Investment and uncertainty in the international oil and gas industry

Investment and Uncertainty in the International

Oil and Gas Industry1

by

Bård Misund† and Klaus Mohn†∗

December 2006

Abstract This paper presents a micro-econometric study of corporate investment and uncertainty in a period of market turbulence and restructuring in the international oil and gas industry. We specify and estimate a modified Q model of investment to test for the influence of uncertainty on capital formation. Based on data for 170 companies over the period 1992-2005, we find that Q is a poor investment indicator

for the international oil and gas industry, and that uncertainty measures contribute significantly to the explanation of investment. However, the sign of the investment uncertainty relationship remains

unsettled, as general and industry-specific risk indicators exert a mixed influence on investment rates. The results suggest that financial market volatility creates a bottleneck for oil and gas investment and

production, whereas oil price volatility has a stimulating effect.

JEL classification: C32, G31, L72

Key words: Capital formation, uncertainty, dynamic panel data models

1 The authors would like to thank Frank Asche and Petter Osmundsen for very useful comments. The usual disclaimer applies. † Both authors are research fellows at the University of Stavanger (Department for Industrial Economics), 4036 Stavanger, Norway. Correspondence: [email protected]. Misund’s web-site: http://www5.his.no/kompetansekatalog/visCV.aspx?ID=08326&sprak=ENGELSK Mohn’s web-site: http://www5.his.no/kompetansekatalog/visCV.aspx?ID=08332&sprak=ENGELSK ∗ Corresponding author.

2

1. INTRODUCTION Over the last 15 years, the international oil and gas industry has moved from a period of stability and predictability to an era of uncertainty, instability and constant change. A wide range of new developments has contributed to a more unforeseeable business environment. Globalisation has progressed rapidly, with accelerating technology diffusion, reduced costs of information and communication, and increasing integration of financial markets. Deregulation and privatisation have caused a stronger focus on financial and operational efficiency in all industries – including oil and gas (Osmundsen et al. 2006). Historically, the management of oil and gas resources has been strongly influenced by national political strategies (e.g., Claes 2001). But over the last two decades, business principles have gradually gained ground in the oil and gas industry. Competition among international oil companies – for increasingly scarce oil and gas resources – has become more aggressive than ever (Weston, Johnson and Siu 1999). A growing mistrust developed between oil companies and the capital markets over the 1990s, as the management of the companies had offered inferior investment returns for several years (Antill and Arnott 2002). Triggered by the temporary oil price collapse in 1998, pressures for operational and financial performance in the short term were coupled with demands for increased dividends to shareholders. A combined result of these developments was a wave of mergers and acquisitions that erased former prominent independent names such as Elf, Fina, Mobil, Amoco, Arco, YPF, Texaco, Phillips, Lasmo – and recently also Unocal. The international oil and gas industry entered a new stage towards the end of the 1990s, with heavy focus on production growth, cost-cutting, operational efficiency and short-term profitability. Huge cashflows were building when the oil price started increasing at the turn of the century, and a large share was returned to shareholders, through comprehensive share buyback programmes and extraordinary dividends. A reflection of the above events and trends is that international oil and gas industry were faced with increasing risks over the 1990s. The mirror image can be traced in relevant commodity and financial markets. As illustrated in Figure 1, a period of oil price stability was succeeded by increasing volatility from the mid 1990s, culminating with the oil price slump caused by the Asian economic crisis in 1998. On average, oil price volatility has been significantly higher over the last 8 years than in the early 1990s. A corresponding development took place in financial markets. In the right-hand panel of Figure 1, we see that a period of stable returns was succeeded by accelerating returns and higher average volatility towards the end of the 1990s. Share prices were inflated by New Economy euphoria (Shiller 2000), which again increased the pressures on under-performing industries and companies. The subsequent restructuring and improvement efforts across the oil and gas industry may well have served as a choke on the accumulation of reserves and production capacity. Indirectly, this econometric study tests how these event may have affected investment, production and oil supply from the international oil and gas companies in the aftermath of the Asian economic crisis.

3

FIGURE 1. OIL AND FINANCIAL MARKET VOLATILITY

Oil price - and price volatility

0

20

40

60

1992 1995 1998 2001 20040

20

40

60Oil price Volatility (rhs)

Stock market - and volatility

0

400

800

1200

1600

1992 1995 1998 2001 20040

10

20

30S&P 500 Volatility (rhs)

Volatility: Annualised standard deviation of daily price/index change (250 days moving data window). See Section 5 for details. Source for oil price and stock market data: Reuters Ecowin (http://www.ecowin.com). This paper presents a study of the impact of uncertainty on corporate investment in a period of market turbulence and industrial restructuring. Investment theory is combined with modern econometric procedures to test the impact of industrial turbulence on total investment expenditures in the oil and gas industry. Applying modern panel data estimators on a data set covering 170 companies over the period 1992-2005, we study the characteristics of investment behaviour in the oil and gas industry. Our approach draws on recent empirical research of the relation between investment and uncertainty. 2 A Q model of investment is derived and augmented with measures for both financial market risk and industry-specific risk. One of our key results is that Q is a poor indicator of investment in the oil and gas industry. In our econometric specification, the Q ratio does not offer the exhaustive explanation for capital formation predicted by theory. The naïve Q model is improved by the inclusion of oil price volatility and stock market volatility. Still, our results are not clear-cut on the sign of the investment uncertainty relationship. General financial market risk is negatively related to investment among the companies in our sample. On the other hand, oil price volatility variable takes a highly significant and positive parameter value. This suggests that the positive convexity effect of uncertainty on investment (e.g., Oi 1961; Hartman 1972; Abel 1983) is dominating for our industry-specific uncertainty indicator, whereas the negative irreversibility effect (e.g., Cukierman 1980; Bernanke 1983; Dixit and Pindyck 1994) is most relevant for overall uncertainty. The paper is organised as follows. Section 2 gives a brief review of previous research on investment and uncertainty, focusing especially on empirical studies. Based on a theoretical Q model of investment in Section 3, our econometric specification is outlined in Section 4. A brief tour of the data set is the subject of Section 5. Estimation procedures and results are presented and discussed in Section 6, before some concluding remarks are offered in Section 7. 2 See Bond et al. (2005) for a recent overview.

4

2. LITERATURE REVIEW The sign of the investment uncertainty relationship has occupied interest among theorists and empirical researchers for many years. A traditional way of thinking has its origin in the properties of neoclassical production technology. Early theoretical contributions (e.g., Oi 1961; Hartman 1972; Abel 1983) stress the implications of convexity of the profit function. This means that any price variation may be exploited for profit maximisation. Accordingly, any increase in uncertainty will raise the marginal valuation of investment, establishing a positive link between capital-accumulation and uncertainty. Further academic interest in theories of investment behaviour was spurred by theoretical work in the early 1980s, when Cukierman (1980), Bernanke (1983), McDonald and Siegel (1985) studied the implications of irreversibility and waiting options for investment decision-making.3 Common for these contributions was the idea that investment could not be reversed. This irreversibility provided the firms with a real option to defer investment. Any increase in the uncertainty around future profitability will increase the value of this waiting option. Accordingly, this strand of literature suggests that investment will respond negatively to increased uncertainty. Following path-breaking theoretical studies of irreversible investments in the early 1980s, applications for natural resources emerged promptly (e.g., McDonald and Siegel 1986; Brennan and Schwartz 1986). However, these early studies were mostly concerned with valuation of development options (e.g., Ekern 1988; Paddock et al. 1988). Econometric studies of historical data appeared somewhat later, both for manufacturing industries, and to some extent for the oil and gas industry. Leahy and Whited (1996) is one of the early, comprehensive econometric studies of uncertainty and investment based on the theory of irreversible investment. In a panel of US manufacturing firms, Leahy and Whited establish a negative link between investment and various uncertainty measures. According to Carruth et al. (2000), subsequent studies confirm a quite robust negative link between investment and uncertainty, with somewhat more clear-cut results for studies of micro data than for aggregate data. The availability of modern panel data techniques has also stimulated a variety of new empirical studies, and most of them are supportive of a negative impact of uncertainty on investment. See Bond et al. (2005) for a recent overview. A central component of empirical studies of the investment uncertainty relationship concerns the measurement of uncertainty. A range of options is available, and no consensus is yet obtained for the appropriate way to proxy uncertainty in empirical models. One approach is to measure uncertainty in terms of dispersion in surveys or published forecasts. Fuss and Vermeulen (2004) applies business sentiment surveys to address price and demand uncertainty, as perceived by company managers in Belgium. Guiso and Parigi (1999) retrieve their measures for demand uncertainty

3 It deserves mentioning that the same irreversibility issues in investment under uncertainty were touched upon even earlier by Henry (1974).

5

from an investment survey among Italian business leaders. Bond et al. (2005) derive proxies for uncertainty from the observed spread in analysts’ earnings forecasts. An alternative method is to incorporate risk measures embedded in the term structure of relevant prices, e.g. interest rates or option prices (e.g., Ferderer 1993). All these methods rely some sort of indirect measurement. A more direct address of the uncertainty indicators facing industry and business leaders has become increasingly popular. The perhaps most common procedure to describe uncertainty is in terms of the historical volatility in explanatory variables, often based on higher frequency data than applied in the econometric investment model. For example, the annual average volatility of stock market returns based on daily data may serve as an uncertainty indicator for company investment in econometric studies on annual company data (e.g, Bond and Cummins 2004). Expected volatility was common for the early studies (e.g. Leahy and Whited 1996, Price 1996), whereby expectations were approached through predictions from ARCH or GARCH models estimated on high-frequency data (e.g, Price 1996). The strength of this approach lies in the explicit address of expected volatility. However, as pointed out Carruth et al (2000), these two-stage strategies also introduce model uncertainty in two stages. Volatility predictions become sens itive to model choice and exhibit substantial variation across alternatives. The validity of the model specification become momentous to the validity of the predictions.4 Furthermore, ARCH and GARCH models estimated on high-frequency data will usually also imply a low persistence of shocks for our purpose. The majority of recent work therefore relies more directly on observed uncertainty measures, and so will our modelling approach. A few econometric studies have also addressed the development of oil and gas fields in the context of irreversible investments, but the empirical findings are mixed for the influence of uncertainty. In a study of UK oil and gas fields, Favero, Pesaran and Sharma (1992) conclude that uncertainty plays an important role for the appraisal lag. Even more interestingly, their results imply a non- linearity in the relation between oil price volatility and the development lag of oil and gas fields. At low oil prices, increased volatility will reduce the development lag. On the other hand, a positive relationship is established at high oil prices. In other words, the irreversibility effect dominates at low oil prices, whereas the convexity effect rules the relationship at high oil prices. Hurn and Wright (1994) find no statistical significance of oil price variability in their investment equations for oil and gas fields in the same region. On the other hand, Mohn and Osmundsen (2006) uncover a negative influence on investment from both oil price uncertainty and geological hazard in an econometric study of exploration activity on the Norwegian Continental Shelf.5 We are not aware of any microeconometric studies of company data that address investment and uncertainty specifically for the oil and gas industry. 4 See also Engle (1983) for instructive reflections on this point. 5 In addition to geological and financial uncertainty indicators, Mohn and Osmundsen (2006) also estimate asymmetric effects from explanatory variables, which is an alternative avenue for studies of investment and uncertainty. When investments respond abruptly to oil price reductions, but only gradual to price increases, capital accumulation (and oil supply) will be enhanced by oil price stability.

6

3. A Q THEORY OF INVESTMENT Tobin (1969) gave rise to one of the currently most popular models of modern empirical investment research. The attractiveness of this specification of investment behaviour has at least two sources. First, the model is simple and has an intuitive appeal. Second, theory implies that Q represents a sufficient statistic for investment, a property that has been tested in a range of empirical applications. The most prominent include studies of financial constraints (Schiantarelli 1996) and uncertainty (e.g. Carruth et al. 2000) in investment behaviour. The focus of this study is the role of uncertainty in oil and gas investments. As a point of departure for our theoretical model, consider the profit function of a representative firm, given by: 6

)],,([)(),,( ttttIttttttt eKIGIpKFpeIK +−=Π [1]

F(Kt) is a well-behaved neoclassical production function, Kt represent the stock of fixed capital, and It is gross investment. Variable inputs are suppressed for simplicity of exposition. The firm puts its product at the market at the price pt. G(It, Kt,et) is an adjustment cost function, pt

I is the price of investment goods, and et is a stochastic shock to the adjustment process. The Q model is subject to a range of standard assumptions of economic theory. The validity of these assumptions are questioned and discussed by Bond et al (2004). To summarise, our model assumes that the firm’s only quasi- fixed input is a single homogenous capital good and no adjustment costs associated with labour and current inputs. We also assume that the absence of financial policy issues and that the firm’s only objective is to maximise returns to risk-neutral shareholders. Moreover, our firm operates in competitive input and output markets with symmetric information, and is allowed to issue unlimited amounts of equity at exogenously given required rates of return. 7 Finally, we assume separability between real and financial decisions (Modigliani and Miller 1958).8 Under these circumstances, the firm’s dynamic optimisation problem can be stated as:

( )

∏= ∑

∞

=+++

0

,,maxi

itititi

tt eIKEV β [2]

6 See Bond and Van Reenen (2005) for a recent overview of microeconometric models of employment and investment. 7 The implication is that internal sources of finance (retained profits) and external sources finance (new share issues) are perfect substitutes. 8 The validity of this assumption is the subject of another influential strand of the empirical investment literature over the last 10 years. See Schiantarelli (1996) for an overview of the early contributions on the investment cash-flow relation. More recent econometric studies include Bond et al. (2003), Baum et al. (2005) and Jin and Jorion (2006).

7

where Vt is firm value, r+= 11β is a constant discount factor,9 and Et[.] denotes the

expected value conditional on information available in period t. Firm value as defined by Equation [2] is maximized subject the usual constraint on capital accumulation:

ttt IKK +−= −1)1( δ [3] where d is an exogenous and fixed rate of depreciation for capital. First-order conditions are given by:

tt

t

Iλ−=

∂Π∂

[4]

1)1( +−−=

∂Π∂

tttt

t EK

λδβλ [5]

where the Lagrange multiplier ( )

111

−∂∂

−=t

tKV

t δλ represents the shadow value associated with capital accumulation as described by the constraint in Equation [3]. Equation [4] sets this shadow value equal to the marginal cost of acquiring additional units capital in period t. The first-order condition [5] describes the relation between shadow values in period t and future shadow values. Linear homogeneity of the profit function implies:

t

tt

t

ttt I

IK

K∂∏∂

+∂∏∂

=∏ [6]

The first-order conditions [4]-[5] can now be combined with Equation [3] to yield:

[ ]tttttt KEK )1()1( 11 δλβδλ −+∏=− +− [7] Solving this equation forward we obtain for the value of the firm:

ti

iti

ttt VEK =

∏=− ∑

∞

=+−

01)1( βδλ . [8]

We now define marginal qt as the ratio of the shadow value of capital to purchase cost

It

t

ptq λ≡ . Hayashi (1982) demonstrates that under linear homogeneity of the profit

function marginal q equals average q.10

9 Irreversibility options and uncertainty over price and demand could imply an increase in the required rate of return. Thus, both the irreversibility theory of investment (e.g., Abel and Eberly 1999) and general CAPM principles (e.g., Mossin 1966) can be embedded in our Q theory of investment. 10 This is equivalent to constant returns to scale, a classical assumption for competitive market equilibria. As pointed out by Bond et al. (2004), failure of the Hayashi conditions will threaten the validity of the average Q model. With imperfect competition or non-convex adjustment costs, Q will no longer be a

8

A useful implication of the Hayashi conditions is that the unobservable shadow value of capital can be related to an observable average q ratio, namely the ratio of total asset value to its replacement cost:

1)1( −−

==t

It

tIt

tt Kp

Vp

qδ

λ [9]

Equation [9] presents marginal qt as the ratio of Vt to the replacement cost value of the capital stock from the previous period. The numerator (Vt) represents the sum of discounted cash-flows from the existing capital stock, and captures changes in current and expected product prices, production plans, unit costs and discount rates. The denominator represents the cost of replacement for the capital stock. This ratio is known as Tobin’s q (Tobin 1969), and suggests that companies with superior investment returns (high q) attract more capital and spend more on capital investment than underperforming firms. We follow the mainstream approach to adjustment costs, (e.g., Bond et al. 2005), and assume a quadratic adjustment cost function:

( ) ttt

tttt Kea

KIb

eKIG2

2,,

−−

= , [10]

where b is a coefficient to represent the importance of adjustments costs. Equation [10] states that the unit cost of capital adjustment is a convex function of the investment rate, that adjustment costs kick in as soon as some investment threshold (a) is passed, and that the investment process is disturbed by stochastic shocks (et). With this specification, the first-order condition in Equation [4] can be modified into a simple relation between investments and the q ratio in company i:

,1

ititit

eQb

aKI

++=

[11]

where Qit = qit-1, and a and b are parameters of the adjustment cost function. 4. ECONOMETRIC MODEL SPECIFICATION Following Bond and Van Reenen (2003), we now include fixed effects (ηi) and time-specific error-components (ζt) in the error-term according to the following structure: eit = ηi + ζt + ε it, yielding for the equation to be estimated:

sufficient statistic for investment rates. For international oil and gas companies, we find these assumptions reasonable, and therefore abstract from further sophistication. Bond et al. (2004) provides a more thorough discussion, as well as suggestions for mitigation.

9

,1

ittiitit

Qb

aKI

εζη ++++=

[12]

According to our theoretical model, Qit should contain all the information that is relevant for investment decisions. Consequently, investment theory implies that Equation [12] represents an exhaustive explanation for capital formation. Following Bond et al. (2005), our econometric approach implies an augmentation of Equation [12], to test the validity of the theoretical model on our data. Implicitly, we question the predicted sufficiency of the Q ratio for investment behaviour in the international oil and gas industry. We also suspect that uncertainty plays a role that is not captured by our theoretical model. The first uncertainty indicator is the volatility of overall stock market returns, measured as the annualised standard deviation of daily returns on the S&P 500 index (σt

m). We will refer to this measure as general market risk – or extrinsic risk. To capture industry-specific or intrinsic risk, we also include the corresponding volatility measure the crude oil price (σt

p).11 For simplicity of exposition, the econometric model is outlined with only one of these volatility measures (σt).12 The modified investment equation is now given by:

,1

ittititit

Qb

aKI

εζηφσ +++++=

[13]

With company data over 14 years, it is hard to believe that the idiosyncratic, time-varying error-component (ε it) is serially uncorrelated. We therefore assume that these shocks follow an AR(1) process of the form: ε it = ρε it-1 + νit, with νit representing white noise. With serially correlated residuals, there are no valid moment conditions to obtain consistent estimates of the model parameters. Taking account of these properties of the error structure, we manipulate Equation [13] to obtain the following dynamic specification of our model:

,)1(

1)1(

11

11

itttitt

itititit

v

Qb

QbK

Ia

KI

+−+−+−+

−+

+−=

−−

−−

ρζζηρρφσφσ

ρρρ

[14]

For econometric purposes, Equation [14] is more conveniently written as:

,**154132

110 ittittitit

itit

vQQKI

KI

+++++++

+=

−−−

ζησπσπππππ [15]

11 In preliminary estimations, we also included a sector-specific indicator of stock market volatility. However the volatility of the S&500 Oil and Gas Index did not contribute favourable to our results. Its coefficients were small and unstable. Moreover, the inclusion of this variable had a disturbing effect on the statistical quality other parameter estimates, and on the general model diagnostics. 12 This may be interpreted as if we collect the volatility measures in the 2x1 vector σt, with φ as the corresponding parameter vector.

10

where:

.,)1(

,,,)(

,,,)1(

**

541

3

1210

tttii

b

ba

ρζζζηρη

ρφπφπρπ

πρπρπ

−=−=

−===

==−=−

−

While the error-term of Equation [13] is serially correlated at all lag- lengths, the error-term of Equation [15] (νit) is now serially uncorrelated. We may therefore obtain consistent estimates of the unrestricted parameter vector π = (π0 , π1, π2, π3, π4, π5). In our estimations, we test a range of model specifications, from the simple static form without uncertainty indicators to the rich dynamic form of Equation [15]. In line with the panel data literature (e.g., Arrellano 2003), we treat the fixed-effects part of the residual (ηi

*) of our model as a stochastic variable. This means that fixed effects will be correlated with the lagged dependent variable ((I/K)it-1). At the same time, we assume that the true disturbances (vit) are serially uncorrelated. The implication is a correlation between the lagged dependent variable and the residual (ηi

* + vit ) which is quite robust to sample size. Standard OLS estimates are therefore

inconsistent – with an upward bias (Bond 2002). The fixed-effects estimator removes this inconsistency through a normalisation of variables around its sample mean for each unit in the data set, whereby the fixed effects and its influence is eliminated from the estimation. OLS is then normally used to estimate these transformed equations. However, a non-negligible correlation prevails between lagged dependent variable and the error term also in the transformed fixed-effects equation, 13 implying that this estimator is also inconsistent – with a downward bias. 14 A variety of instrumental variable techniques have been developed to handle the endogeneity bias of dynamic panel data models. Anderson and Hsiao (1981) propose a first-differenced 2SLS estimator, with instruments that are correlated with the transformed lagged dependent variable and orthogonal to the differenced error term. Lagged levels of the lagged dependent variables are valid instruments to obtain consistent estimates of the dynamic model, according to this approach. In a first-difference GMM framework, Arellano and Bond (1991) show that deeper lags of the dependent variables are also available for the preferred instrument matrix. A challenge with the original Arellano Bond framework is that lagged levels of the dependent variable is a poor instrument for first differences of persistent time series,

13 Consider the simple AR(1) model: yt = αyt-1+(ηi+vt). For this case, the transformed lagged

endogenous variable is )......( 1111

1 −−− ++++− TitiiTti yyyy , whereas the transformed error term is

)......( 1211

TitiiTti vvvv ++++− −− . The component 1−−T

yit of the first term will now be correlated

with vit of the second term. In a similar fashion, 11

−− −

Tv ti of the second term will correlate with yit-1 in the

first term. Other positive correlations will be dominated by these two negative, leading correlations, implying that the correlation between the lagged endogenous variable and the error term in the transformed model is negative (Nickell 1981, Bond 2002). 14 OLS and fixed-effects results for Equation [15] are presented in Appendix 1.

11

and especially for variables that are close to a random walk.15 In meeting this challenge, Arellano and Bover (1995) demonstrate that additional moment conditions may be provided by the inclusion of the original level equations in the system of estimated equations, with significant efficiency gains. This estimator is referred to as the system GMM estimator, and was further developed by Blundell and Bond (1998), and Blundell, Bond and Windmeijer (2000). We take advantage of these developments in the estimation of Equation [15]. See Bond (2002) for a nice guide to modern micro data practice and methods for dynamic panel data models. TABLE 1. DESCRIPTIVE STATISTICS FOR DATA SAMPLE

Obs. Mean St. dev. Min. Max.

(Ι/Κ)it 778 0.183 0.116 0.000 1.108

Qit 778 0.817 1.080 0.000 9.212

σtm 14 15.411 5.971 7.887 26.081

σtp 14 36.16 10.378 20.149 52.573

Source: JS Herold (http://www.herold.com). 5. DATA The estimation sample consists of an unbalanced panel of oil and gas companies for the 1992-2005 period drawn from John S. Herold Company’s (JS Herold) oil and gas financial database.16 We utilise the following data items denoted in USD; market capitalisation, total assets, long term debt, and capital expenditure. In line with previous literature we use the sum of market capitalization of equity and long term debt divided by total assets as our Q measure. Investment (Iit) is measured by capital expenditure and total assets represent capital (Kit). 17 The JS Herold database consists of financial and operating data from annual financial statements of 542 publicly traded energy companies worldwide. To focus on exploration and production (E&P or upstream) activities, we exclude 120 companies involved in domestic pipelines, power production, refining, marketing, chemicals, midstream limited partnerships (MLPs), drilling and oilfield services, coal and other mining. Furthermore, we exclude 193 firms that have ceased to exist due to acquisitions, mergers or other forms for reorganisation. This leaves us with 229 oil and gas companies mainly engaged in exploration and production, resulting in 229x14 = 3206 potential firm years. On the original sample, we apply a number of additional sample selection criteria. First, we exclude non-positive values of total 15 Random walks in differences imply I(1) in the levels. 16 Founded in 1948, John S. Herold Inc. is an independent research firm that specialises in the analysis of companies, transactions, and trends in the global energy industry. Herold serves a global client base with analyses and key financial and operational data on the valuation, performance, and strategy of more than 400 oil and gas companies (http://www.herold.com/ ). 17 We assume that the book value of long term debt is equal to its market value.

12

assets, investment and Q. Second, we leave out observations where firms have undergone substantial changes due to mergers, acquisitions or divestments. To this end, we use the growth rate in total assets as a signal of restructuring, and remove observations where companies show annual growth in excess of 33 per cent. After this screening procedure, the extent of our data set is down to 778 firm years. The application of lags and differences shaves off another 335 observations, taking the number down to 443 for the estimation of the dynamic model. Descriptive statistics for the variables used in the analysis are presented in Table 1. The average investment rate for our sample is about 18 per cent. This is somewhat higher than for other comparable studies, and serves as a clear illustration of the capital- intensive nature of the oil and gas industry. On the other hand, average Q at 0.817 is relatively low compared to prior studies. Our sample consists of international oil companies of a range of sizes. Table 2 reports the descriptive statistics for the explanatory variables for a total of 10 groups in our sample. The groups are defined by JS Herold, and coincide with financial analysts’ definitions (see e.g. USB Warburg, 2001; Deutsche Bank, 2004). TABLE 2. SAMPLE DESCRIPTION BY INDUSTRY GROUP

K I/K Q

N Mean St.dev. Mean St.dev. Mean St.dev.

1 Global Integrated majors 51 119651 45493 0.086 0.019 0.547 0.346

2 International integrated

69 32460 25419 0.109 0.037 0.732 1.187

3 US integrated 27 15788 25164 0.148 0.050 0.300 0.253

4 US Super E&Ps 97 8524 7493 0.189 0.067 0.518 0.344

5 Canadian integrated

39 7088 3536 0.140 0.064 0.678 0.526

6 Large US E&Ps 51 1453 1118 0.212 0.110 0.691 0.572

7 International E&Ps

88 1280 2286 0.143 0.099 1.365 1.416

8 Canadian E&Ps 104 638 895 0.207 0.164 0.979 1.169

9 Mid-sized US E&Ps 75 617 489 0.208 0.107 0.622 0.587

10 Small US E&Ps 177 198 297 0.240 0.127 0.954 1.444

Source: JS Herold (http://www.herold.com). With an average capital base of USD 120 bn, the super majors (Exxon, BP, ChevronTexaco, Total and Royal Dutch/Shell) are by far the largest companies in the oil and gas industry. International integrated and US Integrated companies also belong to the billion dollar club. The dispersion in investment rates (I/K) is relatively modest, ranging from approx. 0.09 to 0.24. The variability in average Q is also quite mild. We find the lowest average Q among US integrateds (0.3), while international E&P companies display the largest ratio of market value to replacement cost of capital (1.4).

13

Carruth et al (2000) survey the variety of approaches to the measurement of uncertainty in empirical investment studies. To address general financial market risk as well as industry-specific risk, we calculate two different measures of aggregate uncertainty. The first is the volatility of the S&P 500 index of the US stock market. The underlying index value is adjusted for divident payments, to portray total investment returns. The second is a corresponding volatility measure for the oil price.18 We illustrate the methodology for the oil price. Based on daily price (pkt) data for the brent blend quality for each of the last 14 years, we calculate annualised standard errors of daily returns (rkt=∆pkt, k = 1, 2 . . N, t = 1992-2005):

( ) ,)(1

1

2

∑=

−=N

kktkt

pt rEr

Nσ [16]

Where N is the annual number of trading days (~ 250) and the average daily change in each year is used as a proxy for E(rkt). Corresponding volatility measures are calculated for S&P 500 index, to produce two annual volatility variables for our model. Daily calculations of these volatility measures with a rolling window of 250 trading days are illustrated in Figure 1. This methodology is according to financial market practice, and in line with previous studies (e.g., Paddock et al. 1988; Hurn and Wright 1994). Oil price volatility has increased from an annualised level of 20-30 per cent during the early 1990s to 40-50 per cent during 1998-2002, when the oil industry witnessed substantial restructuring and a severe oil price drop during the Asian economic crisis (Weston et al. 1999). Since then, oil price volatility has fallen to levels just below 40 per cent, which is slightly higher than average levels from the 1990s.19 From moderate levels around 10 per cent, the volatility of US stock market returns increased gradually towards the 1990s, to levels around 20 per cent at the turn of the century. This volatility increase coincided with the “New Economy” optimism, which created a pricing bubble in large parts of the equity market at the turn of the century (Shiller 2000). After a temporary peak at approx. 25 per cent in 2003, stock market volatility has fallen sharply over the last couple of years, to levels more typical for the 1990s.

18 An alternative is to apply an estimate from an ARCH specification of oil price volatility. However, as argued by Baum et al. (2006), the quality of such measures depends crucially on the validity of that specific empirical model (Engle 1983). 19 A standard result from the literature is that the oil price responds positively to increases in volatility, due to increased demand for storage (Pindyck 2001; Geman 2005). At historically high prices, the recent drop in oil price volatility may therefore seem like an oxymoron. One possible interpretation is that the oil market has gone through a structural shift. In that case, the established new long-term price level may have broken the historical connection between price level and price volatility.

14

6. ESTIMATION AND RESULTS We now arrive at the actual econometric estimation, where the model of Section 4 is confronted with our sample of company data. As a reminder, the purpose of our research design is twofold. First, we set out to establish empirical relations for investment behaviour in the international oil and gas industry, based on Q theory of investment. This implies econometric testing of the validity of our theoretical model on micro data for a range of companies. Second, we study if and how the simple Q model can be improved to provide a better understanding of the investment process. To this end, we regress the investment ratio ((I/K)it) against Q, and augment the relation with uncertainty indicators, in a dynamic model specification as described by Equations [13] and [15]. To capture trend- like investment variation, we also include time dummies for each of the years in our sample.20 We acknowledge that there are potential endogeneity challenges even in our simple framework. First, an endogeneity bias for the lagged endogenous variable of dynamic panel data models is firmly established in the literature (e.g., Arrellano 2003), as discussed above. Second, there may well be endogeneity challenges involved for the Q variable as well. For the last 10 years or so, Osmundsen et al. (2006) argue that oil and gas companies have been rewarded by the equity market for strict capital discipline. Pressures from financial markets have probably put a lid on investments to support the short-term key performance indicators in the oil and gas industry, especially return on average capital employed (RoACE). As market valuation is the numerator of our Q, this variable might therefore also not have all the desired exogeneity properties. Consequently, both the lagged endogenous variable and the Q ratio should be treated as potentially endogenous variables in our estimation framework. The problem of endogeneity has been discussed and addressed in a wide range of areas of the literature on accounting and capital markets, but a consensus on how to address the problem is not yet reached. Nikolaev and van Lent (2005) argue that there is no clear-cut statistic or diagnostic instrument available to test for endogeneity. The general advice from the econometrics literature is to apply introspection (Wooldridge 2002) and reasonableness (Greene 2000; Kennedy 2003) as a way to determine whether there is an endogeneity problem. The standard textbook solution to endogeneity is to apply additional exogenous variables (which by assumption are uncorrelated with the error term) to instrument the suspected endogenous predictor. Unfortunately, independent instrumental variables are usually hard to find for the typical microeconometric study of company data, and our analysis is no exception. However, Arellano and Bond (1991) show that a range of instruments is available in the lagged differences and levels of endogenous and predetermined variables. We therefore rely on the GMM approach suggested by 20 The significance of the time dummies vary from year to year, but collectively they contribute favourably to the quality of the other parameter estimates, as well as general model diagnostics. Interestingly, the time dummies for 1999 and 2000 are negative and highly significant in statistical terms. This suggests that oil and as investments dropped in the aftermath of the Asian Economic Crisis, for reasons not captured by the explanatory variables of our model.

15

Arellano and Bond (1991), and refined by a range of subsequent contributions. See Arellano (2003) and Bond (2002) for an updated overview. As pointed out by Bond (2002), the instruments available for the equations in first differences are likely to be weak when the individual time series have near unit root properties. Differenced unit root variables are close to random walks, offering limited information as instrumental variables. The original Arellano Bond estimator therefore requires autoregressive parameters to be significantly less than one in simple autoregressive specifications. Consequently, the autoregressive structure of the variables of our model is of interest for the choice of estimator. Before we proceed to the estimation of the dynamic Q model, we therefore stop for a moment to check the dynamic properties of our series. We regress a simple AR(1) specification for each of the variables in our model, using the OLS levels and the fixed effects (FE) estimators. Coefficient estimates and p-values for the lagged dependent variable are reported in Table 3. TABLE 3. AR(1) ESTIMATES FOR MODEL VARIABLES

Depvar/ Estimator (I/K)it Qit m

tσ ptσ

OLS 0.523*** (0.000)

0.762*** (0.000)

0.684*** (0.010)

0.302 (0.261)

FE 0.121*** (0.001)

0.436*** (0.000)

*) Significant at 90, **) 95 and ***) 99 per cent confidence level, respectively. p-values in brackets. There is no cross-sectional variation in the volatility indicators, and the FE estimator is therefore not appropriate for these variables. Observe that the number of available observations for the estimation of the autoregressive structure of the volatility indicators is only a small fraction (T=14) of the amount of data available for the company-specific variables I/Kit and Qit (NT = 778). The persistence of the series of our model is highly significant, but the results do not suggest the presence of an exact unit root for any of the variables. Based on these results, we will therefore apply the system version of the GMM estimator (Arellano and Bover 1995). System GMM implies the inclusion of the original level equations in the estimated system of equations. Applying an instrument matrix with deeper lags of both first-differences and levels of the predetermined variables, this procedure provides improved efficiency to the original Arellano Bond framework. The Arellano Bond procedure involves a two-step GMM estimator, whereby a consistent estimate for the weight matrix from a first step is applied to obtain the asymptotically efficient estimator in the second step.21 However, Bond (2002) shows that an asymptotically equivalent estimator can be obtained in one step, using an crude approximation of the two-step weight matrix. The two-step estimator involves

21 All the presented estimations are performed with robust estimators for the covariance matrix.

16

TABLE 4: ESTIMATION RESULTS, DYNAMIC Q MODEL All variables in differences

One-step System GMM Two-step System GMM

Estimated parameters

(I/K)it-1 0.507*** (0.000)

0.507*** (0.000)

0.514*** (0.010)

0.513*** (0.000)

Qit 0.026 (0.158)

0.026 (0.158)

0.025 (0.189)

0.025 (0.189)

Qit-1 -0.018 (0.205)

-0.018 (0.205)

-0.019 (0.211)

-0.019 (0.211)

mtσ

-0.050*** (0.002)

-0.048***

(0.008)

mt 1−σ

0.035*** (0.002)

0.034***

(0.009)

ptσ

-0.010*** (0.007)

-0.010** (0.022)

pt 1−σ

0.032*’* (0.003)

0.030*** (0.008)

Intercept -0.089*** (0.000)

-0.703*** (0.003)

0.086** (0.000)

-0.679** (0.011)

Model diagnostics

Wald χ2 149.10*** (0.000)

149.10*** (0.000)

147.90*** (0.000)

147.90*** (0.000)

Wald χ2 (td) 41.88*** (0.000)

33.57*** (0.000)

42.08*** (0.000)

34.58*** (0.000)

Hansen J 83.25 (0.731.)

83.25 (0.731)

83.25*** (0.472)

83.25 (0.472)

AB AC(1) (0.001) (0.001) (0.001) (0.001)

AB AC(2) (0.146) (0.146) (0.143) (0.143)

Firms 115 115 115 115

Obs (#) 443 443 443 443

*) Significant at 90, **) 95 and ***) 99 per cent confidence level, respectively. p-values in brackets.

efficiency gains in the presence of heteroskedasticity. On the other hand, Arellano and Bond (1991) recommend one-step results for inference on the parameter estimates, as two-step standard errors has a downward bias, especially in small samples. However, Windmeijer (2005) proposes a finite-sample correction of the two-step covariance matrix, showing substantial accuracy gains in Monte Carlo simulations. We present both one-step and two-step results for our model, and the Windmeijer methodology is applied on our two-step covariance matrix estimator.22 22 There is somewhat more to the discussion of strengths and weaknesses of the one-step versus the two-step estimator. For a discussion, the interested reader is referred to Bond (2002).

17

Table 4 presents estimates for four different specifications.23 The results include lagged levels of the endogenous variables, as well as current and lagged levels of all explanatory variables. The Wald χ2 statistic is a test for joint signficiance of all model parameters and Wald χ2 (td) tests the joint significance of the time dummies. Tests for the validity of the over- identifying restrictions are always relevant for GMM estimation of dynamic panel data models. Arellano and Bond (1991) recommended the Sargan statistic to test the exogeneity properties of the instruments as a group, with a null of invalidity. However, the Sargan statistic is sensitive to heteroskedasticity and autocorrelation, and tends to over-reject in the presence of either. We therefore follow follow Rodman’s (2005) advise and report the Hansen J statistic, 24 which is robust. AB AC(n) is the p-value for the Arrellano-Bond test for nth-order autocorrelation in the differenced residua ls, with a null of no autocorrelation. Non-rejection of 1st order autocorrelation is as expected, and not critical for the validity of the differenced equations. 2nd order autocorrelation in the residuals of the differenced model would be more troublesome, as it would imply a breach of the assumption of well-behaved residuals in the level representation of our model, as specified by Equation [15]. Finally, we have included the number of firms and observations for the various estimators. All the four specifications show satisfactory general performance. Tests for joint parameter significance are highly significant and the inclusion of time dummies is strongly justified. Moreover, the difference between the two Wald χ2 statistics suggests that our explanatory variables add vastly to the explanation of the data-generating process. According to the Hansen J statistic, the validity of our over-identifying restrictions can not be rejected for any of the four models. The Arrelano Bond test statistics suggests 1st order autocorrelation in the residuals of our differenced equations, but 2nd order autocorrelation does not seem to be a problem. Observe that the Q ratio does not contribute significantly to the explanation of investment rates in any of the 4 model specifications. The parameter for the current Q ratio takes the right sign, and a plausible value – comparable to similar previous studies.25 However, our estimated models produce estimates for the lagged Q ratio with the opposite sign. This suggests that any impact from Q is partly temporary. However, none of the estimated coefficients are statistically significant, and neither is their sum. Consequently, our results suggest that the Q ratio is a poor indicator for investment in the international oil and gas indus try. On the other hand, both volatility measures take highly significant parameter estimates. Theory is not conclusive on the expected sign of these parameters. The immediate investment response to an increase in stock market volatility is negative, according to our results. However, this negative effect is also dampened over time, due to the positive coefficient of lagged stock market volatility. In contrast to the

23 The models are estimated in Stata 9.0, using the xtabond2 estimator developed by Rodman (2005). 24 The Hansen J statistic is the minimised value of the two-step GMM criterion function (Hansen 1982). 25 See Bond and Van Reenen for an updated overview.

18

results for the Q ratio, the sum of the two coefficients for stock market volatility is negative, and highly significant. For the two-step results, the cumulated effect of an increase in stock market volatility can be calculated to –0.014, with a p-value of 0.008. The implication is that an increase in our annualised volatility measure of 1 percentage point will reduce average the average investment rate by 1.4 percentage points. Thus, we establish a material and highly significant negative effect between oil and gas investments and extrinsic risk. The results are quite different for oil price volatility, our intrinsic (industry-specific) risk indicator. In the short-term, investment rates respond negatively to an increase in oil price risk. More precisely, an increase in oil price volatility of 1 percentage point will reduce the average investment rate by 1 percentage point in the short run. However, the lagged effect of oil price volatility takes a positive sign, and is higher in magnitude than the current effect. The sum of the two effects is 0.022, with a p-value of 0.005. This implies that the longer-term effect of an increase in oil price volatility by one percentage point is an increase in the average investment rate by 2.2 percentage points. Oil price shocks are usually temporary, and oil companies may therefore consider any increase in oil price volatility as a transitory phenomenon. In that case, peaking oil price volatility will usually be followed by a period of stimulating relief, resulting in a positive lagged effect of oil price volatility changes. This is a possible interpretation for the estimated lag structure of the oil price volatility effects on oil and gas investment. Our results suggest that the persistent connection between oil price risk and oil and gas investment is positive, and highly significant in statistical terms. This is is at odds with the bulk of the empirical literature, which reports a negative investment-uncertainty relationship (Carruth et al. 2000; Bond et al. 2006). Traditionally, the oil and gas industry has a prominent place in the investment literature on irreversible investments, and irreversibility is usually connected to a negative relationship between investment and uncertainty. 26 Favero et al. (1994) and Mohn and Osmundsen (2006) are examples of previous econometric studies that have detected a negative influence from uncertainty on investment in the oil and gas industry. However, these studies are devoted to specificl subtypes of investment (field development and exploration, respectively). They also apply aggregated regional data for the United Kingdom Continental Shelf and the Norwegian Continental Shelf, respectively. In our study of aggregate investment among international oil and gas companies, a robust result is that the irreversibility effect of increased uncertainty (e.g., Cukierman 1980; Bernanke 1983; Dixit and Pindyck 1994) is dominated by the more traditional convexity effect (e.g., Oi 1961; Hartman 1972; Abel 1983). Caballero (1991) and Abel and Eberly (1994) stress that uncertainty may not necessarily lead to lower investment, even in the presence of irreversibility. According to Abel and Eberly, irreversibility and uncertainty may work in two

26 Examples include McDonald and Siegel (1986), Brennan and Schwartz, Ekern (1988), Paddock et al. (1988), Favero et al. (1992), Hurn and Wright (1994), Dixit and Pindyck (1994).

19

separate directions. Irreversibility and uncertainty may increase the user cost of capital, which in turn will reduce the desired capital stock. Alternatively, a hangover may pull in the opposite direction. Increased uncertainty tends to increase the expected long-run capital stock because the firm is constrained by previous irreversible investments. Since neither user cost effect nr the hangover effect dominates globally, Abel and Eberly argue that irreversibility may increase or decrease capital accumulation. 7. CONCLUSION At current market conditions, oil plays a crucial role for the prospects of the world economy. Security of supply at acceptable prices is vital for political stability and for continued economic growth – in developing countries as well as in the industrialised part of the world. Accordingly, general public interest in the factors behind oil price formation has increased. Capital formation in the oil and gas industry is important for supply side dynamics of the oil market. This study explores the economic, financial and industrial upheaval in the late 1990s, and their effects on capital formation in the oil and gas industry. Our study may also be looked upon as an empirical test of the different strands of the theoretical literature on fixed investment. Theories of irreversible investment suggest a negative relation between investment and uncertainty, whereas the original neo-classical theories conclude that price variability raise expected profits, and therefore stimulate capital formation. The sign of the investment/uncertainty relationship is therefore left for empirical assessment. We specify and estimate a dynamic Q model of investment, augmented with various uncertainty indicators. The model is estimated on a panel of company data over the period 1992-2005, applying modern GMM techniques for dynamic panel data models. The model performs satisfactory, in terms of parameter significance, specification tests, and general model diagnostics. The insights are twofold. First, the simple Q model does not offer a satisfactory empirical explanation for oil and gas investments. The estimated long-term effect has the expected sign and a plausible magnitude, but is not significant in statistical terms. Thus, our results suggest that the Q ratio is a poor indicator for investment in the international oil and gas industry. The theoretical prediction of Q as a sufficient investment indicator is therefore not supported for our data set. Second, we establish a robust link between investment and two sources of uncertainty. General financial market risk is negatively related to investment among the companies in our sample. On the other hand, oil price volatility takes a highly significant and positive parameter value. A possible interpretation of this result is the revitalisation Oi’s (1961) argument for “the desirability of price instability . . .”. This argument stems from the convexity of the profit function, implying that risk-neutral competitive firms will be able to exploit any of uncertain price outcomes to increase profits compared to a situation with full information on future prices. We study a period of fundamental restructuring and change in the international oil and gas industry. It therefore might seem far- fetched to assume that the same model

20

would hold for all of the 14 years of our sample. One route for further research could be to test for structural shifts during our sample period. Another idea is to look at the impact of volatility on investment in various price situations. In periods of oil price (stock market) increase, volatility may have another effect than in periods with falling oil price (stock market). In a similar fashion, the influence of oil price volatility may also depend on the level of the oil price (stock market index). The presented study is not supportive of the Q model for our data sample. Consequently, the above ideas should preferably be tested in alternative specifications of the underlying investment model.

21

APPENDIX A: ESTIMATION RESULTS, TRADITIONAL ESTIMATORS Dependent variable: (I/K) it

OLS OLS FE FE

Estimated parameters

(I/K)it-1 0.458*** (0.000)

0.477*** (0.000)

-0.007 (0.888)

0.014 (0.792)

Qit -0.007 (0.160)

-0.006 (0.261)

0.007 (0.352)

0.008 (0.242)

Qit-1 0.011** (0.048)

0.009 (0.129)

0.024*** (0.002)

0.022*** (0.004)

mtσ

0.000 (0.669)

0.002**

(0.044)

mt 1−σ

-0.001 (0.187)

-0.001 (0.162)

ptσ

0.001 (0.063)

0.000 (0.477)

pt 1−σ

-0.001* (0.063)

-0.001* (0.006)

Intercept 0.087*** (0.000)

0.087*** (0.000)

0.152*** (0.000)

0.176*** (0.000)

Modell diagnostics

F(*) 55.52*** (0.000)

26.44*** (0.000)

5.33*** (0.001)

4.33*** (0.000)

R2 0.275 0.299 0.047 0.299

Obs (#) 443 443 443 443

*) Significant at 90, **) 95 and ***) 99 per cent confidence level, respectively. p-values in brackets.

22

LITERATURE Abel, A. B. 1983. ”Optimal investment under uncertainty”. American Economic

Review 73: 228-233. Abel, A. B. and J. C. Eberly. 1994. “A Unified Model of Investment Under

Uncertainty”, American Economic Review 84: 1369-1384. Anderson, T. W. and C. Hsiao. 1981. „Estimation of Dynamic Models With Error

Components“. Journal of the Americal Statistical Association. 76: 598-606. Antill, N. and R. Arnott. 2002. Oil Company Crisis, Managing Structure, Profitability and Growth. SP 15. Oxford Institute for Energy Studies. Arellano, M. 2003. Panel Data Econometrics. Oxford University Press. Arellano, M. and S. R. Bond. 1991. “Some Tests of Specification for Panel Data:

Monte Carlo Evidence and an Application to Employment Equations.” Review of Economic Studies 58: 277-297.

Arellano, M. and O. Bover. 1995. “Another Look at The Instrumental Variable Estimation of Error-Component Models”. Journal of Econometrics 68: 29-51. Baum, C.F., Caglayan, M. and O. Talavera. 2006. "Firm Investment and Financial

Frictions". Mimeo, Department of Economics, Boston College. Blundell, R. W. and S. Bond. 1998. “Initial Conditions and Moment Restrictions

in Dynamic Panel Data Models”. Journal of Econometrics 87: 115-143. Blundell, R. W. and S. Bond. 2000. “GMM Estimation with Persistent Panel Data:

An Application to Production Functions.” Econometric Reviews 19: 321-340. Blundell, R. W., Bond, S. and F. Windmeijer (2000).”Estimation in Dynamic Panel

Data Models: Improving on The Performance of The Standard GMM Estimator”. In: Baltagi, B. (ed.) Advances in Econometrics. Vol. 15. Nonstationary Panels, Panel Cointegration, and Dynamic Panels. JAI Elsevier. Amsterdam

Bond, S. 2002. “Dynamic Panel Data Models: A Guide to Methods and Practices”. Portuguese Economic Journal 1 (2): 141-162.

Bond, S., Elston, J. A., Mairesse, J. and B. Mulkay. 2003. “Financial Factors and Investment in Belgium, France, Germany and the UK: A Comparison Using Company Panel Data.” The Review of Economics and Statistics 85 (1): 153-165.

Bond, S. and J. G. Cummins. 2004. “Uncertainty and Company Investment: An Empirical Model Using Data on Analysts’ Profits Forecasts.” Mimeo, Institute for Fiscal Studies.

Bond, S., Klemm, A., Newton-Smith, R., Syed, M. and G. Vlieghe. 2004. “The Roles of Expected Profitability, Tobin’s Q and Cash-Flow in Econometric Models of Company Investment.” Working Paper 222, Bank of England.

Bond, S., Moessner, R. Mumtaz, H. and M. Syed. 2005. “Microeconometric Evidence on Uncertainty and Investment”. Mimeo. Institute for Financial Studies. February 2005.

Bond, S. and J. Van Reenen. 2005. “Microeconometric Models of Investment and Employment”, forthcoming in Heckman, J. and E. Leamer (eds.) Handbook of Econometrics. Vol. 6. Elsevier North Holland.

Bernanke, B. 1983. “Irreversibility, Uncertainty, and Cyclical Investment”. Quarterly Journal of Economics 98: 85-106

23

Brennan, M. J. and E. S. Schwartz. 1985. “Evaluating Natural Resource Investments”, Journal of Business 58 (2): 135-157.

Caballero, R. J. 1991.“On the Sign of the Investment-Uncertainty Relationship“. American Economic Review 81 (1): 279-287.

Carruth, A., Dickerson, A. and A. Henley. 2000. “What Do We Know About Investment Under Uncertainty?“. Journal of Economic Surveys 14: 119-153.

Claes, D. H. 2001. The Politics of Oil-Producer Cooperation. Westview Press. Oxford. UK.

Cukierman, A. 1980. „The Effects of Uncertainty on Investment under Risk Neutrality with Endogenous Information“. Journal of Political Economy 88 (June 1980): 462-475.

Damodaran, A. 2002. Investment Valuation, Wiley Finance. Deutsche Bank. 2004. Major Oils. Annual Assessment of Strategies and

Valuation of The World’s Largest Integrated Oil Companies. Dixit, A. and R. S. Pindyck. 1994. Investment under Uncertainty. Princeton

University Press. Ekern, S. 1988. “An Option Pricing Approach to Evaluating Petroleum Projects”.

Energy Economics 7: 91-99. Engle, R. F. 1983. “Estimates of The Variance of US Inflation Based Upon

The ARCH Model.” Journal of Money, Credit and Banking 15: 286-301. Favero, C. A., Pesaran, M. H. and S. Sharma. 2002. “Uncertainty and Irreversible

Investment: An Empirical Analysis of Development of Oilfields on the UKCS”. Working Paper (EE17), Oxford Institute for Energy Studies, Oxford, UK.

Ferderer, C. A. 1993. ”The Impact of Uncertainty on Aggregate Investment Spending: An Empirical Analysis.” Journal of Money, Credit and Banking 25: 825-849.

Fuss, C. and P. Vermeulen. 2004. "Firms' Investment Decisions in Response to Demand and Price Uncertainty". ECB Working Paper No. 347.

Geman, H. 2005. Commodities and Commodity Derivatives. John Wiley & Sons. England.

Gordon, M. 1962. The Investment, financing and valuation of the corporation. Homewood, IL: Irwin

Guiso, L. and G. Parigi. 1999. “Investment and Demand Uncertainty.”Quarterly Journal of Economics 114: 185-227.

Greene, W. H. 2002. Econometric Analysis. 5th edition. Prentice Hall. Hansen, L. 1982. “Large Sample Properties of Generalised Method of Moments

Estimators”. Econometrica 50 (3): 1029-1054. Hartman, R. 1972. “The Effects of Price and Cost Uncertainty on Investment.”

Journal of Economic Theory 5: 258-266. Hayashi, F. 1982. "Tobin's Average q And Marginal q: A Neoclassical

Interpretation". Econometrica 50: 213-224. Henry, C. 1974. “Investment Decisions Under Uncertainty: The Irreversibility

Effect”. American Economic Review 64 (6): 1006-1012. Hurn, A. S. and R. E. Wright. 1994. “Geology or Economics? Testing Models of

Irreversible Investment Using North Sea Oil Data”. The Economic Journal 104 (423): 363-371.

24

Jin, J. and P. Jorion. 2006. “Firm Value and Hedging: Evidence From US Oil and Gas Producers”. Journal of Finance 61 (2): 893-919.

Kennedy, P. 2003. A Guide to Econometrics. 5th edition, MIT Press. Leahy, J. V. and T. M. Whited. 1996. “The Effect of Uncertainty on Investment:

Some Stylized Facts.” Journal of Money, Credit and Banking 28 (1): 64-83. McDonald, R. and D. Siegel. 1986. “The Value of Waiting to Invest”, Quarterly Journal of Economics 101 (4): 707-727. Modigliani, F. and M. H. Miller. 1958. “The Cost of Capital, Corporation Finance,

and The Theory of Investment. American Economic Review 48 (3): 261-297. Mohn, K and P. Osmundsen. 2006. “Investment, Uncertainty and Asymmetry:

An Application to Oil and Gas Exploration”, Working Paper (October), University of Stavanger.

Mossin, J. 1966. “Equilibrium in a Capital Asset Market”, Econometrica 34: 768-783

Nickell, S. 1981. “Biases in Dynamic Models With Fixed Effects”. Econometrica 49: 1417-1426

Nikolaev, V. and L. van Lent (2005). “The endogeneity bias in the relation between cost-of-debt capital and corporate disclosure policy”. European Accounting Review 14 (4): 677-724.

Oi, W. 1961. “The Desirability of Price Instability Under Perfect Competition”. Econometrica 29 (1): 58-64.

Osmundsen, P., F. Asche, B. Misund and K. Mohn. 2006a. “Valuation of Oil Companies.” Energy Journal 27 (3): 49-64. Osmundsen, P., K, Mohn, B. Misund, B. and F. Asche. 2006b. “Is Oil Supply Choked by Financial Markets?” Energy Policy 35: 467-474. Pindyck, R. S. 2001. “The Dynamics of Commodity Spot and Futures Markets:

A Primer”. Energy Journal 22 (3): 1-29. Price, S. 1996. “Aggregate Uncertainty, Investment and Asymmetric Adjustment in

the UK Manufacturing Sector”, Applied Economics 28: 1369-1379. Paddock, J. L., Siegel, D. R. and J. L. Smith. 1988. “Option Valuation of Claims on

Real Assets: The Case of Offshore Petroleum Leases”. Quarterly Journal of Economics 103 (3): 479-508.

Rodman, D. 2005. “xtabond2: Stata Module to Extend xtabond Dynamic Panel Data Estimator.” Center for Global Development, Washington. http://econpapers.repec.org/software/bocbocode/s435901.htm .

Schiantarelli, F. 1996. "Financial Constraints And Investment: A Critical Survey of The International Evidence". Oxford Review of Economic Policy 12 (2): 70-89.

Shiller, R. 2000. Irrational Exuberance. Princeton University Press. Tobin, J. 1969. "A General Equilibrium Approach to Monetary Theory." Journal of

Money, Credit, and Banking 1: 15-29. UBS Warburg. 2003. Global Integrated Oil Analyzer. Quarterly Assessment of The

Strategies and Valuation of The World’s largest Integrated Oil Companies. Weston, J. F., Johnson, B. A., Siu, J. A. 1999. “Mergers and restructuring in the world oil industry.” Journal of Energy Finance and Development 4: 149-183. Windmeijer, F. 2005. “A Finite Sample Correction for the Variance for Linear Efficient

Two-Step GMM Estimators. Journal of Econometrics 126: 25-51.

25

Wooldridge, J. M. 2002. Econometric analysis of cross section and panel data, MIT Press, Cambridge.