This article seeks to make a contribution to the recent discussion of investment

behavior of firm under uncertainty. It conducts an empirical study in the case

of China with a focus on the links between the uncertainty and firm investment w
ith uncertainty measured as the volatility of daily stock market returns. It is
based on a panel data set constructed by the author of all Chinese listed firms
for the time period from 1994 to 2005. This data set is constructed from CCER da
ta and is similar to the COMPUSTAT and Data Stream databases in the USA. Utilizi
ng this data set, I investigate the effects of the uncertainty on firm-level inv
estment. The basic empirical findings are: On one hand, the uncertainty is found
to have a highly significant and positive influence on firm’s investment decisi
ons. In addition, the results of sample splits into manufacturing and non-manufa
cturing and large and small firms show that, the uncertainty has particular stro
ng positive impact on non-manufacturing and large firms. In fact, the real effec
ts of uncertainty are found to be even more pronounced for large firms and they
have greater potential to grow as compared to smaller firms. On the other hand,
as the reported estimation results show that a greater efficiency gain is achiev
ed by applying the System GMM estimation which combines the equations in levels
form, in addition to, differenced transformed equation.

Chapter 1
Introduction
This dissertation is an empirical study of the link between uncertainty and inve
stment in the real economy of China. It seeks to make a partial contribution to
the on going discussions on the role of uncertainty in investment by focusing on
one particular country and Reporting the firm-level evidence on the link betwee
n uncertainty and long-run investment and, also how does uncertainty effect the
investment decisions. Despite the economic significance of listed Chinese firms
and their reliance on the Stock market for funding investment needs and corporat
e growth, there have not been any systematic studies of the real consequences of
uncertainty on firm level Investment. In an effort to bridge this gape, this di
ssertation carries out the first systematic empirical analysis of the uncertaint
y effect on real investment in China. Particularly, it is one of the few studies
using information from transition or developing economies. The empirical result
s are obtained by using an unbalanced panel data set of Chinese companies listed
in Shanghai and Shenzhen stock markets. The time period of our study is from 19
94 to 2005. This data set is constructed from the CCER data base which is the u
nique and most reliable data set, it provides detail and comprehensive informati
on about listed companies’ financial and accounting structure. Appendix provided
in the end contains a detail description of variables used in the study.
Investment decisions of a firm dependent on number of factors and the uncertaint
y about future events is one of them. In uncertain environment investors face di
fficulties for future anticipations because the environment may be highly volati
le and it would be hard to make predictions from the available information. Thus
, uncertainty may be a crucial factor determining investment in developing and t
ransitional economies. The benefits of financial globalization aside, the growin
g interdependence of trade, investment, and finance have arguably increased the
developing countries’ exposure to volatility (Prasad et al., 2002). A salient ma
nifestation of this was the decade of 1990’s when several emerging market econom
ies witnessed a series of macroeconomic crises. These financial crises culminat
ed in an abrupt economic decline, a deeper understanding of such extreme cyclica
l downturns remains a high priority for policy makers and academicians alike. An
y standard macroeconomics textbook would stress the simple insight that investme
nt is the most volatile component of GDP, and is therefore, central to understan
ding the degree and nature of output fluctuations. To the extent that economies
suffering from recent financial crises also experienced a dramatic investment co
llapse, a careful examination of the investment behavior of firms is doubly usef
ul.
A substantial theoretical advance in the literature of investment during the las
t decade pertains to the role of uncertainty as a critical determinant of invest
ment. In particular, the real options literature in investment has underscored t
hat uncertainties facing a firm’s environment can deter investment, as forcefull
y argued by Dixit and Pindyck (1994). These new insights have sparked much theor
etical work on the relation between uncertainty and investment, but a central fe
ature of this work is the considerable ambiguity on the long-run effects of unce
rtainty on investment. Particularly, there is little consensus on whether uncert
ainty deters or promotes investment.
In the face of such theoretical ambiguities, empirical evidence on the links bet
ween uncertainty and investment remains surprisingly limited. Theoretical litera
ture provides an extensive discussion about the relationship between firm invest
ment decisions and uncertainty. Yet, it concludes that a number of factors inclu
ding; model specification and the underlying assumptions, market competition and
investor attitude, the technology involved and the production function may caus
e uncertainty to either raise or lower the investment. Most empirical studies in
vestigating the relationship find that uncertainty reduces investment. Recent re
views on investment have alluded to this dearth of evidence. As Bond and Jenkins
on (1996) note: Agreement on the effects of uncertainty on the level of investme
nt remains elusive. In contrast to the literature on financial constraints, empi
rical work in this area is still notably scarce. Similarly, Pindyck (1991) obser
ved: “The existing literature on these effects of uncertainty and instability is
a largely theoretical one...the gap here between theory and empiricism is distu
rbing.”
Whatever the limited evidence is available in this area relates to developed cou
ntries, with hard econometric evidence notoriously lacking for developing and em
erging market economies. This is a serious omission given that developing countr
ies suffer from greater output fluctuations and, if anything, uncertainty is lik
ely to be a much more important concern in these economies. These economies are
generally more volatile by nature and information problems in these economies ar
e more prevalent due to deficient markets and institutions. Usually less diversi
fied macro and micro activities increase the likelihood of more adverse shocks.
Moreover, information problems in these economies are more prevalent due to defi
cient markets and institutions. Additionally, because of the rapid transition to
a market economy these economies are introducing many policy changes, both at m
acro and micro level. The studies related to the developing countries focus on a
ggregate relationships. For instance, recent evidence on aggregate investment ha
s established that socio-political instability can be an impediment to aggregate
investment (Campos and Nugent, 2003). Similarly, a greater policy uncertainty a
s manifested in a higher volatility of real effective exchange rates, government
expenditures, and nominal money growth hinders private investment in developing
countries (Aizenman and Marion, 1999). Despite this mounting evidence on the ef
fects of various types of uncertainty on aggregate investment levels, the corres
ponding microeconomic evidence at the level of the firm remains insufficient.
A couple of studies using the data from developing or transition economies are P
attillo (1998), Lensink and Sterken (2000) and Bo and Zhang (2002). Yet, investi
gating the investment-uncertainty relationship seems to be particularly relevant
in the context of these economies. The main goal of this research is to fill th
is empirical deficit by gauging the investment effects of uncertainty in an impo
rtant emerging market economy during a period of its economic boom. To be specif
ic, this dissertation establishes a few stylized facts on the relation between u
ncertainty and investment using a panel of Chinese firms during the period (1994
-2005). Given the significant growth of emerging market economies in the last tw
o decades, a special focus on China seems to be relevant. Another practical adva
ntage of solely focusing on China is the availability of company accounting info
rmation for a large number of stock market quoted firms. Importantly, high frequ
ency data on share prices, a critical ingredient in computing a measure of uncer
tainty is now more easily accessible for Chinese firms.
China, world’s 4th biggest economy and the 3rd largest exporter, has an active s
tock market. If compared in terms of stock market capitalization, it is the 2nd
largest stock market in the Asia as Japan being the 1st. Currently; there are mo
re than 1300 publicly listed firms and approximately 67 million individual inves
tors participating in both of the stock markets. Trading has been tremendously r
ational and annual turn over ratio for our sample period (1994-2005) averages ap
prox 500%, the highest in comparison to other major stock markets in the world.
Provided, such stock market volatility reflects some fundamental influences, the
corresponding variation in share prices can be meaningfully used to construct a
n informative measure of uncertainty. With the dramatic economic developments si
nce early 1990’s, the estimation period covered in this study is independently i
nteresting. In contrary to other East Asian economies, China is still witnessing
a period of high economic growth and pronounced lending boom. Almost for two de
cades, it has maintained her annual real income growth above ,,,,,%.
Based on this particular data set, I explore two contingent issues. Firstly, the
uncertainty effect on firm level investment by using a dynamic investment model
. In other words, whether the stock market volatility deters or spurs the invest
ment decision? Secondly, the econometric issue of estimation in the presence of
lagged dependent variables and the explanatory variables which are not strictly
exogenous thus combating with the endogeneity issue.
To investigate the uncertainty impacts on investment, we specify an error correc
tion formulation, introduced in the investment literature by Bean (1981) and dev
eloped more systematically for the microeconomic literature on investment in Bon
d et al. (2003). In order to overcome the endogeneity problem study then applies
both GMM (DIFF) GMM (SYS) techniques.
The basic empirical findings are: On one hand, the uncertainty is found to have
a highly significant and positive influence on firm’s investment decisions. In a
ddition, the results of sample splits into manufacturing and non-manufacturing a
nd large and small firms show that, the uncertainty has particular strong positi
ve impact on non-manufacturing and large firms. In fact, the real effects of unc
ertainty are found to be even more pronounced for large firms and they have grea
ter potential to grow as compared to smaller firms. On the other hand, as the re
ported estimation results show that a greater efficiency gain is achieved by app
lying the System GMM estimation which combines the equations in levels form, in
addition to, differenced transformed equation. One should be careful as it proli
fically generates the number of instrument variables to be used but the validity
of instruments used can be tested by Sargan (1958) test.
These findings could be used to strengthen the view of how investors make judgme
nts in the face of uncertainty. However, to summarize results from this current
study of the firm level investment in China, I find that, the Chinese stock mark
et volatility has an increasingly important role in allocating investable resour
ces; the current investment attitude in China is likely to produce inefficient a
llocation of resources in some particular sectors and to cause the harmful effec
ts on the real economy.
The study provides a number of important policy implications. First, at
the firm level, and too much reliance on the stock market should be discouraged,
so that, firm can make more prudent investment decisions. Second, at least in t
he very near future, the stock market in China is likely to continue its importa
nt role in allocating resources; therefore, sensible regularity measures must be
taken in order to contain the detrimental real effects of an unregulated stock
market expansion. Third, at the aggregate level, too much faith in the positive
economic contribution of the stock market might be misguided.
The rest of this dissertation is organized as follows; A literature review is pr
ovided in Chapter 2. I begin with an overview of the general relationship betwee
n uncertainty and investment. I provide some background information of micro str
ucture, economic reforms, emergence and development of stock market in China. I
also reviewed relative research efforts on understanding the Chinese stock marke
t. In chapter 3, I provided a detail literature about panel data and different
estimation techniques, especially dynamic panel data and System GMM and Diff GMM
estimation techniques. In Chapter 4, I gave an introduction about Investment Mo
dels. In Chapter 5, I first provided the theoretical frame work upon which the s
ubsequent empirical estimations are drawn. Second, I presented econometric speci
fications and discussed the estimation methodology. Then, I provided a detail a
bout key important variable i.e measure of uncertainty. In order to fully examin
e the uncertainty dynamics, econometric estimations are carefully conducted base
d on a panel data set I collected and constructed of Chinese listed firms. At th
e end of Chapter 5, I provide an introduction to this data set and then list som
e descriptive statistics of the key regression variables Detailed econometric an
alysis of main issue are carried out in section 5.? .The last Chapter 6, conclud
es.
Chapter 2: literature Review
The standard neo-classical models of investment have produced rich insights but
they have fared rather poorly in empirical practice. In particular, the standard
determinants of investment emphasized by the neo-classical investment models se
em to have met with limited empirical success. This has inspired a series of imp
ortant extensions to the study of investment behavior, principal among which are
the role of internal finance considerations and uncertainty in influencing inve
stment behavior. A prominent line of investigation in this regard relates to the
connection between uncertainty, irreversibility and investment. It argues that
uncertainty in a firm’s environment, whether it is due to volatility of aggregat
e demand or unpredictability of prices and costs, can impede investment in fixed
capital. The study of investment under uncertainty has been a subject of much a
cademic interest recently. This chapter will briefly survey the key theoretical
and empirical insights of this literature.
In the view of Hartman (1972) and Abel (1983) a greater uncertainty in the situa
tion when the marginal product of capital is convex spurs the investment. A high
er out put variance with fixed mean boosts the expected profitability and thus l
eading to a higher investment. Given the labor to be more flexible than the capi
tal, firms may adjust labor in response to price variations and amending the lab
or-capital ratio and thus changing the marginal product of capital more rapidly
than the price movement. Traditionally, it has been contended that uncertainty
in a firm’s environment is reflected in the behavior of Q, and consequently once
the Q effects are controlled for, uncertainty has no additional effect on inves
tment (Abel, 1983). A major challenge to this view comes from the influential th
eoretical work of Dixit and Pindyck (1994), who demonstrated that uncertainty co
uld be a significant deterrent to a firm’s decision to invest. These effects ari
se in a situation when investments in fixed capital involve considerable sunk co
sts which can only be partially recovered through disinvestment. Implicit in thi
s is the idea of an inherent asymmetry in the adjustment of capital stock: downs
ide adjustment of capital stock is easier than the upward adjustment. Investment
in this sense is considered to be irreversible. Bar-Ilan and Strange (1999) rep
ort that even if the investment is considered to be irreversible the uncertainty
can have a positive effect and it can be the situation when the intensity of in
vestment is considered. Generally the higher the degree of irreversibility the g
reater is the likelihood of a negative relation between investment and uncertain
ty.
Under sunk costs and irreversibilities, firms facing an uncertain environment ca
n be reluctant to invest. In the jargon of real options literature, firms that a
re faced with uncertain rewards have an option to invest now or delay the invest
ment until they receive new information on prices, output, and costs. Under irre
versibility, firms expanding today consider the impact of future uncertainty, in
 particular, the likelihood that they may be stuck with excess capital or low re
turns. An investment opportunity is thus viewed as an option. Investing today el
iminates this option and imposes an additional cost. It is suggested that the ne
t-present value should then be modified to account for this opportunity cost the
cost of forgone option. The effect of this uncertainty is to increase the optio
n value of waiting (opportunity cost); with the possibility that firms facing an
uncertain environment may decide to postpone their investment decisions. The up
shot is that in the presence of irreversibilities, uncertainty is predicted to h
ave adverse effects on capital formation. Extensive reviews of this and related
literature are provided by Pindyck (1991), Hubbard (1994), Serven (1996), and Bo
nd and Van Reenen (2003). In the spirit of Greenwald and Stiglitz (1990), Ghosal
and Loungani (2000) relate the negative link of investment and uncertainty to t
he extent of capital market imperfections.
Recent theoretical innovations, however, tend to cast doubt on some of these ass
ertions. In an important contribution, Caballero (1991) argues that asymmetric a
djustment costs alone are insufficient to produce a negative relationship betwee
n uncertainty and investment. Instead, the critical factors in signing this rela
tionship are the assumptions concerning market structure and returns to scale. M
ore specifically, it is mainly under decreasing returns to scale and imperfect c
ompetition that higher uncertainty deters investment. Thus, even under irreversi
bility higher uncertainty can lead to a rise in investment, provided the market
structure is one of constant returns to scale and perfect competition.
Even if the short-run impact of uncertainty is to reduce investment, the long-ru
n impact remains ambiguous. Theoretically, it is not clear if the relationship b
etween uncertainty and investment should be negatively signed. Depending on the
modeling environment and the specific assumptions, uncertainty may lead to eithe
r a fall or a rise in investment. In principle, it is possible to envision a sit
uation where a higher level of uncertainty is associated with a higher capital s
tock on average (Abel and Eberly, 1999 and Caballero, 1999). Abel and Eberly (19
99) discern between short-run and long run effects of uncertainty when investmen
t is irreversible. The short-run effect is termed as the ‘user cost effect’ and
the long-run effect is termed as the ‘hangover effect’. On the one hand, uncerta
inty increases the user cost of capital, which implies a negative effect of unce
rtainty on investment. On the other hand, when facing unfavorable states of natu
re, e.g. a fall in demand, the firm is not able to disinvest because of irrevers
ibility. Therefore, the hangover effect implies a higher level of capital stock
in the long run. Since the user cost of capital and hangover effect work in oppo
sing directions at the same time the overall effect of uncertainty on investment
is ambiguous. If there is greater capital accumulation during good times, firms
may be stuck with too much capital during bad times, with the result that the a
verage capital may be high despite irreversibilities.
The extant theoretical literature is not only ambiguous in signing the relations
hip between uncertainty and investment; it is also less precise in pinning down
the channels that mediate this relationship. Recent research has, however, under
scored the role of such factors as the degree of irreversibility, market structu
re, and risk aversion. An important explanation for a negative link between unce
rtainty and investment relates to the idea that capital market imperfections cre
ate non-linearities in the investor’s inter-temporal budget constraint. For indi
vidual investors, the presence of a credit ceiling can deter investment in good
times without mitigating the drop in bad times. As a result, higher volatility c
ould be associated with lower investment (Aizenman and Marion, 1999).
The theoretical ambiguities on the link between investment and uncertainty under
line the need for robust empirical evidence. On the empirical side, testing for
the effects of uncertainty on investment is fraught with several challenges, how
ever. For one, the analytical complexity of theoretical models does not easily l
end itself to empirical testing. In particular, there is little consensus on the
appropriate empirical framework for testing the link between uncertainty and in
vestment. Another difficulty relates to constructing a meaningful measure of unc
ertainty that corresponds well to its theoretical construct. Notwithstanding the
se challenges, whatever limited evidence is available in this area tends to poin
t towards a negative relationship between uncertainty and investment.
In evaluating the effects of uncertainty on investment, empirical work in this a
rea has tended to rely on micro data from firm-level establishments. Using data
on seven thousand US manufacturing plants, Caballero et al. (1995) establish the
non-linear pattern of adjustment of capital stock to its optimal level. The ava
ilability of panel data on individual companies has facilitated a more systemati
c exploration of the theoretical predictions on uncertainty. Using micro data on
US corporations and an asset returns-based measure of uncertainty, Leahy and Wh
ited (1996) demonstrate that uncertainty has an adverse effect on investment onl
y in the absence of Tobin’s Q. These results support the assertion that there ar
e no independent effects of uncertainty on investment over and above the effects
of Q. Nilsen and Schiantarelli (2003) employ a discrete hazard model on Norwegi
an data to establish a convex adjustment cost function and indicate the presence
of fixed costs of investment. The adverse impact of firm-level uncertainty on i
nvestment has been similarly established in the Italian context by Guiso and Par
igi (1996).
Using panel data on Ghanian manufacturing firms, Pattillo (1998) estimates inves
tment thresholds and confirms a negative relationship between uncertainty and in
vestment. Employing a simple reduced form erorr-correction model, Bloom, Bond an
d Van Reenen (2001) investigate the link between investment and uncertainty usin
g panel data for UK firms. The main empirical result of this study is that deman
d shocks have a smaller effect on current investment in the presence of higher l
evels of uncertainty. A key methodological limitation of the reduced-form models
is their failure to systematically control for expected future profitability. B
ond and Cummins (2004) offer a possible remedy by offering a better control for
expected future profitability. In particular, they replace average Q with a more
direct measure based on analysts forecasts. Based on annual data for publicly t
raded US companies and a Q-measure based on analysts’ forecasts, Bond and Cummin
s (2004) show that a higher level of uncertainty is associated with lower invest
ment levels.
Economic theories could not succeed in predicting the exact functional relation
between uncertainty and investment. Most of the studies report a negative while
few evidences support a positive relation but there are others who suspect that
it can be nonlinear. Recently Hong and Robert (2005) report the uncertainty and
investment relation to be nonlinear of inverted U type shape by using a panel of
Dutch firms. A lower level of uncertainty up to a threshold level my increase t
he investment but a higher level of uncertainty after the threshold level my red
uce the investment. As is clear from this short review, apart from a handful of
studies, there have been relatively few attempts at empirically assessing the re
lation between uncertainty and investment in developing countries. Against this
background, we will explore a connection between uncertainty and investment usin
g a sample of Chinese firms.

2.3 Economic Reform and Stock Market Development in China

This chapter provides some background information on the stock market developmen
t in China and a description of the microstructure of the Chinese stock market.
In addition, this chapter reviews the related research
2.3.1 Economic Reform
China initiated economic reforms in 1978. The reforms process was slow,
gradual and consisted of different phases, aimed to improving the efficiency of
micro-management institutions by providing some incentive mechanisms. On the int
ernational front, the government adopted an open-door policy to promote trade an
d attract foreign investment. Markets for goods and resources were established a
nd prices were liberalized. More market-oriented reforms were introduced to limi
t the scale of planning in the economy and to increase the degree of autonomy of
the micro-economic units.
Virtually there was no privatization in China before the official commencement o
f reform and opening. However, since the mid 1990s, China has been straying fro
m its centrally planned and state-directed policies. The government officially a
bandoned the concept of planned economy and declared the establishment of a full
-fledged “socialist market economy.” This notion of “socialist market economy” l
egitimized the privatization of the state-owned sector. The government has stepp
ed up its privatization effort and has increasingly implemented extensive market
dominated reform measures. Since the reforms in 1978, the financial system has
been increasingly diversified and open to competition. Enterprises were granted
more autonomy in making production and investment decisions. Nevertheless, evide
nce seems to suggest that this policy shift since the mid 1990s might have backf
ired. Prior to reforms, China’s financial system was a nonbanking practice with
the state-owned and controlled People’s Bank of China (PBC) virtually handling a
ll financial operations. Interest rates were partially liberalized and in 1994 t
he Chinese currency was made convertible on the current account. In 1998, state-
owned banks ceased allocating credits according to the credit quota system set b
y the State Planning Commission.
China’s peaceful and phenomenal transition to the privatization and capitalism k
ept on going and by the end of year 2001, it got the membership of the WTO (Worl
d Trade Organization). It gave a further impetus and pledging more openness to i
nternational trade and further liberalization of the financial sector. With the
membership in WTO the foreign banks were allowed to have business in the mainlan
d China.
2.3.2 Emergence and Development of Stock market in China
The emergence and establishment of the stock market is one of the most important
institutional changes in China’s financial system. The socialistic China decide
d to take the full advantage of all forms of institutions including the stock ma
rket while overcoming its negative effects. This prompted China to ‘march forwar
d’ and ‘make socialist use of stocks and bonds.’ The Shanghai Stock Exchange (SH
SE) was officially established in December 1990 and the Shenzhen Stock Exchange
(SZSE) opened on July 1991. Since then, the government has been increasingly pro
moting the stock market, and as a result, the stock market has been expanding ra
pidly in terms of the number of firms listed, total market capitalization, and t
he turnover ratio. The successful endorsement of the stock market proved to be d
oubly useful. First, the promotion of the stock market, which is often considere
d as an ultimate symbol of capitalism, reflects the overall policy shift to endo
rse more privatization and implement more market-based reforms. Second, the stoc
k market proved to be a handy source of raising funds for SOE’s, an expedient fo
r improving their governance structure and efficiency, and for effectively direc
ting investment incentives.
The growth and development of stock market in China has been very profound and
rapid. The market has enlarged in terms of the firms listed; as compared to the
10 companies in

Table 2.3.1 Main indicators of the Chinese Stock Market (1994-2005)

year
Number of
Firms
Nominal
GDP Total Market
Capitalization Market Capitalization of
Negotiable Shares Total Trading Volume
(Turnover)
Volume % of GDP Volume Volume
1994 287 48197.9 3690 7.89 969 8128
1995 311 60793.7 3474 5.94 938 4036
1996 514 71176.6 9842 14.50 2867 21332
1997 720 78973.0 17529 23.44 5204 30722
1998 825 84402.3 19521 24.52 5745 23527
1999 922 89677.1 26471 31.82 8214 31320
2000 1060 99214.6 48090 53.79 16088 60827
2001 1140 109655.2 43522 45.37 14463 38305
2002 1213 120332.7 38329 37.43 12484 27990
2003 1277 135822.8 42457 36.38 13179 32115
2004 1363 159878.3 37055 27.14 11689 42334
2005 1381 183084.8 32430 17.8 10630
Note: (RMB Yuan: 100 Million)
1990 the total number of listed companies has reached up to 1381 by the end of 2
005. Only the best performing, well functioning and big companies are permitted
by CSRC to be listed. Manufacturing firms constitute a majority of all listed
firms (a little over 60%) as is also evident from our sample . The total market
capitalization has been very encouraging. If compared in terms of market capital
ization then China ranks second in the Asia as Japan being the first (Economist,
2nd June, 2003). It exceeds the size of the Hong Kong market, is twice as large
as that of Taiwan, and it triples the size of that of South Korea. The total ma
rket capitalization in the year 2007 reached up to 21 trillion Yuan (2.77 trilli
on U.S dollars) in comparison to 3 trillion Yuan in the year 2005. The trading h
as been extremely rational but rather highly volatile; the annual turn over rati
o in most of the years has been in excess to the other major stock markets in th
e world. The annual average turn over ratio for our sample period of 1994-2005 i
s approx 500%, the highest in the World.
Table 2.3.2.2 Turnover Rates(Percent) Of Major Stock Exchanges ,(1994-2005)
Stock Exchange 1994 1995 1996 1997 1998 1999 2000 2001
2002 2003 2004 2005 Average
Shanghai 1,135 529 913 702 454 471 493 269
214 251 289 274 500
Shenzhen 583 255 1,350 817 407 424 509 228
198 214 288 316 467
New york 53 59 52 66 70 75 88 87
95 90 90 99 77
Tokyo 25 27 27 33 34 49 59 60 68
83 97 115 56
London 77 78 58 44 47 57 69 84 97
107 117 110 79
Hong Kong 40 37 44 91 62 51 61 44
40 52 58 50 53
Singapore 28 18 14 56 64 75 59 99
54 74 61 48 51
Currently the stock market is hosting more than 1,600 trading seats, the SHSE is
said to have the largest trading floor in the Asian and Pacific area. The stock
market has also allured an enormous amount of investors. In 1992, there were 2
million registered investors. By the end of 1999, the number shot up to 45 milli
on but in the year 2007 the number crossed the barrier of ??????? A high percen
tage of annual turn over ratio can be related to such a great number of market p
articipants who keep the stock for a very short period of time. The stock mania
in China has also caused the market index to witness certain ups and downs in it
s limited time period . The stock fever in China has resulted in unprecedented h
igh levels of market valuations as well as excessive price volatility. Although
sock prices overall have appreciated significantly, the market has already exper
ienced many boom-and-bust cycles. Most of the time the market has been driven th
rough arbitrage, short term and frequent trading which caused the share prices t
o be volatile. Our measure of uncertainty, calculated as the annual standard dev
iation of daily returns reveals the story and as shown in the figure (?????) and
also in the table (????) the average volatility of share prices is highest amon
g all the variables.
2.3.3 Research efforts related to China
Despite of the significant growth and rapid development of Chinese stock market
there have mot not been much studies about its empirical consequences, especiall
y in reference to determining the firm level investment. Internationally there h
ave been only few efforts, perhaps because of no information disclosure or some
other issues and language barrier is also of them. Most of the available researc
h material is in Chinese language, but they have also considered some other conc
erns and do not provide any tangible evidences.
Especially the research efforts regarding the firm level investment are limited
because even in the recent past it was managed by the government. Jefferson, Hu
and Singh (1999) make an impressive effort to study the investment activity of C
hinese industrial firms. They focus on two aspects of firm behaviour. First, the
y study whether investment decisions of Chinese firms are guided by the profit-s
eeking and the value-maximizing principle. Second, they investigate whether the
investment behaviour varies across firms of different ownership types (i.e. sta
te-owned, collectively owned or joint-ventures). Using a data set from the World
Bank (for the time period of 1986-1990) and another data set collected from Chi
nese Statistical Yearbook (for the period from 1986 to 1995), they find evidence
of considerable efficiency in industrial investment in China. They report that
regardless of ownership types, firms make investment decisions in a manner consi
stent with neoclassical profitability and availability of internal funds. In add
ition, interestingly, they find the state –owned enterprises in fact respond to
profitability and cash flow to a larger degree than firms of other ownership typ
es.
Some studies discuss the early developments of stock market and the prevailing s
peculations (Hertz, 1998; Li and Wong, 1997). For instance, Li and Wong (1997) d
escribe the development of stock exchanges and investors trading activities from
the start to 1995. They observe that Chinese investors tend to bet on governmen
t policies instead of making informed judgment based on the fundamentals.
Given a unique kind of ownership structure some studies have been made to invest
igate its effects on the performance of listed companies. Xu and Wang (1997) stu
dy the performance of listed companies over the time period of 1993 to 1995. The
y define ownership structure as both the ownership mix and the ownership concent
ration and, introduce three??????????????? measure of performance, i.e., the mar
ket to book value of equity, returns on assets (ROA; after-tax profits divided b
y book value of equity). While controlling for other determining factors of perf
ormance such as sales, profit growth, and debt-to-asset ratios, they report a si
gnificant positive correlation between firm performance and institutional shares
, while on the other hand there is no significance between A shares and firm pro
fitability. More over, state ownership has a negative effect on firm performance
. So the policies must be implemented to restrict the state shares and to encou
rage the portion of institutional sharers in order to increase the efficiency an
d healthy profits. Similarly, a more recent study by Chen and Lin (2000) investi
gates the effect of ownership mix on firm performance (as measured by earnings p
er share as well ROE) for the year 1997???????. They also reported that there is
a positive link between the proportion of institutional shares and firm perform
ance while tradable shares do not have significant impacts on firm profitability
.
Among the studies analyzing the macroeconomic economic aspects of Chinese stock
market, Wang (2002) employs the cross-country regression methodology of King and
Livine (1993c). His analysis suffers because of short period sample (1993-1999)
, any how he could not report any significant relation between growth indicators
and stock market development. Some studies used the CAPM and APT model to magni
fy the risk and return structure in Chinese stock market (Su and Fleisher, 1998;
Yuenan and Amalia, 2007). Nicolaas et al. (2003) studied the efficiency of the
Chinese stock market and the role of the banks for the period 1992–2001 by using
efficient markets hypothesis. Ying et al. (2004) applied the new risk managemen
t tool, Value at Risk (VaR) methodology, to the stock market in China. These stu
dies however, do not provide any conclusive results as either the Chinese stock
market is significantly efficient or not. Cheng and Oliver (2000) examined empir
ical contemporaneous and causal relationships between trading volume, stock retu
rns and return volatility in China’s four stock exchanges and across these marke
ts.
For instance, Leung and Wong (1997) study the performance of daily stock indices
for the period from September 1992 to August 1994. After conducting tests for a
utocorrelation, cointegration, and causality, they conclude that the Chinese sto
ck market is inefficient in the week form. On the other hand, using the GARCH-M
(1,1) specification, both Song, Liu, and Romilly (1998) and Su & Fleisher (1998)
also study the effect of government policies on stock market volatility. They r
eport that “spikes” of volatility are associated with changes in government regu
lations and therefore they conclude, “Government’s market intervention policies
have affected stock-market volatility in China” (Su and Fleisher, 1998: 239). A
number of studies tried to unveil the speculative characteristic of Chinese stoc
k market and others tried to differentiate between A shares and B shares ( Ferna
ld and Rogers, 2000: Li, 2001; Yang,2001)
Recently (Wang, Y. et al, 2 006) conducted an empirical study to investigate the
link between firm-level investment and stock market valuations in China. They r
eport that the Chinese stock prices contain very little information about future
operating performance over the current fundamentals. A possible reason of defic
ient information is that Substantial proportions of the shares of many listed fi
rms are owned by the state, and they cannot be traded freely. The quality of lis
ted firms is poor, which is reflected by poor profitability and poor corporate g
overnance (Allen et al., 2005a,b). Market manipulation, including trading based
and information-based manipulation, is severe because of the weak legal system (
Chen and Zhou, 2002).
In fact the studies conducted to the date portray only anecdotal and tangential
aspects rather than any fascinating or rigorous views.

Chapter 3: Panel Data

Panel data refers to the pooling of observations on a cross-section of household
s, countries, firms or individuals over several time periods thus providing mult
iple observations on each individual in the sample.
The use of panel data in economic research provides several major benefits over
conventional cross-sectional or time series data sets, see Hsiao (1985,1986). Th
ese include controlling for individual heterogeneity, availability of more infor
mative data and the improved ability to study adjustment dynamics.
A key econometric problem arising in empirical studies is the inability to contr
ol for individual specific effects (unobserved or mismeasured) which may be corr
elated with other included variables in the specification of an economic relatio
nship. As panel data utilizes information both on the intertemporal dynamics and
the individuality of the entities being investigated, it allows more scope to c
ontrol for the effects of missing or unobserved variables.
Through the availability of a large number of data points, in micro data, the de
grees of freedom are increased and the collinearity among explanatory variables
is reduced, leading to an improvement in the efficiency of econometric estimates
. Time-series studies on the other hand, often tend to suffer from multicollinea
rity. Also, as panel data are gathered on micro units, like individuals, firms a
nd households, some variables may be more accurately measured at the micro level
. However biases resulting from aggregation over individuals or firms are elimin
ated.
Apparently stable cross-sectional distributions conceal a number of adjustment d
ynamics in the model which can be better studied using panel data. For example,
in economic models, unemployment spells, job turnover, residential and income mo
bility, panels would be much more suited to the study of the duration of economi
c states of unemployment or poverty. Panel data sets are necessary for the analy
sis of intertemporal relations, overlapping generations and lifecycle models thu
s enabling research on more complicated behavioral models.
Applied econometric research is increasingly making use of panel data sets partl
y because it is becoming more widely available and partly because computer techn
ology makes it relatively easy to do things now that were relatively difficult t
o do some years ago due to difficulties of handling big datasets.
One of the useful applications of these data sets is that they can be used to es
timate dynamic models. Bond (2002), notes that dynamic models are of interest in
a range of economic applications. They can be used for example in Euler equatio
ns for household consumption, adjustment cost models for firm’s factor demand an
d empirical growth models. In some cases allowing for dynamics can be critical i
n recovering consistent estimates of other parameters in the model even when the
coefficients on lagged dependent variables are themselves of primary interest.
A commonly used method of estimating dynamic models particularly in panels where
N is large and T is small is to apply instrumental variable estimators of the ki
nd suggested by Anderson and Hsiao (1981), Arellano and Bond (1991) and Blundell
and Bond(1998). These methods are well known to have desirable asymptotic prope
rties and a number of studies have looked at their small sample properties using
Monte Carlo methods, including Arellano and Bond (1991) and, Blundell and Bond
(1998). Almost without exception, these Monte Carlo studies have used rather art
ificial designs which have not been closely related to any particular empirical
context or empirical applications.
3.1 Benefits and Limitations of panel data
3.1.1 Benefits of using the Panel data.
(1) Controlling for individual heterogeneity. Panel data suggest that indivi
duals, firms, states or countries are heterogeneous. Time-series and cross-secti
on studies not controlling this heterogeneity run the risk of obtaining biased r
esults, e.g. see Moulton (1986, 1987).
(2) Panel data give more informative data, more variability less collinearit
y among the variables, more degrees of freedom and more efficiency. Time-series
studies are plagued with multicollinearity; for example in the case of demand fo
r ……………….. In fact, the variation in the data can be decomposed into variation b
etween different cross sections of various/different sizes and characteristics,
and variation within cross sections/states. The former variation is usually is u
sually bigger. With additional, more information data one can produce more relia
ble parameter estimates. Of course, the same relationship has to hold for each c
ross section/state, i.e. the data have to be poolable and this is a testable ass
umption.
(3) Panel data are better able to study the dynamics of adjustment. Cross-se
ctional distributions that look relatively stable hide a multitude of change. Pa
nel data are also well suited to study the duration of economic states like unem
ployment and poverty , and if these panels are long enough.
(4) Panel data are better able to identify and measure effects that are simp
ly not detectable in pure cross-section or pure time-series data.
(5) Panel data models allow us to construct and test more complicated behavi
or models than purely cross-section or time-series data .For example technical e
fficiency is better studied and modeled with panels (see Baltagi and Griffen, 19
98b ; Cornwell, Schmidt and Sickles, 1990; Kumbhakar and Lovell, 2000; Baltagi,
Griffin and Rich, 1995; Koop and Steel, 2001). Also fewer restrictions can be im
posed in panels on a distributed lag model than in a purely time series study (s
ee Hsiao, 2003).
(6) Micro panel data gathered on individuals, firms and households may be mo
re accurately measured than similar variables measured at the macro level. Biase
s resulting from aggregation over firms individuals may be reduced or eliminate
d (see Blundel, 1988; Klevmarken, 1989). For specific advantages and disadvantag
es of estimating and disadvantages of estimating life cycle models using micro
panel data , see Blundell and Meghir (1990).
(7) Macro panel data on the other hand have a longer time series and unlike
the problem of nonstandard distributions typical of unit roots test in time-seri
es analyst,
3.1.2 Limitations of Panel data
(1) Design and data collection problems. For an extensive discussion of prob
lems that arise in designing panel surveys as well as data collection and data m
anagement issues see Kasprzyk et al. (1989). These include problem of coverage (
incomplete account of the population of interest), non response (respondent not
remembering correctly), frequency of interviewing, interview spacing, reference
period, the use of bounding and time-in-sample bias (see Bailar, 1989).
(2) Distortions of measurement errors. Measurement errors may arise because
of faulty response due to unclear questions, memory errors, deliberate distortio
n of response (e.g. prestige bias), inappropriate informants, misrecording of re
sponses and interviewer effects (see Kalton, Kasprzyk and McMillen, 1989)
(3) Selectivity problems. These include;
a. Self-selectivity. People choose not to work because the reservation wage
is higher than the offered wage. In this case we observe the characteristics of
these individuals but not their wage. Since only their wage is missing, the sam
ple is censored. Information from this truncated sample introduces bias that is
not helped by more data, because of the truncation (see Hausman and Wise, 1979).
b. Nonresponse. This can occur at the initial wave of the panel due to refu
sal to participate, nobody at home, untraced sample unit, and other reasons. Ite
m (or partial) nonresponse occurs when one or more questions are left unanswered
or are found not to provide a useful response.
c. Attrition. While nonresponse occurs also in cross-section studies, it is
a more serious problem in panels because subsequent waves of the panel are stil
l subject to nonresponse. The degree of attrition varies depending on the panel
studies; see Kalton, Kasprzk and McMillen (1989) for several examples. In order
to counter the effects of attrition, rotating panels wave to replenish the sampl
e. A special issue of the Journal of Huaman Resources, Spring 1998, is dedicated
to attrition in longitudinal surveys.
(4) Short time-series dimension. Typical micro panel involve annual data cov
ering a short time span for each individual. This means that asymptotic argument
s rely crucially on the number of individuals tending to infinity. Increasing th
e time span of the panel is not without cost either. In fact, this increases the
chances of attrition and increases the computational difficulty for limited dep
endent variable panel data models.
(5) Cross-section dependence. Macro panels on countries or regions with long
time series that do not account for cross-country dependence may lead to mislea
ding inference.
Panel data is not a panacea and will not solve all the problems that a time seri
es or a cross section study could not handle. Collecting panel data is quite cos
tly, and there is always the question of how often one should interview responde
nts. Pay off from panel data is over long time periods, five years, ten years, o
r even longer.
Griliches (1986) argued about economic data in general, the more we have of it,
the more we demand of it. The economist using panel data or any data for that ma
tter has to know its limitations.

3.2 The static panel data model

The identification of time-series parameters traditionally relied on notions of
stationarity, and uncorrelated shocks. The identification of cross-sectional par
ameters appealed to exogenous instrumental variables and random sampling for ide
ntification.
By combining the time-series and cross-sectional dimensions, panel data sets hav
e enriched the set of possible identification arrangements . A static panel data
model can be used to take account of heterogeneity across individuals and/or ti
me.
 (3.1)
Where the numbers of observations are , and the data is stacked first over tim
e periods for each individual i,
(3.2)
and then stacked again across N individual,
(3.3)
This allows the model to be rewritten as follows:
(3.4)
3.2.1 Fixed effect model
An initial interpretation for the model in (3.1) would be to treat as a dummy
variable to allow for the effects of those omitted variables that are specific t
o individual cross-section units but stay constant over time. Assuming T to be l
arge, the value of the dependent variable for the unit at time , , would depe
nd on K exogenous variables measured by , that differ among individuals in a cr
oss-section at a given point in time and also exhibit variation through time, as
well as on variables that are specific to the unit and that stay constant ove
r time measured by . With fixed T we need the stronger assumption of strict ex
ogeneity for ordinary least squares estimates discussed below to be consistent a
s . Strict exogeneity assumes that is uncorrelated with and can be character
ized by an independently identically distributed random variable with mean zero
and variance . The above mentioned relationships can be depicted by the followin
g model:
(3.5)
where is a vector of constants and is a scalar constant representing the ef
fects of those variables particular to the individual. The error term repres
ents the effects of the omitted variables that are peculiar to both the individu
al units and time periods.
Given the above mentioned properties of , the ordinary-least squares (OLS) estim
ator of (3.5) is the best linear unbiased estimator (BLUE). The OLS estimators o
f and are obtained by minimizing
(3.6)
Taking partial derivatives of M with respect to and setting them equal to zero,
results in
(3.7)
Substituting (3.7) into (3.6) and taking the partial derivative of M with respec
t to
Gives
(3.8)
Where,
(3.9)
and since in this model the observed values of the variable for the coefficient
takes the form of dummy variables, is the least-squares dummy-variable (LSDV
) estimator . The formulation in (3.5) is referred to as the analysis-of-covaria
nce model and so the LSDV estimator is sometimes referred to as the covariance e
stimator. It is unbiased and also consistent when either N or T or both tend to
infinity.
However, if the true model is fixed effects as in (3.5), OLS on the model where
are unobserved or not explicitly included would yield biased and inconsistent
estimates of the regression parameters in the case where .This estimator referr
ed to as the pooled OLS estimator, given by:
(3.10)
is inefficient even if . This is because the error terms in the model would be
correlated over time periods,
(3.11)
This would occur due to the presence of no matter what restrictions are placed
on
This bias can be eliminated by using techniques involving instrumental variables
that are discussed later or by the classic approach, which is to undertake a Wi
thin transformation.
3.2.2 Within Group estimation
The computational procedure for estimating the slope parameters in the LSDV esti
mation does not require that the dummy variable for the individual effects actua
lly be included in the matrix of explanatory variables. The means of time-series
observations can be calculated separately for each cross-sectional unit, the ob
served variables can be transformed by subtracting out the appropriate time-seri
es means, and then the least-squares method applied. Formally this technique is
called the Within transformation.
For the within transformation define,
, (3.12)
Where are the unweighted averages of the time-series for individual . The mode
l is then redefined in terms of deviations from the above mentioned group means
so that
(3.13)
The Within Groups estimator applies this transformation to both sides of the mod
el.
Since the individual specific effect remains constant over time it is eliminated
from the model and the new formulation is as follows:
(3.14)
OLS can now be applied to this transformed model to get the within groups estima
tor which is given by

(3.15)
Notice that is equivalent to from (3.8). will be consistent even if under
strict exogeneity of with respect to is for all s and t. Since all past dis
turbances enter the transformed error term, the orthogonality between the explan
atory variables and all realizations of the error term are particularly importan
t when T is small. the orthogonality between the explanatory variables and all r
ealizations of the error term are particularly important when T is small.
There could be two drawbacks of using this approach. In the extreme case if ther
e is no time variation in the observed variables the estimates of cannot be i
dentified.
With a little time variation will be inaccurate and there is a cost to using
this transformation. Also there may be variation in the error term which is un
accounted for.
An alternative estimator to the Within is the Between Groups estimator which col
lapses the time-variation in the model by averaging to give a cross-section equa
tion as follows.
where (3.16)
The between-group estimator is given by
(3.17)
Where and are the overall means,
and
Like pooled OLS, , the between groups estimator will be biased and inconsisten
t if .
3.2.3 Random effects model
The fixed effects model is a reasonable approach when differences between indivi
dual units can be confidently viewed as parametric shifts of the regression func
tion. This model could be viewed as applying only to cross-sectional units in th
e study. For example, an inter-country comparison may include the full set of co
untries for which it is reasonable to assume that the model has constant individ
ual specific effects. In other settings where the sampled cross-sectional units
are drawn from a large population it would be more appropriate to view the indiv
idual specific effects as randomly distributed . Mundlak (1978a) criticizes this
formulation as it ignores the correlation between the fixed effects and (any o
f the) observed regressors. With the introduction of a random , the loss of
degrees of freedom in the fixed effects model due to the presence of too many p
arameters is avoided only at the risk of introducing bias. Consider a reformulat
ion of the original model from (3.5) as follows:
(3.18)
The component is the random disturbance characterizing the observation and is
constant through time. Since the residuals in (3.18) now consist of two compone
nts it is often referred to as the error components or the error decomposition m
odel. In this model the following assumptions are made about the errors:
(3.19)
Strict exogeneity of the explanatory variables with both the time varying and th
e individual specific shocks is also assumed, and for all t and s.
3.2.4 Generalized Least Squares estimation
Since the errors are defined as they will be serially correlated because of
.In this case the pooled OLS estimator although consistent will not be efficient
and we can use the generalized-least-squares (GLS) estimator which would be the
best-linear unbiased- estimator (BLUE).
As before, it is useful to view the formulation of the model in blocks of obse
rvations.
For these T observations, let
(3.20)
and
(3.21)
Following Greene (2000), for the T observations for unit i, let .Then
(3.22)
where is a identity matrix and is a column vector of ones as defined in
(3.2). Since observations are independent the disturbance covariance matrix for
the full NT observations is given by
(3.23)
where denotes the Kronecker product .
For GLS is required which is given by . Therefore only is needed,
(3.24)
Where,
(3.25)
Then using theta-differencing the model can be transformed to,
(3.26)
Where and are again the averages across T observations as defined in (3.11).
This transformation coincides with the Within transformation in the special case
when . The next step would be to pre-multiply the model with

and similarly for the rows of . For the entire data set generalized least squar
es is computed by the regression of these partial deviations of on the same tr
ansformations of . If is known and theta-differencing is applied to the model
, so that
then

Greene (2000) shows that the GLS estimator is a matrix weighted average of the
Within and Between Groups estimator
(3.27)
Where
(3.28)
There are some extreme cases to consider. If equals one then generalized lea
st squares is the same as ordinary least squares. This occurs if is zero and a
 classical regression model is applicable. If is zero then the estimator is
the dummy variable estimator discussed in the fixed effects model. This arises a
s for (see (3.25).
3.2.5 Feasible Generalized Least Squares
The method considered above is applicable when the variance components are known
.
In applied work the variances will generally be unknown and will need to be esti
mated. Incorporating this, the approach is now called feasible generalized least
squares (FGLS). FGLS requires consistent estimates of and . This can be done
using the residuals from the Within estimation and the Between Groups estimation
.
Using the Within Groups model from (3.13) and rewriting it as deviations from th
e group means the model is
(3.29)
The within residuals are given by
(3.30)
Since,
(3.31)
using the within estimator a consistent estimator of can be calculated,
(3.32)
The degrees of freedom are N opposed to because N observations are used up in
the Within transformation.
Now using the model with the Between Groups transformation
(3.33)
Where
(3.34)
the between residuals are given by
(3.35)
So a consistent estimate of the variance of the between residuals is

Since , can be written as

(3.37)
Given the above an estimate of can now be inferred,
(3.38)
From (3.32) and (3.38) an estimate for in (3.28) can be derived which will al
low feasible generalized least-squares to be used,
(3.39)
The efficiency gains from using feasible GLS get smaller in comparison to the wi
thin estimator as T gets larger. The two coincide when Hence in the large T ca
se is consistent under the same conditions as .In the small T case the GLS es
timator will be biased if is correlated with any of the variables. With T fi
xed is consistent even if , while is consistent even if . However is
efficient relative to when .
The Hausman test can be employed to check for relative robustness of the GLS and
Within Groups estimators. Let

Under

where is the asymptotic variance and h is the Hausman statistic. In the above c
ase if the null hypothesis is rejected the Within estimator is more robust tha
n the GLS estimator. However, since the relative efficiency statement depends he
avily on the iid assumption of the errors the Hausman test is not robust to hete
roskedasticity.
3.2.6 Individual-specific variables
Generalizing model (3.18) to include a vector of individual specific variables t
hat vary across individual units but do not vary over time would give the follow
ing formulation:
(3.40)
where are observed variables. Applying the within transformation
(3.41)
results in the elimination of the individual specific variables and cannot be
estimated. There are, however, three cases in which consistent estimates of c
an be obtained.
Case 1: The first case assumes that all the observed regressors, and are unc
orrelated with and strictly exogenous with respect to .
(3.42)
Here the GLS can be applied as before, that is, OLS applied to equations after t
heta- differencing, using a consistent estimate of . This method provides estima
tes of coefficients in model (3.40) above but requires very strong assumptions.
Case 2: In the second case assume that all of the regressors are still uncorre
lated with but some or all of the explanatory variables may be correlated wit
h . We maintain the assumption
(3.43)
Consistent estimates of the coefficients can be obtained following a two-step me
thod.
In the first instance is estimated which is consistent regardless of any corre
lation between and .Then the following cross section regression
(3.44)
can be used where the unknown is replaced by the consistent as obtained in
the first step. Applying OLS to this equation would give an estimate for which
is consistent as long as the regressors are uncorrelated with .
Case3: The third case is the most general case where some but not all of the v
ariables and some but not all of the regressors are correlated with .If there
are enough variables that are uncorrelated with these can be used as instru
ments for the variables that are correlated with , and two-stage-least-square
s (2SLS) estimation can be employed. In the first stage the and the variabl
es need to be partitioned so that:
(3.45)
Where the variables are uncorrelated with ,and the variables are correlated
with . Similarly
(3.46)
Where variables are uncorrelated with , the variables are correlated with
and . After partitioning the variables in this way, the first two steps from t
he second case where is obtained and substituted into the cross section equati
on (3.41), can be applied. Then 2SLS can be employed to estimate using as in
struments for . The condition for identification in this case is that is con
sistent for when this condition is satisfied. However, in the over identified
case when is not efficient. More efficient estimators for along these lin
es were developed by Hausman and Taylor (1981) and Amemiya and MaCurdy (1986).
3.3 The dynamic panel data model
Many economic applications are dynamic in character and one of the advantages of
panel data, mentioned earlier is that it allows the researcher to better unders
tand the slow adjustment process underlying such relationships. These interactio
ns in panel data models are often characterized by the presence of a lagged depe
ndent variable among the regressors,
(3.47)
where follows a one-way error component model
(3.48)
and it is assumed that are observable. In the above formulated model
and are independent of each other and among themselves. This regression is c
haracterized by two sources of persistence over time. Autocorrelation due to the
presence of a lagged dependent variable among the individuals. This result’s in
the pooled OLS estimator being biased and inconsistent even if the are not se
rially correlated. The Within transformation removes the but is still biased d
ue to the correlation between and
Where
To show the effects on the standard estimators a simple first-order autoregressi
ve
(AR1) process formulated without the explanatory variables can be used:

(3.49)
where as in (3.48)
3.3.1 Ordinary Least Squares estimation
In the static case in which all the explanatory variables are strictly exogenous
and are uncorrelated with the effects, the error component structure can be ign
ored and OLS applied to the model. The OLS estimator, although less efficient th
an GLS, is still unbiased and consistent. This is not true for the dynamic model
formulated in (3.49).
In this case cannot be strictly exogenous with respect to the since Also s
ince is a function of it immediately follows that will also be a functi
on of . Therefore, is correlated with the error term, unless which would o
nly occur in the case of no unobserved heterogeneity. So, the correlation betwee
n the lagged dependent variable and individual-specific effects seriously biases
the OLS estimator.

(3.50)
The asymptotic bias of the OLS estimator is given by the probability limit of L*
in
(3.50).The probability limits as in Hsiao (2003) for the numerator and denominat
or of
L* are

where and is zero if are assumed to be arbitrary constant, and positive if

are assumed to be generated by the same process as any other . It is evident
that the asymptotic limit of the OLS estimator depends on the two moments, a
nd . Given that the initial values are bounded the
OLS method overestimates the true autocorrelation coefficient when N or T or b
oth tend to infinity. This overestimation is more pronounced the greater the var
iance of the individual effects,
Under the particular assumption of random correlated initial observations as in
Sevestre and Trognon (1985) the asymptotic bias of is given by
(3.51)
(3.52)
As emphasized in (Trognon 1978) the bias is an increasing function of , at an i
ncreasing rate if , constant if and decreasing if .
Hsiao (2003) reports that, the addition of exogenous variables to a first-order
autoregressive process does not alter the direction of the bias in the OLS estim
ator of the lagged dependent variable, although its magnitude is reduced. So the
coefficient on the lagged dependent variable remains upward biased and the coef
ficients of the exogenous variables are biased towards zero.
3.3.2 Within Groups estimation
Consider again the simple AR (1) case in (3.49). By performing a Within transfor
mation the model can be rewritten as
(3.53)
can be obtained by applying OLS to the transformed model:
(3.54)
The Within Groups estimator exists if the denominator of P* is nonzero and is
consistent if the numerator of P* converges to zero. Let
By continuous substitution equation [2.49] can be rewritten as
(3.55)
and the sum of over all t is given by (see Hsiao [2003])

(3.56)
Since is independently and identically distributed and is uncorrelated with,
using (3.56) Hsiao (2003) shows that as N tends to infinity,
(3.57)
and since are random drawings from a normal distribution when T is fixed, a
s N can be replaced by expectations, E. Thus obtaining

(3.58)
The denominator of P*, B can be shown to converge to the following. This result
is derived from Nickell (1981),
(3.59)
As T tends to infinity A converges to zero and B converges to a nonzero constant
Therefore from (3.54) is a consistent estimate of a. If T is fixed then A is
a nonzero constant and is an inconsistent estimate no matter how large
N is. The asymptotic bias is
(3.60)
The bias for is due to the elimination of the unknown individual specific effe
cts from each observation, which results in a correlation of order between th
e explanatory variable and the residuals in the model after applying the Within
transformation. This is a reflection of the fact that is not strictly exogeno
us with respect to the
For reasonably large values of T, Nickell (1981) shows that
(3.61)
Therefore the bias goes to zero as . While for small values of T
(3.62)
and
(3.63)
This shows that not only is the bias always negative if it gets larger as T be
comes smaller and it does not go to zero even if goes to zero
It is useful that we can sign the respective biases for the pooled OLS estimator
and the Within Groups estimator of . We can contrast the two cases and note
that the pooled
OLS and Within estimators are biased in opposite directions and give an upper an
d lower bound for . Thus, we would expect that a consistent estimator will lie s
omewhere between the OLS and Within Groups estimates, or at least not be signifi
cantly higher or lower than or a respectively.
The random effects GLS estimator is also biased in a dynamic panel data model. T
his is because after theta-differencing, will be correlated with When T tend
s to infinity the GLS estimator becomes equivalent to the Within estimator. Then
the GLS estimator is consistent when both N and T tend to infinity. When T is f
ixed and N tends to infinity the asymptotic bias of the GLS estimator is positiv
e, which shows the positive correlation between and .
3.3.3 Instrumental Variables estimation
An alternative transformation that eliminates the individual effects is the firs
t difference transformation. Consider again a simple first-order autoregressive
AR (1) process as in (3.49) formulated as follows
(3.64)
where again as in [3.48] but here we assume is unobserved. The first diff
erence transformation removes the individual specific effects from the model to
give:
(3.65)
In this case it is easier to deal with the correlation between the predetermined
explanatory variables and the remaining error terms. This is particularly diffe
rent from the Within transformation in that it does not introduce all realizatio
ns of the disturbances in the residuals of the transformed equation. Considering
(3.64) at first sight we have . Applying OLS to this would give a biased and
inefficient estimator. The bias of the OLS estimator in this case is given by
3.3.4 Anderson and Hsiao estimators
Anderson and Hsiao (1981) suggested first differencing the model to get rid of t
he and then using or as instruments which are correlated with and orthogo
nal to .These instruments are not correlated with as long as the are not ser
ially correlated. The instruments can be used in a 2SLS type estimation procedur
e that requires for the levels instrument and for The Anderson and Hsiao
estimates of a for each one of the two cases are given by
(3.66)
(3.67)
Both and are consistent when or or both. Comparing
asymptotic variances of the two estimators Anderson and Hsiao (1981) show that i
f there is prior belief that successive observations are positively correlated t
hen should be used and should be used if successive observations are negativ
ely correlated.
This instrumental variable (IV) estimation method leads to consistent bu
t not necessarily efficient estimates of the parameters in the model because it
does not make use of all the available moment conditions (see Ahn and Schmidt, 1
995) and it does not take into account the differenced structure of the residual
disturbances
Arellano (1989) finds that for simple dynamic error component models the estimat
or that uses differenced instruments has a singularity point and very large vari
ances over a significant range of parameter values. On the contrary, the estimat
or that uses instruments in levels has no singularities and much smaller varianc
es. Arellano and
Bond (1991) proposed a generalized method of moments (GMM) procedure that is mor
e efficient than the Anderson and Hsiao (1982) estimator.
3.3.5 First differenced GMM estimation
Arellano and Bond (1991) discuss that additional instruments can be obtained in
a
dynamic panel data model by utilizing the orthogonality conditions that exist be
tween lagged values of and the disturbances This can be illustrated by consi
dering the simple dynamic AR(1) model from (3.64). The basic assumptions of this
model are as follows:
(3.68)
 (3.69)
(3.70)
It assumes that the error terms have lack of serial correlation but not necessar
ily independence over time. With these assumptions, values of y lagged two perio
ds or more are valid instruments in the equation in first differences (3.65). Th
e first differenced equation for t = 3 is
(3.71)
Here is a valid instrument since it is correlated with and not correlated wi
th given assumption (3.69). Continuing the same type of analysis for
we have the first differenced equation as follows
(3.72)
In this case, and are valid instruments for , since both are not correla
ted with the error term, Continuing in this way and adding one extra valid ins
trument with each period the set of valid instruments for period T would be
Bond (2002) notes that, since the model is over identified with
and the first differenced error term has a MA (1) with unit root process and
the are serially uncorrelated, 2SLS would still be inefficient even with the
complete set of instruments and homoskedasticity of the error terms, .
Asymptotically efficient estimators can be obtained by using the generalized met
hod of moments (GMM) framework, developed by Hansen (1982). For the dynamic AR (
1) model first differenced GMM estimators were developed by Holtz-Eakin, Newey a
nd Rosen (1988) and Arellano and Bond (1991).
These methods use assumptions (3.68), (3.69) and (3.70) to imply the following m
moment restrictions that are sufficient to identify and estimate for :
(3.73)
and These moment restrictions can be expressed in matrix notation as
(3.74)
where is the matrix given by
(3.75)
and is the vector The GMM estimator based on the moment restrictions minim
izes the quadratic distance for some weight matrix , where is the matr
ix and is the vector This results in the following GMM estimator:
(3.76)
where the and are the vectors given by

and (3.77) and and are the vectors st

acked across individuals in the same way as (see Arellano and Bond (1991)). Dif
ferent values of the weight matrix result in a set of estimators all of which
are consistent as . In contrast to the within estimator this results holds true
for fixed values of T also.
GMM distribution theory implies that a consistent estimator of
(3.78)
where
(3.79)
Based on the assumptions mentioned above, can be replaced with
(3.80)
where are consistent estimates of obtained from
(3.81)
for some consistent estimate of . Now setting
(3.82) gives the optimal GMM estimator with the smallest asymptotic variance in
this class of estimators that are efficient for a given set of moment conditions
.
3.3.6 Two-step GMM estimator
This leads to the introduction of the two-step GMM estimator with
(3.83)
The first step is to calculate a consistent estimator, as in (3.76), use it to
compute the residuals as in (3.80) and use the residuals to determine a new wei
ght matrix as in (3.82). The next step is to apply GMM employing the new weight
matrix in (3.76).
Asymptotic efficiency results for the estimators are true in large samples but d
o not necessarily hold in finite samples. In practice the initial choice of the
weight matrix in the initial phase of GMM2 (two-step GMM estimation) affects the
final result. The small sample properties for a are likely to better if a rea
sonable choice of is used to evaluate . Under the additional assumption of homo
skedasticity of disturbances,
(3.84)
Arellano and Bond (1991) report a sensible choice of in the initial step is
(3.85)
where
(3.86)
This leads to an estimator that is asymptotically as efficient as for the s
pecial case of homoskedasticity. Therefore, is an appropriate choice not only
for initial stage of calculating but also for applied work. Simulation studies
have shown very modest efficiency gains from using the two-step estimation, eve
n in the presence of heteroskedasticity. (see Arellano and Bond (1991), Blundell
and Bond (1998) and Blundell, Bond and Windmeijer (2000)). Also the dependence
of the two step weight matrix,
(3.87) on estimated parameters makes the usual
asymptotic distribution less reliable for . It has been shown in more si
mulation studies that asymptotic standard errors tend to be too small for in
sample sizes where the equivalent tests based on the one-step estimator are quit
e accurate (see Bond and Windmeijer (2002)). It has been shown by Windmeijer (20
00) that the downward bias of the asymptotic standard errors of the two-step GMM
estimator is due to the presence of estimated parameters in the weight matrix.
This extra variation in for small samples can be calculated and Windmeijer (20
00) also provides a finite sample corrected estimate of this variance.
These methods can be extended to higher order autoregressive models and to model
s with limited moving-average serial correlation in the disturbances given a mi
nimum value of T required to identify the parameters. For example in the case wh
ere is an MA(2) process (results if disturbance term is MA(1)) would not b
e a valid instrument in the first differenced case but and longer lags still r
emain in the set of available instruments, and can be identified using this se
t provided The GMM estimators introduced are for the case where first differen
cing is used to eliminate the individual specific effect while not introducing d
isturbances for periods earlier than into the transformed error terms. There e
xists a wide class of alternative transformations that can be used to achieve th
e same result. Arellano and Bover (1995) show that optimal estimators are invari
ant to the transformations used to remove time-invariant individual effects.
3.3.7 Ahn and Schmidt conditions
Ahn and Schmidt (1995) show that under the standard assumptions used in a dynami
c panel data model, (3.68), (3.69) and (3.70), there are additional moment condi
tions that are ignored by the estimators suggested by Anderson and Hsiao (1981),
Holtz- Eakin, Newey and Rosen (1988) and Arellano and Bond (1991). They find th
e following non-linear moment conditions which are not implied by the linear
moment conditions from (3.73).
(3.88)
These moment conditions are expected to improve efficiency. Asymptotic efficienc
y comparisons reported in Ahn and Schmidt (1995) report that these moment condit
ions are particularly informative when is close to unity and/or when is high.
There are still more moment conditions that can be added to the set of linear an
d nonlinear conditions from (3.73) and (3.88). This work is also attributed to A
hn and Schmidt (1995). These conditions, however, rely on imposing more assumpti
ons.
Including the assumption that the disturbances, , are homoskedastic through time
, results in the addition of useful linear orthogonality restrictions. This is
given by
(3.89) and implies linear orthogonalit
y restrictions
(3.90) and allows a further columns to be added to the
instruments matrix in
(3.75).The additional columns can be written as
(3.91)
The homoskedasticity through time restrictions also makes available one more non
linear moment condition which can be written as
(3.92) where
(see Ahn and Schmidt (1995)). Therefore with the inclusion of homoskedasticity t
hrough time (3.89) to the original collection of (3.68), (3.69) and (3.70), the
complete set of linear and non-linear moment condition is (3.73), (3.88), (3.90)
and (3.92).
3.3.8 System GMM estimation
There are also cases where first differences of can be used as instruments in
levels equations. This was proposed by Arellano and Bover (1995)/Blundel and Bon
d (1998) and became increasingly popular. Consider the additional assumption
(3.94) which is a restriction on the ini
tial conditions process generating . To guarantee
(3.94) we require the initial conditions restriction
(3.95)
which is satisfied through the assumption that the initial conditions satisfy me
an stationarity so the series , have a constant mean for each individual .Thi
s is given by:
(3.96)
There are two immediate implications of (3.94). First, it follows that

(3.97)
because the AR(1) process implies
(3.98)
Second, it also makes the following linear moment conditions available,
(3.99)
where as before.
These conditions imply the quadratic conditions given in (3.88) because
(3.100)
thus making them redundant. Without introducing homoskedasticity, but assuming m
ean stationarity the complete set of moment conditions now is the m conditions f
rom (3.73) and the conditions from (3.99) which can be implemented as a linear
GMM estimator. This forms the basis for a system GMM (GMM-SYS) estimator propos
ed by Blundell and Bond (1998). A greater efficiency gain is achieved with more
instruments available now. The GMM-SYS combines the orthogonality conditions in
the first differenced GMM and levels GMM (GMM-LEV) estimators introduced below.
GMM-LEV is based on the following m moment conditions
(3.101)
which are only related to the equation in levels and can be written in vector no
tation as
(3.102)
where is the as follows

(3.103)
and is the vector . Given these moment conditions GMM estimators can then
be computed as before.
The GMM-SYS estimator uses the complete set of linear moment conditions (3.73) a
nd (3.101), and can be set up as a stacked system comprising both the equations
in first differences and the equations in levels where instruments are observ
ed. The moment conditions for the GMM-SYS can be given as follows
(3.104)
(3.105)
where (see Blundell, Bond and Windmeijer (2000)).The moment conditions in vecto
r notations are given by
(3.106)
Where
(3.107)
(3.108)
where is the same as defined in (3.75) and is the non-redundant subset of
.
Blundell, Bond and Windmeijer (2000) show that the GMM-SYS estimator,
(3.109)
where is based on , p is based on and
(3.110)
is equivalent to the following linear combination
(3.111)
In (3.111) is the first-differenced and is the levels estimator utilizing
only the moment conditions as given in (3.105). is given by
(3.112)
where and are the OLS estimates of the initial stage regression coefficients
of the GMM estimation. The implication of this is that as the level of persiste
nce increases, all the weight for the system estimator comes from the levels m
oment conditions (3.101)11. Blundell and Bond (1998) attribute the bias and the
poor precision of the first-difference GMM estimator found in simulation studies
where increases towards 1, to the problem of weak instruments as described in
Nelson and Startz (1990) and Staiger and Stock (1997).
Ahn and Schmidt (1995) show that in the case where the initial conditions satisf
y (3.97) and the disturbances, are homoskedastic through time, the condition
s from (3.88) and the one condition from (3.92) can be replaced by a set of
moment conditions
(3.113)
which are all linear in the parameter . Hence, all the moment conditions, which
are made available through the initial assumptions given by (3.68), (3.69) and
(3.70), the added assumptions of homoskedasticity through time (3.88) and the in
itial condition restriction (3.97), can now be implemented in a linear GMM estim
ator.
Blundell and Bond (1998) show that the additional moment conditions (3.105) rema
in informative when approaches unity and becomes large, in contrast to the
moment conditions (3.104) for the first differenced case. With a persistent seri
es, thus, the addition of these assumptions gives better results. It has been sh
own that there is an increase in efficiency and a decrease in the finite sample
bias and the gains are greater the smaller the value of N.
3.3.9 Autoregressive distributed lag models
The GMM estimators for simple autoregressive models can be extended to apply to
autoregressive-distributed lag models of the form
(3.114)
Where as before and for simplicity is a scalar. Let
which allows the model in (3.114) to be written as
, (3.115)
where
Different moment conditions are available depending on the assumption regarding
the correlation between and the two components of the error term, and .T
he analysis can be divided into two cases, where
(3.116)
(3.117)
Furthermore the availability of valid instruments would also depend on the assum
ptions about the correlation of with the time varying shock, .
Case A: Consider, first the assumption where is endogenous, so that it is corr
elated with current and past , for , allowing for contemporaneous correlatio
n and feedbacks:
(3.118)
(3.119)
Here is treated symmetrically with the dependent variable , so that lagged v
alues , and longer lags will be valid instrumental variables in the first diffe
renced equations for periods . Thus the new moment condition would be
(3.120)
These will be added to the moment conditions from (3.73) to obtain the new momen
t conditions that can be written as
(3.121)
where is the new instrument matrix as follows
(3.122)
and If, (a) is the matrix of observations on stacked in the usual way an
d (b) is the vector of observations on and (c) is the instrument matr
ix, then
(3.123)
in accordance with the GMM estimation described before and is an arbitrary wei
ght matrix.
Making the assumptions (3.118) and (3.119) stronger so that there is no contempo
raneous correlation

(3.124)
(3.125) and the series is pred
etermined allows the use of as a valid instrument in the first-differenced equ
ation for period t. In this case there will be additional moment conditions gi
ven by
(3.126)
and these would result in the replacement of by the vector in forming each
row of the matrix in
(3.122).
Now making the strongest assumption and letting the process be strictly exogen
ous, so that
(3.127)
the following moment conditions can be added to the original set
(3.128)
This results in the complete time-series being valid instrumental variables in
each of the first-differenced equations. In this case is replaced by the vec
tor in forming each row of the matrix in (3.122).
If the strict exogeneity assumption is true there is an increase in precision, b
y using the instrument matrix that contains the complete time-series of and t
he estimator is asymptotically more efficient. Under the strict exogeneity assum
ption of the variable, there is no need to assume that the disturbances are se
rially uncorrelated, and in order to identify and . In this case the opti
mal GMM2 based on
(3.129)
(3.130)
can be used to estimate coefficients in the model. This coincides with the gene
ralized three-step least squares estimator proposed by Chamberlain (1984).
In case A it can be seen that with stronger assumptions the size of has increas
ed. This is not a problem asymptotically but in small samples it can lead to an
overfitting bias in the direction of the Within Groups estimator. There is a tra
de off between asymptotic efficiency and a finite sample bias; the latter can be
reduced at the expense of the former by reducing the size of the instrument mat
rix systematically.
Case B: Assuming that is uncorrelated with the unobserved fixed effects mak
es more moment conditions available. With this assumption there are now valid in
strumental variables for the untransformed levels equations also. Efficient esti
mation combines the set of moment conditions, those already discussed above for
first- differenced equations and those implied for the levels equations under th
e current assumption.
If is endogenous with respect to , non-redundant additional moment conditio
ns
(3.131)
are added to moment conditions for first-differenced equations. This allows the
use of as an instrumental variable in the levels equation for period t.
If is predetermined or strictly exogenous T non-redundant moment conditions of
the form
(3.132)
and
(3.133)
3.4 Summary
The estimations display the expected properties in all reported estimates. Ordin
ary Least
Squares estimates are biased upwards; the bias depends on the relative variance
of the error components . In section (3.3.1) we discussed why we expect to obser
ve these effects. The Within estimate (discussed in the AR (1) context in Sectio
n (3.3.2) is expected to be biased downwards. For IV estimation theory predicts
that all of the instrumental variable estimators we have looked at should be con
sistent and indeed we see essentially negligible bias for all of them in the est
imation results.
Additionally we expect that the System GMM estimator would be more precise than
the Diff GMM estimator and this theoretical prediction is also confirmed. It is
true that efficiency gains are rather modest and that is also in line with the T
able ??? results where we get a very small gain in precision from including more
instruments.

Chapter 4: Dynamic investment models

While discussing investment models we primarily focus to dynamic investment mode
ls. Most of the empirical structural models of investment can be derived from mi
croeconomic optimization exercises. It is nearly about four decades before that
the first micro econometric dynamic models were introduced in the investment lit
erature Jorgenson (1963), Eisner and Strotz (1963) and Gould (1968)). The dynami
c models of investment under uncertainty can be classified in two categories. On
e category includes the models assuming investment as reversible while other one
 assumes it to be at least partly irreversible. The models of at least partial i
rreversibility fall in the category of real option literature. Among the numerou
s investment models the most popular one is the Q model, and under certain assum
ptions it equates the shadow value of capital to the market-to-book value or ave
rage q ratio. In the special case of symmetric adjustment costs the complicated
quadratic functional form is simplified to linear form and linking investment to
observable average q. Despite of its wide spread use, the poor empirical perfor
mance related to it has made a less restricted adjustment cost model more attrac
tive, such as Euler equations and approach by Abel and Blanchard (1986) and Abe
l (1980) more attractive in the literature of investment.
4.1 The Q model
In order to explain the Q model (Tobin, 1969) of investment we us assume that
firm has a single (quasi-fixed) homogenous input, labor and other current inputs
has zero adjustment cost and the model is fully flexible to employ any number
of inputs needed.
The firm’s net revenue function is given as (4.1)
Assuming the markets to be perfectly competitive and by solving the first order
conditions of optimization we can get the shadow value of capital as
(4.2)
Here is the rate of depreciation and is the discount factor which discounts b
ack the revenue of period to the period .The shadow value of each extra unit o
f capital ( ) is an estimated forecast value of the current and expected future
value of marginal revenue product of capital. In the case of static factor deman
d model the optimal capital stock is characterized by or by , and the ratio of
shadow value of capital to the cost is known as marginal q. If the adjustment c
osts are strictly convex then the marginal adjustment costs are an increasing fu
nction of current gross investment and the investment is an increasing function
of the deviation between the actual value of marginal q and the desired value in
the absence of adjustment costs. Another striking finding is that all the usefu
l information about expected profitability of the current investment is containe
d in the marginal q.
As is a common practice in the most applications of the Q model, we assume that
the cost of capital adjustment is symmetric and is quadratic regarding some stan
dard rate of investment, not necessarily related to the depreciation rate. Princ
ipally the, the Q model necessitates the adjustment cost function to be homoge
nous and linear in and in addition to the production function to be constant
return to scale. A widely used functional form having all these characteristics
is proposed by Sumer (1981)
(4.3)
Where the parameter represents the adjustment cost. The linear form can be obt
ained as
(4.4)
Hayashi (1982) under certain assumptions sets up a unique characteristic of Q mo
del, the equality between unobservable marginal q and the measurable average Q.
The necessary condition is that the net revenue function is homogenous and li
near and the sufficient condition is that both production function and adjustmen
t cost function exhibit constant return to scale, in addition to the firm to be
a price taking.
(4.5)
(4.6)

The basic Q investment model after substituting the average q in place of margin
al q is as
(4.7)
All the information is carried through the share prices and if these are not aff
ected by noise or bubble then the market valuations well reflect the prospects o
f future earnings and the predictions made on the basis of the Q model are suffi
cient for investment.
Another point worthy to be noted is that the parameters are identified only by
assuming the perfect competition and constant return to scale without any speci
fic form of the production function. It can have different consequences dependin
g on the framework. It may be advantageous if the aim is to measure the magnitud
e of the parameters and to test the robustness of specification to the changing
functional form of the production function. While it may be a shortcoming or mig
ht not be enough to estimate the adjustment parameters only if the intensions ar
e to estimate the investment response to the policy change or other aspects of a
djustment cost.
Although the Q model requires a restrictive structure to equate the marginal q a
nd the average q in order to get a linear relation between investment and margin
al q, yet it has advantages over the reduced form models. It can explicitly mode
l expectations influence on the current investment, the parameters obtained are
the technological parameters robust to any structural change in the process gene
rating function and interest rate. It provides an additive advantage to test the
hypothesis of perfect capital markets.
Nonetheless there are many ways to show that how a Q model might be falsely spec
ified and is not astonishing that the empirical performance of the Q model has b
een unsatisfactory. In most of its applications the specification in (4.7) yield
s low values of thus overstating the marginal cost of adjustment and doubtful
ly slow adjustment of capital stock. The empirical tests reject the predictions
based on the specifications as being insufficient more over the inclusion of o
ther variables in the model like sales and cash flow further weakens the explana
tory power of the variable. Numerous efforts to control the measurement error
in the variable have not been successful to avert the basic criticism faced by
the Q specification however Cummins et al. (1994) report that variable calculat
ed during the periods of tax changes can improve the results of the Q model to p
rovide a satisfactory description of investment behavior. By incorporating usefu
l measure of controlling the effect of fade and bubbles in the share prices Eric
kson and Whited (2000) and Bond and Cummins (2001) have noticed some encouraging
results. The above stated facts illustrate that the variation in the measured
is mostly accompanied by bubbles and noise in the equity prices in addition to t
he other factors contributing to misspecification of Q model.
4.2 The Abel and Blanchard model
If the necessary assumptions needed to equate the marginal q and average q are d
ifficult to attain or if the average q is infected because of stock market devia
tions from fundamentals in this case an alternative is to try to calculate the m
arginal q directly and estimating the equation (4.4) directly.
An alternative suggested by Abel and Blanchard (1986) is to employ an auxiliary
regression to calculate the shadow value of capital. It requires an explicit for
m of marginal revenue product of capital in the form of observable variables and
a model to forecast these variables. This forecasting model is merely needed to
obtain an estimate of marginal revenue in future rather than to provide exact p
redictions. How far well does it perform in given econometric specification is n
ot known.
This method does not require any information on share price and given a suitable
form of marginal revenue product of capital it relaxes the assumption of market
s to be perfectly competitive and the production function to be constant return
to scale. The specific form of marginal revenue product of capital and the forec
asting model replaces to these assumptions. The linearity of investment model de
pends a lot on the conditions of symmetric and quadratic cost of adjustment.
4.3 The Euler equation
Abel (1980) introduced the Euler equation approach; it relaxes the use of share
price and the net revenue function to be linear and homogenous. In particular, t
he approach does not require the parameterization of expectation-formation proce
ss. It is obtained by using the first-order condition for investment
(4.8)
Substituting the shadow value of capital from the Euler equation
(4.9)
and estimating the Euler equation directly. For a single capital input it reduce
s to
(4.10)
Using a specific functional form of net revenue and assuming the markets to be c
ompetitive enough it gives
(4.11)
Here is the real discount factor, is the user cost of capital
By a careful examination of equation (4.11) we find that the right hand side of
equation almost contains the information same as that of marginal q. Particularl
y by using the difference between user cost of capital and the marginal product
of capital, the information pertinent to expected future profitability can be su
mmed up with the one-step ahead forecast
(4.12)
of discounted marginal adjustment. In order to apply this model, the actual val
ues of these variables at time t+1 are used instead of the one step ahead foreca
st; it will create forecast error that is orthogonal to the information availabl
e at time t in the presence of rational expectations. In the case of constant re
turn to scale, the marginal product of capital without any specific parametric f
orm of production function can be substituted, as used by Bond and Meghir (1994)
. Otherwise, the model with a specific form of production function can be implem
ented, as by Abel (1980). It is more explicit, as the substituted parameters can
be identified because of Euler equation contrary to the case of Q model. The Eu
ler equation can be further extended to the production function with decreasing
return to scale and imperfectly competitive product markets.
4.4 Multiple quasi-fixed factors
The models discussed so far assume the capital as a single quasi-fixed input, an
d the cost less adjustment of other inputs. A more specific model of capital inp
uts can explain that how do more than one quasi-fixed factors affect these model
. In order to explain let us consider a firm which can substitute between two ty
pes of capital inputs and which are subject to capital adjustment costs. We get
the form of an investment model for the first kind of capital input
(4.13)
a similar kind of investment specification can be obtained for the second capita
l input. It shows that the basic Q model is mis-specified in the case of multipl
e quasi fixed inputs. Although this system of equations might be estimated, but
the emphasis is given on the restrictions in the form of adjustment cost that pe
rmit a single equation to estimate the total investment. The particular conditio
ns have been derived by Hayashi and Inoue (1991) under which the basic structure
of Q model is preserved.
Abel and Blanchard approach and the Euler equation approach further extend to th
e case of additively separable adjustment costs. A specification for the shadow
value of every additional unit of each capital can be obtained and might be used
to get an estimate of marginal q for the corresponding capital unit. It will re
sult to a system of equations for each type of capital input. These specificatio
ns can further be extended to incorporate the interrelated adjustment costs give
n that a specific form of adjustment cost function is provided. A detail about E
uler equation system with interrelated adjustment costs is provided by Shapiro (
1986).
4.5 Non-Convex adjustment costs
Numerous concerns initiated the researchers and economists to think of investmen
t models with non-convex costs of adjustments. The models of strictly convex adj
ustment cost were primarily introduced only because of analytical simplicity alt
hough the long time series data on investment raised the doubts on the validity
of these assumptions. The diagnostic models with non-convex adjustment costs are
popular and preferred by researchers to analyze the investment dynamic and thei
r empirical performance is well consistent as compared to those models with stri
ctly convex adjustment costs.
The evidences prove, generally the production data is more disaggregated than th
e companies accounting data. Doms and Dunne (1998) by using the data of 12,000
US plants for the period 1972-1989, explain numerous features of the ‘lumpy’ cap
ital adjustment. The data shows that about half of the all units have at least o
ne year during which the capital stock increased by 35% or more. Similar finding
s are reported about Norwegian plant data by Nilsen and Schiantarelli (1998), in
the view of their findings almost 30% of the plants in Norway have an average y
ear during which zero investment is being made, this ratio declines to 6% in the
case of major production units or for an aggregate of units of the same firm. A
bel and Eberly (1996) by using a large sample of US listed firms report that the
investment series distribution is positively skewed and two hikes in investment
rates are found in successive years for almost half of the sample firms.
The non convex adjustment costs can better explain the irregular and lumpy cap
ital adjustment. If the gross investment is characterized to be non-negative the
n there can be complete irreversibility. Where as partial irreversibility may ta
ke place if the used capital be sold at a price less than their actual cost valu
e. This price wedge creates a linear cost of adjustment and is kinked at the po
int of zero gross investment. This kink is sufficient enough to explain the zer
o investment although the firm does not attain the desired level of capital stoc
k but does not reveal about actual lumpy adjustment.
The literature on investment has been emphasizing on the behavior of irreversibi
lity and linear adjustment costs since a long time (Arrow (1968) and Nickell (19
78)). However, the influential work of Dixit and Pindyck (1994) and Bertola and
Caballer (1994), have sparked the implications of these adjustment costs and the
se models have a major impact on the empirical literature. Abel and Eberly (1994
, 1996) further adorned the Q model with the non-convex adjustment costs to get
a structural model of investment. Assuming the constant return to scale and all
markets to be perfectly competitive, the adjustment cost function is as
(4.14)
Here is the fixed cost of adjustment, is adjustment cost, is convex adjust
ment cost, and parameters can have different values according to the sign of gro
ss investment. The investment rates are as given
(4.15)
Here is the shadow value of capital, and are upper and lower threshold val
ues respectively. It would not be worthwhile for a firm to carry out a positive
investment below the or a disinvestment below . Since the adjustment cost fun
ction is homogenous and linear the model can be executed by equating marginal a
nd average q in the usual manner. The specification describes non-linear and mon
otonic link between average q and investment, and a zone with zero gross investm
ent.
Caballero and Leahy (1996) disapproved the Abel and Eberly specificati
on, they prove that the perfect competition and constant return to scale are the
vital assumptions only in the case of fixed cost of adjustment. The linear homo
geneity of net revenue function is not only necessary for equating the marginal
and average q, but also to helpful to attain a monotonic link between investment
and marginal q in the case of fixed adjustment costs. They view both the perfec
t competition and constant return to scale assumptions to be completely inapprop
riate, especially in the case of big firms. It departs from the Lucas (1967) poi
nt of view, who considers these conditions as underpinnings of the q theory. Fir
m can have optimal size only in the case of linear homogeneity of net revenue fu
nction and it cannot attain optimal size without linear homogeneity. It may reco
ncile the model with the theories arguing that the changes in firm size are stoc
hastic and not easy to predict. Most of the structural models of investment wit
h zero fixed costs use these assumptions. It may be difficult to maintain the li
near homogeneity, and the only advantage that it simplifies the characterization
of optimal investment in the case of fixed adjustment cost is not worthwhile. A
s a matter of fact, any decisive evidence to discard the assumption of linear ho
mogeneity is not available in the literature to the date.
Abel and Eberly (1996) and Barnett and Sakellaris (1998) have reported signific
ant non-linearities in investment rate to average q in the case of US companies
data. Eberly (1996) observed an S-shaped sigmoidal and Barnett and Sakellaris (
1998) reported a concave relationship.
It seems to in contrast to the predictions of model with non-convex adjustment c
osts, which means a zero response of investment to average q in the zone of inac
tion and a relevant response as the average q fall outside of the zone. Abel an
d Eberly (1996) suggest that this observed paradoxical finding can be explained
by aggregation over different types of capital goods. The basic idea is that wit
h non-convex adjustment costs, adjustment occurs at both the expensive margin (t
he decision to invest or not to invest) and the intensive margin (how much to in
vest or disinvest given that the adjustment is worthwhile). At intermediate valu
es, a rise in average q will trigger adjustment at the extensive margin for some
types of capital, with appending on those types of investment increasing from z
ero to some level that is large enough to justify paying the corresponding fixed
cost. At higher values, a further rise in average q produces adjustment only at
the intensive margin for all types of capital, so that the response of total in
vestment spending to the change in average q may be flatter.
The idea is nice, but the non-linearity results reported by Abel and Eberly (19
96) might be induced due to censoring and measurement error problem, although th
e true relation is likely to be linear. The measurement errors in the variables
are likely to introduce serious bias and change the shape of the actual relation
ships between the variables Chesher (1991). It has been observed repeatedly, a s
evere problem with average q is the measurement error. As suggested by Bomd and
Cumminsd (2001), the share prices inflated by noise are a likely to create non-
linearity especially when the distorted market valuations are used to calculate
the average q.
Caballero et al. (1995) used an ad hoc approach for testing the implications of
fixed adjustment costs. Their empirical analysis includes, a model with fixed
cost of adjustments without any linear or convex components, any capital adjustm
ent if likely to happen will keep the capital stock at the same target level ir
respective of the fact that either this adjustment will take place from upper or
lower side. Given fixed cost levels, an excess value of the ratio given a thre
shold level will cause a positive investment and lower value of the may cause
a disinvestment. If the levels of actual fixed cost are random then we can say t
hat the likelihood of happening an investment or disinvestment must be an in inc
reasing function of the absolute value of
Recently Cooper and Haltowangert (2000) used the plant level of data
of US in order to investigate the dynamic of capital stock adjustment. They emp
loyed an indirect inference for the estimation of parameters rather than a direc
tly estimating the investment equation. In fact, it inquires about a particular
form of adjustment cost to be consistent with the non-linear relationship betwee
n investment and profitability present in the data. Not astonishing, they sugges
t a common specification of adjustment cost with a combination of convex and non
-convex components with irreversibility in order to fit this relation ship. Thei
r results, in the line of empirical findings about the plant level investment da
ta reported in earlier studies as mentioned before, uncover doubts on the relian
ce of structural models of investment involving simple specifications. It still
remains a major challenge to develop a structural specification of investment mo
del with the dynamics of adjustment costs, but nor acquiring the linearity and h
omogeneity of net revenue function.
4.6 Reduced form models
Actual adjustment process is some what complicated; especially when the firm lev
el investment data is an aggregate of different kinds of capital inputs and is c
omprised of several units/plants. The structural models introduced in the litera
ture of investment seem to be not much successful in illustrating the adjustment
process; perhaps they did not pay much concern to the issues like aggregation a
nd non-convexities. Nonetheless it is obvious that capital adjustments incur som
e cost and is not instantaneous, so it would be better not to rely on convention
al static models . The dynamic econometric specifications are best alternatives
to be employed since they have not been obtained as optimal adjustment behavior
for some particular structure of adjustment costs. The unique characteristic of
the reduced form models is that they employ the empirical approximation which ca
n even explain the complex data generating process. A likely limitation of reduc
ed form models is that they compound the parameters of adjustment process with t
hat of expectations-formation. Any how these models are worth using and would be
better to be aware of their shortcoming.
The most popular approach used in the literature of investment is the fi
rst differencing transformation. An investment equation can be obtained by takin
g the first difference of a static factor demand model, as given
(4.16)
Where is the logarithm of the optimal capital stock, given by
(4.17)
A log-linear version of in the level of output directs to the famous acceler
ator model in which investment is linked to output growth. In fact, in response
to the changes in the desired level the actual capital stock may not adjust com
pletely and quickly, the accelerator models include the lags of and . The incl
usion of the distributed lagged terms provide a dynamic specification as
(4.18)
Where and are polynomials in the lag operator A number of accelerator mode
ls of such type have been estimated by Eisner (1977) and Mairesse and Dormont (
1985) by using firm level data.
A partial adjustment specification in the level of capital stock is ano
ther possible alternative, such as
(4.19)
A fixed fraction of the deviation of actual levels of capital stock from that
of the desired levels is closed in each period. It seems to be less flexible, ra
ther a less restrictive dynamic adjustment model is one that combines the partia
l adjustment and accelerator models in an error correction specification, for ex
ample
(4.20)
Where and are polynomials in the lag operator. Bean (1981) introduced the Er
ror correction models in the literature of investment and later Bond et al (1999
) and Bond et al. (2003) estimated these models by using the firm level data.
The logical link between error correction specifications and the co-inte
gration techniques made them widely used and well known in the context of adjust
ment models. Although the cointegration techniques are newly developed in econo
metric literature as compared to the error correction specifications. As a matte
r of fact an error correction model is a special case of an ADL model. Consider
an ADL (1,1) model
(4.21)
and is re-parameterized as
(4.22)
Where the long-run proportionality restriction is
In order to implement the error correction model we may require an expli
cit form of the target level of the capital stock. As observed the target and ob
served level of capital stocks are generally not the same as the earlier is subj
ect to adjustment costs while the later is not. Using a static factor demand mod
el
(4.23)
(4.24)
The long-run elasticity of the capital stock can be estimated by using these mod
els with the given fact that capital adjustment does take place instantaneously.
Theoretically, these models are flexible to the inclusion of non-linear and asy
mmetric adjustments, the parameters and may take different values conditional
on the sign or the magnitude of .
Another interesting link between the reduced form and adjustment cost mo
dels is that reduced form models are empirical while adjustment cost models are
theoretical generalization of static factor demand model. As described in Nickel
l (1978, Chapter 11), for symmetric and quadratic adjustment cost function, the
level of investment is given by
(4.25)
where
(4.26)
and is the level of capital stock that the firm may decide when there are no ad
justment costs. Hence, the investment choices are to some extant explained by a
partial adjustment mechanism. More over the ‘target’ level of capital is defined
as a function of current and expected future levels of the static optimum. Nic
kell (1985) defines more strict conditions for the process. With these restric
tions the process can either generate a partial adjustment process or an error
correction model linking investment to . The probable benefits of reduced form s
pecification can be explained as following. Let us assume a simple structural a
djustment process defined as
(4.27)
is a particular case of (4.25) including the expected values only up to the per
iod and the static optimum level of capital stock is proportional to output. Al
so the expected value of future out put is given by
(4.28)
where the vector includes some other variables which help to forecast the fut
ure production. By solving the (4.27) and (4.28) simultaneously we get the inves
tment equation in reduced form
(4.29)
This equation demonstrates that in the reduced form specification the structural
parameters are mixed with those of the expectations-formation process . Empiri
cally the blend of these two different kinds of parameters has some consequences
. First, with the constant parameters of structural adjustment process, the unst
able parameters of expectations-formation process may cause the parameters of re
duced form investment equation also to be unstable. For example, any structural
break in the out put generating process, caused by firm’s decision or by any cha
nge in macro policy may induce the parameters of the reduced form investment mod
el to be instable. It explains that the reduced form models have to face the Luc
as (1976) critiques. Unlike the reduced form models, the Structural investment m
odels do not face such criticism since there parameters are deep; explain the ad
justment cost and robust to the changes in the out put generating process. Pract
ically the authenticity of these claims depends on the correct specification of
these models. The second limitation is that the variables in vector might be s
ignificant in the case of reduced form equations as (4.29), although they might
not have any significance in the structural model of investment, and their speci
fic role in the context of reduced form equation is to facilitate to forecast th
e values of the basic determinants of investment. As a matter of fact it is a se
rious challenge if the objective is to obtain any inference about the behavior o
f fundamental structural model. The significant coefficients of financial variab
les in the reduced form investment equation do not necessarily imply that either
these are also significant in the structural equation or just facilitate to for
ecast the yield of profitability.

Chapter 5:
5.1 Econometric Specification
For testing the uncertainty effects, we specify a simple version of reduced form
model for investment in this section. Since our primary interest lies in signin
g the relationship between uncertainty and investment, adopting an empirical spe
cification that allows for simple investment dynamics may be a useful starting p
oint. Structural models, such as a Q model or an Euler Equation framework, are l
ess amenable to testing the links between uncertainty and investment since these
models assume away the real options effects from the outset. Even when modified
to test for the presence of non-convex adjustment costs, the structural models
remain highly sensitive to the choice of particular assumptions and do not even
encompass the plethora of theories on this issue.
In order to assess the empirical importance of uncertainty, we specify an error
correction formulation, introduced in the investment literature by Bean (1981) a
nd developed more systematically for the microeconomic literature on investment
in Bond et al. (2003). The linear error correction models represent long-run evo
lution of the capital stock while allowing the short-run investment dynamics to
be directly estimated from the data. In the spirit of Bond et al. (2003), we pro
pose the following error-correction specification:
(5.1)

where and : The specification in (5.1) relates the current investment rate to
lagged investment rates , current and lagged terms for sales growth ; lagged
error correction term . A key benefit of working with panel data is the possibi
lity of controlling for common time shocks as well as firm-specific effects. The
se effects are controlled for by including a time dummy, and explicitly allowi
ng for unobserved firm-specific effects : Finally, is an idiosyncratic error
term in (5.1). We then augment equation (5.2) with cash flow term to capture th
e liquidity constraints or the profitability aspects unexplained by the sales gr
owth variable. (5.2)
Finally to investigate the effect of uncertainty, we include our proposed measur
e of uncertainty and additionally, we also include an interaction term between
uncertainty and sales growth . The inclusion of this interaction term is motiv
ated by the prediction that firms’ response to demand shocks may be lower at hig
her levels of uncertainty (Bloom et al., 2001). With these additional uncertaint
y terms, our general error correction specification becomes:
(5.3)
In keeping with our empirical predictions, we should expect to see a positive ef
fect of sales growth terms and cash flow on investment. The coefficient on the e
rror correction term, which captures the long-run adjustment of capital stock, i
s likely to be negatively signed , implying that if capital is less than its “de
sired level” future investment will be higher and conversely. The sign on uncert
ainty term is theoretically ambiguous, since the long-run effect of uncertaint
y on the level of capital- sales ratio can be either positive or negative. In or
der to be consistent with the theoretical prediction in Bloom et al. (2001), we
require the coefficient on the interaction between sales growth and uncertainty
to be less than zero. A negative coefficient on this interaction would signify
that firms respond less to demand shocks at higher levels of uncertainty.
We experiment with different variations of (5.3) by including and excluding the
differences, lags, and squares of some of the included terms. In particular, we
investigate if the data supports the inclusion of the square of sales growth, ch
anges in cash flow, changes in uncertainty, and lags in the uncertainty terms.
5.2 Measuring uncertainty
A central issue in implementing the empirical model is the measurement of uncert
ainty. It is not an easy procedure to obtain a specific measure of the uncertain
ty affecting a firm. The important thing to note at the outset is that there are
conceptually different sources of uncertainty, ranging from prices, wages, and
profits, to macroeconomic uncertainty. Despite of the China’s tremendous economi
c growth and development as the previous experiences reveal that the market anti
cipates a great level of uncertainty. The transition to market economy, fledglin
g government policies and reforms, the consideration of trading state owned sha
res lost the investor confidence, are sometimes fatal as has been observed in th
e past and leading to a thin market. High frequency data for constructing an emp
irically meaningful measure of uncertainty is hard to obtain. To analyze a large
panel of firms, obtaining this measure would be daunting, if not impossible tas
k. In the face of these data limitations, we use a more general approach adopted
in the literature, namely using the daily share price data to construct a measu
re of uncertainty. In the spirit of Leahy and Whited (1996) and Bloom et al. (20
01), we consider a measure of uncertainty based on the volatility of stock marke
t returns of publicly traded corporations in China. Volatility is measured as th
e standard deviation of recordings of stock market returns over the relevant acc
ounting year for each firm. The daily stock market returns provided by the CCER
account for capital gain on the stock, dividend payments, rights issues, and sto
ck dilutions. Our measure is adjusted for gearing, since as Leahy and Whited (19
96) note, return variance is expected to increase with the leverage of the firm.
However, regardless of whether or not we allow for gearing adjustment, the resu
lts are qualitatively similar.
Measuring uncertainty by the variance of stock returns is based on the p
resumption that information contained in share prices has some correspondence to
the firms underlying fundamentals. The advantage of this measure is that it cap
tures the total uncertainty that is relevant to the firm in a single variable.
To the extent that stock market returns capture investors’ perception of a firm’
s overall environment, the volatility of these returns could provide a useful me
asure of uncertainty. However, one potential source of concern with using a stoc
k returns-based measure of uncertainty is that movement in stock prices may refl
ect influences other than the fundamental valuation of firms. In particular, the
movement in stock prices could be influenced by noise, bubbles and fads. Pindyc
k (1991) reasons that increasing volatility in the product markets are well depi
cted by rising volatility in the stock market. Leahy and Whited (1996) argue tha
t stock returns capture the changing aspects of a firm’s environment that is imp
ortant to investors. Additionally, Berk , Green and Naik (1999) show that time v
ariation in asset returns are driven by the firm’s expectations about its return
s from assets in place and from potential growth options. Simply put, common sto
ck represents claims on the future profits of a firm. Innovations to a firm’s st
ock return are reactions to news about the firm’s future profitability and futur
e investment opportunities. If markets are efficient, news about asset fundament
als and the firm’s future prospects are priced by the market, hence the volatili
ty of a firm’s stock returns should reflect these variations and provide an adeq
uate measure of the uncertainty that a firm faces. In the empirical analysis tha
t follows, use volatility a (and hence standard deviations) to measure uncertain
ty. This is consistent with theoretical models of irreversible investment that i
dentify uncertainty with the volatility of innovations to a firm’s profit functi
on.
If the stock market is excessively volatile because of irrational exuberance,
we may not be able to infer meaningful information about changes in the firms ex
pected future profitability. Attempts have been Recent research has attempted to
circumvent this measurement issue by using survey based data on manager’s expec
ted distribution of future variables (Guiso and Parigi, 1999) or a dispersion ac
ross the analysts’ forecasts of future profits (Bond and Cummins, 2004).Using I/
B/E/S data for UK firms, Bond et al. (2005) report that stock returns volatility
is positively correlated with both the within-year variability of analyst earni
ngs forecast, and with the cross-section dispersion across forecasts made by dif
ferent analyst for the same firm. In the latter case, the measure based on analy
sts profits forecasts is shown to be highly correlated with the stock market vol
atility. These measurement considerations aside, using an imperfect proxy for un
certainty that is based on data on stock returns can be a useful first step in l
earning about the firm’s investment behaviour in a period of considerable uncert
ainty. Recent evidence suggests that stock market returns have significant power
in predicting output growth in emerging market economies (Mauro, 2003). Recall
that theoretical literature generally refers to a relationship between idiosyncr
atic uncertainty and investment, not the uncertainty that affects all firms in g
eneral. The key advantage of using an asset returns based measure is the availab
ility of high frequency data, which is critical for constructing an efficient me
asure of uncertainty that varies across time and firms.
5.2.1 Decomposition of uncertainty
Total firm uncertainty for firm in year is measured as the volatility of th
e firm’s equity returns. This is calculated as the annualized standard deviation
of the firm’s daily returns in that fiscal year and this measure should capture
the fundamental variations in the firm’s expected future profitability.
Even though the level effect of uncertainty identifies the variation in uncertai
nty that is not accounted by time or firm-specific effects, the same could not b
e said for the interaction term, however. In order to explore the further effec
ts of uncertainty, we decompose our measure of uncertainty ( ) in three componen
ts-a macroeconomic component, common to all firms in a particular year ( ); a ti
me-invariant firm-specific component ( ) and a residual component for each firm
year ( ).
5.3 Data, Sample and Descriptives
The accounting information for investment expenditures, capital stock, sales, pr
ofits, and cash flow is obtained from the CCER database, which is known for its
standardized presentation of investment portfolios and its good coverage of hist
orical data. In order to construct the uncertainty variable, we use daily share
price data from CCER. Further details on this data are available in the appendix
. By merging the accounting information with the stock market data from CCER, ou
r sample is reduced to about 934 non-financial firms out of which 530 are liste
d in Shanghai and 404 are listed in Shenzhen stock exchange. We construct an unb
alanced panel of firms over the period 1994-2005. We require that each firm shou
ld have a minimum of five years of complete data. The core variables used in our
estimation are flows of investment, cash, and sales.
Table: 1 Basic Descriptives of the regression variables
Variables Min Lower quartile
Median Mean Upper quartile Max S.D

-0.796 0.002 0.098 0.162 0.242 11.042 0.368

-0.512 -0.245 0.0003 0.0158 0.041 1.585 0.1112
-7.910 -0.026 0.125 0.1298 0.291 5.995 0.454
0.093 0.810 1.265 1.7089 2.101 16.136 1.440
0.089 0.863 1.073 1.2420 1.403 16.002 0.728
Note: Total observations: 8506; number of firms: 934; time period: 1994-2005.
In order to obtain more robust results the outliers are removed from each variab
le so that our estimations might not be effected through extreme values. After e
xcluding the outliers and missing observations we are left with 8506 observation
s for regression analysis.
Table 1 summarizes the detailed distribution of key variables used in the estima
tions. The average investment rate over the entire period is about 16.2%. Regard
less of how we measure it, average level of uncertainty during the sample period
hovers around 1.7%. Compared to similarly estimated measures for developed coun
tries, our estimate of 1.7% indicates a generally higher level of uncertainty. A
mong all the regression variables the average and standard of uncertainty is the
highest.
Figure 1 plots the average of standard deviation of daily returns for firms duri
ng the period, 1994-2005. The pattern of evolution of this measure looks remarka
bly sensible. Consistent with our priors, the constructed measure of uncertainty
displays ups and downs in the following years. The evolution of our uncertainty
measure is also broadly in line with that of figure 2, though the latter is bas
ed on an index and is generally more spiky. As figure 2 shows, In 2001 there is
a sharp decline in stock index and even till 2005 it further dip to the lowest i
n spite of gaining little boom in 2004. These figures reveal that: stock market
volatility in China is of a different order of magnitude than that in the US or
Japanese stock markets.

Figure . 1
Figure .2

5.4 Estimation of Investment Specification

This section will provide the main empirical findings of the basic specification
in (5.3) and its different variations. Before presenting the results, a brief d
iscussion of selected estimation preliminaries is in order. Through out the esti
mations, we characterize all the right-hand side variables, including the uncert
ainty variables, as endogenous allowing the use of values lagged two periods or
more as potential instruments. The validity of these instruments and the moment
conditions underlying our models is tested using the Sargan test of over identif
ying restrictions (p-values reported) and the tests for serial correlation in th
e differenced residuals (test-statistics reported). The results emanate from mod
els that are estimated in first differences to remove the unobserved firm-specif
ic fixed effects. A full set of time dummies is also included (though not report
ed) in the estimations. Owing to its superior performance in the presence of fin
ite sample bias and short panels, we shall generally emphasize the System GMM r
esults, although the OLS, WG, and differenced-GMM estimates are also reported fo
r comparison .
We begin our estimations in this section by picking up our preferred specificati
on in the previous section and adding the proposed measures on uncertainty. The
first panel in Table ??? documents the main empirical results for the four alter
native estimators. In each model, we have included the level of uncertainty ( ),
an interaction term between sales growth and uncertainty ( ), levels ( ) and la
gged ( ) term of sales growth and cash flow ( ) term as controls. The results su
ggest the presence of error correction behavior, as manifested in a negative and
significant error correction term. In addition the cash flow term has a signifi
cant positive coefficient. To test for the prediction of Bloom et al. (2001) the
term ( ) is also included in each model but the term has no significant effect.
Although the coefficient of ( ) is not informative but the coefficient of ( ) i
s significantly positive in all regressions. Controlling for these investment dy
namics, the level term of uncertainty has a significant positive coefficient and
is a consistent feature of all the estimators, from the OLS to the System GMM.
Other things being constant, a higher level of measured uncertainty is associate
d with higher level of investment in the short-run. Since the estimations contro
l for unobserved effects, it is worth noting that the level effect is capturin
g the uncertainty for individual firms. It is the piece of uncertainty that is n
ot identified by the unobserved time and firm effects.
In the second panel of Table ???,we successively introduce a number of non-linea
r terms. Columns (1)–(3) find that the positive coefficient on the uncertainty t
erm remains robust. Column (1) explores a non-linear effect of sales growth by e
ntering the square of but it has no significant effect. Finally, columns (2) a
nd (3) represent the difference of uncertainty term ( ). This additional term is
statistically significant and does not affect the robustness of uncertainty ter
ms. In all of these specifications, the validity of the instrument set is not re
jected at conventional levels of significance, as suggested by the p-values for
the Sargan test and the test statistics for the serial correlation tests (as rep
orted in the relevant table).
Table 5.4.1 Uncertainty and investment in a panel of Chinese firms (1994–2005)
Dependent Variable -
Panel.1 Panel.2 System Estimates
OLS WG GMM System(1) 1 2 3
-.01205
(0.562) -.199
(0.0001) -.402
0.002) -.0997
(0.118) -.235
(0.192) -.217
(0.122) -.226
(0.117)
.1914
(0.0001) .2167
(0.0001) .173
(0.230) -.054
(0.788) -.0336
(0.843) .072
(0.715) .131
(0.490)
.0928
(0.0001) .190
(0.0001) .271
(0.009) .135
(0.014) .174
(0.023) .153
(0.013) .166
(0.017)
-.046
(0.0001) -.2097
(0.0001) -.257
(0.031) -.136
(0.020) -.154
(0.022) -.135
(0.021) -.148
(0.038)
1.2295
(0.0001) 1.13
(0.0001) 1.156
(0.021) 1.890
(0.010) 1.752
(0.015) 2.028
(0.001) 1.945
(0.0001)
.0195
(0.0001) .0145
(0.001) .112
(0.041) .112
(0.025) .128
(0.010) .150
(0.006) .152
(0.007)
.036
(0.014) .0447
(0.001) .084
(0.235) .149
(0.356) .134
(0.273) .0992
(0.484) .0692
(0.621)
-.004
(0.908) .0104
(0.793) .015
(0.694)
.0403
(0.006) .0421
(0.005)
.027
(0.765)
Diagnostic Tests
m1 -2.62 -5.60 -3.82 -4.42 -4.21
m2 -1.85 0.20 -0.66 -0.41 -0.43
Sargan 0.865 0.651 0.787 0.840 0.90
Panel 1 Instrume
nts
GMM
System(1) plus lagged
differences of
Panel.2 :System Estimates Instruments
Column 1 plus lagged differences of these instruments.
Column 2 plus
lagged differences of .
Column3 plus lagged differences of these instruments.
One-step coefficients, robust to autocorrelation and heteroscedasticity are repo
rted. Numbers in brackets are t-values and in parentheses are p-values. The numb
er of observations is 8506 by using an unbalanced panel of 934 firms over 1994-2
005. For the serial correlation tests, test statistics for the null of no first-
order and second-order serial correlation in the first differenced residual are
reported (m1 and m2). All regressions include a full set of year dummies. Instru
ment validity is tested using the Sargan test of overidentifying restrictions; p
-values of the null hypothesis of valid specification are reported.
Table 5.4.2Uncertainty and investment in a panel of Chinese firms (1994–2005)
Dependent Variable -
1 2 3 4 5 6
-0.186
(0.166) -0.235
(0.192) -0.224
(0.174) -.0997
(0.118) -0.188
(0.192) -0.152
(0.364)
0.036
(0.846) -0.033
(0.843) 0.083
(0.559) -.054
(0.788) 0.187
(0.252) 0.219
(0.051)
0.213
(0.062) 0.174
(0.023) 0.194
(0.026) .135
(0.014) 0.154
(0.039) 0.182
(0.032)
-0.144
(0.026) -0.155
(0.022) -0.174
(0.027) -.136
(0.020) -0.140
(0.061) -0.184
(0.024)
1.495
(0.027) 1.752
(0.015) 1.385
(0.075) 1.890
(0.010) 1.914
(0.0001) 1.636
(0.004)
0.134
(0.018) 0.129
(0.01) 0.135
(0.010) .112
(0.025) 0.119
(0.035) 0.099
(0.054)
0.169
(0.159) 0.135
(0.273) 0.061
(0.595) .149
(0.356) 0.004
(0.968)
0.041
(0.98) -0.004
(0.908) 0.002
(0.945) -0.002
(0.961)
0.042
(0.005) 0.031
(0.055) 0.032
(0.024)

0.029
(0.707) 0.043
(0.524)
Diagnostic Tests
m1 -3.57 -3.82 -3.70 -5.60 -3.56 -3.47
m2 -0.27 -0.66 -0.75 0.20 -0.33 -0.02
Sargan 0.893 0.787 0.635 0.651 0.828 0.765
Instruments
Column (1) plus
lagged differences of these instruments.
Column (2) plus lagged
differences of
Column (3) plus lagged differences of these instruments.
Column (4) plus
lagged differences of these instruments.
Column (5) plus
lagged differences of these instruments.
Column(6) plus lagged differences of these instruments.
One-step coefficients, robust to autocorrelation and heteroscedasticity are repo
rted. Numbers in brackets are t-values and in parentheses are p-values. The numb
er of observations is 8506 by using an unbalanced panel of 934 firms over 1994-2
005. For the serial correlation tests, test statistics for the null of no first-
order and second-order serial correlation in the first differenced residual are
reported (m1 and m2). All regressions include a full set of year dummies. Instru
ment validity is tested using the Sargan test of overidentifying restrictions; p
-values of the null hypothesis of valid specification are reported
5.5 Sensitivity Analysis
This section considers the robustness of our general result on uncertainty to a
number of different variations. First of all, we explore the possibility of nonl
inearity and threshold effects of uncertainty and the effect of inclusion of Tob
in’s Q. To test for non-linear effects, we add a quadratic term to our baseline
specification. The quadratic term of uncertainty does not provide any significa
nt information (the results are not being reported) while the level of uncertain
ty still remains positive and statistically significant. In addition, we test th
e robustness of different components of uncertainty, which remain significant ir
respective of their separate inclusion or all at the same time. More over we als
o test the uncertainty effects based on different sample splits, manufacturing v
s non-manufacturing and large vs small firms.
5.5.1 Inclusion of Tobin’s Q:
The earlier literature on investment proposes the marginal revenue product of ca
pital or marginal Q as a possible channel mediating the effect of uncertainty on
investment. Consistent with this prediction, Leahy and Whited (1996) find that
their measure of uncertainty only matters as long as average Q the conventionall
y used proxy for marginal Q is not entered in the model. Once the average Q is i
ncluded in the regression, the effect of uncertainty disappears. Despite the wel
l known pitfalls of traditionally defined measures of average Q, it could be a u
seful exercise to enter them as a possible control for expected future profitabi
lity. To explore the effects of controlling for average Q, we add a measure of
average Q in columns (5) and (6) table 5.4.2. As the results suggest, the Q meas
ures have a modest positive on investment, though not significant. What is more,
far from knocking the uncertainty effect out, the coefficient on remains posi
tive and statistically significant at 1% level of significance. This rules out t
he simple possibility of proxying for omitted Q effects. But clearly this is n
ot a rejection of the Q theory, given that the prevailing view regards the conve
ntional Q measures as generally weak.
5.5.2 Robustness of Volatility
We next turn to exploring the effects of different components of uncertainty mea
sure( ). In table 5.5.2, we decompose the uncertainty term ( ) into three key co
mponents: a time-invariant firm-specific effect ( ), a macroeconomic component (
), and a residual component for each firm-year ( ). In columns (2) to (4), we s
eparately add the three components of uncertainty interaction. When entered on t
heir own, both the firm and time-specific pieces are positive and statistically
significant.
A more interesting result is obtained in column (5) where we enter all the three
components together. There is a clear suggestion that the most informative comp
onent is the time invariant firm specific ( ). Noticeably in column (5), the un
certainty components are more informative in terms of having significant positiv
e effect and are also greater in magnitude as compared to the separate inclusion
of these components through column (2) to (4).
Table 5.5.2 Separating uncertainty ( ) into firm ( ), time ( ) and residual ( )
components
Dependent Variable -
 1 2 3 4 5
-0.068
(0.128) .031
(0.842) -0.025
(0.86) -0.157
(0.167) -.032
(0.518)
0.127
(0.181) 0.128
(0.264) 0.143
(0.166) 0.208
(0.024) 0.060
(0.550)
0.136
(0.013) .149
(0.085) 0.162
(0.041) 0.191
(0.005) .122
(0.030)

-0.144
(0.018) -.176
(0.023) -.183
(0.009) -0.201
(0.002) -.129
(0.039)
1.906
(0.008) 2.218
(0.0001) 2.142
(0.0001) 2.11
(0.0001) 1.985
(0.007)
0.126
(0.005) _ _ _ _

0.214
(0.012) _ _ .247
(0.008)

0.090
(0.0001) _ 0.109
(0.0001)
0.093
(0.032) 0.106
(0.032)
Diagnostic Tests
m1 -5.43 -4.07 -4.25 -4.64 -3.82
m2 0.58 1.45 1.20 0.14 -0.66
Sargan 0.590 0.268 0.481 0.317 0.787
Instruments
Column (1)
plus lagged differences of these instruments for level equation.
Column (2)
plus lagged differences of these instruments for level equation.
Column (3)
plus lagged differences of these instruments for level equation
Column (4)
plus lagged differences of these instruments for level equation.
Column (5)
plus lagged differences of these instruments.
One-step coefficients, robust to autocorrelation and heteroscedasticity are repo
rted. Numbers in brackets are t-values and in parentheses are p-values. The numb
er of observations is 8506 by using an unbalanced panel of 934 firms over 1994-2
005. For the serial correlation tests, test statistics for the null of no first-
order and second-order serial correlation in the first differenced residual are
reported (m1 and m2). All regressions include a full set of year dummies. Instru
ment validity is tested using the Sargan test of overidentifying restrictions; p
-values of the null hypothesis of valid specification are reported.
5.5.3 Uncertainty effect on: Manufacturing vs Non-Manufacturing firms
More than half of our sample consists of manufacturing firms, 509 out of 934 ar
e manufacturing and only 325 are non-manufacturing firms, it would be quite info
rmative to explore the uncertainty effects in both of these groups separately. A
s is evident from the table 5.5.3.1, despite of the fact that the total number o
f non-manufacturing firms is far less than the manufacturing firms, not only the
average volatility in non-manufacturing firms is higher also there is overall g
reater variation as is depicted by higher standard deviation.
Table 5.5.3.1 Descriptive statistics of sub-sample of manufacturing and non-manu
facturing firms
Variables Min Lower quartile
Median Mean Upper quartile Max S.D
Manufacturing firms, Number of firms=509
-.770 .005 .098 .153 .2334 11.042 .339
-.512 -.028 2.36*10-13 .013 .041 1.251 .106
-6.001 -.008 .1297 .136 .2867 5.333 .378
.093 .778 1.224 1.651 2.029 13.280 1.370
.088 .843 1.048 1.190 1.265 1.1098 .5738
Non-Manufacturing firms Number of firms=325
-.796 -.0004 .098 .173 .255 9.460 .401

-.494 -.021 3.87*10-13 .019 .043 1.586 .117

-7.910 -.057 .120 .122 .300 5.995 .532
.179 .832 1.030 1.780 2.20 16.137 1.518
.175 .886 1.105 1.305 1.458 16.001
The results based on the sample dichotomy, manufacturing and the non-manufacturi
ng firms are presented in the table 5.5.3.2 and table 5.5.3.3. For both groups t
he uncertainty effect is measured with our earlier preferred measured of uncerta
inty and there is a notable difference between them. The estimated results are r
eported only for System GMM estimator. The uncertainty effect is more obvious in
the case of non-manufacturing firms. The mean volatility for the manufacturing
firms is 1.651 and in our reported results in table 5.5.3.2 the uncertainty effe
ct on manufacturing firms ranges from 0.110 to 0.176. Although the number of non
-manufacturing firms in our sample is far less than manufacturing firms, 325 fir
ms only but surprisingly the uncertainty has much greater effect than those manu
facturing firms. The mean volatility for the non-manufacturing firms is 1.78 and
in our reported estimation results in the table 5.5.3.3 the uncertainty effect
on non-manufacturing firms ranges from 0.181 to 0.203. Even the lowest effect on
the non-manufacturing firms (0.181) is greater than the highest effect of manuf
acturing firms (1.76).

Table 5.5.3.2 Uncertainty effect in a panel of manufacturing firms (1994-2005)

Dependent Variable - (Manufacturing Firms)
1 2 3 4 5 6
-.205
[-1.31]
(0.191) -.232
[-1.09]
(0.275) -.326
[-2.40]
(0.017) -.130
[-.62]
(.538) -.1.84
[-1.39]
(.163) -.107
[-.83]
(.405)
.255
[1.22]
(0.22) .395
[1.83]
(0.67) .488
[2.26]
(0.024) .440
[1.48]
(.139) .377
[1.54]
(.123) .354
[2.40]
(.016)
.224
[1.94]
(0.052) .222
[1.88]
(0.060) .248
[2.07]
(0.038) .021
[.17]
(.863) .220
[1.99]
(.047) .141
[1.91]
(.056)
-.249
[-2.09]
(0.037) -.264
[-2.14]
(0.032) -.290
[-2.35]
(0.019) -.022
[-.17]
(.864) -.238
[-2.26]
(.024) -.131
[-1.64]
(0.100)
1.950
[1.90]
(0.058) 1.876
[2.29]
(0.022) 2.156
[5.38]
(0.000) 2.664
[6.18]
(.000) 2.00
[2.13]
(.033) 1.06
[1.78]
(.074)
.110
[2.45]
(0.014) .123
[2.50]
(0.012) .128
[2.42]
(0.015) .1767
[1.92]
(.055) .120
[2.80]
(.005) .162
[2.44]
(.015)
.113
[.64]
(0.521) -.017
[-.10]
(0.922) -.081
[-.48]
(0.629) -.231
[-1.42]
(.156)
-.059
[-.67]
(0.502) -.067
[-.90]
(0.369) -.089
[-1.07]
(.285)
.023
[1.38]
(0.169) .033
[.89]
(.375) .018
[0.41]
(0.679)
.017
[.12]
(.901) 0.033
[0.41]
(0.680)
Diagnostic Tests
m1 -3.80 -4.58 -4.90 -3.57 -3.24 -2.67
m2 -0.49 -0.78 -1.54 -0.85 1.44 0.44
Sargan 0.685 0.777 0.886 0.961 0.817 0.972
Instruments
Column (1) plus
lagged differences of these instruments.
Column (2) plus lagged
differences of
Column (3) plus lagged differences of these instruments.
Column (4) plus
lagged differences of these instruments.
Column (5) plus
lagged differences of these instruments.
Column (6) plus lagged differences of these instruments.
One-step coefficients, robust to autocorrelation and heteroscedasticity are repo
rted. Numbers in brackets are t-values and in parentheses are p-values. The numb
er of observations is 8506 by using an unbalanced panel of 934 firms over 1994-2
005. For the serial correlation tests, test statistics for the null of no first-
order and second-order serial correlation in the first differenced residual are
reported (m1 and m2). All regressions include a full set of year dummies. Instru
ment validity is tested using the Sargan test of overidentifying restrictions; p
-values of the null hypothesis of valid specification are reported.
Table 5.5.3.3 Uncertainty effect in a panel of manufacturing firms (1994-2005)
Dependent Variable - (Non-Manufacturing Firms)
1 2 3 4 5 6
-.059
[-1.09]
(0.277) 0.048
[0.34]
(0.731) 0.041
[0.28]
(0.779) 0.025
[0.18]
(0.856) -0.144
[-1.97]
(0.049) -0.088
[-0.65]
(0.517)
.091
[0.47]
(0.638) -0.023
[-0.12]
(0.908) 0.311
[1.38]
(0.169) 0.258
[1.27]
(0.206) 0.220
[2.00]
(0.045) 0.201
[1.76]
(0.078)
0.043
[0.67]
(0.503) .023
[0.27]
(0.785) 0.088
[1.07]
(0.282) 0.056
[0.69]
(0.492) 0.126
[1.73]
(0.084) 0.090
[1.15]
(0.249)
-.012
[-0.15]
(0.878) -0.003
[-0.03]
(0.974) -0.099
[-1.22]
(0.222) -0.057
[0.69]
(0.492) -0.112
[-1.35]
(0.176) -.068
[-0.83]
(0.409)
1.289
[1.77]
(0.077) 1.050
[1.81]
(0.071) 1.318
[1.58]
(0.114) 1.358
[1.85]
(0.064) 1.309
[1.75]
(0.080) 1.047
[1.83]
(0.067)
0.203
[3.33]
(0.001) .193
[3.45]
(0.001) .183
[2.86]
(0.004) 0.201
[2.90]
(0.004) 0.193
[3.41]
(0.001) 0.181
[2.32]
(0.020)
0.014
[0.11]
(0.915) 0.119
[1.00]
(0.315) -.055
[-0.38]
(0.705) -0.087
[-0.71]
(0.477)
0.041
[1.02]
(0.309) 0.057
[1.31]
(0.190) 0.025
[0.89]
(0.371)
.0504
[2.69]
(0.007) 0.045
[2,24]
(0.025) 0.004
[0.11]
(0.912)
-0.061
[-1.51]
(0.131) 0.016
[0.22]
(0.822)
Diagnostic Tests
m1 -4.08 -3.54 -3.46 -3.18 -4.05 -2.67
m2 -0.14 0.79 1.33 1.08 0.04 0.05
Sargan 0.260 0.674 0.588 0.503 0.520 0.824
Instruments
Column (1) plus
lagged differences of these instruments.
Column (2) plus lagged
differences of
Column (3) plus lagged differences of these instruments.
Column (4) plus
lagged differences of these instruments.
Column (5) plus
lagged differences of these instruments.
Column (6) plus lagged differences of these instruments.
One-step coefficients, robust to autocorrelation and heteroscedasticity are repo
rted. Numbers in brackets are t-values and in parentheses are p-values. The numb
er of observations is 8506 by using an unbalanced panel of 934 firms over 1994-2
005. For the serial correlation tests, test statistics for the null of no first-
order and second-order serial correlation in the first differenced residual are
reported (m1 and m2). All regressions include a full set of year dummies. Instru
ment validity is tested using the Sargan test of overidentifying restrictions; p
-values of the null hypothesis of valid specification are reported.
5.5.4 Uncertainty effect on: Large vs Small size firms.
In order to further magnify the uncertainty effect with respect to the firm size
 . The whole sample of 943 firms is split into two sub-samples according to the
percentage of tradable A shares. The sample split is based on the median value
of tradable A shares. The firms having the average tradable shares more than the
sample median value of A shares are declared to be large firms and the firms wi
th the average tradable shares less than the sample median of A shares are in th
e category of small firms. The descriptive statistics of the main regression var
iables for these two groups are reported in table 5.5.4.1. These statistics repo
rt every best possible aspect of larger firms, having more average investment an
d higher market valuation (as depicted by Q values) in response to greater mean
volatility.
Table 5.5.4.1: Descriptive statistics of subsample of large and small firms
Variables Min Lower quartile
Median Mean Upper quartile Max S.D
Panel A: Big firms Number of firms=786
-.796 .006 0.104 .170 .252 11.042 0.382
-.512 -0.025 3.34*10-13 0.017 0.042 1.586 .111
-6.009 -.019 0.129 0.136 0.30 5.996 0.436
0.093 0.776 1.209 1.674 2.052 16.137 1.453
0.089 0.870 1.073 1.248 1.404 16.001 0.736
Panel B: Small firms Number of firms=158
-0.736 -0.019 0.073 0.118 0.198 2.842 .283
-.494 -.025 2.37*10-13 .039 1.365 .110
-7.910 -.080 .104 0.099 .281 2.983 .537
.174 1.041 1.544 1.891 2.294 12.255 1.356
.171 .823 1.078 1.210 1.402 8.402 0.688

The results are reported only for system GMM estimator because of its superiori
ty.The estimated results for these two groups are reported in the table 5.5.4.2
and table 5.5.4.3. However, it is obvious from the reported estimation results
of these two groups that the uncertainty is more a statistically significant det
erminant for the large firms. The average volatility for the large firms is 1.67
4 and the uncertainty effect on large firms as reported in our estimation result
s ranges from 0.137 to 0.170. Nevertheless the uncertainty has a significant pos
itive effect on the smaller firms also, but is less firm than the larger firms.
The average volatility for the small firms is 1.891 and the uncertainty effects
on the small firms as reported in estimation results ranges from 0.079 to 0.120.
A more vigorous effect on the large firms indicates that lager firms are consid
ered to be more reliable and have brighter chances of further investment. Althou
gh, the average volatility of smaller firms is greater than the larger firms (1.
891>1.671) but the highest positive effect in less than the least positive effec
t on the large firms (0.129<0.137) and this finding reveals that the less volati
le larger firms are more appreciated than the high volatile smaller firms. In sh
ort, the greater the amount of tradable shares the more the prospects of investm
ent.

Table 5.5.4.2 Uncertainty effect in a panel of large firms (1994-2005)

Dependent Variable - (Large Firms)
 1 2 3 4 5 6
-.131
[-.91]
(.364) -2.06
[-1.534]
(.125) -.167
[-1.18]
(0.238) -.255
[-1.84]
(.065) -.170
[-1.25]
(.210) -.147
[-1.01]
(0.313)
-.065
[-.28]
(.783) -.078
[-.43]
(.669) -.072
[-0.40]
(.691) .204
[1.16]
(.245) 0.044
[0.27]
(.786) .334
[2.50]
(0.012)
.144
[1.22]
(.221) .101
[1.16]
(.244) .057
[0.86]
(.391) .180
[1.98]
(.047) .058
[0.91]
(.361) .187
[2.32]
(0.021)

.145
[-1.96]
(.050) -.115
[-1.80]
(.072) -0.065
[-1.54]
(.124) -.157
[-1.73]
(0.084) -.053
[-1.21]
(.228) -.182
[-2.16]
(0.031)
1.313
[1.83]
(.067) 2.209
[3.47]
(0.001) 2.157
[3.03]
(0.002) 1.960
[3.02]
(0.003) 1.90
[2.69]
(0.007) 1.151
[1.34]
(0.179)
.137
[1.95]
(.051) 0.149
[2.35]
(0.019) 0.170
[2.53]
(.011) 0.153
[2.58]
(0.010) 1.64
[2.70]
(0.007) .154
[2.48]
(0.013)
.240
[1.85]
(.064) .151
[1.38]
(.167) 0.107
[1.04]
(0.300) .026
[.23]
(.821) .066
[0.78]
(.438)
.063
[.80]
(.426) .066
[1.32]
(.188) 0.064
[1.15]
(0.249) .016
[0.24]
(0.812)
.046
[2.40]
(.017) -0.022
[-0.76]
(0.446) 0.017
[0.52]
(0.605)
0.044
[0.42]
(.671) 0.027
[0.64]
(0,522)
Diagnostic Tests
m1 -3.55 -4.59 -4.94 -4.41 -3.47 -3.19
m2 0.53 -0.71 -0.88 -1.32 -1.05 -0.34
Sargan 0.970 0.832 0.878 0.917 0.881 0.888
Instruments
Column (1) plus
lagged differences of these instruments.
Column (2) plus lagged
differences of
Column (3) plus lagged differences of these instruments.
Column (4) plus
lagged differences of these instruments.
Column (5) plus
lagged differences of these instruments.
Column (6) plus lagged differences of these instruments.
One-step coefficients, robust to autocorrelation and heteroscedasticity are repo
rted. Numbers in brackets are t-values and in parentheses are p-values. The numb
er of observations is 8506 by using an unbalanced panel of 934 firms over 1994-2
005. For the serial correlation tests, test statistics for the null of no first-
order and second-order serial correlation in the first differenced residual are
reported (m1 and m2). All regressions include a full set of year dummies. Instru
ment validity is tested using the Sargan test of overidentifying restrictions; p
-values of the null hypothesis of valid specification are reported.
Table 5.5.4.3 Uncertainty effect in a panel of small firms (1994-2005)
Dependent Variable - (Small Firms)
1 2 3 4 5 6
-0.681
[-2.77]
(0.006) -.610
[-2.54]
(0.011) -.619
[-2.75]
(.006) -.577
[-2.86]
(0.004) -.675
[-3.12]
(.002) -0.576
[-2.95]
(0.003)
0.230
[1.45]
(0.147) 0.386
[2.35]
(0.019) .383
[2.32]
(.020) .399
[2.40]
(0.016) .224
[2.26]
(.024) 0.149
[1.83]
(0.067)
0.285
[3.80]
(.000) .277
[2.15]
(.032) .277
[2.13]
(.033) .235
[2.39]
(0.017) .355
[3.38]
(.001) 0.259
[2.61]
(0.009)
-.183
[-3.71]
(0.000) -.153
[1.65]
(.099) -.155
[-2.00]
(.045) -0.114
[-2.09]
(0.036) -.187
[-4.39]
(.000) -0.126
[-2.26]
(0.024)
.613
[1.69]
(0.090) .585
[1.65]
(.099) .572
[1.55]
(.121) 0.681
[2.02]
(0.043) .552
[1.41]
(.158) 0.658
[1.95]
(0.051)
0.080
[1.76]
(0.079) .112
[2.55]
(.011) .114
[2.50]
(.012) .120
[2.67]
(.008) .079
[2.03]
(.042) 0.082
[2.45]
(0.014)
-0.005
[-0.07]
(0.948) -.086
[-1.20]
(.231) -.082
[-1.18]
(.237) -.094
[-1.09]
(0.277)
.034
[1.44]
(.151) .035
[1.55]
(.122) 0.040
[1.56]
(0.120)
-.006
[-.20]
(.841) -.021
[-.71]
(0.478) 0.005
[0.39]
(0.696)

-0.044
[-.84]
(0.399) -0.056
[-1.18]
(0.238)
Diagnostic Tests
m1 -0.94 -1.72 -1.48 -1.65 -1.24 -1.69
m2 -1.27 -1.31 -1.44 -1.19 -1.63 -1.47
Sargan 0.256 0.200 0.278 0.536 0.376 0.240
Instruments
Column (1) plus
lagged differences of these instruments.
Column (2) plus lagged
differences of
Column (3) plus lagged differences of these instruments.
Column (4) plus
lagged differences of these instruments.
Column (5) plus
lagged differences of these instruments.
Column (6) plus lagged differences of these instruments.
One-step coefficients, robust to autocorrelation and heteroscedasticity are repo
rted. Numbers in brackets are t-values and in parentheses are p-values. The numb
er of observations is 8506 by using an unbalanced panel of 934 firms over 1994-2
005. For the serial correlation tests, test statistics for the null of no first-
order and second-order serial correlation in the first differenced residual are
reported (m1 and m2). All regressions include a full set of year dummies. Instru
ment validity is tested using the Sargan test of overidentifying restrictions; p
-values of the null hypothesis of valid specification are reported.
Chapter 6: Concluding Remarks
This dissertation contributes to the existing and on-going research/debate on th
e relationship between uncertainty and investment by conducting an empirical stu
dy focusing specifically on the impact of uncertainty on the firm level real inv
estment in the case of China. An unbalanced panel data (1994-2005) which contai
ns financial and accounting information of Chinese companies listed in Shanghai
and Shenzhen stock markets is being used. Based on this panel data set, the unce
rtainty effects on the firm level investment are investigated. We empirically as
sess this relationship in a simple error correction formulation, using a composi
te measure of uncertainty, constructed as the standard deviation of daily stock
returns. We find that a higher level of uncertainty is associated with a greater
pursuit for investment. It provides an empirical evidence to support the idea t
hat uncertainty my have positive link with invest whereas most of the studies re
port a possible negative link between uncertainty and investment. Regardless of
how we estimate it, we get a positive and statistically significant effect for m
easured uncertainty. This positive effect of uncertainty is obtained after contr
olling for short- and long-run investment dynamics and is robust to a host of di
fferent alterations .
In particular, we experiment with a number of different interactions to ensure t
hat our uncertainty variable is not an artifact of omitted investment dynamics.
Contrary to Bloom et al (2001) and Bond and Cummins (2004), there is no evidence
to support the suggestion that firms display a weaker response to demand shocks
at higher levels of uncertainty. As a matter of fact this interaction term is n
ot significant in either direction. Some studies suggest that uncertainty effect
may fade out after the inclusion of average Q in the model, but our observed po
sitive relation is unaffected even when we include it in our preferred specifica
tion. As in preliminary estimations, we found a positive and statistically signi
ficant effect of measured uncertainty with investment the same is observed when
we decompose the uncertainty measure. Of what it’s worth, a decomposition of thi
s term into firm, time, and residual components reveals that it is the firm comp
onent that is the most informative. The same message is delivered when all three
components are included at the same time, the coefficient on firm component rem
ains pronouncedly significant and greater in magnitude.
Since the firm specific component of uncertainty is found to be highly significa
nt, the uncertainty effect on different firm groups e.g. manufacturing vs non-ma
nufacturing and large vs small will be more informative. The uncertainty effect
on sample split between manufacturing and non-manufacturing firms indicates tha
t it holds a pronounced/strong statistical significant effect on non-manufacturi
ng firms rather than manufacturing firms. However, the uncertainty effect on the
sample split of large and small firms reveals that it has more vivid relation w
ith larger firms. Although, the sample average volatility of small firms is grea
ter than the large firms but the estimations indicate that the highest effect on
small firms is less then the least effect on the large firms. The big firms hav
e further chances of expansion and are considered to be more reliable.
In conclusion, uncertainty, at least as measured by the volatility of stock mark
et returns, has a unremitting/pronounced effect on company investment in China.
And the effect is quantitatively important: [TBA: short- and long-run effect of
uncertainty...a 10% increase in our uncertainty measure leads to a ...% rise in
investment]. Clearly, these quantitative effects must be interpreted with cautio
n since they are based on estimations from a reduced-form model.
The empirical results in this dissertation are generally supportive of theories
that predict a positive relationship between uncertainty and investment. Pedagog
ically, convexity of marginal revenue product of capital ///////////////////////
////////////////f the marginal revenue prodc The observed positive relation bet
ween investment and measured uncertainty is consistent with the theories suggest
ing that the marginal revenue product of capital is convex in terms of out put a
nd thus provides and incentive to invest. The relationship is positive if the ma
rginal revenue product of capital is convex, while it is negative if the margina
l revenue is concave. It is, however, conceptually consistent with other stories
as well, such as the those involving risk-averse behavior on the part of manage
rs. But unfortunately, discriminating these interpretations remains outside the
scope of this paper but constitutes a potentially fruitful area for future resea
rch. In the spirit of Hong and Robert (2005) it might that in the case of china
the uncertainty is less than the threshold level that is why it has a positive
effect on investment. But the exact threshold level is difficult to measure afte
r which the relation becomes negative. As a matter of fact “Excess of every thin
g is bad”
The inability to pin down a precise source of uncertainty is partly a consequenc
e of using a more general measure of uncertainty based on stock market returns.
But a composite measure of uncertainty is both a strength as well as a weakness.
If apriori we lack an adequate knowledge of the source of uncertainty most rele
vant for investment, arbitrarily choosing a specific notion of uncertainty may b
e problematic. Not to say that constructing a narrower and more specific measure
of uncertainty remains an uphill task. In this sense, by capturing the market’s
perception of a firm’s value, the composite measure tries to capture all source
s of uncertainty, and is thus a strength. This can be particularly useful when o
ur primary interest lies in establishing the direction of the uncertainty-invest
ment relationship.
Once it is established that, other things being equal, a higher level of uncerta
inty is associated with greater investment, it is pertinent to ask what kinds of
uncertainty matter more than others. This would be particularly relevant for ev
aluating different theories and for guiding effective policy advise. From this p
erspective, a composite uncertainty measure remains a weakness, and offers a fir
st approximation, at best. Constructing more refined measures of uncertainty and
developing convincing empirical frameworks for testing the investment-uncertain
ty links therefore remains an important challenge for future research. Moreover,
a key path forward is to define innovative criteria for cross-sectional heterog
eneity.
A final issue relates to introducing appropriate controls for expected future pr
ofitability. In a reduced-form setting, like the one used in this study, it may
be difficult to fully disentangle the effects of uncertainty from those of expec
ted future profitability. In particular, it is likely that higher uncertainty is
simply proxying for greater pessimism about expected future profitability. Adeq
uately controlling for such expectations remains an important task for future re
search.
The empirical finding reported in this study can be informative for policy makin
g and for more prudent investment. The current risk affection attitude in China
is likely to subdue the potential of real investment and firms might be stuck up
with too much capital. As is evident from the failure stories of other countrie
s a too much investment made without sufficient risk consideration is irreversib
le. A positive effect discovered in our sample might indicate that there in an o
ver-investment in risky projects relying on the government implicit loss protect
ion during the exuberant investment period. An over-investment and quantitative
expansion without sufficient risk consideration may result/lead/ at least a stag
nant growth or at most to some economic crisis. implementation/………….. First to m
ake the best possible use of exact/pricse information at the right/proper time t
o have investment in the right direction. Care must be taken while making too mu
ch positive investment in the case of higher level of investments. The failure e
xample of other east Asian countries is a great lesson for policy maker. Policy
must be made to make investment more prudent and
(Korea) The firms are going /moving making quantitative expansion being unaware
of future risks. This finding suggest that enhancing financial soundness by corp
orate restructuring could reduce the negative uncertainty-investment relationshi
p. The Chinese might not be trapped in over-investment without sufficient risk c
onsideration.
(Xiao)Lastly, even though the uncertainty has a significant positive effect irre
spective of any sample split or sector, but there is no significant in either di
rection in response to change in sales. It can be related to market imperfection
s that it is difficult to obtain reliable significant effect of sales growth.
However, more careful studies must be undertaken to evaluate the demand/sales re
sponse in an interaction combination to see the investment effect.
There are still some puzzles to be solved about the uncertainty and investment d
ynamics in the case of China. Considering how significant positive impact the un
certainty might continue to have on real economic activity in the future. (Dutch
, 2005) A fruitful area for future research may consist of testing the robustnes
s of our results, especially with respect to the construction of the uncertainty
 measure.

Table 2: Description of sample companies

Industry Category Code No of companies
1 Agricultural, forestry, livestock, fishing Industry A 18
2 Mining Industry B 11
3 Manufacturing Industry C 509
4 Electricity, Coal and water Industry D 37
5 Construction E 14
6 Transportation and warehousing F 35
7 Information Technology G 33
8 Wholesale and Retail H 82
9 Real Estate J 28
10 Socila Service Industry K 34
11 Mass Media and Cultural Industry L 8
12 Combined Industry M 67
13 Others 58
Total 934

He also told me how to overcome difficulties and believe in my ideas and myself.
I could not have completed this dissertation without his encouragement and advi
ce.

ACKNWLEDGMENTS
[All the praises and thanks be to Allah, Who has guided us to this, and never co
uld we have found guidance, were it not that Allah had guided us] The Holy Quara
n, Chapter 7, Verses 43.
My thanks first and foremost is to my God, Allah, for his countless blessings th
at He (The Almighty) bestows on me. I ask Allah to help me stay on his straight
path throughout my life and be a thankful slave.
I would like to express my sincere thanks to my thesis supervisor, Prof Dr Wang
Shaoping, who gave his valuable, thoughtful advice and guidance through the cou
rse of my work.
I started this dissertation about three years ago when I was attending a class t
aught by Professor Dr Wang Shaoping. I thank to him for getting me interested in
the topic and methodology .He made me realize that econometrics is not a mere c
omplication of statistical techniques. Without him, I couldn’t have learned how
we should meet reality with econometric theory in such a dynamically unified min
d. He set an excellent example as a researcher and a teacher, provided encourage
ment, and pointed to the right direction. No words describe what a great advisor
Prof Wang has been to me. It would be impossible to ever repay him help in my
life. He is every thing to me.
I am greatly indebted to Prof Dr Xu Chang Xeng , the chairman school of economic
s. I have always been impressed by his pedagogical techniques, responsibility a
nd enthusiasm for academic studies. His class lectures were very informative and
influential and I can never forget.
My special thanks to Dr Adeel Malik for his kind help in solving my research pro
blems and kind help and valuable suggestions from remote place. My deepest thank
s and good wishes for Mr Shamyl Imam, who motivated me for PhD studies.
Thanks go to Mr Kang Ya for his enormous help with data collection; otherwise it
could have been very difficult to start my research.
Special thanks to the staff of New library for providing me substantial support
and some extra privileges in information access.
I would like to express my warmest thanks to those who have given their continuo
us encouragement, including ??????????????????I would not forget my days at Huaz
hong University of Science & Technology and those experiences will combine us fo
rever. My graduate study here has greatly shaped both my intellectual and my per
sonal developed.
Finally, my deepest gratitude goes to my, parents for their unrelenting patient
and support. I will never be grateful to my parents, for their unwavering suppor
t, both moral unconditional love and support. Through out these years, they have
been always available, helped me keep everything in perspective and managed to
keep my spirit up in crises. They have always been a role model of mine.
I would also like to show my highest appreciation to Professor ???????? for his
priceless discussions, comments, suggestions, and, in particular, his strong rea
ding of the dissertation. He himself has tried to prove each analytical result a
s well as main theorems. With his readings, the dissertation becomes more fruitf
ul as well as more rigorous.
I only remain possible for all errors.
Last but not least
.
I want to extend my appreciation to Dr Oyang Hongping for his many insightful c
omments and suggestions.
I would like to all of my friends and colleagues who have helped me stay focused
and have encouraged me to move forward. Much of appreciation is due to Mr ……..
and
In the last my deepest thanks to the Govt and HEC (Higher Education Commission)
of Pakistan for the financial support of graduate study, provided me an opportun
ity to study and enhance my knowledge. Hope to serve the country on my successfu
l return to home.
“East or West, Home is the best”
References
[1] Abel, A. and J. C. Eberly. “A unified model of investment under uncertai
nty,” American Economic Review, 1994, 84, 1369-84.
[2] Abel, A. Optimal investment under uncertainty. American Economic Review,
1983, 73, 228-33.
[3] Abel, Andrew B. and Blanchard, Olivier J.“The Present Value of Profits a
nd Cyclical Movements in Investment” Econometrica, March 1986, 54(2), pp. 249-73
.
[4] Abel, A. and J. C. Eberly, “Investment and q with fixed costs: an empiri
cal anlysis”, mimeo, The Wharton School, University of Pennsylvania, 1996.
[5] Adam, C.S. “Implementing Small-sample Bias Corrections in Dynamic Panel
Data Estimators using STATA.” Mimeo, Center for the Study of African Economies,
University of Oxford, 1998.
[6] Aggarwal, R., C. Inclan and R. Leal (1999): .Volatility in emerging stoc
k markets,. Journal of Financial and Quantitative Analysis, 34, 33-55.
[7] Aggarwal, R., C. Inclan and R. Leal (1999): .Volatility in emerging stoc
k markets,. Journal of Financial and Quantitative Analysis, 34, 33-55.
[8] Agenor, P. (2000): The Economics of Adjustment and Growth, Academic Pres
s, San Diego.
[9] Aizenman, J. and N. Marion.“Volatility and investment: interpreting evid
ence from developing countries,”Economica,1999, 66, 157-79.
[10] Allen, Franklin and Gale, Douglas. Comparing Financial Systems, Cambridg
e, MA: MIT Press, 1999.
[11] Anderson, W. and Hsiao, Cheng. “Estimation of dynamic models with error
components.” Journal of the American Statistical Association, 1981, 76 (375), pp
. 598 -606.
[12] Arellano, M. “A Note on the Anderson-Hsiao Estimator for Panel Data.” Ec
onomics Letters, 1989, 31, pp. 337-341.
[13] Arellano, M. and Bond, S. “Some tests of specification for panel data: M
onte Carlo evidence and an application to employment equations.” Review of Econo
mic Studies, 1991, 58, pp. 277-97.
[14] Arellano, M. and Bover, O. “Another look at the Instrumental-Variable Es
timation of Error-Components Models.” Journal of Econometrics, 1995, 68, pp. 29-
52.
[15] Arellano, M. and S. R. Bond (1998): .Dynamic panel data estimation uing
DPD98 for Gauss: a guide for users., mimeo, Institute for Fiscal Studies, London
(http://www.ifs.org.uk/sta¤/steve_b.shtml).
[16] Arrow, K.J.(1968), “Optimal capital policy with irreversible investment”
, in: J.N. Wolfe, ed., Value, Capital and Growth: Papers in Honor of Sir John H
icks ( Edinburgh University Press).
[17] Atje, Raymond and Jovanovic, Boyan. “Stock Markets and Development,” Eur
opean Economic Review, April 1993, 37 (2/3), pp. 632-40.
[18] Baltagi, B. H and J.M. Griffen, A general index of technical change, Jou
rnal of Political Economy, 1998b (96), 20-41.
[19] Baltagi, B. H , J.M. Griffen and D. Rich,. Airline deregulation: The cos
t pieces of the puzzle International Economic Review, 1995 (36), 245-258.
[20] Bar Ilan, A. and Strange, W. C. The timing an intensity of investment. J
ournal of Macroeconomics, 1999, 21, 57-77.
[21] Barro, Robert. “The Stock Market and Investment.” Review of Financial St
udies, 1990, 3(1), pp. 115-31.
[22] Barnett, S.A. and P. Salellaries,. “Nonlinear response of firm investmen
t to Q: testing a model of convex and non-convex adjustment costs”, Journal of M
onetary Economics, 1998, 42: 261-288.
[23] Bean, C. R.: “An econometric model of manufacturing investment in the UK
,” Economic Journal, 1981, 9 1, 106-121.
[24] Bertola, G. and R.J. Caballaro, “Irreversibility and aggregate investmen
t”, Review of Economic Studies, 1994, 61: 223-246.
[25] Bloom, N., S. Bond, and J Van Reenen: “The dynamics of investment under
uncertainty,” Institute for Fiscal Studies, London, IFS Working Paper,2001, 01/0
5,.
[26] Blundell, R. and Bond, S. “Initial Conditions and Moment Restrictions in
Dynamic Panel Data Models.” Journal of Econometrics, 1998, 87, pp. 115-143.
[27] Blundell, Richard, Bond Steven R., Devereux, Michael P., and Schicntarel
li, Fabio. “Investment and Tobin’s Q: Evidence from company panel data.” Journal
of Econometrics, 1992, 51, pp. 233-257.
[28] Bond, S. and C. Meghir, “Dynamic investment models and the firm’s financ
ial policy”, Review of Economic Studies, 1994, 61:197-222.
[29] Bond, S. and T. Jenkinson: “Investment performance and policy,” Oxford R
eview of Economic Policy,1996, 12 (2), 1-29.
[30] Bond, S. and J. Van Reenen: “Microeconometric models of investment and e
mployment,” Handbook of Econometrics, forthcoming, 2003.
[31] Bond, S. and J. G. Cummins, “Noisy share prices and the Q model of inves
tment”, Working Paper no. W01/22, 2001. The institute for Fiscal Studies, London
.
[32] Bond, S. and J. G. Cummins: “Uncertainty and company investment: an empi
rical investigation using data on analysts’ forecasts,” Unpublished draft, 2004.
[33] Bo, H. and Zhang, Z. (2002). ‘The impact of uncertainty on firm investme
nt: Evidence from machinery industry in Liaong province of China’, Economic Syst
ems, 2002(26), pp. 335–52.
[34] Bond, S., J. A. Elston, J. Mairesse, and B. Mulkay : “Financial factors
and investment in Belgium, France, Germany, and the U.K.: A comparison using com
pany panel data.” Review of Economics and Statistics, 2003, 85(1), 153-165.
[35] Caballero, R.J, “On the sign of investment-uncertainty relationship,” Am
erican Economic Review, 1991, 81(1), 279-288.
[36] Caballero, R.J, E. Engle and J. Hatiwanger, “Plant-level adjustment and
aggregate investment dynamics”, Brooking Papers on Economic Activity 1995 (2):1-
54.
[37] Caballero, R.J and J.V. Leahy, “Fixed costs: the demise of marginal q”,
National Bureau of Economic Research Working Paper no.5508, 1994.
[38] Campos, N. F. and J. B. Nugent (2003): “Aggregate investment and politic
al instability: an econometric investigation,” Economica, 70(279), 533-549.
[39] Chakravarty, Sugato, Sarkar, Asani, and Wu, Lifan. “Information Asymmetr
y, Market Segmentation and the Pricing of Cross-listed Shares: Theory and Eviden
ce from Chinese A and B Shares.” Working paper, Social Science Research Network
(http://www.ssrn.com), July 1998.
[40] Chen, Zhilan and Lin, Shaogong. “State-owned enterprise reform, ownershi
p structure and managerial incentives,” in P. J. Lloyd and Xiao-guang Zhang, ed.
, China in the Global Economy, Cheltenham, UK: Edward Elgar, 2000.
[41] China Statistical Yearbook (various issues*). Beijing: State Statistical
Bureau of the PRC.
[42] Cheng F. Lee, and OliverM. Rui: Does Trading Volume Contain Information
to Predict Stock Returns? Evidence from China s Stock Markets. Review of Quanti
tative Finance and Accounting, 14 (2000): 341-360.
[43] Chesher, A. “The effect of a measurement error”, Biometrica, 1991, 78: 4
51-462.
[44] Cooper, R. W and J.C. Haltiwanger, “On the nature of capital adjustment
costs”, National Bureau of Economic Research Working Paper no.7925, 2000.
[45] Corbett, J. and Jenkinson, T. “The financing of industry, 1970-89: an in
ternational comparison.” Discussion Paper No. 948, Center for Economic Policy Re
search, London, 1994.
[46] Cornwell, C., P. Schmidt and R. C. Sickles,. Production frontiers with c
ross-sectional and time-series variation in efficiency levels, Journal of Econom
etrics, 1990 (46),185-200.
[47] Cummins, J.G, K.A. Hassett and R.G. Hubbard, “A reconsideration of inves
tment behavior using tax reforms as natural experiments”, Brooking Papers on Eco
nomic Activity 1994(2):1-59.
[48] Davidson, J.E.H., D.F. Hendry, F. Srba and S. Yeo. “Econometric modeling
of the aggregate time-series relationship between consumers’ expenditure and in
come in the United Kingdom”, Economic Journal, 1978, 88: 661-692.
[49] Dixit, A. and R. S. Pindyck (1994): Investment under Uncertainty. Prince
ton University Press, Princeton, New Jersey.
[50] Doms, M and T. Dunne, “Capital adjustment patterns in manufacturing plan
ts”, Review of Economic Dynamic, 1998,1(2):409-429.
[51] Enagle, R.F and C.W.J. Granger, “Cointegration and error correction: rep
resentation, estimation, and testing”, Ecnometrica, 1987, 55:251-276.
[52] Eisner, R and Strotz, R. “Determinants of Business Investment,” in Impac
ts of Monetary Policy, Englewood Cliffs: Prentice-Hall, 1963, 59-337.
[53] Elton, Edwin J. and Gruber Martin J. Modern Portfolio Theory and Investm
ent Analysis, NY: John Wiley and Sons, 1995.
[54] Erickson, T. and T.M. Whited, “Measurement error and the relationship be
tween investment and q”, Journal of Political Economy, 2000, 108:1027-1057.
[55] Fernald, John and Rogers, John H. “Puzzles in the Chinese Stock Market.”
Working Paper No. 13, Federal Reserve Bank of Chicago, 2000.
[56] Ghosal, V. and Loungani, P. The differential impact of uncertainty on in
vestment in small and large business. Review of Economics and Statistics. 2000,
82,338-49.
[57] Gould, John P., “Adjustment Costs in the Theory of Investment of the Fir
m,” Review of Economic Studies, January 1968, 35, 47-55.
[58] Greenwald, B., J. E. Stiglitz, and A. Weiss : “Informational imperfectio
ns in the capital markets and macroeconomic fluctuations,” American Economic Rev
iew,1984, 74 (1), 194-99.
[59] Greenwald, B., Stiglitz, J. Macroeconomic models with equity and credit
rationing. In G. Hubbard (ed), Asymmetric Information, Corporate Finance, and I
nvestment. Chicago: University of Chicago Press, 1990, pp.15-42.
[60] Greene, Willaim H. Econometric Analysis, 4th edition, New Jersey: Prenti
ce Hall, 2000.
[61] Guiso, L. and G. Parigi : “Investment and demand uncertainty,” Quarterly
Journal of Economics, 1999,114, 185-227.
[62] Hartman, R. The effects of price and cost uncertainty on investment. Jou
rnal of economic Theory, 1972, 5, 258-66.
[63] Hayashi, Fumio, “Tobin’s average q and marginal q: a neoclassical interp
retation”, Econometrica, 1982, 50: 213-224.
[64] Hayashi, Fumio. “Corporate Finance Side of the Q Theory of Investment,”
Journal of Public Economics, August 1985, 27(3), pp. 261-80.
[65] __________. “Tobin’s Marginal q and Average q: A Neoclassical Interpreta
tion,” Econometrica, January 1982, 50(1), pp. 213-24.
[66] Hermes, Niels and Lensink, Robert. “Banking Reform and the Financing of
Firm Investment: An Empirical Analysis of the Chilean Experience, 1983-1992.” Jo
urnal of Development Studies, Feb. 1998, 34(3), pp. 27-43.
[67] Hertz, Helen. The Trading Crowd: An Ethnography of the Shanghai Stock Ma
rket, UK: Cambridge University Press, 1998.
[68] Hsiao, Cheng. Analysis of Panel Data , New York: Cambridge University Pr
ess, 2003.
[69] Hubbard, R. G. (1994). “Investment under uncertainty: keeping one’s opti
ons open,” Journal of Economic Literature, 32 (4), 1816-1831.
[70] Jorgenson, D.W, “Capital theory and investment behavior”, American Econo
mic Review, 1963, 53:247-259.
[71] King, Robert G. and Levine, Ross. “Financial Intermediation and Economic
Development,” in Colin Mayer and Xavier Vives eds., Financial intermediation in
the construction of Europe, Londin: Centre for Economic Policy Research, 1993a,
pp.156-89.
[72] Kiviet, J. F. “On Bias, Inconsistency, and Efficiency of Various Estimat
ors in Dynamic Panel Data Models.” Journal of Econometrics, 1995, 68, pp. 53-78.
[73] Klevemarken, N.A,. Panel studies: What can we learn from them? Introduct
ion, European Economic Review, 1989 (33), 523-529.
[74] Koop, G. and M.F.J. Steel,. Bayesian analysis of stochastic frontier mod
els, Chapter 24 in B.H. Baltagi, ed., A Companion to Theoretical Econometrics (B
lackwell Publishers, Massachusetts), 2001, 520-537.
[75] Kumbhakar, S. C and C. A. K. Lovell,. Stochastic Frontier Analysis, Camb
ridge University press, 2000.
[76] Leahy, J. V. and T. M. Whited : “The effect of uncertainty on investment
: some stylized facts,” Journal of Money, Credit, and Banking, 1996, 28(1), 64-8
3.
[77] Lensink, R. and Sterken, S. ‘Capital market imperfections, uncertainty
and corporate investment in the Czech Republic’, Economics of Planning, 2000 (33
), pp. 53–70.
[78] Li, Jing. “The Stark Secrets of Investment Funds.” http://www.homeway.co
m (originally published in Finance and Economics), Beijng, 2001.
[79] Li, Kui-Wai and Wong, Kwok-Fai. “The Stock Markets in China,” in Kui-Wai
Li, ed., Financing China Trade and Investment, London: Praeger, 1997.
[80] Li, Xunlei. “Bubbles on the Chinese Stock Market (*).” http://www.cninfo
.com.cn (previously published in Securities Market Herald , August 2001, 8, Shen
Zhen: ShenZhen Stock Exchange).
[81] Lucas, R.E., “Adjustment costs and the theory of supply”, Journal of Pol
itical Economy, 1967, 75:321-334.
[82] Mairesse, J. and B. Dormont, “Labor and investment demand at the firm le
vel: a comparison of French, German and US manufacturing, 1970-79”, European Eco
nomic Review, 1985, 28:201-231.
[83] Moulton, H.R., Random group effects and precision of regression estimate
s, Journal of Econometrics, 1986 (32), 385-397.
[84] Moulton, H.R., Diagnostics for group effects in regression analysis, Jou
rnal of Business and Economic Statistics, 1985 (5), 275-282.
[85] Nickell, S. J. The Investment Decisions of Firms. Cambridge University P
ress, Camgridge, 1978.
[86] Nickell, S. J. “Error correction, partial adjustment and all that: an e
xpository note”, Oxford Bulletin of Economics and Statistics, 1985, 47(2):119-12
9.
[87] Pattillo, C: “Investment, uncertainty and irreversibility in Ghana,” IMF
Staff Papers, 1998, 45(3), 522-553.
[88] Pindyck, R. S. : “Irreversibility, uncertainty, and investment,” Journal
of Economic Literature, 1991, 29, 1110-1148.
[89] Sargan, J. D. “The estimation of economic relationship using instrumenta
l variables.” Econometrica, 1958, 26, pp. 393-415.
[90] Sargan, J. D, “Wages and prices in the UK: a study in econometric method
ology”, in: P.E. Hart, G. Mills and J.K. Whitaker, eds., Econometric Analysis fo
r National Economic Planning (Butterworths, London, 1964).
[91] Sarkar, S. (2000): “On the investment-uncertainty relationship in a real
options model,” Journal of Economic Dynamics and Control, 24, 219-225.
[92] Serven, L. “Irreversibility, uncertainty and private investment: analyti
cal issues and some lessons for Africa,” 1996.
[93] Shaw, E.S. Financial Deepening in Economic Development, New York: Oxford
University Press, 1973.
[94] Shapiro, M.D., “The dynamic demand for capital and labor”, Quarterly Jou
rnal of Economics CI, 1986, 513-542.
[95] Su, Dongwei and Fleisher, Belton M. “Risk, Return and Regulations in Chi
nese Stock Markets.” Journal of Economics and Business, 1998, 50, pp. 239-56.
[96] Summers, Lawrence H. “Taxation and Corporate Investment: A q theory appr
oach.” Brookings Papers on Economic Activity, 1981, 1, pp. 67-127.
[97] Tobin, J., “A general equilibrium approach to monetary theory”, Journal
of Money, Credit and Banking, 1969, 1:15-29.
[98] Wang, Y. et al (2008) Does the stock market affect firm investment in Ch
ina? ..., Journal of Banking & Finance, 2008,VV.XX.
[99] Wong, Michael C.S., Leung, Winson S.C., and Wong, Kwok-Fai. “Risk, Retu
rn, and Price Linkage in China Stocks,” in Kui-Wai Li, ed., Financing China Trad
e and Investment, London: Praeger, 1997.
[100] Wu, Xiaoqiou. “Controversies on the Chinese Capital Markets (*).” Workin
g paper, http://www.cninfo.com.cn, August 27th, 2001.
[101] Xiao, Feng. “Stock Market Development and Financing Patterns of Chinese
Listed Firms.” Working Paper, Economics Department, California State University
at Fullerton, CA, 2003.
[102] Xu, Xiaonian and Wang, Yan. “Ownership Structure, Corporate Governance,
and Corporate Performance: The case of Chinese Companies.” Working Paper, World
Bank, Washington D.C., 1997.
[103] Ying Fan*, Yi-Ming Wei, Wei-Xuan Xu. “Application of VaR methodology to
risk management in the stock market in China” Computers & Industrial Engineering
46 (2004) 383–388.
[104] Yuenan Wang and Amalia Di Iorio. “Are the China-related stock markets se
gmented with both world and regional stock markets?” International Financial M
arkets, Institutions & Money xxx (2007) xxx–xxx.
[105] Zhu, Andong, Ash, Michael, and Pollin, Robert. “Stock Markets and Econom
ics Growth: a Critical Appraisal of the Ross Levine and Sara Zervos Model.” Work
ing Paper, Political Economy Research Institute, University of Massachusetts, MA
, 2002.