1 Introduction. Mahdi Rastad

Transcription

1 Do Oil Producers Extract More as They Become Richer? A Natural Experiment Using Unanticipated Reserve Discoveries (Work in progress, please do not cite) Mahdi Rastad Abstract The relationship between oil reserves and oil extraction has recently become of interest to research community. Current empirical literature lacks an identification strategy that addresses endogeneity problem. This study takes advantage of (arguably) exogenous reserves discoveries to identify the effect of size of reserves on oil production. Our findings show that although in general oil producing countries increase their production after one year of new reserve discovery, OPEC countries in particular prefer to leave the new stock under the ground possibly because of real option value of the new reserves. 1 Introduction Oil extraction ratio, that is the ratio of oil production to the remaining reserves, is one of the key choices of oil producing countries. The extraction ratio in aggregate determines the world supply of oil which clearly affects both spot and expectations about future oil prices. The inverse of this ratio, oil reserves to annual oil production, shows the number of years a country has until it runs out of oil given its current reserves and production levels. As Pickering (2008) points out, whilst literature on optimal extraction rate of natural resources has a long history going back to Hotelling (1931) the main focus of that literature has been on characterizing pricing path rather than the extraction rate. Pickering (2008) shows in his two-period model with no uncertainty that production, q, is a linear function of remaining reserves, R, regardless of market structure and extraction cost. His empirical analysis seems to support his theoretical finding where he uses panel data of all oil producing countries to estimate the slope of this relationship. His analysis indicates a 1

2 smaller slope for OPEC countries than for non-opec oil producing countries. However, his empirical analysis lacks an identification strategy which is the goal of this paper. The main caveat with estimating extraction slope using time series and cross sectional variation in production and reserves is the endogeneity problem which makes it hard to have a causal interpretation for the relationship. To see this problem, notice that remaining reserves,r is a function of previous production levels, Q, new reserves discoveries, RD, and initial proved reserves R. Therefore the same omitted variable X (say access to technology or political institutions, etc.) that affects previous and current production choices could also affect remaining reserves thorough this relationship as (1) or equivalently (2) illustrates 1 : R it = R it 1 + RD it Q it (1) R it = R it0 + t 1 s=t 0 RD is t 1 s=t 0 Q is (2) This study tries to identify the relationship between oil reserves and its production by exploiting (arguably) exogenous variations in proved oil reserves by unanticipated oil discoveries as in (1). The main questions to answer are (a) how does oil exporting countries production choice is affected by their reserves? Is reserve extraction ratio an increasing or decreasing function of the size? And (b) whether or not this relationship is the same for OPEC and non-opec countries. 2 Empirical Strategy In our analysis we use oil field discoveries events as a natural experiment. Treatment in this experiment is the discovery of a sizeable (for example above 10%) additional oil fields in a country. Figures 5.5 and 5.5 depict the 1 Similarly one can argue for reverse causality problem which again leads into endogeneity issue. Given that reserves discoveries are a function of level of mining and exploration activities which itself depends on investment level in oil sector, one can argue as follows: more oil production increases resources available for investment in oil sector which induces exploration activities and therefore will lead to higher reserves discoveries. We will get back to this point in section 5 2

3 nature and truncated distribution of reserves discovery shocks. We consider the group of countries who enjoy new field discoveries as the treatment group and the ones who did not have any discovery as control group. The identification comes from a difference in difference analysis by comparing the production choice of the two groups before and after discovery of the new reserves.to see if OPEC countries production choice is different than fringe countries we split the sample to OPEC and non-opec and repeat the previous exercise for each of the sub-samples. Statistical comparison of the two estimated slopes then concludes the result. Notice that treatment event may happen in a different point of time for each country.therefore, we normalize the discovery year as t = 0 and compare production choice between the two groups in years before t = 1, 2,..., T with years after t =1, 2,..., T. To put the Diff-in-Diff method into regression framework one can think of the following model specification for oil extraction: ( T T ) log (Q it ) = α s D is + β s D is RD it + γrd it + δx it + µ i + λ t + ɛ it RD it = T i log R it (3), where Q it is the oil production of country i in year t, D is s are dummy variables that switch on s year(s) after new reserves discovery, where s { T,..., 1, 0, 1,..., T }, RD it is interaction of a dummy variable, T i, that switches on if country i had oil discovery at some point of time (is in treatment group) with logarithm of reserves at time t for country i, X it is a vector of other control variables, and finally µ i and λ t are country and year fixedeffects. The interaction form for RD it is chosen to address the intensity of treatment which will be discussed in 5.1) The main parameter(s) of interest in (3) are β s s. If a country adjusts its production linearly to the new level of reserves we would expect to see β s > 0 for s S. In contrast, when s<0 If reserves discoveries are totally unpredicted we would expect β s = 0. However, if discoveries were predictable before it actually happens, then some of the β s s may be significant and positive. Notice that the way we set up D is enables us to deal with series of treatments for an individual country. In that case D is variables may overlap in 3

4 some years for that country. For example consider two consecutive discoveries that happen in 1985 and 1987 for country i. In 1987 both D i0 and D i2 switch on. While D i0 contributes to estimation of β 0, which is the immediate effect of oil discovery on extraction rate, D i2 would contribute to the estimation of β 2 that is the effect of oil discovery in two years ago on current extraction. 3 Estimation 3.1 Data The main source of the date is BP Statistical Review of World Energy which is available on their website. We use annual oil production and reported proven oil reserves for 49 countries from 1980 to Reserves numbers are the remaining oil reserves at the end of each year (net of production in that year) which are with reasonable certainty economically recoverable 2. Out of total 186 reserves shocks for OPEC countries there are 113 positive shocks of which 26 are above 10%. Non-OPEC countries had 213 positive shocks out of total of those discoveries added more than 10% to their reserves. 3.2 Multiple (Repeated) Treatments In cases where there are more than one reserve discovery for a country during the 30 years of sample period we need to find a way to distinguish the effects of each treatment separately. Using a linear model in levels requires a structure for discovery dummy variable to be zero before event and stays on afterwards. So if there is a second discovery this variable is already on and can not capture the effect of the new event unless we add a second dummy variable. One way to solve this problem is to estimate the model in differences instead of levels as in (4). Difference model makes this easy as all we need for the event dummy variable is to switch on in the event year. 2 Current reserve numbers are net of production. We need to correct for this by adding back production numbers to find the gross reserves numbers in the next version of the paper 4

5 log (Q it )= T α s D is + ( T β s D is ) RD i + γ RD i + δx it + µ i + λ t + ɛ it, where RD i is changes in proven reserves if reserves growth is above a threshold (say 10%), and otherwise zero. This variable in fact is an interaction of a dummy variable for discovery of new reserves, and the size of discovery ( RD i = T i log R it ). Note that in a this model fixed effect terms, µ i s, are interpreted as country specific time trends. As mentioned in previous section, the main parameters of interest here are β s s. A similar difference model in (5) is used to answer the second question that is to compare extraction behaviour for the OPEC countries to non- OPEC countries. Notice that country fixed effect term, µ i, is deleted here as it conflicts with OPEC dummy variable. Here the main focus is on θ s that shows the marginal effect of being an OPEC member on reserve extraction. (4) log (Q it )= T α s D is + + ( T ( T 3.3 Choice of Control Group β s D is ) RD i + γ RD i + φop EC+ (5) θ s D is ) RD i OP EC + δx it + λ t + ɛ it For the first model, there are two options for selection of control group: first to use those countries that never had any significant reserve discovery, and second to use the ones that did not have a reserve discovery in the past T years. Unfortunately, the first choice is not possible in here given the sample size: among 49 countries in our sample, there are only 4 and 6 countries with no reserve discovery above 10% and 20%. Therefore we use oil producing countries with no deserve discovery in the past two years as control. For the second model, equation 5, there are three possibilities: first to use non-opec countries who did or did not have a shock, second to use non-opec countries that also had a shock and third to use non-opec 5

6 countries who did not have a shock in that year. We choose the closest control group in this study that is the second one and therefore our focus will be on θ s terms in (5). When RD i > 0, similar to the basic model in (4), reserves extraction rate is measured by β s, and therefore θ s only captures the marginal change in this rate for OPEC versus non-opec shock receivers. 4 Results Estimation results for model (4) is reported in Table 1. The first two columns in this table reflect estimation result with no restriction on sign of reserves shocks whereas in columns three and four only observations with positive shocks in reserves are used in regression. Among the three β s reported in the table β 1 turns out to be significant at 1% level and β 0 is marginally significant at 10% level in the first two columns. They are both positive that is consistent with our a-priori expectations. β 1, the elasticity of reserves extraction with respect to reserves, for positive discoveries is estimated as , that is for every 1% expansion in reserves, oil production increases by 0.73% after one year. Table 2 summarizes the estimation results for equation (5). It shows that OPEC members have both lower production levels as well as lower extraction rates. When sample is split in two parts, we observe that non-opec countries have higher extraction rate of 0.895% comparing to Table 1 and OPEC members have much lower extraction rates. In fact net extraction rate of new reserves for OPEC countries is very close to zero if not negative. This is an important finding that shows reserve extraction behaviour is asymmetric between OPEC and non-opec countries. A reserve discovery in fringe countries causes a significant increase in production after one year, whereas in OPEC countries it has no impact on production even after three years 3. This may be explained (a) by measurement error in reserves data which will be discussed in 5.2, (b) by log-term optimization of OPEC to control oil prices that calls for theoretical research on reserve extraction in a cartel, (c) or by other observable or unobservable characteristics of OPEC 3 In an unreported regression model the original results is found to be robust when the dummy interaction terms for up to three lag years are added to the model. The fourth year interaction terms however shows marginal response for OPEC members to reserve discovery after four years at 10% level 6

7 countries that makes them different from the rest of the sample such as less availability of extraction technology in OPEC countries. The main result of the second part that cartel membership causes lower reserves extraction after controlling for observables which is in line with Adelman (1990) conclusion that lower depletion was the result of collusion to maintain prices. He argues that OPEC countries look at an undeveloped oil reserve is as a real option to the owner which is always better to be postponed under certain condition. This finding is in contrast to the empirical literature on OPEC market power which concludes OPEC does not act as a cartel. (For example see Loderer (1985), Griffin (1985), Gulen (1997) and the response by Kaufmann et al. (2004) ) 5 Discussion on Next Steps 5.1 Treatment Intensity Another challenge for this study is to incorporate the effect of size of the discovery, namely, intensity of treatment (also known as continuous treatment as opposed to binary treatment) that is more of a concern for (3). Here we use RD variable that is an interaction of a dummy variable for a reserve discovery event with size of the discovery. Alternatively, we can follow Angrist and Imbens (1995) 2SLS procedure to correct for estimation of average causal response (ACR). 5.2 Measurement Error In 1986 conducted a study to change the formula for quota. OPEC decided to allocate quotas based on a formula that contained proven reserves as a key variable 4. This gave raise to overreporting incentives among OPEC members to push for a larger quota. Sharp responses in the data can be observed: Iran s reserves went from 50 to 93 in 1986, Iraq s reserves increased from 72 to 100 in 1987 and UAE reserves jumped from 33 to 97 in However, dropping those observations does not change the main estimation result. Other researchers have contributed rapid increase in OPEC reserves 4 Alsalem et al. (1997) conclude that proven reserves, productive capacity, energy factors and domestic investment needs are the main factors in setting quota as a result of 1986 analysis 7

8 between 1986 and 1989 to misreporting of the members to fight for larger quota. However, they fail to compare the pattern to similar reserve expansion in Non-OPEC countries in that period. To investigate whether the results in Table 2 is due to overstatement of reserves by OPEC countries after changes in quota allocation formula, we can compare discovery reporting pattern among OPEC and non-opec countries. Figure? shows a very similar pattern for reserve expansion for OPEC and Non-OPEC countries which is inconstant with quota motivated overreporting story. Any systematic difference in pattern that relates reserves discovery to market conditions can be a sign of inaccuracy of the data. This can also be done before and after changes that happened between 1986 and 1988 in a Diff in Diff format for OPEC vs fringe countries to see if changes in quota setting system in OPEC had any effect on reporting of the new reserves. Table 3 shows the result for this exercise. If there are any systematic differences in reserve growth pattern we expect to observe a positive and significant coefficient for the interaction term between OPEC dummy and after 1986 (or 1988) dummy on the third line. The result clearly is not in support of any difference in reserve growth pattern: none of the interaction terms are statistically significant and in fact 5 out of 6 models are not even economically significant (negative sign for interaction term instead o positive). Therefore we conclude that reserve overstatement story is not supported by the data. 5.3 Matching Estimation The ideal experiment for the first part would be random assignment of the treatment, the oil reserves discovery, to a subset of oil producing countries. However, new oil field discovery is potentially a function of observable country characteristics such as investment in oil sector, availability/cost of technology, geographical location or proximity to oil fields, and so on. While country and time fixed effect terms can control for the factors that are time invariant or have no cross sectional variation, the model still potentially suffers from selection problem. To compensate for this effect we can use propensity score matching procedure using observable factors affecting a new discovery. However, given that here treatment is continuous the matching procedure must use a generalized propensity score method (GPS) as suggested by Hirano and Imbens (2004), Imbens (1999) and Imai and van Dyk (2004) to incorporate intensity of treatment. 8

9 Similarly for the second model, the last point in part 4 is suggestive of using a matching approach to see if there are other underlying drivers for lower extraction behaviour among OPEC countries. In order to answer the question of whether cartel membership causes lowers extraction rates, we need to make sure that this effect is only coming from cartel membership and no other sources. In another words, we need to correct for non-random assignment of the cartel membership. Using a matching estimator is one solution in this case in order to control for the effect of omitted observables that may be the original contributor to asymmetry in behaviour between OPEC members and fringe countries. Some of these factors are pointed out in the literature. For example Mason and Polasky (2005) show that remaining reserves and level of domestic oil consumption determine membership decision of the countries. Matching OPEC countries to oil producing countries outside OPEC with close reserves sizes and domestic consumption then improves the bias in estimation that is due to self-selection problem. 5.4 Dose Response Function: QTE instead of ATE Departing from linear case, one can think of reserves extraction behaviour to differ either for large vs. small producers (in levels or rates) or for different dosages of treatment. For the latter, general propensity score method can be used to estimate a does-response function to discover heterogeneity in production response to reserves discoveries (dosage of treatment). Alternatively for the former, one can look at treatment effect at each quantile, namely, quantile treatment effect (QTE) instead of average treatment effect (ATE). This can be done using quantile regression for extraction level or rate on the same set of covariates as in equation (3). 5.5 Controls Current version of estimation results in Table 1 and 2 do not include any control (variable X it in equation 3 and 5). Although µ i and λ t capture the effect of unobservable time invariant factors and time fixed effects, we still need to include other observable covariates that might affect extraction in those equations. The only factor that affects the slope of the relationship between reserves and extraction in Pickering (2008) s two period model is the discount rate. 9

10 Discount rate that is the weight countries put on their current period vs. next period profits in their optimization problem when deciding over how much to extract now vs. in future. Intuitively, higher discount rates correspond to higher current levels of hunger or equivalently lower levels of patience to wait for future income. Unless this weight is constant over time (and therefore is captured by country fixed effects in estimation) we need controlling variables for it. Possible proxies for discount rate are budget deficit to GDP, external debt to GDP, per capita income level, relative size of oil income in GDP and changes in political structure (for example democracies maybe argued to be less patient). Appendix An alternative specification for model (3) is illustrated in (6). This specification is in particular useful to answer the second part of the first question: the relationship between reserve extraction rate and the reserves size. Q it R it = T α s D is + RD i ( T β s D is ) + γrd i + δx it + µ i + λ t + ɛ it (6), where Q it and R it are the oil production and oil reserves of country i in year t, D is s are dummy variables that switch on s year(s) after new reserves discovery, where s { T,..., 1, 0, 1,..., T }, RD i is a dummy variable that switches on if country i had oil discovery at some point of time (is in treatment group) and remains off if it never had any discovery (is in control group), X it is a vector of other control variables, and finally θ i and λ t are country and year fixed-effects. The main parameter(s) of interest in (6) are β s s. If a country adjusts its production linearly to the new level of reserves and therefore reaches the same extraction rate as before in S years after discovery, one expects to see β s = 0 for s S. In contrast, we we expect β s < 0 for s<sas there will be a mechanical drop in extraction rate in years before S due to increase in denominator of the dependent variable. 5 5 Of course we could have a situation where β s = 0 for s<0. This would be the case if discoveries were predictable. In this case production choices were adjusted to the upcoming level of reserves even before actual discovery happens. 10

11 References M. A. Adelman. Mineral depletion, with special reference to petroleum. The Review of Economics and Statistics, 72(1):pp. 1 10, Ahmad Saleh Alsalem, Subhash C. Sharma, and Marvin D. Troutt. Fairness measures and importance weights for allocating quotas to opec member countries. The Energy Journal, 18(2):1 22, Joshua D Angrist and Guido W Imbens. Two-stage least squares estimation of average causal effects in models with variable treatment intensity. Journal of the American Statistical Association, 90(430): , James M. Griffin. Opec behavior: A test of alternative hypotheses. American Economic Review, 75(5):954, Salih Gurcan Gulen. Is opec a cartel? evidence from cointegration and causality tests. The Energy Journal, 17(2):43 57, K. Hirano and Guido Imbens. The propensity score with continuous treatment. a chapter for missing data and bayesian methods in practise: Contributions by donald rubins statistical family, John Wiley, Inc, Harold Hotelling. The economics of exhaustible resources. Journal of Political Economy, 39:137 75, Kosuke Imai and David A. van Dyk. Causal inference with general treatment regimes: Generalizing the propensity score. Journal of the American Statistical Association, 99: , January Guido W. Imbens. The role of the propensity score in estimating doseresponse functions. NBER Technical Working Papers 0237, National Bureau of Economic Research, Inc, April Robert K. Kaufmann, Stephane Dees, Pavlos Karadeloglou, and Marcelo Sanchez. Does opec matter? an econometric analysis of oil prices. Energy Journal, 25(4):67 90, Claudio Loderer. A test of the opec cartel hypothesis: Journal of Finance, 40(3): , July Charles F. Mason and Stephen Polasky. What motivates membership in non-renewable resource cartels?: The case of opec. Resource and Energy Economics, 27(4): , ISSN

12 Andrew Pickering. The oil reserves production relationship. Energy Economics, 30(2): , March

13 Graphs by region Reserves(Billion Barrels) Year Algeria Angola Indonesia Iran Iraq Kuwait Libya Nigeria Qatar Saudi Arabia United Emirates Venezuela 300 Algeria Angola Indonesia Iran Reserves(Billion Barrels) Iraq Qatar Kuwait Saudi Arabia Libya United Arab Emirates Nigeria Venezuela Graphs by region Year Figure 1: Changes of oil reserves in billion barrels for OPEC countries from 1980 to Data Source: BP Statistical Review of World Energy Density Growth in Reserves 0 5 Density Growth in Reserves.2.4 Figure 2: Distribution of % change in oil reserves that is Gr(R) = Rt Rt 1 R t 1 truncated at 0.4 and 0.4.Data Source: BP Statistical Review of World Energy. 13

14 Reserves(Billion Barrels) Year Non-OPEC Reserves(Billion Barrels) Reserves(Billion Barrels) Year OPEC Non-OPEC Figure 3: OPEC (blue line on left vertical axis) and Non-OPEC (red line on right vertical axis) proven oil reserves trend from 1980 to Data Source: BP Statistical Review of World Energy. 14

15 Dependent Variable: Difference in Log (Oil Production) All Shocks Only Positive Shocks D(t) (.94) (.66) (.48) (.33) Reserves Discovery (.21) (.17) (.60) (.31) D(t)*Reserves Discovery (.10) (.10) (.51) (.27) D(t-1) (.08) (.11) (.67) (.82) D(t-1)*Reserves Discovery (.00) (.00) (.00) (.00) D(t-2) (.04) (.30) (.09) (.46) D(t-2)*Reserves Discovery (.33) (.51) (.17) (.31) Constant (.00) (.35) (.00) (.05) Year Fixed Effect No Yes No Yes Country Fixed Effect Yes Yes Yes Yes N R 2 between withing overall Table 1: Estimation Results for the Effect of Reserves Expansion on Changes in Production. Dependent variable is first difference in Log(Production). D(t s) is a dummy variable which switches on in s years after reserves discovery and off in all other years. Reserves Discovery is equal to first difference in Log(Reserves) if reserves growth is above 10% and otherwise zero. The first two columns in this table reflect estimation result with no restriction on sign of reserves changes whereas in columns three and four only observations with positive changes in reserves are used in regression. Numbers in parenthesis are p-values 15

16 Dependent Variable: Difference in Log (Oil Production) All Shocks Only Positive Shocks D(t) (.27) (.06) (.06) (.02) Reserves Discovery (.00) (.01) (.59) (.22) OPEC (.19) (.09) (.01) (.01) D(t)*Reserves Discovery (.00) (.00) (.41) (.13) D(t)*OPEC (.71) (.40) (.56) (.33) Reserves Discovery*OPEC (.41) (.55) (.96) (.81) D(t)*Reserves Discovery*OPEC (.40) (.48) (.91) (.75) D(t-1) (.19) (.06) (.36) (.73) D(t-1)*Reserves Discovery (.00) (.00) (.00) (.00) D(t-1)*OPEC (.67) (.83) (.16) (.43) D(t-1)*Reserves Discovery*OPEC (.07) (.07) (.02) (.04) Constant (.00) (.61) (.00) (.45) Year Fixed Effect No Yes No Yes Country Fixed Effect No No No No N R 2 between withing overall Table 2: Estimation Results for Comparison of Reserve Extraction of OPEC and Non-OPEC countries. Dependent variable is first difference in Log(Production). D(t s) is a dummy variable which switches on in s years after reserves discovery and off in all other years. Reserves Discovery is equal to first difference in Log(Reserves) if reserves growth is above 10% and otherwise 16 zero. OPEC is a dummy variable for OPEC membership. The first two columns in this table reflect estimation result with no restriction on sign of reserves changes whereas in columns three and four only observations with positive changes in reserves are used in regression. Numbers in parenthesis are p-values

17 Dependent Variable: Growth in Reserves Before and After 1986 Before and After 1988 All Positive > 10% All Positive > 10% OPEC Dummy (.93) (.44) (.80) (.71) (.57) (.66) 1986 Dummy (.87) (.74) (.36) OPEC*1986 Dummies (.84) (.84) (.88) (.40) (.61) (.61) 1988 Dummy (.74) (.60) (.42) Constant (.04) (.01) (.02) (.01) (.00) (.01) N R 2 between withing overall Table 3: Estimation Results for Comparison of Reserve Growth Pattern of OPEC vs. Non-OPEC countries Due to Changes in Quota Allocation Formula. Dependent variable is growth rate in reserves. Dummy 1986 and Dummy 1988 are indicator variables that switch on after each year. OPEC is a dummy variable for OPEC membership. The first three columns in this table reflect estimation result before and after start of changes in OPEC quota system in Similarly, columns 4 to 6 show the results for year Columns 1 and 4 have no restriction on sign of reserves changes, columns 2 and 5 include only observations with positive changes in reserves, and columns 3 and 6 restrict the sample to only sizeable reserve growth (larger than 10%). Numbers in parenthesis are p-values 17