a long run relationship without cointegration may exist between two series using
unobserved components model. As the correlation method, first differencing, which is often
used alternatively to render data stationary for VAR estimation, loses valuable information
about the data and again confounds the role of permanent and transitory shocks.
The third empirical method uses a dynamic factor model (Gregory, Head, and
Raynauld 1997; Forni, Hallin, Lippi, and Reichlin 2002; Forni and Reichlin 2001, Kose,
Otrok, and Whiteman 2003). These models typically assume the existence of a common
factor or factors to capture the cross-country correlation. This assumption may affect the
75
results. Again, these models are often applied to first-differenced data, losing information in
a similar way as for the other two methods.
The two-country correlated unobserved components model applied in this paper does
not require any prior transformation or detrending of the data and places fewer restrictions
among the series. We thus avoid the above problems in simple correlation, VAR, and
dynamic factor methods. In particular, our method combines the detrending and correlation
estimation into a single stage which improves both the estimates of the trend and cycle as
well as the estimates of the correlations. The model is an extension of the univariate
correlated unobserved components model which has been applied to the output fluctuation
analysis of the US and Canada (Basistha 2007, Morley, Nelson, and Zivot 2003). Similar
multivariate models have been applied to macroeconomic variables within single economies
such as the US and Canada (Basistha 2007, Morley 2007, Sinclair 2009), and cross countries
study for G7 countries (Mitra and Sinclair 2009). Furthermore, this model nests many of the
common detrending methods (Trimbur and Harvey, 2003) and is thus more general than
selecting a more restrictive model.
2.2 Studies on the Relationship of Macroeconomic Fluctuations of the US and China
with Other Countries
The US, as the largest economy in the world, is no doubt influential on the rest of the
world. Research on the relationship of macroeconomic fluctuations of the US with other
countries is rich and has generally focused on the correlations across industrialized countries,
mainly among G7 countries and OECD countries. The literature has documented a high
degree of correlation of the US business cycle with other industrialized countries in key
macroeconomic variables (e.g. Kose, Otrok and Whiteman, 2003). Empirical studies on the
76
relationship of the US economic fluctuations with developing countries, concentrated on
Latin American countries, show unsurprisingly strong linkages given the heavy dependence
of these countries on the US economy and the large commodity or tourism trade, as well as
capital and labor flows (e.g. Samuel and Sun 2009). On the trend of the business cycle
correlations, Heathcote and Perri (2003) examined the correlations of HP filtered, first
differenced and high-band pass filtered macroeconomic time series between the US and the
other 15 developed countries. Their study documents that the US economy has been less
synchronized with the fluctuations of the rest of the developed world since 1960 due to
change in the nature of real shocks and the increase of global financial integration.
China, as the largest developing and transitional economy, has been studied mostly
with the Asia and Pacific economies in terms of business cycle synchronization. These
studies are based on the economic integration of the region and the discussion of Optimal
Currency Area (OCA) for the region (Genberg, Liu and Jin, 2006). Trade has been
recognized as the major determinant of the output fluctuation correlation of China with other
East Asian and Pacific economies (Sato and Zhang 2006, Shin and Sohn 2006). Beyond the
region, Calderon (2007) finds increasing output co-movement of China’s output fluctuation
with Latin America countries along with the growing trade integration among the countries.
2.3 Studies on the Relationship of Macroeconomic Fluctuations of the US and China
Among the limited literature that addresses the US and China output fluctuation
correlations, Fidrmuc and Batorova (2008), using quarterly CPI deflated GDP data from
1992-2006, analyses the dynamic correlations of China’s business cycles with selected
OECD countries under different cyclical frequencies. They find that the US has a positive
correlation with China in both long run cycles (over 8 years) and short run cycles (less than
77
1.5 years). Qing (2002) and Chen (2004) 47, using classical correlation techniques, document
the business cycle correlations of China with the US, Japan and select European developed
countries and find positive weak correlation between the output fluctuations of the US and
China, while the correlations between China and Japan and the European countries are
negative. Zhang (2006) investigates the correlations over different sample periods and finds
that the US and China business cycle correlation is stronger during the recent years. Ren and
Song (2004) and Keidel (2008) find there is no correlation between the US and China after
1990 and China’s economic growth has been motivated mainly by domestic factors. In
addition to connections through aggregate output, there are increasing discussions
theoretically on the linkages of the two economies in macroeconomic variables such as
savings and consumptions, trades, finance and money supply (Ferguson and Schularick 2007;
Yang, Askari, Forrer and Teegen 2004; and Johansson 2009).
2.4 Contribution of this paper
This paper is the first study that applies the multivariate correlated unobserved
components model, a more general model with less restrictions and priors than the simple
correlation and VAR approaches, to investigate economic relationships of two economies at
different development levels and with more divergent economic structures. The relationship
between the macroeconomies of the US and China is for the first time viewed through the
lens of permanent and transitory components in the fluctuations of real output of the two
countries through our model. First, we present new properties of the permanent and transitory
US output fluctuations with information from China’s output movements which may carry
information not well studied and understood and different from the information provided by
47 Published in Chinese.
78
developed countries. Second, this paper also contributes to the limited literature on empirical
studies on properties of China’s macroeconomic fluctuations with a reasonably long sample
of quarterly data.
79
III The Model
This paper applies a two-country correlated unobserved components model similar to
Sinclair (2009) and Mitra and Sinclair (2009) to distinguish the correlation of the permanent
shocks to output of US and China, separately from the correlation of the transitory shocks.
The model simultaneously decomposes each output into a stochastic trend, or permanent
component, and a stationary transitory component. The trend, or permanent component, is
assumed to be a process of random walk with drift (Stock and Watson 1988) in order to
capture the steady-state level or long term potential output of the economy. The transitory
component, defined as real GDP deviations from the permanent trend, is assumed to be
stationary following a second order autoregressive process, or AR (2). The two-country
approach enables us to: 1) identify the correlation of the shocks to permanent and transitory
components of real output for each economy with information of dynamics of the other in
order to examine the linkages of permanent shocks and transitory shocks between the two
economies, and 2) obtain new estimates of the permanent and transitory components for each
country using the information of the other country.
Note that the transitory component captures transitory deviations (Morley and Piger
2009) from the permanent or steady state level, which may be fundamentally different from
the traditionally defined business cycle. The traditional business cycle is often isolated from
the series with a filter such as the Hodrick-Prescott (HP) or Band-Pass (BP) filter. In this
paper, we follow a more general definition of permanent and transitory components, which is
associated with the Beveridge and Nelson (1981) decomposition and the Harvey (1985) and
Clark (1987) unobserved components models. The permanent component, or the trend,
follows a stochastic process (a random walk with drift in the model) rather than a fixed or
80
pre-determined path. The transitory component is stationary and deviated from the stochastic
trend, rather than the traditional “alternating-phases” defined (Morley and Piger 2009)
cyclical component. The notion is more general than the traditional definition in that it avoids
any prior determination of appropriate business cycle frequencies. This is particularly
important for macroeconomic fluctuations of developing countries such as China, which may
not experience typical traditional business cycle fluctuations. Under the “transitory-
deviation” definition, the permanent and transitory components of the economic fluctuations
can be directly formulated in structural time series models (Harvey 1993), cast in state space
form and estimated using the Kalman filter or smoother.
The measurement equation of our model is:
ititit cy +=τ , 2,1=i , (1)
where τit is the unobserved trend component and cit is the unobserved cycle component for
country i.
The transition equations are:
ititiit u ηττ ++= −1 , (2)
ititiitiit ccc εφφ ++= −− 2211 , (3)
where itη and itε are assumed to be normally distributed (i.i.d) with mean zero. There are no
restrictions on the correlations between any of the contemporaneous shocks, i.e. no
restrictions are imposed on the variance-covariance matrix, which allows us to estimate all
potential contemporaneous correlations within and across series.
81
The variance-covariance matrix is:
==Σ
2
2
2
2
ccuscccus
cusususcusus
ccuscccus
cusususcusus
εεεεηεη
εεεεηεη
εηεηηηη
εηεηηηη
σσσσσσσσσσσσσσσσ
(4)
We cast equations (1)-(3) into state space form and estimate the unobserved
components and the parameters of the model using the Kalman filter and maximum
likelihood in GAUSS. The unobserved components are estimated with the Kalman
smoothing algorithm, which uses information from the whole sample period, i.e. the future
data as well as the past data. In the results, we will show that the smoothed components are
different from filtered estimates.
IV The Data
The model is estimated with quarterly real GDP data of the US and China from
1978q1 to 2009q4. The Chinese data are from the National Bureau of Statistics of China
(NBS), the nation’s statistical authority. For quarterly real GDP before 1992, when quarterly
real GDP data were not published officially, the data are disaggregated from annual data
using the Chow-Lin (Chow-Lin ,1971) related series method based on Abeysinghe and
Rajaguru (2004)48. The output data for the United States are seasonal adjusted quarterly real
GDP from the Bureau of Economic Analysis of the US Department of Commerce.
48 The disaggregation uses money supply and international trade data, both available at the monthly frequency.
Abeysinghe and Rajaguru’s Chinese disaggregation method is the only published temporal disaggregation estimation of china’s real GDP data for the period. This essay was drafted and submitted before the first Chapter, when A&R was the only available estimation for China’s quarterly real GDP from 1978-1991. In the first Chapter, I present that, although it is less efficient than the multivariate unobserved components disaggregation estimations provided by chapter 1, the A&R estimation is valid. The year 2000 is chosen as the base year because the inflation rate (CPI inflation) was close to zero during that year, which will minimize the distortion from inflation on the quarterly data within the base year.
82
Starting Date:
Although longer history would make our study more robust, the analysis of this paper
focuses on the output fluctuations starting from 1978 due to China’s economic institutional
structure change and the limitation of Chinese data availability. We choose the first quarter
of 1978 as the starting point for the following reasons. First, in 1978, Deng Xiaoping, the
former head of China’s Communist Party after the Cultural Revolution, initiated the market-
oriented economic reform and openness in China. Although the changes did not happen
overnight, the structure of the underlying economic institutions started to change in 1978.
The economy prior to 1978 was generally an autarky and centrally planned, and the
economic growth was interrupted by the political turmoil of the Great Leap Forward
movement and the Cultural Revolution. Along with the launch and implementations of
economic reforms, the post-1978 economy is increasingly market-oriented and open to the
rest of the world. The economic institutions after the start of the reforms has much greater
influence on China’s economic growth pattern now and in the foreseeable future than
economic institutions prior to these reforms. Secondly, the methods applied in this paper
require high frequency macroeconomic data, which are not available before 1978. Due to the
institutional problem mentioned above, we also cannot apply the same disaggregation
method to the period before 1978. Thirdly, the economic growth after 1978 shows an obvious
cyclical pattern (Liu, Zhang and Zhang 2005) which allows us to investigate the dynamics of
the trend and cycle with advanced econometric techniques that have been applied to the
output fluctuations of developed countries.
83
V. Estimation Results
Table 1 presents the classical correlations of the Hodrick-Prescott (HP) and Band-
Pass (BP) cycles and the growth rates of real GDP of the US and China over the entire
sample period. As documented in most of the existing studies, the cycles and growth rates of
the two economies are significantly and positively correlated through the sample period.
Note that the relatively high correlations of HP and BP cycles may be due to spurious cycles
generated by the detrending methods.
Table 2 reports the parameters of the maximum likelihood estimation of our two-
country correlated unobserved components model for the entire sample period, as well as the
parameters estimates from the related univariate model (MNZ model) for comparison.
5.1 Parameter Estimates
Estimates of the drift terms and autoregressive parameters for both countries are all
significant based on our two-country model. With information from the other economy, the
estimated parameters values for both countries are similar to the estimates from the
comparable univariate models.
5.1.1 The Drift Terms
Since each series is in logs and multiplied by 100, the estimated drift term multiplied
by 4 can be interpreted as the average annualized growth of the permanent component, or
trend of the real output in percentage within the sample period.
According to our two-country correlated model, the average annual real growth rates
of the US GDP is estimated as 2.5%, While China’s average permanent real growth rates is
as high as 9.0% annually.
84
We tested for structural breaks in the drift terms for each country using the Quandt-
Andrews unknown date Breakpoint tests (Andrews 1993), but we did not find any significant
structural breaks in our sample period.
5.1.2 The Autoregressive Parameters
The estimated autoregressive coefficients, which reflect the dynamics of the
transitory components, are similar across the different models. The sum of the autoregressive
coefficients, which provides a measure of persistence of the transitory components, shows
that China and the US both have relatively persistent transitory components, with a sum for
each country around 0.80.
5. 2 The Estimated Permanent and Transitory Components
Figure 1 shows the estimated permanent and transitory components of the real GDP
of the US and China based on our two-country uncorrelated UC model. We will discuss each
of these estimated components in the following subsections.
5.2.1 The Permanent and Transitory Components
As MNZ (2003) pointed out, additional information introduced by the real output of
the other country does affect the estimates of permanent and transitory components of each
country in the two-country model. The influences of the information of the other country
appear clearly in the transitory components.
Figure 2-1 compares the estimated US transitory component of the two-country
model with the univariate MNZ estimate. The transitory movements of the US real GDP
better correspond to the NBER-dated recessions (shaded areas of Figure 2-1) than the MNZ
cycle. China’s economic fluctuations introduce very different information for the US output
85
transitory movements than any of the real GDP of G7 countries, with information of which
the US transitory components do not change much. (Mitra and Sinclair, 2009)49.
The official dated economic slowdowns for China, which are represented by the
shaded areas50 in Figure 1-2, appear to correspond mainly to the significant downward
movement of the permanent component. Adding information from the US economic
fluctuation does not visibly change the amplitudes and movement pattern of the transitory
component of China (Figure4). China’s transitory economic fluctuations are not influenced
or forecasted (we do not discuss causality here) by the US real output fluctuations during the
sample period.
Note that China’s transitory movements shift to the left from the MNZ filtered
transitory component, which is equivalent to the Beveridge and Nelson decomposition (MNZ
2003)51. This is due to the Kalman smoothing method we apply in estimating the permanent
and transitory components52. Beveridge and Nelson and MNZ decompositions use the
Kalman filter to estimate the components. The Kalman filter is based on historic information
available up to time t. The Kalman smoothing used here is based on all available information
in the sample. With information from the future, the turning points for China’s transitory
component are estimated to occur earlier than when only information up to time t is used to
estimate the components.53
5.2.2 The Permanent and Transitory Standard Deviations 49 In an unpublished manuscript, Mitra and Sinclair have examined the role of information from a set of Latin
American countries and a set of Emerging Asian economies, and found that the estimated transitory component for the US does not change substantially with the inclusion of information from these countries.
50 To be consistent with the data used in chapter two, the shaded area in chapter two is based on the A&R data, which are slightly different from the MUC estimates in 1984-1985.
51 MNZ (2003) show that their model is equivalent to the Beveridge and Nelson decomposition in the univariate case. Sinclair (2009) shows that this equivalence no longer holds true in the multivariate case.
52 When using basic filter, the gaps between the tuning points disappear. 53 MNZ find that the smoothed and filtered estimates are qualitatively similar for their univariate model applied to US
real GDP.
86
Presented in Table 3, based on the estimates of the two-country model, the standard
deviation of permanent shocks is larger than the standard deviation of the transitory shocks
for both countries, which is consistent with the result from the univariate MNZ models. The
result implies that the trend or permanent components for both countries are much more
variable than the traditional HP and BP smoothed trends. Permanent shocks are relatively
more important than the transitory shocks for both countries. The volatility of China’s real
output fluctuations are higher than that of the US in both permanent and transitory
components.
Figure 2-1 and Figure 2-2 compare the transitory components of the two countries
from our model with the cycles from the HP filter, with λ=1600 for quarterly data. The
transitory components from our model are larger than HP cycles in magnitude for both
countries. It is possible in our case to have both more variable permanent components and
more variable transitory components, because allowing for correlation opens up the
possibility that there may be offsetting movements between the two components (if the
correlation is negative, as we find for both countries in our study).
With information from the other country, the ratio of standard deviations of
permanent shocks over that of transitory shocks are smaller than the univariate MNZ model
results for both countries, especially for the US. This finding is consistent with Cochrane’s
(1994) argument that if we include a series which provides information that increases the
long-horizon forecastability of another series, then we will find larger transitory variation
when we include that information.
5.2.3 Correlations between the Permanent and Transitory Shocks within Economy
87
Based on our two-country correlated UC model, the correlations between the
permanent and transitory shocks with-in economies of the US and China are both
significantly negative, -0.89 for the US and -0.97 for China (Table 4). The estimates are
consistent in the sign with the univariate MNZ model results but with smaller absolute value
for both countries. Note that the correlation of permanent and transitory shock for China is
nearly perfectly negative based on both models. Negative correlated permanent and transitory
shocks have been interpreted as due to slow adjustment of the actual output of the economy
to the permanent shocks on the output. As Stock and Watson (1988) and MNZ (2003)
explained, strongly negative correlation of the permanent shocks with the transitory shocks
implies that the economic fluctuations are driven mainly by permanent shocks, while the
permanent shocks immediately shift the long term path of the output, the short run
movements may include adjustments toward the shifted trend.
5.3 The US- China Relationship—Permanent and Transitory Correlations
Table 4 shows the estimates of the correlations of the permanent-permanent shocks,
the transitory –transitory shocks cross country and the permanent-transitory cross-
correlations. The correlations are estimated simultaneously with the components. We find
that the real GDP of US and China are positively correlated in both permanent shocks (0.56)
and transitory shocks (0.60). The two giants are closely related in both long run and short run
economic fluctuations and share about half of the permanent and transitory shocks. The
values of the correlations are higher than correlations for the US with Japan, Italy, Germany
and France, and only smaller than the US with UK and Canada based on similar multivariate
models (Mitra and Sinclair 2009).
88
5.3.2 Information Carried by the Chinese Real Output to the US Transitory
Components
Figure 3 shows that the bivariate model estimation of the US transitory components
with the real GDP of China are very different from that with other economic variables. The
magnitude of the movement of the US transitory components is enlarged and the turning
points correspond much more directly to the NBER-dated recessions as compared to the
univariate result. In other studies, such as Mitra and Sinclair (2009), Morley (2007), and
Sinclair (2009), the estimated transitory component for US real GDP changes little when
other variables are included in the model. We apply the same bilateral model of US real
GDP with real GDP of Canada, the biggest trade partner of the US, and do not find larger
transitory components for US (Figure 3-2). Following Cochrane’s (1994) argument, the
Chinese real output appears to carry information relevant for forecasting US real GDP which
is not in the GDP data of developed economies such as the G7 (Mitra and Sinclair, 2009) or
in other US data series such as the unemployment rate (Sinclair, 2009) or consumption
(Morley, 2007).
Hamilton (2008) suggests that the US economic fluctuations are mainly driven by the
changes of oil price, which influenced by the increasing energy demand from rapidly
growing China. Estimating a bivariate correlated UC model with the US real GDP and the
world oil price for the same period, we get larger transitory movements for the US real GDP
but the effects are not as big as that from China.
One exception to the finding of a small transitory component for US real GDP is
Basistha and Nelson’s (2007) correlated unobserved components model of GDP, inflation,
and the unemployment rate. Their finding, when compared to the finding of Sinclair (2009)
89
which includes just GDP and the unemployment rate, suggests that inflation may provide
additional forecasting information for US real GDP. Therefore, we estimate another
bivariate model of inflation (measured as the US GDP deflator) with US real GDP. In this
case, the transitory component of US real GDP is smaller in magnitude than the estimation
with oil price, and therefore much smaller than when we use the Chinese data.
Figure 3-1 compares the different estimated transitory components of US real GDP
from four different models: 1) a bivariate model with Chinese real GDP, 2) a bivariate model
with the oil price, 3) a bivariate model with inflation, and 4) a univariate model.
5.3.3 Stability of China’s Transitory Component ---Comparing with Other Bivariate
Models
As discussed by Cochrane (1994), transitory variation, which is mean reverting, is the
forecastable component of the series. The permanent component, which is assumed to follow
a random walk with drift in our model, is the unforecastable component.
Similar to our exercise for the US, we next explore relevant alternative series and
estimate three additional bivariate models with China’s real GDP. Figure 4 compares the
estimated transitory components of China’s real GDP from bivariate models with 1) the US
real GDP; 2) China’s export to the US; 3) real GDP of Hong Kong; 4) Oil Price54. We also
include the estimated transitory component from the MNZ univariate model. None of the
additional series appears to change the magnitude of the transitory variation of China’s real
GDP from the univariate MNZ model, which uses information from China’s lagged real GDP
only. Among the series, China’s transitory component generated with real GDP of Hong
Kong is the most similar to the univariate transitory component.
54 Data resources of the series are: Direction of Trade, International Monetary Fund(China’s export to the US); Census
and Statistic Department of Hong Kong Government (Real GDP of Hong Kong); Wall street Journal (Oil Price)
90
Possible interpretations for the stability of China’s transitory components across
different bivariate models55 could be: first, most of the external shocks are permanent shocks
to China which are not forecastable and thus do not change the transitory components;
secondly, Domestic factors such as domestic demand or monetary policy may be the major
sources of China’s real GDP fluctuations, thus external information sets do not provide much
forecasting information; thirdly, China’s macroeconomic controls or adjustment policies
could have largely isolated the external shocks from greatly influencing the macroeconomic
performance of the country.
5.4 Where are the “G2” now? ----the Recession since 2007
Based on our estimates, both China and the US experienced a large (in absolute
value) negative permanent shock in 2007 which lowered their respective trends. The real
output levels of the two countries at the end of 2008 are both above the permanent trend
(positive in the transitory components) and on the way to converge down to the permanent
path. Since the transitory components are the differences between the series and the
permanent component, the slow adjustment of the actual real GDP levels to the trend after
the big negative shock leaves the transitory components peaking at the beginning of the
recession.
VI Conclusion
In this paper, we estimated a two-country correlated UC model for the real GDP of
the US and China with quarterly data from 1978 through 2008. Our model permits us to
examine both the within-country long term and short term properties of the output
55 We do not apply domestic information sets because: first, availability of quarterly data of domestic economic
indicators for our sample period are very limited , and second, the data construction of the data before 2000 has used the total international trade and money supply--the only quarterly available series.
91
fluctuations of the two countries and the cross-country relationship of the two giant
economies simultaneously. The estimation result also reveals the relative importance of
permanent versus transitory movements in the relationship.
We find that the economic fluctuations of the US and China, are significantly
positively correlated for both permanent and transitory shocks. The two countries share about
half of the shocks both in the long run trend and short run movements. The US transitory
components estimated with China data are very different from that estimated with other
information sets such as inflation, GDP of other developed countries and the oil price.
Estimates of China’s permanent and transitory components do not change too much with
information from the US and alternative external information sets as well, which suggests
that domestic factors may be the major drivers of China’s real GDP fluctuations.
92
Tables and Figures
Table2-1. Correlations of cycles of the US and China real GDP with HP, BP
decomposition and the growth rates Quarterly Data, 1978.1 – 2008.4
Growth Rates* HP Cycles
(lamda=1600)
BP Cycles
(cycle periods
6-32)
YOY
growth
rates**
0.12 0.39 0.44 0.32
*The growth rate is defined as the first difference of the log of real GDP for the US and China. **YOY growth rates: Year on Year growth rate is defined as log changes from same quarter the previous year,
which is often used by literatures published in Chinese. 100)log( ×= realGDPyt Year on year growth
rates 4−−= ttt yyg
Table 2-2. Estimation Results
Model 1 Univariate MNZ
US
(SE)
China
(SE)
US MNZ
(SE)
China MNZ
(SE)
Drift 0.6773
(0.0996)
2.2599
(0.1715)
0.7112
(0.1006)
2.2200
(0.1665)
phi1 1.2520
(0.0394)
1.2610
(0.0806)
1.3601
(0.0983)
1.3240
(0.0798)
phi2 -0.4081
(0.0331)
-0.4612
(0.0632)
-0.6160
(0.0404)
-0.5324
(0.1362)
Log
Likelihood: -288.127 -134.589 -173.023
93
Table 2-3. Standard Deviations of Shocks
Model 1 US MNZ China MNZ
US Permanent 1.0795 (0.0507)
1.1160 (0.2261)
China Permanent 1.8844 (0.0876)
1.8517 (0.4870)
US Transitory 0.9648 (0.0612)
0.7947 (0.1274)
China Transitory 0.7947 (0.1274)
1.1925 (0.6346)
US Ratio Perm/Trans
1.1189 1.4043
China Ratio Perm/Trans
1.4981 1.5529
Table 2-4. Correlations of Permanent and Transitory Shocks
Model 1 US MNZ China MNZ
Permanent shocks China – US
0.5554 (0.2156)
Transitory shocks China – US
0.5972 (0.1038)
Permanent US with Transitory China
-0.6994 (0.1673)
Permanent China with Transitory US
-0.5492 (0.1023)
Permanent US with Transitory US
-0.8859 (0.0747)
-0.9738 (0.1195)
Permanent China with Transitory China
-0.9690 (0.0040)
-0.9999 (0.0001)
94
Figure 2-1: Estimated permanent and transitory components.
Figure 2-1-1 : The US
Note: Shaded areas are NBER-dated recessions.
-6
-4
-2
0
2
4
6
840
860
880
900
920
940
1980 1985 1990 1995 2000 2005 2010
US Transitory Component US Permanent ComponentUS Ln(Real GDP) *100
95
Figure2- 1-2: China
Note: Shaded areas are economic growth slowdown periods recognized by China’s Academy of Social Science
based on annual real growth rates. (Liu 2004) The periods start at the time with peak high growth rate and end
at trough. The quarterly point of peak and troughs are based on the A&R data used in the estimation.
-8
-4
0
4
8
700
750
800
850
900
950
1,000
1980 1985 1990 1995 2000 2005 2010
Transitory ComponentPermanent ComponentChina ln(RealGDP)*100
96
Figure 2-2 Transitory Components Comparison
2-2-1 US Transitory Component: Comparing with HP Cycle
2-2-2 China Transitory Components: Comparing with HP Cycle
-6
-4
-2
0
2
4
6
1980 1985 1990 1995 2000 2005 2010
US Transitory componentUS HP cycle
-8
-6
-4
-2
0
2
4
6
8
1980 1985 1990 1995 2000 2005 2010
China Hp cycleChina Transitory Component
97
Figure2- 3. US Transitory Component Comparing Different Information Sets
2-3-1 US Transitory Components Comparing: Univariate. with China, with Inflation and
with Oil price
-6
-4
-2
0
2
4
6
1980 1985 1990 1995 2000 2005 2010
US Transitory Component MNZUS Transitory Component with InflationUS Transitory Component with OilUS Transitory Component with China
98
2-3-2 US Transitory Component Comparing: with China vs. with Canada
-6
-4
-2
0
2
4
6
1980 1985 1990 1995 2000 2005 2010
US Transitory Component with ChinaUS Transitory Component with CanadaUS MNZ Transitory Component
99
Figure2- 4 China Transitory Components with Different Information Sets
-10.0
-7.5
-5.0
-2.5
0.0
2.5
5.0
7.5
10.0
1980 1985 1990 1995 2000 2005 2010
China Transitory Component with USChina Transitory Component with China export to USChina Transitory Component with HKChina Transitory Component with Oil PriceChina Transitory Component MNZ
100
Chapter3. Permanent and Transitory Macroeconomic Relationships between China
and the Developed World56
1. Introduction
Although research on business cycles and economic growth has traditionally focused
on developed countries, there is increasing interest in the economic fluctuations of
developing countries. In particular, policymakers and researchers have focused on the
growing importance of China, the largest developing country, within the global
macroeconomic environment. In chapter two I explored the connection between the macro-
economic fluctuations of China and the US. This paper extends that analysis to examine the
relationships between the real GDP of China and that of developed countries more generally.
In terms of the discussion about China’s modern role in the global economy, much of
the focus has been placed on China’s connection with the US, given that they are the largest
developing and developed economy respectively, and on China’s connection with
neighboring Asian and Pacific economies. Most research in terms of business cycle
synchronization has focused on the relationships of China with Asian and Pacific economies.
These studies are based on regional economic integration and the discussion of the possibility
of an Optimal Currency Area (OCA) for the region (Genberg, Liu and Jin, 2006). Trade has
been recognized as the major determinant of the output fluctuation correlation of China with
other East Asian and Pacific economies (Sato and Zhang, 2006, Shin and Sohn, 2006).
Beyond the region, Calderón (2007) finds increasing output co-movement of China’s output
56
The third chapter is based on the joint work with Tara Sinclair prepared for the CESifo Venice Summer Institute workshop on “The Evolving Role of China in the Global Economy” and to be published in a conference volume by MIT press.
101
fluctuation with Latin America countries along with the growing trade integration among the
countries.
Much has also been made of the “special relationship” between China and the US,
with terms such as “G-2” and “Chimerica” (Ferguson and Schularick 2007). China is,
however, also tightly connected with developed countries other than the US. For example,
although the US has been China’s largest single country trade partner since the 1990s, Japan,
South Korea, and Germany are also large trade partners with China. In total, developed
countries comprise the majority of both China’s export and import sources, but the US
comprises less than 25%. According to the IMF direction of trade database, the US averaged
only 20% of China’s export market between 2000 and 2009, but the remaining six countries
of the G7 were another 22% of China’s export market and the remaining members of the
developed OECD countries[57] were another 10% (OECD other countries account for 7%).
In terms of imports, the US on average supplies only 8% of China’s imports, whereas
the remaining countries of the G7 supply an additional 24% and the remaining developed
OECD members another 7%. There is limited literature that addresses the output fluctuation
correlations between China and developed countries. Fidrmuc and Batorova (2008), using
quarterly CPI deflated GDP data from 1992-2006, analyze the dynamic correlations of
China’s business cycles with selected OECD countries under different cyclical frequencies.
They find that despite the increasing trade and financial links between China and other 57 The developed OECD countries include the 25 OECD members in the aggregate data: Australia, Austria, Belgium,
Canada, Denmark, Finland, France, Germany, Greece, Iceland, Ireland, Italy, Japan, Luxembourg, Mexico, Netherlands, New Zealand, Norway, Portugal, Spain, Sweden, Switzerland, Turkey, United Kingdom and United States. The developing OECD members include: the Czech Republic, Hungary, Korea, Poland, the Slovak Republic, and Mexico. The data do not include Chile, Slovenia and Israel, new members that joined the OECD after May of 2010.
102
economies, China’s business cycle behaves differently from most other economies. Non-
European OECD countries such as the US, Korea, Australia, and Japan, which have more
intensive economic linkage with China, show relatively high positive correlation of long run
cycles (over 8 years). In general, the dynamic correlations tend to increase in more recent
years. The US has a positive correlation with China in both long run cycles (over 8 years)
and short run cycles (less than 1.5 years). Qing et al (2002) and Chen et al. (2004), using
classical correlation techniques, document the business cycle correlations of China with the
US, Japan, and select European developed countries and find positive weak correlation
between the output fluctuations of the US and China, while the correlations between China
and Japan and the European countries are negative. Zong (2007), using a VAR model on
annual data of China’s GDP, G7 countries aggregate GDP and China’s FDI, reports that G7
GDP Granger-caused the fluctuation of China’s FDI and China’s GDP, while there is no
evidence for an effect in the opposite direction. Lowe (2010) shows that the rolling
correlation of real quarterly growth of China and Australia outpaces the correlation between
growth of the US and Australia since 2000. Fidrmuc and Korhonen (2010) show that
business cycle correlations between China and developed countries are zero on average.
Given the increased emphasis on China’s role in the global economy, it is important
to further investigate the nature of the relationships between China and the developed
countries. In particular, this paper focuses on China’s relationship with two different
aggregate measures for developed economies, the G7 and the OECD. The model employed
in this paper is based on the two-series correlated unobserved components (UC) model
employed in the second chapter (Jia and Sinclair, 2009) which was applied to examine the
relationships between China and the US. The model was developed in Sinclair (2009) as a
103
two-series extension of the correlated unobserved component model proposed by Morley,
Nelson and Zivot (2003, hereafter MNZ). Similar multivariate UC models have been applied
to macroeconomic variables within single economies such as the US (Morley 2007, Sinclair
2009) and Canada (Basistha 2007) and for an aggregate of the euro-zone countries (Berger,
2011). The model has also been applied for a cross-country study of the real output
fluctuations of the G7 countries (Mitra and Sinclair, forthcoming). The model specifically
allows us to distinguish cross-country correlations driven by the relationships between
permanent shocks, caused by real shocks such as changes in technology and economic and
social institutions, from those between transitory or cyclical movements, caused by changes
in aggregate demand or monetary shocks. The model also allows us to explore the role of
information from the dynamics of each series in identifying fluctuations in the other series.
The correlated unobserved components model applied in this paper does not require any prior
transformation or detrending of the data and places fewer restrictions among the series than
other models. In particular, our method combines the detrending and correlation estimation
into a single stage which improves both the estimates of the trend and cycle as well as the
estimates of the correlations. Furthermore, this model nests many of the common detrending
methods (Trimbur and Harvey, 2003) and is thus more general than most other methods.
Two different estimates are presented: one with quarterly real GDP data for China
with aggregate real GDP for the G7 countries and the other with quarterly real GDP data for
China with aggregate real GDP for the 25 OECD member countries. Both models are
estimated with quarterly data from 1978 through 2009.[58] We also compare these estimates
58 We also estimated two additional models for robustness. One was a model for a 30-country OECD aggregate. The
other was for a subsample from 1992-2009 to consider only the officially reported quarterly real GDP for China. The estimates were both quantitatively and qualitatively similar to those reported in this chapter. These estimates are available from the authors upon request.
104
with those based on a univariate unobserved components model of Chinese real GDP as well
as a trivariate model of the real output of China, the US, and Japan. To preview the results,
we find that China has little connection with the developed world aggregate. We cannot
reject that there is no cross-series correlation, and the estimates of the components for both
China and the developed world aggregates are not substantially different from the findings
based on univariate models. The results are similar whether we use the G7 or the OECD
aggregate.
The structure of the rest of the paper is as following: Section 2 presents the
econometric model and estimation method. Section 3 discusses the data used in this paper.
Section 4 presents the results of the model estimation. Section 5 concludes.
II. the Model
This paper applies a two-series correlated unobserved components model similar to
Sinclair (2009) and Jia and Sinclair (2009) to distinguish the correlation of the permanent
shocks to output of China from permanent shocks to aggregate developed country output (in
one model measured as an aggregate of OECD countries and in the other measured as an
aggregate of the G7 countries), separately from the correlation of the transitory shocks. The
model simultaneously decomposes each output series into a stochastic trend, or permanent
component, and a stationary transitory component. The trend, or permanent component, is
assumed to be a random walk with drift (Stock and Watson 1988) in order to capture the
steady-state level or long term potential output of the economy. The transitory component,
defined as real GDP deviations from the permanent trend, is assumed to be stationary
following a second order autoregressive process, or AR (2). The two-series approach enables
us to: 1) identify the correlation of the shocks to permanent and transitory components of real
105
output for each series with information from the dynamics of the other, in order to examine
the linkages of permanent shocks and transitory shocks between the two economies, and 2)
obtain new estimates of the permanent and transitory components for each series using the
information of the other series.
This model is general enough to be applied to cointegrated series, but it does not
require cointegration or common trends. The model allows any amount of correlation
between permanent shocks to the series, from zero correlation to a common trend. If the
series do share a common trend, then cointegration can be imposed in this framework to
improve the efficiency of the estimates (Morley, 2007). Johansen Cointegration tests were
applied to our series for both models and we cannot reject the null of no cointegration
allowing for either a constant or a linear deterministic trend in our data. We thus do not
impose cointegration in the model.
It is important to note that the transitory component captures transitory deviations
from the permanent or steady state level, which may be fundamentally different from the
traditionally defined business cycle (Morley and Piger, forthcoming). The traditional
business cycle is often isolated from the series with a filter such as the Hodrick-Prescott (HP)
or Band-Pass (BP) filter. In this paper, we follow a more general definition of permanent and
transitory components, which is associated with the Beveridge and Nelson (1981)
decomposition and the Harvey (1985) and Clark (1987) unobserved components models. The
permanent component, or the trend, follows a stochastic process (a random walk with drift in
the model) rather than a fixed or pre-determined path. The transitory component is stationary
and is defined as the deviation from the stochastic trend, rather than the alternative definition
of a cycle that captures alternating phases. The notion is more general than the alternating-
106
phases definition in that it avoids any prior determination of appropriate business cycle
frequencies. This is particularly important for macroeconomic fluctuations of developing
countries such as China, which may not experience typical traditional business cycle
fluctuations. Under the “deviation from trend” definition, the permanent and transitory
components of the economic fluctuations can be directly formulated in structural time series
models (Harvey, 1993), cast in state space form, and estimated using the Kalman filter for
maximum likelihood estimation (MLE) of the parameters using prediction error
decomposition.
The measurement equation of our model is:
ititit cy +=τ , (1)
where τit is the unobserved trend component and cit is the unobserved cycle component for
series i (where i=DW represents the aggregate for the developed world and i = C represents
China).
The transition equations are:
ititiit u ηττ ++= −1 , (2)
ititiitiit ccc εφφ ++= −− 2211 , (3)
where itη and itε are assumed to be normally distributed with mean zero. There are no
restrictions on the correlations between any of the contemporaneous shocks, i.e. no
restrictions are imposed on the variance-covariance matrix, which allows us to estimate all
potential contemporaneous correlations within and across series.
The variance-covariance matrix is:
107
=Σ
2
2
2
2
ccDWcccDW
cDWDWDWcDWDW
ccDWcccDW
cDWDWDWcDWDW
εεεεηεη
εεεεηεη
εηεηηηη
εηεηηηη
σσσσσσσσσσσσσσσσ
(4)
We cast equations (1)-(3) into state space form and estimate the unobserved components and
the parameters of the model using the Kalman filter and maximum likelihood in GAUSS.
The unobserved components are estimated with the Kalman smoothing algorithm, which uses
information from the whole sample period, i.e. the future data as well as the past data.[59]
III. the Data
The model is estimated with quarterly real GDP data for China and a developed
country aggregate from 1978 through 2009. The Chinese data are from the National Bureau
of Statistics of China (NBS), the nation’s statistical authority.[60] Our study focuses on the
real output fluctuations since 1978, when China embarked on the market-oriented and
openness economic reform. Our data include the most recent official revisions for 2005
through 2009 based on the information collected through the second economic census
completed at the end of 2009. For quarterly real GDP before 1992, when quarterly real GDP
data were not published officially, the data are disaggregated from annual data using the
Chow-Lin (1971) related series method based on Abeysinghe and Rajaguru (2004).[61] Their
disaggregation uses money supply and international trade data, both available at the quarterly
59 The smoother does produce different estimates of the components as compared to the filter, particularly for Chinese real
GDP. The cycle based on the smoothed estimates is substantially larger. Results for the filtered estimates are available from the authors upon request. 60
The official data are published as cumulated year on year growth rate at comparable price. Data from 1992-2005 are from
the publication of National Bureau of Statistics of China: Historical Data on China Quarterly GDP Estimator 1992-2005, 2/2008 China Statistics Press ISBN/ISSN 9787503753565 61
The year 2000 is chosen as the base year because the inflation rate (CPI inflation) was close to zero during that year,
which will minimize the distortion from inflation on the quarterly data within the base year.
108
frequency, as related series. Abeysinghe and Rajaguru estimate the quarterly growth rates of
real GDP for 1978 through 1994 based on the estimated relationship of annual real GDP
growth rates and the related series from 1978 through 1996.[ 62 ] The results of the
disaggregation are tested by the authors through model fitting and out-of-sample forecast
evaluation. The Abeysinghe and Rajaguru estimates are the only published estimates of
quarterly real GDP data for China for this period. The data allow us to investigate the
relationship of the Chinese economy with the developed world since it started to integrate
with the world economy. This longer time series provides more information on China’s
macro-economic fluctuations and improves the efficiency of the estimation. To investigate
the possible irregularity caused by the difference of data sources and the robustness of the
result, the model was also estimated with official Chinese real output data from 1992 through
2009. We find that the results are remarkably similar to the full sample results.
The Chinese real output data are seasonally adjusted using the X-12 ARIMA method.
The X-12 ARIMA (2, 1, 2) and Tramo/seat (Time series Regression with ARIMA noise,
Missing Values and Outliers/Signal Extraction in ARIMA Time series) methods give similar
results. The finding is consistent with Blades (2007), who performed similar tests on current
price quarterly GDP of China. The seasonal pattern of China’s quarterly real GDP is regular
and predictable. The method is consistent with the one applied by the OECD for the
developed world data.
For the developed countries data, we focus on two measures: real GDP for the G7
countries and real GDP for 25 OECD countries (although a model of 30 OECD countries
yielded equivalent results). The data come from the OECD and are measured as millions of
62 We only use Abeysinghe and Rajaguru’s data through 1991 and then use the NBS data.
109
US dollars, volume estimates, fixed PPPs, OECD reference year, annual levels, seasonally
adjusted.[63] The 25 OECD countries included in the OECD aggregate are: Australia, Austria,
Belgium, Canada, Denmark, Finland, France, Germany, Greece, Iceland, Ireland, Italy,
Japan, Luxembourg, Mexico, Netherlands, New Zealand, Norway, Portugal, Spain, Sweden,
Switzerland, Turkey, United Kingdom, and United States (the 30-country aggregate adds the
Czech Republic, Hungary, Korea, Poland, and the Slovak Republic).[64] The G7 countries are
Canada, France, Germany, Italy, Japan, United Kingdom, and United States. It is important
to note that all of the G7 countries are also included in the OECD aggregate.
3.1 Chinese Data Quality
Along with the increasing interest in China’s economic performance, the quality of
Chinese official macroeconomic statistics, including the GDP data,[65] has been repeatedly
questioned by a number of researchers and media reports. Despite the efforts made by NBS
to explain and improve the GDP estimates over years, confidence in the accuracy of official
data is still low. The data quality still remains a problem that must be addressed for empirical
research on Chinese macroeconomic issues.
In the early 2000s, heated discussions on the quality of Chinese macro data generated
a large number of publications on this issue. The criticisms of China’s official data are based
on evidence from alternative GDP calculations (Maddison, 1998; Wu, 2000; Young, 2003),
63 The data were extracted on September 29, 2010 from OECD.Stat. 64 The 30-country aggregate was the largest available OECD aggregate at the time of the writing of this chapter. According to OECD.stat, “Chile became a member of the OECD on 7 May 2010, Slovenia on 21 July 2010 and Israel on 7 September 2010 and data for them now appears in the list of OECD member countries. Nevertheless, Chile, Israel and Slovenia have not yet been included in OECD area aggregation in the quarterly national accounts database for technical and timing reasons.” The estimates using the 30-country aggregate are available from the authors upon request. 65 The Economist (2008) cited Goldman Sachs’ ranking of the reliability of Chinese statistics from high to low as: Foreign trade, Money supply, Industrial production, consumer prices, GDP, retail sales, fixed investment, Employment, Average earnings, Unemployment, where GDP is in the middle.
110
comparison with energy and transportation consumption data (Rawski, 2001), and suspects
of data falsifications, especially on the local level, under the non-democratic political
system.[66] In the media, people are also concerned about the quick publication, only two
weeks after the end of reporting periods, of the preliminary national account data for such a
big economy.[67] This criticism persists even though before 1988 the Bureau of Economic
Analysis released real GDP estimates for the US just 15 days after the end of the quarter
(Young, 1993).
Refutations to the criticisms show the alternative data series constructed or corrected
by researchers have not been proved to be more precise or reliable (Holz, 2006). Many
researchers find that GDP data problems are unlikely to be unique to China and the evidence
is not robust for a conclusion of data manipulation or systematic data falsification (Holz,
2005 and 2006; Chow, 2006; Klein and Ozmucur, 2003). Chinese statistical authorities
explain most of the questions as lack of understanding of China’s transitional statistical
system and the nature of a transitional economy. Some problems have gained
acknowledgement from the authorities (Xu, 2002 and Xu, 2004) and efforts have been made
to improve the data quality. The data are compiled and revised based on the information
gained from recently established regular surveys and economic censuses, revised financial
statement reports for enterprises and the more sophisticated data sources system.
Manipulating statistics to meet political objectives, as the most usual concerns, are much
66 See Holz (2006) for a detailed survey of the literature. 67 The most recent official announcement on the timing of revisions of the quarterly data has become more cautious and leaves more time for the first and final revisions of the number.
111
harder at the national level. Xu Gao of the World Bank provides evidence of the consistency
of data from different government institutes for recent years in his official blog.[68]
After carefully reviewing the literature on Chinese data quality and their national
accounting system, and comparing different data resources and data construction methods,
we agree with many researchers and most international organizations (OECD, IMF[69]) that
although there are weaknesses or short-comings in the statistical system that provides
Chinese national accounts estimation, the Chinese official macroeconomic data after 1978
are not proved to be politically manipulated or systematically biased. The official data can
serve as “a reliable guide” to the level and growth pattern of GDP, even though the margins
of error are “certainly larger than that of the most developed countries” (Lequiller and
Blades, 2006).
IV. Results
Table 1 presents the classical correlations of the Hodrick-Prescott (1997) and Baxter-
King (1999) cycles and the growth rates of real GDP of China with the G7 and the OECD
aggregates over the entire sample period.[70] Note that the correlations of Hodrick-Prescott
and the Baxter-King cycles may be due to spurious cycles generated by the detrending
methods (Cogley and Nason, 1995, and Murray, 2003). Compared with the correlations
68
http://blogs.worldbank.org/eastasiapacific/are-chinese-statistics-manipulated 69
The World Bank criticized the Chinese national account statistics and revised their GDP estimation for China upward for
34% from the officially reported number in 1993. In 1996, the World Bank accepted China’s reformed statistical system and the official GDP number again. But the World Bank revision and method of estimation are also questioned by many researchers.
70 The quarterly growth rate is defined as the first difference of the log of real GDP. The year-on-year growth rate is defined
as log changes from the same quarter of the previous year, which is often used by articles published in Chinese, i.e.
100)log( ×= realGDPyt Year on year growth rates .4−−= ttt yyg
112
between the US and China as reported in the second chapter, the pattern is similar but in all
cases the correlations are lower between the G7 and the OECD with China than between the
US and China. Depending on the choice of method to address the nonstationarity that is
present in the real GDP series the conclusion about the tightness of the relationship between
China and the developed world differs substantially. In general it appears that China and the
developed world share somewhere between less than 10% and 25% of their fluctuations. This
lack of clear conclusion suggests that further investigation is warranted.
4.1 Correlated Unobserved Components Model Parameter Estimates
Tables 2 – 5 report the parameters of the maximum likelihood estimation of our two
correlated unobserved components models for the entire sample period. The results are
strikingly similar for China when we use either aggregate, although the standard errors
suggest that the results based on the larger OECD aggregate are more precisely estimated
than for the model using the G7 aggregate. The estimates for both aggregates are similar as
well, and are consistent with estimates for developed countries individually, such as those
reported in MNZ for the US and in Mitra and Sinclair (forthcoming) for the G-7 countries.
4.1.1 Drift Terms
Since each series is in logs and multiplied by 100, the estimated drift term multiplied
by 4 can be interpreted as the average annual growth of the permanent component.
According to our estimates, China’s average permanent real growth rate is 9.6% annually
whereas for the G7 it is 2.2% and for the OECD it is 2.3%. These estimates are similar to
other estimates reported in the literature.
4.1.2 Autoregressive Parameters
113
The estimated autoregressive coefficients, which reflect the dynamics of the
transitory components, are similar across the different models. The sum of the autoregressive
coefficients, which provides a measure of persistence of the transitory components, suggests
that China has a more persistent transitory component than either the G7 or the OECD
aggregate. Both the G7 and the OECD have persistence measures less than 0.5, whereas for
China it is 0.83.
4.1.3 Permanent and Transitory Standard Deviations
Presented in Table 3, the estimated standard deviations of the permanent and
transitory shocks are similar across models. The standard deviation of the permanent shocks
is larger than the standard deviation of the transitory shocks for both China and the
developed country aggregate for both models. The result implies that the trend or permanent
component for each series is much more variable than the traditional HP and BP smoothed
trends. Furthermore, permanent shocks are relatively more important than the transitory
shocks for each series. Permanent shocks to Chinese real GDP are substantially more
variable than permanent shocks to the developed aggregates. Chinese permanent shocks have
almost twice the standard deviation of the developed world permanent shocks. For the
transitory components, the difference is even more dramatic, with transitory shocks for China
having almost three times the standard deviation as transitory shocks to the developed world.
Thus, although the absolute magnitudes of both the transitory and the permanent standard
deviations are higher for China than for the developed world aggregates, as might be
expected given China’s higher average growth rate, the ratio of permanent to transitory
variability is less for China than the developed world aggregates. In both cases they are
greater than one, however, suggesting an important role for permanent shocks for all series. It
114
is possible in our case to have both more variable permanent components and more variable
transitory components, because allowing for correlation opens up the possibility that there
may be offsetting movements between the two components.
4.1.4 Within Series Correlations
Based on our two-series correlated UC model, the correlations between the permanent
and transitory shocks within the economies of China and the developed world are all
significantly negative (Table 4). In fact the correlation of permanent and transitory shocks for
all series is nearly perfectly negative based on both models. Negatively correlated permanent
and transitory shocks are a common finding for real GDP. These results are consistent with
prior research that has examined the correlation between permanent and transitory shocks for
the real GDP of the U.S. (MNZ; Sinclair, 2009), Canada (Basistha, 2007), the U.S. and the
U.K. (Nagakura, 2008), and the G-7 countries (Nagakura, 2007; Mitra and Sinclair,
forthcoming). The negative correlation has been interpreted as due to slow adjustment of the
actual output of the economy to the permanent shocks to output. As Stock and Watson (1988)
and MNZ (2003) explained, strong negative correlation of the permanent shocks with the
transitory shocks may be interpreted as implying that the economic fluctuations are driven
mainly by permanent shocks, while the permanent shocks immediately shift the long term
path of the output, the short run movements may include adjustments toward the shifted trend.
4.1.5 Cross Series Correlations
Table 5 shows the estimates of the correlations of the permanent-permanent shocks,
the transitory –transitory shocks cross country and the permanent-transitory cross-
correlations. The correlations are estimated simultaneously with the components. We find
that for the G7 aggregate we cannot reject the null that there is no cross-series correlation. A
115
likelihood ratio test with four restrictions results in a chi-squared statistic of 3.45 which has
p-value of 0.49. Similarly, for the OECD aggregate, the likelihood ratio test statistic is 4.51
with a p-value of 0.34. This finding is consistent with the finding of Fidrmuc and Korhonen
(2010) that business cycle correlations between China and developed countries are zero on
average.
4.2 Estimated Permanent and Transitory Components
Figure 1 shows the estimated permanent and transitory components of the real GDP
of China based on our two different bivariate models as well as the estimated components for
the G7 and the OECD aggregates. These estimates suggest that the transitory components for
the developed-world aggregates are small and noisy, similar to previous findings for
estimates of the developed countries individually (for example see Mitra and Sinclair,
forthcoming, for the G7 countries). The permanent components appear very similar to the
series themselves. For China, however, there appears to be more substantial transitory
movement. Some of this more substantial transitory movement is simply due to the larger
size of fluctuations more generally as compared to the developed countries. Recall from
Section 4.1.3 that the transitory fluctuations for China are almost three times as variable as
those of the developed world. The permanent component for Chinese real GDP still appears
quite similar to the series itself.
The role of the information of the other countries is presented in the estimated
transitory components in Figures 2 and 3. In Figure 2 we compare the estimated transitory
component from two different models – the bivariate model with China and a developed
country aggregate (the G7 and the OECD aggregate provide cycle estimates for China that
are indistinguishable from each other) and the univariate correlated UC model applied to
116
China alone. We see that the estimated components are broadly similar. Figure 3 shows that
separating out two of the key members of the G7, i.e. the US and Japan, to create a trivariate
model does not substantially change the estimated transitory component for China either. We
also estimated a model with a larger OECD aggregate which included the real GDP of 30
OECD Member countries: Australia, Austria, Belgium, Canada, Czech Republic, Denmark,
Finland, France, Germany, Greece, Hungary, Iceland, Ireland, Italy, Japan, Korea,
Luxembourg, Mexico, Netherlands, New Zealand, Norway, Poland, Portugal, Slovak
Republic, Spain, Sweden, Switzerland, Turkey, United Kingdom, and United States. The
results were unchanged.
Result of chapter two showed that adding information from US economic fluctuations
does not visibly change the amplitudes and movement pattern of the transitory component of
China as compared to the univariate results. They further show that adding other alternative
external information sets such as the real GDP of Hong Kong or the oil price does not change
this result. A further investigation of the China’s real GDP fluctuation with China’s
international trade variables, using the bivariate model shows similar result (Figure 4). Here
we show that even a large aggregate of developed world GDP provides little new information
for China’s real output fluctuations. Possible interpretations for the stability of China’s
transitory components across different bivariate models could be: first, most of the external
shocks are permanent shocks to China which are not forecastable and thus do not change the
transitory components; secondly, domestic factors such as domestic demand or monetary
policy may be the major sources of China’s real GDP fluctuations,[ 71 ] thus external
information sets do not provide much forecasting information; thirdly, China’s
71 We do not consider domestic information sets because: first, availability of quarterly data of domestic economic
indicators for our sample period are very limited, and second, the data construction of the data before 2000 has used the total international trade and money supply--the only quarterly series available.
117
macroeconomic controls or adjustment policies could have largely isolated the external
shocks from greatly influencing the macroeconomic performance of the country.
The result is in agreement with the finding of chapter one. Using a global vector
autoregression model (GVAR), chapter one shows that supply side shocks and domestic
factors play an important role in China’s real output movements, while none of the foreign
variables, such as the trade weighted aggregate real output, interest rates and equity price of
the rest of the world and the world oil price, appears to be significant for China’s real output
fluctuations.
4.3 The “Great Recession”
From 2007 through 2009 most of the world experienced the “Great Recession.”
Although China did not experience an absolute decline in real GDP, according to most
sources, including the Economic Cycle Research Institute (ECRI),[72] China experienced a
growth cycle peak in May of 2007 and a trough in December of 2009. Similarly, the G7 and
OECD countries all experienced business cycle peaks and troughs during this period.
Therefore, we next investigate what the model suggests about this important episode in our
sample. Figure 5 presents a “zoom-in” on Chinese real GDP and our estimates for the
permanent component based on three different models for the period 2007-2009. The
estimates show that, although the estimates are broadly similar, if we relied on a univariate
model to estimate the permanent component for China that we would assume that the
permanent component moved substantially below the series between the second quarter of
2007 and the third quarter of 2008. According to the bivariate model, however, the
permanent component remained much closer to the series. By contrast, the estimates for both
72 www.businesscycle.com
118
the G7 and the OECD aggregates suggest that there was substantial downward movement in
their permanent components during this recession (Figure 6).[73]
V. Conclusion
In this paper, we presented the estimates of two different bivariate correlated UC
models for the real GDP of China with aggregate measures of developed country real GDP
with quarterly data from 1978 through 2009: one with a G7 country aggregate and one with
an OECD country aggregate. Our model permits us to examine both the within-country long
term and short term properties of the output fluctuations of the two series and the cross-series
relationships of the two series simultaneously. The estimation results also reveal the relative
importance of permanent versus transitory movements in the relationship. We find that
although China and the developed world share substantial trade connections, we cannot reject
that there are no cross-series correlations between Chinese real GDP and an aggregate of
developed world GDP measured by either the G7 countries or the OECD countries.
Although there seems to be little correlation between the real output fluctuations of
China and the developed world in terms of the permanent and transitory shocks and also little
evidence of additional information for each other’s fluctuations, there remain interesting
similarities between China and the developed world. Like the findings for both individual
developed countries reported in the literature as well as for the developed country aggregates
reported here, we find that China has significant negative correlation between permanent and
transitory shocks to its real GDP. We also find that China has an important role for
73 Comparing the smoothed estimates reported here with the filtered estimates (available from the authors upon request)
does suggest that hindsight improves our understanding of the role of permanent versus transitory shocks particularly for China in this episode. The filtered estimates suggest a much larger drop in the permanent component for China (more similar to the estimates reported for the developed country aggregates) as compared to the smoothed estimates.
119
permanent shocks in its real GDP fluctuations, which is similar to the finding for the
developed countries. China does, however, have a much larger drift term, such that
permanent shocks are substantially larger on average. These larger permanent shocks drive
China’s faster growth rate. Consistent with this faster growth rate, both the permanent and
transitory shocks are substantially more variable for Chinese real GDP than those of
developed countries. The similarities suggest that similar macroeconomic policies may be
appropriate for China as for developed countries, although the lack of correlation and the
greater size and variability of shocks may mean that different timing and size of policy may
be necessary. The small correlation of China’s output fluctuations with the developed world
indicates that domestic factors such as economic reforms, domestic demand, and economic
policies may be the major drivers of China’s macro economic fluctuations.
120
Tables and Figures
Table 3-1:Correlations of Cycles for China and the Developed Country Aggregates
Developed
Country
Aggregate
Quarterly
Growth Rates
Year-on-
Year Growth
Rates
Hodrick
Prescott
Cycles
(lamda=1600)
Baxter-King
Cycles (cycle
periods 6-32)
G7 0.09 0.18 0.28 0.21
OECD 0.11 0.16 0.24 0.14
Table 3-2: Estimation Results
China and G7 China and OECD
Log
Likelihood: -251.16 -247.13
China
(SE)
G7
(SE)
China
(SE)
OECD
(SE)
Drift 2.40
(0.18)
0.56
(0.09)
2.39
(0.18)
0.58
(0.09)
phi1 1.31
(0.04)
0.56
(0.25)
1.31
(0.05)
0.56
(0.15)
phi2 -0.48
(0.04)
-0.07
(0.20)
-0.48
(0.05)
-0.10
(0.17)
121
Table 3-3: Standard Deviations of Shocks
China and G7 China and OECD
Developed
Permanent
1.04
(0.68)
0.99
(0.05)
China
Permanent
1.97
(0.96)
1.97
(0.08)
Developed
Transitory
0.59
(0.61)
0.62
(0.08)
China
Transitory
1.43
(0.09)
1.43
(0.03)
Developed Ratio
Perm/Trans 1.76 1.60
China Ratio
Perm/Trans 1.38 1.38
122
Table 3-4: Within Series Correlations of Shocks
Table 3-5: Cross Series Correlations of Shocks
China and G7 China and
OECD
Permanent Developed with
Transitory Developed
-0.99
(0.03)
-0.97
(0.02)
Permanent China with
Transitory China
-0.99
(<0.01)
-0.99
(0.01)
G7 OECD
Permanent China with
Permanent Developed
0.07
(0.17)
0.07
(0.04)
Transitory China with
Transitory Developed
0.03
(<0.01)
-0.02
(0.01)
Permanent Developed with
Transitory China
0.07
(0.19)
0.07
(<0.01)
Permanent China with
Transitory Developed
-0.16
(0.02)
-0.11
(0.06)
123
Figure3- 1: Estimated permanent and transitory components.
1.a. China Based on Bivariate Model with G7
1.b. China Based on Bivariate Model with OECD
124
1.c. G7 Based on Bivariate Model with China
1.d. OECD Based on Bivariate Model with China
125
Figure 3-2: Comparing the Different Filtered Cycle Estimates: Univariate and Bivariate
Models
Figure 3-3: Comparing the Different Cycle Estimates: Univariate, Bivariate, and
Trivariate Models
126
Figure 3-4: Comparing the Cycle Estimates: DW aggregate, Exports and Trade Balance
Figure 3-5: 2007 – 2009 Chinese Real GDP and Permanent Component Estimates
127
Figure 3-6: 2007 – 2009 G7 and OECD Real GDP and Permanent Component
Estimates
128
Reference:
Abeysinghe, T and G. Rajaguru (2004), "Quarterly real GDP estimates for China and
ASEAN4 with a forecast evaluation." Journal of Forecasting 23(6): 431-447.
Andrews, D. W. K (1993), “Tests for Parameter Instability and Structural Change with
Unknown Change Point," Econometrica 61(4): 821-56
Banerjee, A., J. J. Dolado, J. W. Galbraith, and D. Hendry (1993), “Co-integration, Error
Correction, and the Econometric Analysis of Non-Stationary Data”, Advanced Texts in
Econometrics Oxford: Oxford University Press.
Basistha, A. (2007), "Trend-Cycle Correlation, Drift Break and the Estimation of Trend and
Cycle in Canadian GDP." Canadian Journal of Economics 40(2): 584-606
Basistha, A. and C. R. Nelson (2007), "New Measures of the Output Gap Based on the
Forward-Looking New Keynesian Phillips Curve." Journal of Monetary Economics
54(2): 498-511
Baxter, M. and R.G. King (1999), “Measuring business cycles: Approximate bandpass
filters.” The Review of Economics and Statistics, 81(4): 575-93
Bayoumi, T. & B. Eichengreen (1992), "Shocking Aspects of European Monetary
Unification," NBER Working Papers 3949, National Bureau of Economic Research,
Inc.
Berger, T. (2011), "Estimating Europe’s natural rates." Empirical Economics, vol.40 No.2,
pp. 521-536.
Bersch, J. & T.Sinclair (2011), "Mongolia: Measuring the Output Gap," IMF Working
Papers 11/79, International Monetary Fund
129
Beveridge, S. and C. R. Nelson (1981), “A New Approach to Decomposition of Economic
Time Series into Permanent and Transitory Components with Particular Attention to
Measurement of the Business Cycle,” J. of Monetary Economics 7 , 151- 174
Blades, D. (2007), “Decumulating China’s Quarterly National Accounts”,
http://www.oecd.org/dataoecd/27/11/39085954.pdf
Blanchard, O.J. and D. Quah (1989), “The Dynamic Effects of Aggregate Demand and
Aggregate Supply,” the American Economic Review, 79(4), pp. 655-73
Bosworth, B. and S.M. Collins (2007), “Accounting for growth: comparing China and India”
NBER working paper 12943
Burns, A. F., and W.C. Mitchell (1946), “Measuring business cycles” NBER, New York.
Calderón, C. (2007), “Trade, Specialization and Cycle Synchronization: Explaining Output
Comovement between Latin America, China and India”, China's and India's Challenge
to Latin America: Opportunity or Threat? The World Bank Publication ISBN: 0-8213-
7308-0
Canova, F. (1998), “Detrending and business cycle facts”. Journal of Monetary Economics
41, 475–512
Chan, W.S. (1993), "Disaggregation of Annual Time Series Data to Quarterly Figures: A
Comparative Study," Journal of Forecasting, 1993, 677-688
Chen, K., Zhou, Y. and L. Gong (2004), “The fluctuations of China’s Business cycles: Based
on different Filters.” The World Economy, Vol.10 (In Chinese)
130
Chen, L. and H. Liu (2007), “Empirical Investigation on the Asymmetry and Persistence of
Chinese Business Cycle” Economics Research Journal, Vol.42 No.4 April, 2007 (in
Chinese)
Chow, G.C. (1993), “Capital Formation and Economic Growth in China,” Quarterly Journal
of Economics, 108, 809-42.
Chow, G.C. and K.W. Li (2002), “China’s Economic Growth: 1952-2010,” Economic
Development and Cultural Change, 51, 247-56.
Chow, G. (2006), “Are Chinese Official Statistics Reliable?” CESifo Economic Studies,
Vol.52 2/2006, 396-414.
Chow, G., and A. Lin (1971), “Best linear unbiased interpolation, distribution, and
extrapolation of time series by related series.” Review of Economic Statistics, Vol.53,
pp 372–375.
Chinese academy of Social Sciences, 2008-2010, Economics Blue Book of The Peoples
Republic Of China (Volumes of 2008-2010) Analysis And Forecast, The Chinese
Academy of Social Sciences’ Annual Economic Reports (In Chinese).
Christiano, L. J & T. J. Fitzgerald (2003), "The Band Pass Filter," International Economic
Review, vol. 44(2), pages 435-465, 05.
Clark, P. K. (1987), “The cyclical component of U.S. economic activity.” The Quarterly
Journal of Economics, Vol.102, pp.797-814.
Clarida, R. & J. Gali (1994), "Sources of real exchange-rate fluctuations: How important are
nominal shocks?," Carnegie-Rochester Conference Series on Public Policy, Elsevier,
vol. 41(1), pages 1-56, December.
131
Cochrane, J. H. (1994), “Permanent and Transitory Components of GNP and Stock Prices”,
Quarterly Journal of Economics 107, 241-265.
Cogley, T. and J. M. Nason (1995), “Effects of the Hodrick-Prescott filter on trend and
difference stationary time series Implications for business cycle research.” Journal of
Economic Dynamics and Control, Vol.19, No.1-2, pp. 253-278.
Cuche. N. A. Cuche & M. K. Hess (1999), "Estimating Monthly GDP In A General Kalman
Filter Framework: Evidence From Switzerland," Working Papers 99.02, Swiss National
Bank, Study Center Gerzensee.
Curtis, C. C. and N. C. Mark (2010), “Business Cycles, Consumption and Risk-Sharing: How
Different Is China?” forthcoming inYin-WongCheung,Vikas Kakkar and Guonan Ma,
eds., The Evolving Role of Asia in Global Finance.
Croux, C. ,Forni, M. and L. Reichlin (2001), "A Measure Of Comovement For Economic
Variables: Theory And Empirics," The Review of Economics and Statistics, MIT Press,
vol. 83(2), pages 232-241.
Congressional Research Service (2007), “Is China a Threat to the U. S. Economy?” Report
for Congress, Order Code RL 33604
Dees, S., F. di Mauro, M.H. Pesaran, and , L.V. Smith (2007), “Exploring the international
linkages of the euro area: A GVAR analysis”, Journal of Applied Econometrics, 22,
1—38.
Durbin, J. & S.J. Koopman (2001), "Time Series Analysis by State Space Methods," OUP
Catalogue, Oxford University Press, number 9780198523543.
132
Durbin J. and B. Quenneville (1997),“Benchmarking by state Space Models”, International
Statistical Review, 65, 23-48.
Economist. (2008), “An Aberrant Abacus.” The Economist Economics Focus. May 1st, 2008.
Ericsson, N.R., D.F. Hendry, and H.-A. Tran (1994), “Cointegration, seasonality,
encompassing and the demand for money in the United Kingdom” In: Hargreaves, C.
(Ed.) (1994). Non-stationary Time-series Analysis and Cointegration, pp. 179-224.
Oxford: Oxford University Press.
Engle, R.F. and C.W.J. Granger (1987), “Cointegration and error correction: representation,
estimation and testing”, Econometrica 55, pp. 251–276.
Eurostate (1999), Handbook on Quarterly National Account, European communities 1999.
Everaert, G. (2007), "Estimating Long-Run Relationships between Observed Integrated
Variables by Unobserved Component Methods," Working Papers of Faculty of
Economics and Business Administration, Ghent University, Belgium 07/452.
Ferguson, N. and M. Schularick (2007), "'Chimerica' and the Global Asset Market Boom,"
International Finance, 10(3), 215-239.
Fernandez, P. E. G (1981), “A methodological note on the estimation of time series”, the
review of Economics and Statistics 63, pp. 471-478.
Fidrmuc, J. and I. Batorova (2008), "China in the World Economy: Dynamic Correlation
Analysis of Business Cycles," Working Papers RP2008/02, World Institute for
Development Economic Research (UNU-WIDER).
133
Fidrmuc, J. and I. Korhonen (2010),"The Impact of the Global Financial Crisis on Business
Cycles in Asian Emerging Economies." Journal of Asian Economics Vol. 21, pp. 293-
303.
Forni, M., Hallin, M., Lippi, M., and L. Reichlin (2005), "The Generalized Dynamic Factor
Model: One-Sided Estimation and Forecasting," Journal of the American Statistical
Association, vol. 100, pages 830-840.
Genberg H., Liu, L. and Jin. X. (2006), "Hong Kong's Economic Integration and Business
Cycle Synchronisation with Mainland China and the US." Hong Kong Monetary
Authority working paper 0611.
Gerlach, S &, W. Peng (2006), "Output gaps and inflation in Mainland China," China
Economic Review, Elsevier, vol. 17(2), pages 210-225.
Gerlach, S. and M. Yiu (2004), “Estimating Output Gaps in Asia: A Cross-Country Study,”
Journal of the Japanese and International Economies, 18, 115-136.
Granger, C.W.J. (1983), “Forecasting white noise. In: Zellner, A. (Ed.) Applied Time Series
Analysis of Economic Data”, U.S. Government Printing Office.
Gregory, A. W., A. C. Head, and J. Raynauld (1997), "Measuring World Business Cycles,"
International Economic Review, Department of Economics, University of Pennsylvania
and Osaka University Institute of Social and Economic Research Association, vol.
38(3), pages 677-701.
Hamilton, J. (2008), “Understanding Crude Oil Price”. NBER working Papar. No. 14492.
Ha, J, K Fan and C Shu (2003), "The Causes of Inflation and Deflation in Mainland China",
Hong Kong Monetary Authority Quarterly Bulletin, September 2003. 23-31.
134
Harvey, A. C. (1985), “Trends and Cycles in Macroeconomic Time Series.” Journal of
Business and Economic Statistics, Vol. 3, pp.216-27.
Harvey, A.C. (1993), "Structural time series models", Handbook of Statistics, Elsevier
Science Publishers, Amsterdam, Vol. 11.
Harvey, A. C. and R. G. Pierse (1984) “Estimating Missing Observations in Time Series”,
Journal of American Statistical Association 79, pp.125-131.
Harvey A.C (1989), Forecasting, Structural time Series models and the Kalman Filter,
Cambridge university Press.
Harvey A.C, and, C. Chung (2000), “Estimating the Underlying Change in Unemployment in
the UK” Journal of the Royal Statistical Society, A 163, 303-339.
Heathcote J. and F. Perri (2003), “Why has the U.S. Economy Become Less correlated With
the Rest of the World?” American Economic Review 93(2): 63-69.
Heytens, P. and H. Zebregs (2003), “How Fast Can China Grow?” China: Competing in the
Global Economy, International Monetary Fund, Washington DC.
Hodrick, R. and E. C. Prescott (1997), "Postwar U.S. Business Cycles: An Empirical
Investigation." Journal of Money, Credit, and Banking, Vol.29 No.1,pp. 1-16.
Holz, C. (2005), “The institutional arrangement for the production of statistics, OECD—
China governance project.” OECD statistics Working Paper STD.
Holz, C.(2004), "Deconstructing China's GDP statistics," China Economic Review, Elsevier,
vol. 15(2), pages 164-202.
135
Holz C. (2006), "China's reform period economic growth: how reliable are angus maddison's
estimates?," Review of Income and Wealth, Blackwell Publishing, vol. 52(1), pages 85-
119, 03.
Huang, Y.(2011), “ Missinterpreting China’s Economy” Wall Street Journal, August 25,
2011 http://www.carnegieendowment.org/2011/08/25/misinterpreting-china-s-
economy/4v5q.
Johansson, A. C., (2009), "Is U.S. Money Causing China's Output? " Working Paper Series
2009-6, China Economic Research Center, Stockholm School of Economics .
Keidel, A. (2001), "China's GDP expenditure accounts," China Economic Review, Elsevier,
vol. 12(4), pages 355-367.
Keidel, A. (2008), “China’s Economic Fluctuations: Implications for its Rural Economy”
Carnegie Endowment Report.
Klein, L. R. and S. Ozmucur (2003), “ The estimation of China’s Economic growth rate”
Journal of Economic and Social Measurement, Vol. 28, Number 4, 2002/2003 , PP.
187-202.
King, R. G. and S. T. Rebelo (1993), “Low frequency filtering and real business cycles”
Journal of Economic Dynamics and control 17 (1993) 207-231
Koopman, S.J, A.C. Harvey, J.A. Doornik, and N. Shephard (2007), STAMP 8.0 Structural
Time Series Analysis Modeller and Predictor, London: Timberlake Consultants Ltd.
Kose, M. A., C. Otrok, and C. H. Whiteman (2003), “International Business Cycles: World,
Region, and Country-Specific Factors.” American Economic Review 93(4): 1216-1239.
136
Laurenceson, J. & D. Rodgers (2010), "China's macroeconomic volatility -- How important
is the business cycle?" China Economic Review, Elsevier, vol. 21(2), pages 324-333,
June.
Lequiller, F. and Blades, D. (2006), “Understanding National Accounts” OECD Publishing,
France.
Litterman, R. B.(1983), “A random walk, Markov model for the distribution of time series”,
Journal of Business and Economic Statistics 1, pp. 169-173.
Liu, J.Q. and T.G Zheng (2008), “Recognition of Phases of China s Business Cycle and
Forecasting for the Growth Trend”, China Industrial Economics, 2008(1), 32-39. (In
Chinese) .
Liu, S., Zhang, X. and P. Zhang (2005), “ Smoothing the Business cycles at a Moderately
High Altitude” Economic Study, Vol 11, 2005 (In Chinese).
Liu, S.C. (2009), China: 60 years Economic Growth and Fluctuations, Social Science
Literature Press ISBN:9787509708835 (In Chinese).
Lowe, P. (2010), “The Development of Asia: Risk and Returns for Australia” Speech on
Natstats 2010 Conference, Sydney, http://www.rba.gov.au/speeches/2010/sp-ag-
160910.html.
Maddison, A. (1998), “Chinese Economic Performance in the Long Run.” Paris:
Development Centre of the Organization for Economic Co-operation and Development.
Mitra, S. and. Sinclair, T. M. forthcoming, “Output Fluctuations in the G-7: An Unobserved
Components Approach.” Macroeconomic Dynamics.
137
Moauro, F. and, G. Savio (2005) “Temporal disaggregation using multivariate structural time
series models”. The Econometrics Journal, 8: 214–234.
Morley, J. C., C. R. Nelson and E. Zivot (2003), "Why Are the Beveridge-Nelson and
Unobserved-Components Decompositions of GDP So Different?" The Review of
Economics and Statistics 85(2): 235-243.
Morley, J. C. and J. Piger. (2009), “The Asymmetric Business Cycle.” Review of Economics
and Statistics. Forthcoming.
Morley, J. C. (2007), “The Slow Adjustment of Aggregate Consumption to Permanent
Income.” Journal of Money, Credit, and Banking, Vol. 39 pp. 615-638.
Murray, C. J. (2003), “Cyclical Properties of Baxter-King Filtered Time Series.” The Review
of Economics and Statistics, Vol.85, pp. 472-476.
National Bureau of Statistics of China (2008), “Historical Data on China Quarterly GDP
Estimator 1992-2005”. 2nd ed. China Statistical publishing house, Beijing.
Nagakura, D. (2007), “Inference on the Correlation between Permanent and Transitory
Shocks for Unidentified Unobserved Components Models.” SSRN Working Paper
http://ssrn.com/abstract=981646.
Nagakura, D. (2008), "How Are Shocks to Trend and Cycle Correlated? A Simple
Methodology for Unidentified Unobserved Components Models.” IMES Discussion
Paper Series, 2008-E-24.
OECD, (2006), “China’s National Accounts” .
Proietti,T. and F. Moauro (2005), “ Temporal Disaggregation and Seasonal adjustment”,
Eurostat working paper , European Community ISSN 1725-4825.
138
Proietti, T. (2006), “Temporal disaggregation by state space methods: dynamic regression
methods revisitied”, Econometrics Journal 9, pp. 357-372.
Park, G. (1996), “The Role of Detrending Metods in a Model of Real Business Cycles”
Journal of Macroeconomics, Vol. 18 No. 3 pp. 479-501.
Peng, S. and J. Chen (2009), “Study on China-US Business Cycles synchronization: based on
multi macro-indicators” The world economy, Vol 2. 2009 (In Chinese).
Qing, W., Jin, Y. and Pu, Y.( 2002), “Analysis of the correlation of China’s Business cycles
and the Global Business Cycles” In “2002 Chinese Economic Analysis and Forecast”,
Social Science Literature Publishing House, Beijing.
Rawski T. G.(1976), “ On the reliability of Chinese economic data” Journal of Development
Studies, 1743-9140, Volume 12, Issue 4, 1976, Pages 438 – 441.
Rawski T. G and W. Xiao (2001), “Roundtable on Chinese Economic Statistics
Introduction”” China Economic Review, 12(2001) 298-302.
Rawski, T. G. (2001), "What is happening to China's GDP statistics?" China Economic
Review, Vol. 12 No.4pp.: 347-354.
Ren, Zhiixang and Yuhua, Song (2004), “Influence of External Shocks on China’s Economic
Growth Cycles” Economists. (In Chinese)
Samuel, W. and Y. Sun (2009), “ECCU Business Cycles: Impact of the U.S.” IMF Working
Paper. WP/09/71,
Santos S., J. M. C. & F. N., Cardoso (2001), "The Chow-Lin method using dynamic models,"
Economic Modeling, Elsevier, vol. 18(2), pages 269-280, April.
139
Sato, K. and Zhang, Z. (2006), “Real output co-movements in East Asia: Any evidence for a
monetary union?” World Economy, Vol. 29 No.12, pp. 1671-1689.
Scheibe, J. (2003), “The Chinese Output Gap During the Reform Period 1978-2002,”
Department of Economics Discussion Paper No. 179, University of Oxford, December.
Scheibe J. and D. Vines (2005), “ A Phillips curve for China” Centre for Applied
Macroeconomic analysis, the Australian National University, working paper 2/2005.
Shin, K. and Sohn, C. H. (2006), “Trade and financial integration in East Asia: Effects on
Co-movements,” World Economy, Vol.29 No.12pp. 1649-1669.
Siklos, P.. & Yang Zhang (2010), "Identifying The Shocks Driving Inflation In China,"
Pacific Economic Review, Wiley Blackwell, vol. 15(2), pages 204-223, 05.
Sims, C. (1974), “Seasonality in regression” Journal of the American Statistical Association
69, 618-626.
Sinclair, T. M. (2009),“The Relationships between Permanent and Transitory Movements in
U.S. Output and the Unemployment Rate."” Journal of Money, Credit and Banking,
Vol.41No.2-3, pp. 529-542.
Stram, D. O., and W. S. Wei (1986), “A Methodological Note on the Disaggregation of Time
Series Totals”, Journal of Time Series Analysis 7, 293-302.
Stram, D.O. and W.S. Wei, (1990), “Disaggregation of time series models,” Journal of the
Royal Statistical Society, Series B52, pp. 453-467.
Stock, J. H. and Watson, M. W. (1988), "Variable Trends in Economic Time Series." Journal
of Economic Perspectives, Vol.2 No.3, pp. 147-174.
140
Tasdemirm, M. (2008), “Estimating Monthly GDP for Turkey by State-Space Approach”
International Research Journal of Finance and Economics, Issue 17.
Trimbur, T. and Harvey, A. C. (2003), “General Model-Based Filters for Extracting Cycles
and Trends in Economic Time Series.''”Review of Economics and Statistics, Vol.85,pp.
244-255.
Watson, M. W. (1986), “Univariate detrending methods with stochastic trends” Journal of
Monetary Economics 18 (1986) 49-75.
Wang X, and L. Meng (2001). “A reevaluation of China’s economic growth.” China
Economic Review 12: 338–346.
Wei, W. S., and D. O. Stram, (1990), “Disaggregation of Time Series Models”, Journal of
the Royal Statistical Society B 52, PP 453-467.
Woo, W. T. (1998), “Chinese Economic Growth: Sources and Prospects”, in Michel
Fouquin and Francoise Lemoine (ed.), The Chinese Economy, Economica, London.
Wu, H. X. (2000), “China's GDP level and growth performance: alternative estimates and the
implications” Review of Income and Wealth, Series 46, Number 4, December 2000.
Xu, X. (2002), "Study on Some Problems in Estimating China's Gross Domestic Product."
Review of Income and Wealth, Vol.48 No.2 pp. 205-15.
Xu, X, (2004), “China’s gross domestic product estimation” China Economics Review, 15
(2004) 302-322.
Xu, X. (2008), “Some differences in GDP measurements between China’s practice and 1993
SNA guideline” China Economic Review, Vol 19 480-488, Jan. 2008.
141
Xu, X. (2009), “The Establishment, Reform, and Development of China’s System of
National Accounts.” Review of Income and Wealth 55(1):442–65.
Xu, X.( 2010), “China’s Economic Growth and The Inflation Cycles Since Reform And
Opening to the Outside World”. China Economist, Forthcoming
Young, A. (2003), "Gold into Base Metals: Productivity Growth in the People's Republic of
China during the Reform Period," Journal of Political Economy, University of Chicago
Press, vol. 111(6), pages 1220-1261, December.
Zarnowitz, V. and A. Ozyildirim, (2006), “Time series decomposition and measurement of
business cycles, trends and growth cycles” Journal of Monetary Economics, 53 (2006)
1717-1739.
Zheng,Tingguo & Teng, Yujuan & Song, Tao.(2010), "Business Cycle Asymmetry in China:
Evidence from Friedman's Plucking Model," China & World Economy, Institute of
World Economics and Politics, Chinese Academy of Social Sciences, vol. 18(s1), pages
103-120.
Zhang, Yin. & Wan, Guanghua, (2005), "China's Business Cycles: Perspectives from an AD-
AS Model ," Asian Economic Journal, East Asian Economic Association, vol. 19(4),
pages 445-469, December.
Vahid, F. and R. F. Engle (1993), “Common Trends and Common Cycles.” Journal of
Applied Econometrics 8(4): 341-360.
World Economic Forum (2009), “The Global Competitive Report 2009-2010”,
http://www.weforum.org/en/initiatives/gcp/Global%20Competitiveness%20Report/inde
x.htm.
142
Wu, H. X. (2000), “China's GDP level and growth performance: alternative estimates and the
implications.” Review of Income and Wealth, Vol.46 No,4.
Yang, J., Askari, H., Forrer, J. and H. J.Teegen(2004)“US Economic Sanctions Against
China: Who Gets Hurt?.” The World Economy, Vol. 27, No. 7, pp. 1047-1081
Young, A. (2003), "Gold into Base Metals: Productivity Growth in the People's Republic of
China during the Reform Period," Journal of Political Economy, Vol.111 No.6, pp.
1220-1261.
Young, A. H. (1993), “Reliability and Accuracy of the Quarterly Estimates of GDP.” Survey
of Current Business Vol.73 No.10 , pp. 29-43.
Zhang, B. (2006),”Analysis on Synchronization and Transmission Mechanisms of Sino-
US Business Cycles” World Economic Study, Vol. 10 (In Chinese)
Zong, J. (2007), “The economic fluctuations of developed countries on China---Empirical
analysis on G7 and FDI prospective.” Forum of the World Economy and Politics,
Vol.4.
143
Appendix 1-1: Literature review on studies of China’s macro data
quality
The quality of Chinese data did not draw much attention from researchers outside
China until the late 1990s, when China kept growing at exceptionally rapid rates of growth,
averaging over 8% annually. In the early 2000s, heated discussions 74on the quality of
Chinese macro data generated a large number of publications on this issue.
The criticisms75 of China’s official data are based on evidence from alternative GDP
calculations (Maddison 1998, Wu 2000, Young 2003,), and comparison with energy and
transportation consumption data (Rawski 2001). One source of falsifications in the data is
from the local level. For example, local government officials have incentives to report
inflated numbers to meet the targets of five year plans. In the media, people are also
concerned about the quick publication of GDP data, usually only two weeks after the end of
reporting periods. The release of the preliminary national account data for such a big
economy is considered remarkable (Economist 2008)76.
Studies by Rawski (2001) and Maddison (1998) are two of the most influential
publications. Rawski has followed the Chinese data issue since 1976 (Rawski 1976) and is
one of the most cited authors on Chinese economic data problems. Rawski (2001) challenges
the official statistics by checking the quantitative consistency over output, energy use,
employment and price index. Rawski and Xiao (2001) and Wang and Meng (2001) point out
74 Example of the discussion is the collection of papers on Chinese economic statistics in China Economic Review
12(2001), reviewed and summarized by Rawski and Xiao (2001) 75 See Holz (2006) for a detailed survey of recent literature. 76 The most recent official announcement on the timing of revisions of the quarterly data has become more cautious and
leaves more time for the first and final revisions of the number.
144
the possible falsifications at the local level77. Holz (2004) and Keidel (2001) question the
official GDP estimation from the components, especially the household consumption data of
the expenditure accounting approach. Maddison (1998) offers an alternative real GDP for
China from 1952 to 1995, through checking the real growth rates sector by sector. Maddison
gets a 2.5% lower average annual real GDP growth rate than the official growth rate for years
1978-1995. Maddison’s estimates are used in the Penn World Tables (PWT) Version 6 which
was used widely by researchers on cross-country studies. The major differences of
Maddison’s sector growth rates and the official ones are in “other services” and industry.
Maddison uses employment, which has been criticized by Holz (2004) as invalid, as an
alternative indicator for output growth. Holz argues that the assumption of zero labor
productivity growth in Maddison’s estimation is not valid.
Although the Chinese statistical authorities explain most of the questions as lack of
understanding of China’s transitional statistical system and nature of the transitional
economy, they do acknowledge several problems with its GDP statistics (Xu 2002 and Xu
2004). These include lack of tracking and accurate measurements on housing services, fiscal
subsidies and non marketable welfare services provided within economic entities, weakness
in rural small enterprises statistics and livestock products, as well as the possible falsification
at the local level. For example services that used to be provided within state-owned
enterprises or commodities highly subsidized (such as housing and food) during the late
1970s and early 1980s are now mostly evaluated at market values. This can cause
measurement inconsistency problems for the GDP data (Xu 2002).
77 The possibility of a number emanating from the central government mentioned by media reports (Economist 2008) is
considered low in the academic literature. The Chinese commentaries from central government has explicitly recognized local statistics problems (Rawski and Xiao 2001). The NBS has stopped using local data to generate national economic growth data from 1998. (Xu, 1999)
145
The criticism and discussions urged the Chinese statistical authorities to launch a
comprehensive economic census in 2004, after about 10 years since the 1995 tertiary sector
census. The results of the census led to later revisions in the data, sources and methods.
Contrary to most of the studies stating the official growth rates are overstated, the official
real growth rate estimation of year 1995-2004 were revised significantly upward in 2006.
This was based on the information collected from 2004 economic census. The census results
indicate that untracked economic activity, mainly in the services sector, is growing fast and
accounts for a larger share than previously estimated of the economic activity.
Xu (2009), an NBS official, acknowledges several problems still exist in the service
sector, price indices, quarterly GDP estimation and regional GDP estimation. In his other
publication, Xu (2008) lists the current differences between China’s GDP measurements and
the 1993 UNSNA standards. The above problems are considered common in developing and
transitional economies and should not imply that the errors of China’s GDP estimation are
larger than other developing economies.
Although the media still frequently questions Chinese official data, many researchers
in the academic studies in recent years find that Chinese GDP data problems are not unique
to China and there is no robust evidence for concluding there is systematic data manipulation
or data falsification (Holz 2005 and 2006, Chow 2006). The most recent evidence on the
reliability of Chinese data is from Curtis and Mark (2010), who find that China’s economic
fluctuations have not deviated much from the standard business cycle models using official
aggregate and provincial level data, which means the data are not inconsistent with economic
theory.
146
With the establishment of a more scientific statistical system including regular
surveys and better financial statement reports for enterprises, the quality of Chinese macro
statistics continue to improve. Manipulating statistics to meet political objectives, the most
common concern, is more difficult at least at the national level. Xu Gao from the World Bank
in the official blog 78provides evidence of the consistency of data from different government
institutes in recent years. After working with China’s national account statisticians for about
two decades, OECD (2006) was convinced that although there are weaknesses in the system,
data manipulation does not happen at national level.
One of the big concerns for China’s GDP growth is why the Chinese economy grows
mostly near the government target? One specific feature of Chinese economy should be
noted: although pursuing market-oriented reforms for more than 30 years, the level of
government control of the economy is still relatively high. The political system and the
government institutional structure also largely ensure that government investment and
expenses, economic activities of state owned or controlled enterprises follow the goal of
economic growth set by the government. The close to the target economic growth can be
result of these government influenced economic activities.
Caution about the data
Although I agree that the official data are the best available and not systematically
biased, Cautions must be taken when using the official output data based on the following
considerations, in addition to the problems acknowledged by the authorities mentioned
above:
78 http://blogs.worldbank.org/eastasiapacific/are-chinese-statistics-manipulated
147
First, China’s statistical system and national account data compilation system are still
in a transition from a pure reporting system for a centrally planned economy to a system that
follows international standards79. Transition and reforms in the economy and the statistical
system may result in problems of consistency and comparability of data overtime. Note that
the issue has been partly mitigated by more frequent survey and economic census.
Second, the service sector was very limited in breadth and depth before 1978,
whereas it accounts for more than 40% of total output in 2010. This problem of accurate
service sector data sources is highly focused in the national economic surveys. Still, there are
a lot of weaknesses and emerging problems in this sector. For example informal economic
activities may still cause more potential data problems for this sector.
Third, for real GDP data, the reporting system for state owned sectors still has MPS
features and relies largely on enterprise reporting. For example, output in constant prices is
reported by many enterprises at equal rates of nominal and real change over years due to the
difficulty of calculation with correct deflators80(Woo 1998, Xu 2004). Although modern
statistical methods such as regular surveys are being established, inconsistency in valuation
of real output may still exist.
Fourth, although there is no evidence of systematic manipulation of national level
data, the political events and the communist party administrations may cause some
irregularities in the data. It is always necessary to check irregularities of the data in the
model. Significant irregularities in the real GDP series have not been found in the temporal
79 For detailed prescription and comments on the transformation of institutional organizations and data completion
methods in China, see Holz (2006). See Xu 2002 and Xu 2004 for the NBS explanation about data compilation methodology and officially recognized problems. The system for the planned economy was Soviet Material Product System and the standard China’s statistics following now is Standard National Accounts (SNA)
80 Bosworth and Collins (2007) test this reporting problem by using alternative price index constructed by Young (2003) and find the problem may affect only on secondary (industry) sector.
148
disaggregation models, partially due to the stochastic specifications of the components of the
models, and also might due to the “Gradualism” of China’s economic transition.
149
Appendix1-2. the unobserved components decomposion model
As discussed in the data disaggregation part, the measurement equation of the UC models
is:
ttt cy += τ
Where tτ is the unobserved trend component and tc is the unobserved cycle
component.
The data disaggregation estimation shows that the variation of drift term is
insignificant and close to zero, thus here I follow Harvey (1989) and Watson (1986) and
assume a constant drift term for the trend, the model’s transition equations are specified as81:
ttt ηµττ ++= −1
),0( ~ , ct2211 cctttt QNiidccc ηηϕϕ ++= −−
For both models, � ~iid N�0, σ��); and � ~iid N�0, σ�
�);
The correlations between trend and cycle residuals, interpreted as shocks or
innovations to trend and cycle respectively, are assumed as zero82.
81 While the Clark (1987) model specification is:
tttt ηµττ ++= −− 11 and ttt υµµ += −1 c � ��c �� � ��c �� � η
� , η
� ~iid N�0, Qητ)
i.e. the drift of the trend is a random walk. 82 Morley et al. 2003 introduce correlation between permanent and transitory shocks in to the model, which is an
important release on the restrictions of the model. In the authors’ papers on the relationship of China’s economic fluctuations with the US and the aggregate output of developed countries (coauthored with Tara Sinclair), external series, such as the US real output, developed countries real output, oil price and global volume of trade, are added in a bivariate UC model with correlated cross components shocks to help the identification of the correlation of the permanent and transitory shocks for China’s real output. In this paper, to evaluate the contribution of MUC temporal disaggregated data I focus on apply the MUC data to the most commonly used univariate analytic method and compare the results of MUC data with the literature.
150
Appendix1-3. Standard bivariate Blanchard-Quah model and
decomposition
First, an unrestricted VAR model is formed as:
t
p
iitit eYY ++= ∑
=−
10 φφ , VeEe tt ='
∆
∆=
t
tt
yY
π,
=
t
ytt e
ee
π
Where ty∆ is the first difference of log seasonal adjusted real GDP and tπ∆ is the first
difference of the log inflation rate.
The parameters Π , residuals te and the variance-covariance matrix V can be obtained
by the OLS estimation of the unrestricted VAR. The structural VAR then can be set up as:
tt BuAe =
=
2221
1211
BB
BBB
,
=
d
s
tu
uu
,
Where su and
du are the assumed orthogonal or uncorrelated shocks, in this exercise
the supply shocks and demand shocks (they also can be monetary shocks or external shocks,
based on the endogenous variables in the VAR). Thus, IuEu tt ='
and VBB =' . The
variances of the demand and supply shocks are normalized to one.
151
To recover the two different shocks, we need to indentify B. There are four unknowns
in B, While only three restrictions (by assuming the structural shocks are uncorrelated and
the normalization of the variances). Blanchard and Quah (1989) proposed the identification
method on long -run impact of the orthonormal shocks to help identify B. The accumulated
long-run response C to the structural innovations takes the form:
BAC 1−∞
∧
Π= , where ∞
∧
Π is the estimated accumulated responses to the reduced form
shock te .
Imposing a restriction of 012 =C can be explained as the long-run response of the jth
variable to the ith shock is zero83.
The growth in the output gap is given by:
d
i
gapt uBy ∑
∞
=
=∆1
12
In the exercise, I sum over only 40 periods responds (responds over 40 periods are
near zero, thus extending the sum periods won’t make significant difference to the result). To
obtain the levels of output gap, the gap
ty∆ ,are summed up to t and the zero line is closed at
the mean of the gap
ty∆ 84.
83 Note if using Eviews to estimate the SVAR, A is set to be identity matrix. 84
In the literature, the decomposition is called “historical decomposition”
152
Appendix1-4: More results from the GVAR model estimation with
MUC and DdPS data.
Tests of equality of cointegrating coefficients (Table A4-1) shows that the MUC data
estimation results in much stronger effects from all variables than the original DdPS data,
however there are no discrepancies in the direction and significance of the effect.
Table A4-2 presents the short run or error correction model coefficient estimations for
the four endogenous variables in DdPS model for China, based on the two datasets.
Estimations of coefficients of the real GDP based on MUC data are very different from that
based on the DdPS data. The estimated coefficients for equations of other variables do not
change substantially. Real GDP appears to be exogenous based on the MUC data estimation.
Impulse response functions on China’s real GDP (the first column of Figure 4A-1) also
shows difference of the response of China’s real output to shocks from other variables based
on the two datasets. Not only the economy recovers much quicker, but the direction of the
response to shocks from short term interest rates is different85.
85 Further investigation on the data of China’s inflation and interest rates should be done before making any conclusion
on the economic meaning.
153
Table 4A-1. Likelihood ration tests on the equality of cointegrating coefficients
estimated by GVAR modeling for China with MUC data and DdPS original data
Hypothesis
MUC coefficients=
DdPS estimated
coefficients
DdPS coefficients= MUC
estimated coefficients
Variables Chi^2(1) Prob Chi^2(1) Prob
Endogeous variables
China GDP
China inflation 0.538 [0.4633] 0.582 [0.4454]
real exchange rates 0.916 [0.3386] 0.760 [0.3833]
ST interest rate of China 0.583 [0.4453] 0.625 [0.4293]
Exogenous variables
foreign aggregate GDP 2.078 [0.7795] 0.209 [0.6475]
foreign inflation 0.115 [0.7340] 0.329 [0.5662]
foreign real equity price 0.003 [0.9538] 0.014 [0.9074]
foreign ST interest rates 0.519 [0.4715] 0.871 [0.3506]
foreign LT interest rates 0.363 [0.5470] 0.586 [0.4440]
oil price 0.534 [0.4649] 0.796 [0.3724]
TREND 0.087 [0.769] 0.238 [0.6258]
154
Table 4A-2: Short run error correction equation coefficients of GVAR estimated
based on MUC data and DdPS data
MUC data estimates DdPS data estimates
Short run equation for: China Real GDP
Coefficient Std.Error t-value t-prob Coefficient Std.Error t-value t-prob
China real GDP t-1 0.209 0.113 1.850 0.067 0.815 0.064 12.800 0.000
China inflation t-1 0.182 0.150 1.210 0.229 0.152 0.068 2.230 0.029
real exchange rates t-1 0.006 0.026 0.232 0.817 -0.007 0.012 -0.618 0.538
China ST interest rate t-1 -0.025 0.978 -0.025 0.980 0.682 0.449 1.520 0.133
foreign aggregate GDP 0.078 0.281 0.279 0.781 -0.095 0.130 -0.729 0.468
foreign inflation 0.204 0.346 0.588 0.558 -0.074 0.159 -0.465 0.643
foreign real equity price -0.010 0.021 -0.475 0.636 0.004 0.010 0.456 0.650
foreign ST interest rates -0.093 0.660 -0.142 0.888 -0.157 0.294 -0.533 0.595
foreign LT interest rates -0.581 1.517 -0.383 0.702 -0.181 0.698 -0.260 0.796
World oil price 0.001 0.010 0.108 0.914 0.007 0.005 1.440 0.153
ECM term t-1 -0.020 0.014 -1.470 0.145 -0.020 0.006 -3.100 0.003
constant 0.208 0.130 1.610 0.111 0.190 0.060 3.180 0.002
Short run equation for: China inflation
China real GDP t-1 0.136 0.094 1.460 0.149 0.258 0.113 2.280 0.025
China inflation t-1 -0.177 0.125 -1.420 0.160 -0.162 0.121 -1.330 0.186
real exchange rates t-1 -0.031 0.022 -1.430 0.156 -0.032 0.022 -1.500 0.136
China ST interest rate t-1 -1.484 0.813 -1.830 0.071 -1.456 0.799 -1.820 0.072
foreign aggregate GDP 0.274 0.233 1.180 0.243 0.216 0.232 0.933 0.353
foreign inflation 0.634 0.288 2.200 0.030 0.606 0.283 2.140 0.035
foreign real equity price -0.016 0.018 -0.910 0.366 -0.017 0.017 -0.973 0.334
foreign ST interest rates -0.677 0.548 -1.230 0.221 -0.803 0.523 -1.540 0.128
foreign LT interest rates 1.574 1.261 1.250 0.215 1.711 1.240 1.380 0.171
World oil price -0.017 0.008 -2.080 0.041 -0.016 0.008 -1.950 0.054
ECM term t-1 -0.004 0.011 -0.350 0.728 -0.006 0.011 -0.530 0.597
constant 0.033 0.108 0.302 0.763 0.049 0.106 0.462 0.645
Short run equation for: China real exchange rate
China real GDP t-1 -0.352 0.456 -0.772 0.442 -0.248 0.562 -0.441 0.660
China inflation t-1 0.083 0.608 0.136 0.892 0.017 0.602 0.029 0.977
real exchange rates t-1 0.163 0.107 1.530 0.129 0.162 0.107 1.520 0.133
China ST interest rate t-1 -1.968 3.953 -0.498 0.620 -1.932 3.963 -0.488 0.627
foreign aggregate GDP 0.655 1.135 0.578 0.565 0.631 1.149 0.549 0.584
foreign inflation 1.798 1.400 1.280 0.202 1.776 1.405 1.260 0.210
foreign real equity price 0.002 0.086 0.021 0.984 0.006 0.086 0.072 0.943
foreign ST interest rates 2.454 2.668 0.920 0.360 2.896 2.595 1.120 0.268
155
Figure 4A-1 Impulse response function: MUC data and DdPS data
a. Impulse response function, MUC data
b. Impulse response function: DdPS data