Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
A joint initiative of Ludwig-Maximilians University’s Center for Economic Studies and the Ifo Institute for Economic Research Area Conference on Global Economy 11 - 12 February 2011 CESifo Conference Centre, Munich Foreign Exchange Windfalls, Imports and Exports Torfinn Harding and Anthony J. Venables CESifo GmbH Poschingerstr. 5 81679 Munich Germany Phone: Fax: E-mail: Web: +49 (0) 89 9224-1410 +49 (0) 89 9224-1409 [email protected] www.cesifo.de Preliminary, 1st February 2011 Foreign exchange windfalls, imports and exports: Torfinn Harding and Anthony J. Venables University of Oxford Abstract Foreign exchange windfalls, such as those from natural resource revenues or foreign aid, will change non-resource exports, imports, and the capital account. We study the balance between these responses in order to separate out effects on saving (the capital account), on the domestic tradables sector (net non-resource exports), and on whether the impact on tradables occurs via crowding out exports or drawing in imports. Using data on 134 countries for the period 1975-2007 we show that the response to a dollar of resource revenue is, approximately, to save 35 cents, decrease non-resource exports by 50 cents and increase imports by 15 cents. Foreign aid has a less adverse impact on exports and draws in more imports. We investigate the robustness of these results to alternative model specifications and country groupings. Keywords: natural resources, aid, Dutch disease, trade, exports, imports. JEL codes: E21, E62, F43, H63, O11, Q33 * The work was supported by the BP funded Centre for the Analysis of Resource Rich Economies (Oxcarre), Department of Economics, University of Oxford. We would like to thank Silja Baller for research assistance and Roberto Bonfatti, Ghada Fayad, and Rick van der Ploeg for comments. Address for correspondence: Department of Economics, University of Oxford, Manor Road Building, Manor Road, Oxford OX1 3UQ, UK. 1 1. Introduction Flows of foreign exchange from exports of natural resources or from aid affect the economy through many channels, perhaps the most apparent of which is the balance of payments. Payments balance means that ‘windfall’ flows of this type must cause some combination of changes in other components of the balance of payments, i.e. non-resource exports, imports, or the net accumulation of foreign assets. This paper establishes the form that this combination takes. We analyse data for 133 countries over five decades and, to very broadly summarise results, find that $1 of natural resource exports is associated with approximately a 50c fall in non-resource exports, a 15c increase in imports, and a hence a 65c fall in net nonresource exports. For aid, the negative impact on exports is absent, with the response falling more on imports. We illuminate several distinct issues, the first to do with the Dutch disease effects of such foreign exchange flows. Typical analyses of foreign exchange windfalls point to the structural change they induce, primarily by increasing the size of the non-traded sector of the economy. This structural change may change factor prices and, if the tradable sector has learning by doing or other forms of increasing returns to scale, may have an adverse effect on real income.1 The change in non-resource net exports is the trade counterpart to this structural change. However, if there is no change in net exports (for example because the entire windfall has been saved) then the first link in the causal chain of the Dutch disease is broken. Quantifying the change in non-resource net exports is therefore the essential first step to studying the impact of foreign exchange windfalls on the structure of the domestic economy. 2 We will henceforth abbreviate non-resource net exports to NRB, or non-resource balance. The second issue, and our central focus, is the division of the change in NRB between a change in exports and a change in imports. Typical models of the Dutch disease work with an aggregate tradable goods sector, which is adversely affected by a fall in NRB. With a single tradable good the division of the change in NRB between exports and imports is immaterial (and indeed, undefined). However, we think that this ignores an important part of the effect of windfalls in the economy. Export sectors do different things from import 1 There is an extensive literature on the Dutch disease. Key references are Corden and Neary (1982), van Wijnbergen (1994), Sachs and Warner (1997). See van der Ploeg (2010) for a recent survey. 2 In IMF terminology, a windfall is fully ‘absorbed’ if it induces a one-to-one reduction in NRB, see Gupta and Heller (2002) 2 competing sectors. They have different factor intensities, and possibly also different economies of scale and opportunities for learning. Numerous studies have now established that rapid export growth is a key ingredient of rapid economic growth in many countries. It is therefore essential to break down the change in non-resource net exports into (at least) its two main elements, in order to establish the extent to which a windfall crowds-out exports or crowds-in imports. This, we think, is the main contribution of the present study. Additionally, our evidence on the change in the NRB provides an indication of the amount of saving from foreign exchange windfalls. If $1 of natural resource exports causes a 65c fall in the NRB, then the remaining 35c is accounted for by other elements in the balance of payments. This is primarily a change in aggregate net foreign savings which could take the form of government accumulation of a sovereign wealth fund or foreign debt reduction, private lending or borrowing by nationals, or the de-acquisition of domestic assets (portfolio or direct investment) by foreigners. We therefore provide estimates of net savings which complement those of Davis et al (2002). Our analysis proceeds through two stages. First, we work with long averages and the cross-section of the data. Controlling for other determinants of countries’ trade performance and for endogeneity of aid, we find results summarised in the first row of table 1. Natural resource exports crowd out exports (each $1 of natural resource exports reducing other exports by 41cents), draw in imports (22 cents per dollar), implying 35 cents per dollar is saved. Results for aid indicate a much (and implausibly) higher increase in imports ($2.31 per $1) and also in exports ($1.78 per dollar). The use of long averages has the advantage that adjustment to these foreign exchange flows is likely to be slow, involving the entire production structure of the economy. However, there is the obvious problem that we only control imperfectly for other aspects of countries’ trade performance and do not properly account for the dynamics of the effects. We therefore also analyse a dynamic panel. The long-run effects are summarised in the third row of table 1. These results show a fairly robust pattern, with resource exports, although not aid, having the effect of crowding out other exports. 3 Table 1: Summary responses per unit foreign exchange windfall Natural resource exports Exports Imports NRB Foreign aid Exports Cross-section -0.41 0.22 -0.65 1.78 (table 3) Dynamic panel -0.77 0.21 -0.73 -0.01 (Table 4) Δ NRB = Δ exports - Δ imports; Δ foreign assets = 1 + Δ NRB All cases report IV estimates. Imports NRB 2.31 -0.49 0.35 -0.40 The next section of the paper discusses our framework and data, and sections 3 and 4 of the paper undertake the analysis corresponding to each row of table 1. Section 5 presents extensions, allowing for coefficients to differ across country types and time periods, and distinguishing between primary and manufactured goods. 2. Econometric framework The relationships we investigate take the following general form, Yit/GDPit = a.RESit/GDPit + b.AIDit/GDPit + country controls + time controls + country/ time controls + uit The dependent variables, Y, are the components of the balance of payments that we wish to study, and we look at three main variables. Non-resource exports, X, defined as total exports of goods and services minus resource exports; non-resource imports, M, i.e. total imports of goods and services minus resource imports; and net non-resource exports or non-resource balance, defined as NRB = X – M. Estimates of two of the three equations are enough to back out the third one, but we initially choose to present estimates of all three.3 The key parameters we seek to identify are a and b, capturing the effects of foreign exchange windfalls from net resource exports, RES, and aid inflows, AID. We define resource exports as exports of fuel plus net metals and ores; resource imports are defined analogously, and these definitions are also used in X and M. The variable RES is net resource trade, so contains positive and negative elements. Our model therefore encompasses 3 Since we include the exact same control variables in all the tree models, we do not need to estimate them as seemingly unrelated regressions (SUR). Currently we do not restrict the coefficients to sum up. 4 resource exporters and importers, and in section 6 we allow for different responses for exporters and importers. AID is (gross) development assistance received. We set this at zero for donors and where aid is not reported (we do not consider donation of aid as negative windfalls). All of these trade and income flow variables are normalised relative to GDP. To identify the parameters of interest, the windfalls must be exogenous, i.e. uncorrelated with the error term u given the controls we include. We believe this to be a plausible assumption for net resource exports. The cross section of resource production and exports depend on resource endowments given by nature, and much of the time series variation is determined by price movements exogenous to any single country.4 Aid allocation, on the other hand, is likely to depend partly on economic performance of the recipient countries. If non-resource export and non-resource imports are correlated with the factors aid donors consider in their aid allocation, aid is endogenous. To overcome the potential endogeneity of aid we employ an instrumental variable strategy. In estimations on the cross-sections, we instrument aid with the instrument constructed by Rajan and Subramanian (2008), and we extend this approach in following sections. 5 The details of this are set out in the appropriate sections. The model also contains controls to capture other factors that are important in determining trade volumes. Our approach is to embed our framework in the gravity model, the workhorse model of empirical international trade, by using two sorts of variables. One is own country characteristics. In line with gravity models we use country’s area and a dummy variable if a country is an island; own country GDP is included as the denominator of the trade variables (this restricting its coefficient to unity). The other sort of variables we use are constructed measures of each country’s Foreign Market Access (FMA) and Foreign Supplier Access (FSA). These combine information about each country’ trading partners and the bilateral relationship with each partner (see Redding and Venables 2004). FMA captures the size and proximity of export markets, and FSA captures the size and proximity of import suppliers. We derive these measures from estimating a gravity model on bilateral trade data (appendix 2). Since we are not in this paper interested in bilateral trade flows, these summary measures are the appropriate way to capture all information available from a gravity model that is pertinent to aggregate trade flows. We work with up to 133 countries and cover the period 1960-2007. A full definition of variables and description of data is given in appendix 1. In terms of countries, we aim to be 4 5 For discussion of the exogeneity of resource exports see van der Ploeg and Poelhekke (2010). We would like to thank Rajan and Subramanian for making their data available. 5 as general as possible. Since our windfall measures are defined both for resource netimporters and resource net-exporters as well as for aid recipients and aid non-recipients, our sample is not limited due to windfall characteristics. However, we do some minor other adjustments to the sample. In all estimations we exclude Hong Kong, Singapore and Luxembourg because of their special role as re-exporters. When estimating on panel data , we also exclude Aruba, Central African Republic, Djibouti, Ethiopia, Gambia, The, GuineaBissau, Nepal, Sierra Leone, and Turkmenistan, since in some specifications they would turn up with only one observation each. In terms of time periods we present cross section estimates for the periods 1960-2000 and 1980-2000. In the panel-analysis we employ an unbalanced panel, potentially covering the period 1975-2007. We choose to start in 1975 as previous years are thinly covered. 2007 is the last year for which we have data 3. Long run estimates The relationship we seek to capture reflects, in part, the long run economic structures of the economies under study. Many resource rich economies – Saudi Arabia or Gulf States – have had resource revenues for a long period of time, and have never developed significant nonresource export sectors. Because of this long-run aspect of the issue, we start with crosssection analysis based on long-run averages. Table 2 presents OLS estimates for 133 countries based on averages across 1960-2000 and 1980-2000. Since large changes in the markets for resources and international trade patterns occurred in the 1970s and early 1980s, we include both time periods. 6 Table 2: OLS estimates, cross section OLS estimations 1960-2000 1980-2000 Estimation method: OLS OLS OLS OLS OLS OLS (1) (2) (3) (1) (2) (3) Dep. variable: X/GDP M/GDP NRB/GDP X/GDP M/GDP NRB/GDP RES/GDP -0.354*** 0.351*** -0.705*** -0.409*** 0.322*** -0.735*** (0.083) (0.097) (0.084) (0.074) (0.092) (0.076) AID/GDP -0.975*** -0.384 -0.609*** -1.001*** -0.462* -0.564*** (0.286) (0.340) (0.167) (0.257) (0.262) (0.140) FSA/GDP 0.096*** 0.087 0.010 0.076** 0.104** -0.030** (0.033) (0.054) (0.045) (0.036) (0.044) (0.015) FMA/GDP -0.022 0.014 -0.030 0.016 0.000 0.030 (0.060) (0.097) (0.069) (0.070) (0.090) (0.032) Area -0.058*** -0.063*** 0.007* -0.057*** -0.064*** 0.009*** (0.012) (0.014) (0.004) (0.012) (0.014) (0.003) Island 0.018 0.012 0.011 0.018 0.012 0.013 (0.048) (0.056) (0.013) (0.048) (0.056) (0.014) Constant 1.014*** 1.101*** -0.109** 1.010*** 1.120*** -0.139*** (0.157) (0.180) (0.048) (0.158) (0.182) (0.045) Observations 133 133 133 133 133 133 R-sq 0.48 0.44 0.65 0.49 0.43 0.69 Note: Robust standard errors in parentheses. *, ** and *** denote significance at 10%, 5% and 1%. All variables are averages across 1960-2000. Predicted AID/GDP for 1960-2000 from Rajan and Subramanian (2008) is used as instrument for AID/GDP in column (7)-(9). RS sample refers to their sample. Starting with the effect of resource net exports (RES) on non-resource exports (X), the estimated coefficients -0.35 and -0.41 imply that one extra dollar of resource export revenue decreases non-resource exports by 35 or 41 cents. Moving to the effect of net resource exports (RES) on non-resource imports (M), the estimates suggest that one extra dollar of net resource exports increase non-resource imports by 35 and 32 cents. The fall in non-resource exports and rise in non-resource imports correspond to a weakening of the non-resource current account balance of about 70 cents per extra dollar of net resource exports (to be precise, the estimated coefficients for NRB are -0.71 and -0.74). These cross sectional OLS estimates suggest that resource exports crowd-out other exports and draw in imports in approximately equal measure. As will become clear below, the estimated effect on imports is less robust than the estimated effect on non-resource exports. We return to this issue below. The OLS estimates of the effects of aid are quite different, as higher levels of aid correspond to lower non-resource exports and imports. The coefficient on X is estimated at 0.975 and -1.00, more than twice the corresponding effect of resource exports. The estimated coefficient for non-resource imports is negative in both samples (-0.38 and -0.46), but it is 7 significant only for 1980-2000. The effect on the NRB is however still estimated to be negative, with the coefficients -0.61 and -0.56.6 The negative estimated effects of aid on non-resource imports in Table 2 bring to attention an important concern with these OLS estimations. If there is a third factor not included in our model but correlated with aid and non-resource trade, then we may have an endogeneity problem. In particular, aid may be given to countries that have limited capacity for producing exports and generating foreign exchange. There will then be a negative correlation between exports (and correspondingly imports) and aid donations, creating a negative bias in OLS. As expected, the naive OLS estimates between non-resource imports and aid presented in Table 2 suggest a relatively low level of exports and imports in aidrecipients compared to other countries. Rajan and Subramanian (2008) address potential endogeneity of aid in an empirical model of economic growth in a cross section of countries. In our cross sectional estimates, we will use their instrument to overcome the potential endogeneity problem of aid in our specification. The instrument constructed by Rajan and Subramanian (2008) aims to capture the variation in the supply of aid that is driven by bilateral links between the recipient and the donor country, in contrast to the characteristics of the recipient country itself (like income or growth, or more relevant in our case, exports and imports). The instrument is constructed by estimating an aid-supply function on bilateral data, where the explanatory variables are past and present colonial relationships, common language, and the population size of the donor relative to the recipient.7 The predicted values from this aid-supply function contain variation driven only by these (non-economic) bilateral characteristics, which are assumed to be exogenous to the economic conditions in the recipient economy. The final step in constructing the instrument is to aggregate the predicted values across all donors for each recipient country. 6 The estimates are stable if we instead average over 1970-2000 or 1990-2000 (these results are not presented to save space). 7 See table 3 in Rajan and Subramanian (2008) for the functional form of their aid supply model, The dependent variable in their aid supply model is aid/recipient GDP 8 Table 3: IV estimates, cross section Estimation method: Dep. Variable RES/GDP AID/GDP FSA/GDP FMA/GDP Area Island Constant Observations R-sq IV (7) X/GDP -0.409*** (0.149) 1.778 (1.650) -0.234 (0.357) 0.087 (0.211) -0.031** (0.015) 0.088 (0.075) 0.556** (0.221) 69 -0.11 Panel A: IV estimations 1960-2000 IV IV (8) (9) M/GDP NRB/GDP 0.225 -0.650*** (0.171) (0.065) 2.311 -0.493 (1.554) (0.370) -0.784** 0.537*** (0.326) (0.091) 0.635*** -0.533*** (0.191) (0.069) -0.034** 0.006* (0.014) (0.004) 0.088 0.015 (0.070) (0.017) 0.620*** -0.113** (0.207) (0.052) 69 69 -0.10 0.80 IV (7) X/GDP -0.381** (0.187) 3.265 (2.156) -1.330 (1.299) 1.061 (1.102) -0.024 (0.023) 0.149 (0.132) 0.418 (0.343) 71 -0.94 1980-2000 IV IV (8) (9) M/GDP NRB/GDP 0.264 -0.660*** (0.175) (0.052) 4.254** -0.901** (2.158) (0.393) -1.931 0.545** (1.273) (0.274) 1.615 -0.489** (1.082) (0.235) -0.024 0.005* (0.024) (0.003) 0.172 -0.001 (0.139) (0.015) 0.427 -0.081** (0.362) (0.038) 71 71 -1.12 0.84 Panel B: First stage estimation. Dependent variable: AID/GDP Fitted AID/GDP 0.522** 0.522** 0.522** 0.509** 0.509** 0.509** (0.197) (0.197) (0.197) (0.201) (0.201) (0.201) F instr. 7.05 7.05 7.05 6.40 6.40 6.40 Part. R-sq Shea 0.17 0.17 0.17 0.07 0.07 0.07 Underid. test p 0.02 0.02 0.02 0.01 0.01 0.01 Note: Robust standard errors in parentheses. *, ** and *** denote significance at 10%, 5% and 1%. All variables are averages across 1980-2000. Predicted AID/GDP for 1980-2000 from Rajan and Subramanian (2008) is used as instrument for AID/GDP in column (7)-(9). RS sample refers to their sample. Table 3 presents IV results estimated on the two cross-sections 1960-2000 and 19802000. The samples are reduced to 69 countries and 71 countries, due to the availability of the instrument. All countries that have received aid are in the sample (Rajan and Subramanian, 2008). OLS estimates on the sample of Rajan and Subramanian, presented in appendix 3, are similar to the OLS estimates on the full sample. Before discussing the results, let us focus on the validity of the instrument. In principle, the instrument of Rajan and Subramanian (2008) should be a valid instrument also in our setting. First, in their paper they show that it is strongly correlated with AID/GDP and this translates directly to our setting. The relevance of the instrument is confirmed by the significant coefficient in the first stage (presented in panel B of Table 3), which in our case of one instrument maps directly into an acceptable Fstatistics. The Shea R-squared statistic is also reasonable and the underidentification test (based on the Kleibergen-Paap rk LM statistic) is comfortably passed. Second, the exclusion restriction, i.e. that the instrument has an effect on the outcome only through the endogenous regressor given the controls included, should also be satisfied. The variation represented by 9 the instrument is variation in bilateral links between countries, in addition to their relative population size. In the gravity model of international trade, such bilateral links are included and are found to have some explanatory power empirically. Since we include such bilateral links (common official language and colony dummies) in the gravity estimation that forms the basis for FMA and FSA (see the appendix), the trade effects of the variation captured by the instrument should be properly controlled for by FMA and FSA. Since we have got only one instrument, we cannot test the exclusion restriction. Looking first at the effect of resource exports, RES, the IV estimates of table 3 produce broadly similar estimates to the OLS of table 2, suggesting that each $1 RES reduces X by 40c, raises M by 25c and reduces NRB by 65c. The coefficients on M and NRB are somewhat smaller (each falling by 5-10 percentage points) and statistical significance is lost on the import coefficient, in part because the IV-estimator has lower efficiency than the OLSestimator in small samples. Instrumenting has a much larger effect on the estimated aid coefficients, as can be seen by comparing the OLS estimates of Table 2 with Table 3. We expected some increase in the coefficients for X and M, and find that the increase is large. Aid now draws in additional imports, but much more than one-for-one; the coefficient on M goes from -0.46 to 4.25 (for the 1980-2000 period). Correspondingly, the coefficient on exports goes from -1 to 3.26, suggesting that instead of crowding out exports aid is now associated with a dramatic increase in exports. The size of the export and import coefficients may seem unreasonably large, although they correspond to a reasonable effect on NRB, with each dollar of aid leading to an estimated significant decrease in NRB of 90 cents. Taken together, these estimates suggest that about one-third of one extra dollar of resource export revenue is saved internationally. Of the two-thirds of the extra dollar that are spent, at least half is allocated to a decrease in non-resource exports. These results are consistent with standard models, in which the foreign exchange windfall allows the economy to shift resources from sectors producing tradable goods to sectors producing non-tradable goods. Due to the windfall, a certain level of consumption of tradable goods can be maintained with a smaller tradable good production. 4. Short and long run effects In this section we open up the time dimension of the data. This has several advantages. It increases statistical power by dramatically increasing the number of observations, from less 10 than 100 to more than 2500. It also allows us to learn about the importance of the variation across countries versus the variation within countries across time. The cross sectional estimation presented in the previous section shuts down the time-variation, whereas a panel fixed effect estimation relies only on the within-country variation. By comparing the two we can get a better insight on what is driving our results. Perhaps most importantly, the panel setting allows us to control for unobservable characteristics of countries and hence reduces the possibility that our results are driven by omitted variable bias. A final advantage with the panel-setting is that we can investigate the dynamics of the estimated effects. Before we present the panel results, we first present our aid-instrument for the panel setting. Even though we will include country-fixed effects to control for all time-invariant determinants of aid and trade in our panel estimates, the timing of aid inflows can of course be endogenous. We therefore need an instrument representing exogenous variation in aid inflows over time. Since the donor-recipient characteristics included in the aid-supply function of Rajan and Subramanian (2008) are of a cross sectional nature, we construct the following instrument, Zid, for our estimates on panel-data: N zidt sid aidt i 1 sid is the average share of aid from donor d to recipient country i over the period 1960-2008. aidt is aid from donor d to recipient i in year t. zidt is then the aid from donor d to recipient i as is predicted by the sum of aid from donor d in year t and the historical average share allocated to country i. We employ aid data for 22 donors and the final instrument for aid flows to country i in year t becomes: 22 Zit zidt d 1 Since we use bilateral data to construct Zid (sid varies by donor), Zid contains country-specific time-variation. The idea behind this instrument can be laid out as follows. The time-variation in zidt is driven by the overall aid budget of donor d, which is likely to be determined by national budget considerations and economic features of the donor economy rather than features of the recipient economy. The share sid is a long historical average and should be exogenous to 11 yearly changes in non-resource trade. Furthermore, since we always include country fixed effects in our panel regressions, only the time variation in our instrument will be used to identify the parameters. The behaviour of the instrument will be discussed in relation to the estimation results below. On the dynamics between windfalls and non-resource trade, theory suggests that a number of factors will affect the speed at which the economy adjusts to windfalls. One is the extent to which a windfall is expected to be permanent or temporary; if temporary, adjustment might not be to the full annual value of the windfall, but only to its permanent income equivalent. A second concerns the speed with which the exchange rate changes, and with which this feeds into the domestic relative price changes that guide the allocation of resources in the economy. A third is to do with the speed with which the quantity side of the economy can adjust (see Van der Ploeg and Venables 2010). It is certainly the case that we do not expect to see full results come through instantaneously. To estimate dynamics we apply a standard autoregressive distributed lag (ADL) model. We present results for the simplest specification, ADL(1,1): Yit/GDPit = ρYit-1/GDPit-1+ a0RESit/GDPit + a1RESit-1/GDPit-1 + b0AIDit/GDPit + b1AIDit-1/GDPit-1 + country controls + time controls+ country/ time controls + uit i.e. compared to a standard static model, a one-period lagged dependent variable and oneperiod lags of the windfall measures are included. Our results hold through also in specifications with more lags included. 8 The dynamic specification makes us able to separate out short and long run effects. We choose not to present dynamic estimates without country FE, as the dynamic interpretation of interest is the one within country. Before estimating with time-variation we checked the dynamic properties of our variables. In four different panel unit root tests – results not presented to save space – we 8 Harrigan (1997) and Santos-Paulino and Thirlwall (2004) also apply the ADL(1,1) framework with countryfixed effects to estimate dynamic trade models. Nickell-bias (Nickell 1981), where a correlation between the lagged dependent variable and the country fixed effect causes an under-estimation of coefficient of the lagged endogenous, is a concern in short panels. Judson and Owen (1999) compare the performance of different estimators in terms of Nickell-bias. In an unbalanced panel with T=30, they recommend to use the fixed effect model without correction, which is what we do,. 12 could reject a unit root in all of our variables, with the exception of the control variable FSA/GDP, for which country-specific trends were needed to reject a unit root. Given these unit root results we treat our variables as stationary.9 The first three columns in Table 4 present FE estimates of the ADL(1,1) model. Starting out with the long-run coefficients presented in the lower part of panel A, the coefficients are similar to what we found in the static models. Net non-resource exports are reduced by 73c and 61c for an extra dollar of net resource and aid windfall. The response to resource exports is a 67c crowding out of other exports and 24c of additional imports. The response to aid is a 29c (and not significant) crowding out of exports and 94c pulling in additional imports. It is reassuring that the estimates obtained from the cross-sectional data and the dynamic panel are similar. 9 The four panel unit root tests employed are: Levin, Lin & Chu t*, Im, Pesaran and Shin W-stat, ADF - Fisher Chi-square and PP - Fisher Chi-square. For all tests, automatic lag length selection based on AIC criterion, Newey-West automatic bandwidth selection and Bartlett kernel was used. All tests were done with country fixed effects and with country fixed effects and country-specific trends. In terms of adjustment of standard errors, we always estimate with robust standard errors to take heteroskedasticity into account. For robustness we have also re-estimated all the models with standard errors that in addition are robust to serial correlation (option bw(2), in Stata), and it does not change the results. The results from the unit root tests and the robustness checks are available on request from the authors. 13 Table 4 Observations Countries R-sq Panel A: FE and IV estimations (1) (2) (3) (4) FE FE FE IV X/GDP M/GDP NRB/GDP X/GDP 0.818*** 0.772*** 0.652*** 0.833*** (0.031) (0.030) (0.028) (0.028) -0.553*** -0.120** -0.440*** -0.555*** (0.089) (0.057) (0.110) (0.093) 0.431*** 0.175*** 0.186* 0.427*** (0.087) (0.064) (0.108) (0.087) 0.163** 0.399*** -0.250*** 1.209*** (0.063) (0.085) (0.069) (0.351) -0.109*** -0.185*** 0.036 -1.210*** (0.033) (0.067) (0.067) (0.416) -0.004 -0.017*** 0.015** -0.001 (0.004) (0.004) (0.006) (0.006) 0.004 0.057*** -0.058** 0.012 (0.018) (0.020) (0.029) (0.023) 2374 2374 2374 2374 126 126 126 126 0.78 0.71 0.58 0.69 Long run coeff.: Res Long run P-val: Res Long run coeff.: Aid Long run P-val: Aid -0.67*** 0.00 0.29 0.33 Estimation method: Lagged dependent var. RES/GDP L.RES/GDP AID/GDP L.AID/GDP FSA/GDP FMA/GDP 0.24* 0.06 0.94*** 0.00 -0.73*** 0.00 -0.61*** 0.00 -0.77*** 0.00 -0.01 1.00 (5) IV M/GDP 0.836*** (0.038) -0.127** (0.057) 0.161*** (0.058) 1.918*** (0.517) -1.860*** (0.606) -0.013 (0.012) 0.066** (0.029) 2374 126 0.47 (6) IV NRB/GDP 0.678*** (0.034) -0.436*** (0.099) 0.202** (0.093) -0.701*** (0.251) 0.571** (0.254) 0.013 (0.010) -0.061** (0.027) 2374 126 0.53 0.21 0.35 0.35 0.82 -0.73*** 0.00 -0.40 0.23 Panel B: First stage estimation. Dependent variable: AID/GDP MS x Grant Sum/GDP 1.302*** 1.318*** (0.173) (0.173) L.MS x Grant Sum/GDP 0.277 0.234 (0.259) (0.257) R-sq 0.31 0.32 F instr 0 lag 60.54 62.13 Part. R-sq instr 0 lag 0.27 0.26 1.292*** (0.172) 0.248 (0.260) 0.32 59.58 0.26 Panel C: First stage estimation. Dependent variable: L.AID/GDP MS x Grant Sum/GDP 0.234 0.261 0.216 (0.266) (0.261) (0.265) L.MS x Grant Sum/GDP 1.332*** 1.261*** 1.287*** (0.310) (0.302) (0.304) R-sq 0.28 0.30 0.30 F instr 1 lag 35.67 35.91 35.72 Part. R-sq instr 1 lag 0.24 0.23 0.22 Underid. test p 0.00 0.01 0.01 Note: Robust standard errors in parentheses. *, ** and *** denote significance at 10%, 5% and 1%. The sample is an unbalanced panel over 1975-2007.See the text for the construction of the instrument MS x Grant Sum/GDP used to instrument AID/GDP in column (4)-(6). All controls in the second stage are included in the first stage; not shown to save space. The long run coefficient for, say, RES is calculated as: (a0+a1)/(1-ρ). The p-values presented below the long-run coefficients are calculated with the non-linear test procedure “testnl” in Stata, and indicate the level of significance on which we can reject that the long run-coefficient is zero. Country and year fixed effects are included. R-squared refers to within R-squared. 14 To get more clarity on the dynamics, Figure 1 presents impulse response functions based on column (1)-(3) in Table 4. The impulse is a permanent shift in RES/GDP of 0.10, i.e. a permanent 10 percentage points higher resource exports as share of GDP. 90%confidence bands are shown as dotted lines.10 Non-resource exports react relatively quickly to the shift in the resource exports and stables out at its long run effect of -0.67. Non-resource imports are in the very short run negatively affected, but the effect turns positive during the two first years. Although the import effect is too small to be statistically significant, it slows down the effect on the non-resource balance and makes its adjustment somewhat slower than the adjustment in non-resource exports. -.1 -.1 -.08 -.05 -.06 -.04 0 -.02 0 .05 Figure 1 -5 0 5 X/GDP t 10 15 20 -5 M/GDP 0 5 t 10 15 20 NX/GDP Note: NX=NRB=Non-resource balance. Finally in this section we turn to the dynamic IV estimates presented in the three last columns of Table 4. We instrument aid and lagged aid with contemporaneous and lagged values of the instrument we used in the static panel estimation. Panel B and C of Table 4 show that the explanatory power of the instruments in the two first stages is high, with Fstatistics around 60 and 35 for aid and lagged aid. For the long-run coefficients, the same picture as in the cross section based on the longer period (1960-2000) emerges: only the relations between RES and X and RES and NRB are significant in the long run. Their size implies that one extra dollar of RES decreases X and NRB by three-quarters of a dollar. In the short run, all the estimated windfall-coefficients are significant, and the pattern is the same as in the dynamic FE-estimation described above. 10 We employed standard non-parametric bootstrapping with 500 replications to construct these confidence bands (following Imbs et al. 2005) 15 5. How general are the results? In this section we stress-test our results. Given our most general specification, the dynamic panel data model of the previous section, we investigate whether the estimates are stable across different types of countries (classified in terms of their windfalls), two different time periods and two different categories of products (agriculture and manufacturing). We will focus on the long-run effects. Table 5: Are net resource exporters different? Long run coeff.: Res Long run P-val: Res Long run coeff.: Aid Long run P-val: Aid Observations Countries R-sq Long-run estimates Panel A: Net resource exporters (1) (2) (3) (4) FE FE FE IV X/GDP M/GDP NRB/GDP X/GDP -0.67*** 0.39** -0.79*** -0.71*** 0.00 0.04 0.00 0.00 0.02 0.43 -0.46* 0.71 0.97 0.39 0.07 0.49 764 764 764 764 45 45 45 45 0.82 0.70 0.63 0.81 (5) IV M/GDP 0.29 0.14 0.56 0.62 764 45 0.54 (6) IV NRB/GDP -0.74*** 0.00 -0.31 0.56 764 45 0.55 Panel B: Net resource importers (1) (2) (3) (4) (5) (6) FE FE FE IV IV IV X/GDP M/GDP NRB/GDP X/GDP M/GDP NRB/GDP Long run coeff.: Res -1.04** -0.41 -0.53** -0.91* -0.49 -0.53** Long run P-val: Res 0.02 0.21 0.03 0.07 0.51 0.02 Long run P-val: Aid 0.38 0.00 0.00 0.73 0.91 0.12 Long run coeff.: Aid 0.30 1.00 -0.67 -0.62 -0.29 -0.49 Observations 1610 1610 1610 1610 1610 1610 Countries 81 81 81 81 81 81 R-sq 0.77 0.73 0.57 0.59 0.44 0.55 Note: *, ** and *** denote significance at 10%, 5% and 1%. The p-values presented below the long-run coefficients are calculated with the non-linear test procedure “testnl” in Stata, and indicate the level of significance on which we can reject that the long run-coefficient is zero. The sample is an unbalanced panel over 1975-2007.The estimated model is identical to the one estimated in Table 5 and country and year fixed effects are included. R-squared refers to within R-squared. We have not imposed any restriction on the coefficients and the estimated coefficients of X and M. They do not exactly add up the coefficient on NX because the dynamics are different. A country is defined as a net resource exporter if net resource export as a percentage of GDP was on average larger than 1% in the period 2000-2007. The rest of the countries are then defined as net resource importers. The Dutch Disease debate is most often framed in the context of windfall recipients, like oil exporters or countries experiencing large inflows of aid. Table 5 demonstrates, however, that the Dutch disease mechanism is a general phenomenon. Our sample of 126 countries is split into 45 net resource exporters and 81 net resource importers (see the note of Table 5 for our definition of a resource exporter). The estimates for the 45 net resource exporters presented in 16 panel A of Table 5 are very similar to the ones in the full sample. This is not surprising, since these countries count for the greater variation in net resource exports. Three-quarter of an extra dollar of net resource exports is spent rather than saved (the decrease in NRB) and the reduction in non-resource exports counts for all of it. Perhaps more surprising, the estimates on the sample of 81 net resource importers presented in panel B of Table 5 show that the negative effect of net resource exports on non-resource exports is larger than before. An antiDutch disease seems to be present. Let us illustrate the mechanism by a rise in the oil price. For given quantities of oil imported, an oil price increase will raise the import bill and hence the need for foreign exchange for a net oil importer. In the long run, the extra foreign exchange must be made available through a higher non-resource trade balance. Our estimates suggest that net resource importers generate the extra foreign exchange by increasing their non-resource exports. Table 6: Are aid recipients different? Long-run estimates Panel B: Aid recipients (1) (2) (3) (4) (5) (6) FE FE FE IV IV IV X/GDP M/GDP NRB/GDP X/GDP M/GDP NRB/GDP Long run coeff.: Res -0.19 0.36** -0.59*** -0.30 0.29 -0.46* Long run P-val: Res 0.36 0.05 0.00 0.43 0.82 0.09 Long run coeff.: Aid 0.20 0.91*** -0.71*** -0.27 -1.83 -0.06 Long run P-val: Aid 0.41 0.00 0.00 0.82 0.79 0.89 Observations 767 767 767 767 767 767 Countries 55 55 55 55 55 55 R-sq 0.69 0.62 0.50 0.32 -0.06 0.39 Note-: *, ** and *** denote significance at 10%, 5% and 1%. The p-values presented below the long-run coefficients are calculated with the non-linear test procedure “testnl” in Stata, and indicate the level of significance on which we can reject that the long run-coefficient is zero. The sample is an unbalanced panel over 1975-2007.The estimated model is identical to the one estimated in Table 4 and country and year fixed effects are included. R-squared refers to within R-squared. We have not imposed any restriction on the coefficients and the estimated coefficients of X and M. They do not exactly add up the coefficient on NX because the dynamics are different. A country is defined as aid recipient if Aid as a percentage of GDP was on average larger than 1% in the period 2000-2007. The rest of the countries are then defined as non-recipients of aid. Aid recipients are less developed economies and it is interesting to see whether the same mechanisms apply to them as more developed economies. Table 6 presents the long-run estimates for a samples of 55 for which average aid inflows over the period 2000-2007 counted for more than one percent of GDP. Focusing on the effect of RES on NRB, we find that the fraction saved internationally of an extra dollar of RES is, compared to our other estimates, on the high side. This might be surprising, given that the need for domestic investments is higher in the aid recipients. It is, however, consistent with a large literature documenting that capital tends not to flow to poor countries to the extent one would expect from standard economic theory (Lucas, 1990). The other observation we want to emphasize 17 from Table 6 is that the robust effect of RES on X found in our full sample, is absent for the aid recipients. It is plausible that countries in the need of aid have a less developed capacity to produce tradable goods. Thus, they have little non-resource exports to shift over to domestic use, and any increase in the consumption of tradable goods is achieved via higher non-resource imports. Table 7: Are the effects stable across time? Long run coeff.: Res Long run P-val: Res Long run coeff.: Aid Long run P-val: Aid Observations Countries R-sq (1) FE X/GDP -0.61*** 0.00 0.39 0.40 1518 124 0.64 Long-run estimates Panel A: 1990-2007 (2) (3) FE FE M/GDP NRB/GDP 0.19 -0.66*** 0.19 0.00 0.69* -0.29* 0.06 0.08 1518 1518 124 124 0.53 0.42 (4) IV X/GDP -0.67*** 0.00 -0.33 0.84 1516 122 0.45 (5) IV M/GDP 0.29 0.37 -0.49 0.84 1516 122 -0.01 (6) IV NRB/GDP -0.68*** 0.00 0.02 0.96 1516 122 0.31 Panel B: 1975-1989 (1) (2) (3) (4) (5) (6) FE FE FE IV IV IV X/GDP M/GDP NRB/GDP X/GDP M/GDP NRB/GDP Long run coeff.: Res -0.64*** -0.01 -0.61*** -0.68*** -0.02 -0.60*** Long run P-val: Res 0.00 0.89 0.00 0.00 0.85 0.00 Long run coeff.: Aid -0.10 0.95*** -1.00*** -0.08 0.05 -0.14 Long run P-val: Aid 0.54 0.00 0.00 0.91 0.94 0.80 Observations 856 856 856 851 851 851 Countries 87 87 87 82 82 82 R-sq 0.60 0.43 0.52 0.49 0.14 0.49 Note: *, ** and *** denote significance at 10%, 5% and 1%. The p-values presented below the long-run coefficients are calculated with the non-linear test procedure “testnl” in Stata, and indicate the level of significance on which we can reject that the long run-coefficient is zero. The samples are unbalanced panels. The estimated model is identical to the one estimated in Table 4 and country and year fixed effects are included. R-squared refers to within R-squared. We have not imposed any restriction on the coefficients and the estimated coefficients of X and M. They do not exactly add up the coefficient on NX because the dynamics are different Given large changes in trade patterns and resource markets throughout the period under study, it is of interest to inspect the robustness of our estimates across time. Panel A and B of Table 7 present estimates for two periods; 1990-2007 and 1975-1989.11 The coefficients on RES are almost identical in the two periods. About a third of an extra RESdollar was saved in both periods, and the rest was spent on a reduction in X. AID turns up with significant coefficients in the FE-regressions, but as usual we estimate no effect of AID 11 The 1990-2007 period includes all the 126 countries of our panel regressions. Due to fewer countries and missing data, only 87 countries are included in the 1975-1989 period. The number of countries is reduced in the IV regressions to 122 and 82, due to lack of necessary variation in aid for 4 and 5 countries, respectively. 18 in the IV-regressions. There is no striking difference between the AID coefficients in the two periods. Table 8: Are agriculture and manufacturing products different? Long run coeff.: Res Long run P-val: Res Long run coeff.: Aid Long run P-val: Aid Observations Countries R-sq (1) FE Xag/GDP -0.02 0.46 0.02 0.54 2374 126 0.57 Long-run estimates Panel A: Agriculture products (2) (3) (4) FE FE IV Mag/GDP NXag/GDP Xag/GDP -0.00 -0.01 -0.01 0.18 0.60 0.75 0.01 0.02 0.21** 0.28 0.68 0.02 2374 2374 2374 126 126 126 0.67 0.57 0.52 (5) IV Mag/GDP -0.00 0.21 0.05* 0.09 2374 126 0.60 (6) IV NXag/GDP -0.01 0.85 0.17** 0.05 2374 126 0.54 Panel B: Manufacturing products (1) (2) (3) (4) (5) (6) FE FE FE IV IV IV Xma/GDP Mma/GDP NXma/GDP Xma/GDP Mma/GDP NXma/GDP Long run coeff.: Res -0.40* 0.14 -0.45*** -0.45** 0.11 -0.44*** Long run P-val: Res 0.10 0.16 0.00 0.04 0.41 0.00 Long run coeff.: Aid -0.09 0.40*** -0.47*** -0.56 0.23 -0.59 Long run P-val: Aid 0.61 0.01 0.00 0.28 0.79 0.23 Observations 2374 2374 2374 2374 2374 2374 Countries 126 126 126 126 126 126 R-sq 0.87 0.74 0.63 0.86 0.61 0.54 Note: *, ** and *** denote significance at 10%, 5% and 1%. The p-values presented below the long-run coefficients are calculated with the non-linear test procedure “testnl” in Stata, and indicate the level of significance on which we can reject that the long run-coefficient is zero. The sample is an unbalanced panel over 1975-2007.The estimated model is identical to the one estimated in Table 4 and country and year fixed effects are included. R-squared refers to within R-squared. We have not imposed any restriction on the coefficients and the estimated coefficients of X and M. They do not exactly add up the coefficient on NX because the dynamics are different. Trade data on agriculture (ag) and manufacturing products (ma) is from WDI. Trade in agriculture and manufacturing products, together with trade in services, add up to our measure of non-resource trade. A final question of interest is whether non-resource exports in terms of agriculture products react differently than non-resource exports in terms of manufacturing products. Table 8 shows that RES has no effect on agriculture products (Panel A). For manufacturing products (Panel B), the familiar significant and negative coefficients on RES turn up for exports and net-exports, as well as the insignificant ones for imports. The exports and net-exports coefficients are somewhat smaller in absolute value than what we found in Table 4; around 0.45 compared to -0.7. This difference, given that RES has no effect on agriculture products, must be explained by trade in services, which is the residual between our measures of nonresource trade and the sum of trade in agriculture and manufacturing products. Interestingly, for AID we find the opposite pattern. As was the case for total non-resource trade, we find no effect from AID on manufacturing trade in the IV-estimation (Panel B). Unlike before, we 19 find significant effects of AID on exports, imports and net exports of agriculture products (Panel A). Taking Table 8 seriously, AID improves export capacity in agriculture products to an extent that it dominates the extra imports of agriculture products, and hence improves the trade balance in agriculture products. One may be concerned about the role of GDP as denominator. The economic question of this paper is how windfalls of foreign exchange may affect economic structure, i.e. the possible contraction of the non-resource exporting and import-competing sectors. Hence the appropriate measures are non-resource exports and imports as share of GDP. Econometrically, however, it might be that short run GDP shocks turning up on both sides of the equation are affecting our results. To investigate this issue, we re-estimate the models of Table 4 with two alternative GDP trends as denominator. The first GDP trend used (Appendix Table A3) is generated by the Hodrick-Prescott filter (Hodrick and Prescott 1997) with a lambda-parameter equal 6.25. This is the recommended rigidness of the trend for annual data.12 We then re-do the exercise with a more rigid trend, namely the HP-trend with lambda set to 1600 (Appendix Table A4). By comparing Table 4, Table A3 and Table A4 we gain insights in the role of short run GDP shocks for our estimates. Employing the HP-trended GDP with the recommended lambda as the denominator (Table A3) reproduces the results in Table 4: the robust findings are that the non-resource balance and the non-resource exports contract as response to a positive windfall from resource exports. When the more rigid trend is used as denominator (Table A4), we still find that the non-resource balance contracts. The magnitudes of the contraction of NRB are similar in the two cases (about -0.60), and somewhat smaller than in the normal case presented in Table 4. However, with the rigid trend we find that it is an increase in non-resource imports rather than a decrease in non-resource exports that significantly contributes to NRB-contraction. Figure 2 presents the impulse response functions based on the estimates of column (1) and (2) in Table 4, Table A3 and Table A4. The immediate response is similar on all three cases, both for X and M, although it is less pronounced in the case of the recommended HP-trend. Figure 2 should leave no doubt about the existence of immediate reactions in X and M to a resource windfall. How the effects play out over time does, however, depend on the GDP-measure we 12 Lambda=6.25 in the HP-filtering is recommended by Ravn and Uhlig (2002) for annual data. 20 use. In the case with the rigid trend, the adjustment to the long run levels of X and M is rapid compared to the two other cases, which also can be seen from the lower coefficient on the lagged dependent variable. In the case of the rigid trend, the long-run export effect is relatively small compared to the other two cases, while the long-run import effect is relatively large. -.02 -.08 0 -.06 .02 -.04 .04 -.02 0 .06 Figure 2 -5 0 5 X/GDP X/GDP 1600 t 10 15 X/GDP 625 20 -5 0 5 M/GDP M/GDP 1600 t 10 15 20 M/GDP 625 A final robustness check regards the dynamic specification of the model. Appendix figure A1 compares the impulse response function for the ADL(1,1) model presented so far in this section and a more general ADL(2,2) model, in which an extra lag of the dependent variable as well as the resource and aid variables are included. The ADL(2,2) model produces somewhat larger confidence bands, but the estimated responses are very similar in the two models. We have throughout the paper instrumented for AID, due to the obvious endogeneity of AID caused by donors taking into account the state of the economy in potential and actual recipient countries. The case for endogeneity of net resource exports is weaker, as the supply of resources to a large extent is determined by geological factors and the price of resources is determined on the world market. However, one could argue that the imports of resources is endogenous to economic structure (e.g. a large manufacturing sector may generate high import demand for resources as inputs), and we have therefore explored the bias in our estimated resource effects by instrumenting for net resource exports. From disaggregated trade data on resources and corresponding price indexes we have constructed price indexes for each country’s imports and exports of resources. In Appendix Table A5 we re-estimate the models from Table 4 with these price indexes as instruments for net resource exports (see 21 the footnote of the table for details on how the price indexes were constructed). We find that the instruments predict RES well, that the exclusion restriction for the instruments is not invalidated, and that the conclusions of the paper is not altered by taking into account the possible endogeneity of RES. The long run RES-coefficient in the non-resource export equation comes out bigger when instrumenting, while it takes the similar size as in Table 4 in the NRB-equation. We performed many robustness checks with regards to the construction of the resource instrument (different weights) and the specification of the model (both resources and aid instrumented or only instrumented resources included, to accommodate worries about too complex a model). We found that the results presented in Appendix Table A5 are reasonably robust. 6. Concluding comments The possible adverse effects of foreign exchange windfalls on the tradable sector has been a recurring theme of literature on the Dutch disease, on the resource curse, and on the implications of scaling up aid. There are alternative windows through which researchers can get a view on the issues. Such effects should be associated with relative price changes and real exchange appreciation, at least in the short-run, although finding these effects empirically has proved elusive. Variations in production structure are observable, but empirical work is hindered both by data issues and by the myriad factors that shape comparative advantage. The approach of this paper is to look directly at the trade and balance of payments data. This has advantages of simplicity, with some clear structure imposed by balance of payments accounting, and some robust empirical support provided by gravity models of trade. The approach also enables us to divide tradables into imports and exports, activities that are in practise quite different. We have provided results for a range of different approaches, and two sorts of conclusions emerge robustly. Exports of natural resources seem to crowd out non-resource exports, at a rate of around 50 cents to the dollar, while drawing in imports at around 15 cents to the dollar, the remaining 35 cents of revenue going to the capital account. Aid draws in more imports (around 40 cents to the dollar), while only reducing exports marginally. 22 References Adam, C.S. and S.A. Connell (2004). Aid versus trade revisited: donor and recipient policies in the presence of learning by doing, Economic Journal, 114, 150-173. Collier, P., F. van der Ploeg, M. Spence and A.J. Venables (2009). Managing resource revenues in developing economies, IMF Staff Papers. Corden, W.M. and J.P. Neary (1982). Booming sector and de-industrialisation in a small open economy, Economic Journal, 92, 825-848. Davis, J., R. Ossowski, J. Daniel and S. Barnett (2002). Stabilization and Savings Funds for Non-Renewable Resources: Experience and Fiscal Policy Implications, IMF Occasional Paper No. 205, International Monetary Fund, Washington, D.C. Gupta, S. and P.S. Heller (2002). Challenges in expanding development assistance, IMF Policy Discussion Paper 02/5, International Monetary Fund, Washington, D.C. Harrigan, James (1997), “Technology, Factor Supplies, and International Specialization: Estimating the Neoclassical Model.” American Economic Review, 87(4), 475-494. Hayashi, F. (1982). Tobin’s marginal Q and average Q: a neoclassical interpretation, Econometrica, 50, 215-224.Hodrick, R. and E. Prescott(1997). Post-war U.S. business cycles: An empirical investigation. Journal of Money, Credit and Banking, 29(1), 116. Judson, R., and Owen, A. (1999). Estimating dynamic panel data models: a guide for macroeconomists. Economics Letters, 49 , 1417-1426. Kydland, F.E. and E.C. Prescott (1982). Time to build and aggregate fluctuations, Econometrica, 50, 6, 1345-1370. Lucas, R. E., Jr. (1990), Why Doesn't Capital Flow from Rich to Poor Countries?" American Economic Review, Papers and Proc. 80, 92-96. Neary, J.P. and D.D. Purvis (1982). Sectoral shocks in a dependent economy: long-run adjustment and short-run accommodation, Scandinavian Journal of Economics, 84, 2, 229-253. Nickell, Stephen (1981) “Biases in Dynamic Models with Fixed Effects,” Econometrica, 49, (6), 1417-26 Ploeg, F. van der and S. Poelhekke (2010) ‘The pungent smell of red herrings; subsoil assets, rents, volatility and the resource curse, Journal of Environmental Economics and Management, 60,1,44-55 23 Ploeg, F. van der and A.J. Venables (2009). Harnessing windfall revenues: Optimal policies for resource-rich, developing economies, Oxcarre Research Paper 9, University of Oxford. Rajan , R.G. and A. Subramanian (2008), Aid and Growth: What Does the Cross-Country Evidence Really Show?, The Review of Economics and Statistics, 90(4), 643665Ravn, Morten O. and Harald Uhlig (2002). On adjusting the Hodrick-Prescott filter for the frequency of observations. Review of Economics and Statistics, 84(2), 371-376. Sachs, J.D. and A.M. Warner (1997). Natural resource abundance and economic growth, in G. Meier and J. Rauch (eds.), Leading Issues in Economic Development, Oxford University Press, Oxford. Santos-Paulino, Amelia U. U. and Thirlwall, A. P., (2004), The Impact of Trade Liberalisation on Exports, Imports and the Balance of Payments of Developing Countries, Economic Journal, Vol. 114, pp. F50-F72. Torvik, R. (2001). Learning by doing and the Dutch Disease, European Economic Review, 45, 285-306. Turnovsky, S.J. (2009). Capital Accumulation and Growth in a Small Open Economy, Cambridge University Press, Cambridge, U.K. Wijnbergen, S. van (1984). The Dutch Disease: A disease after all? Economic Journal, 94 (373), 41- Appendix 1: Data 24 Data on aid (AID), Gross domestic product (GDP), and aggregate trade are from World Development Indicators (WDI).13 The trade measures used are exports and imports of goods and services (BoP), resource exports and imports (covering fuel, metals and ores), as well as exports and imports of agriculture product (Xag and Mag) and manufacturing products (Xma and Mma). Non-resource exports (X) is defined as: exports of goods and services (BoP) minus exports of fuel, metals and ores. The measure of non-resource imports (M) is defined in the same way. The non-resource balance (NRB) is an abbreviation for non-resource net exports (X-M). Net resource exports (RES) is defined as resource exports minus resource imports. Data on bilateral non-resource exports used in the gravity estimation are from Comtrade. The bilateral country-fixed variables used in the gravity estimations (distance and dummies for contiguity, common official primary language and colonial relationship) and the unilateral country-fixed variables used in the regressions on aggregate data (area in sq. kms and dummies for being landlocked and island) are from CEPII.14 Both the Comtrade data and the trade data from WDI are presented in current USD. We deflate them with the GDP deflator of the U.S. GDP, which is found by dividing U.S. GDP in current USD by U.S. GDP in fixed USD with year 2000 as the base year. Both are from WDI. All our trade data are in other words denominated in fixed 2000 USD. When we measure a variable as share of GDP, the numerator and the de-numerator are deflated with the same price index. FSA (foreign supplier access) and FMA (foreign market access) are used as control variables. Appendix 2 explains how they are constructed. They are denominated in 2000 USD. The instrument for aid used in the cross-section is directly from Rajan and Subramanian (2008). The instrument we construct for the panel data analyses is based on data from OECD’s DAC-ODA database (Table 2A: total grants, Official Aid Disbursements).15 13 See: http://publications.worldbank.org/WDI/ The data can be downloaded at: http://www.cepii.fr/anglaisgraph/bdd/distances.htm 15 For access to the data, go to: www.oecd.org/dac/stats 14 25 Appendix 2: Gravity, foreign market access, foreign supplier access. In each year, the gravity equation takes the form: Tij ci m jbij where Tij is the value of trade between countries i and j, ci is an export country fixed effect, mj is an importer fixed effect, and bij is a set of between-country variables (distance etc), affecting trade with parameters α. We estimate this in logs as, ln Tij ln ci ln m j ln bij eij This equation is estimated per year, covering all countries available in Comtrade with a population larger than 0.5 million. We include all available trade observations, i.e. also trade observations smaller than hundred thousand dollars at the SITC four-digit level. Tij is measured as non-resource exports from country i to country j. bij includes distance and dummies for contiguity, common official primary language and colonial relationship. Results are available on request Foreign market access and foreign supplier access, FMA and FSA are constructed as follows. For each exporter, summing over export markets FMAi j m jbij . For each importer, summing over sources of imports FSAi j c jbij . Summing over markets and sources of supply respectively, we have total exports of country i, Xi, and imports of country j, Mj T c FMA T m FSA Xi j ij Mj i ij i i j j We suppose that ci GDPi . f (county i characteristics ) . The basis of having ci proportional to GDPi is the unit elasticity of trade with respect to income found in many gravity studies. Using this, X i / GDPi f (countr yi characteristics ) FMAi / GDPi . We linearise and take as country characteristics RES/GDP, AID/GDP, area and island status. An exactly analogous approach is followed for imports. 26 Appendix 3: Extra figures and tables -.1 -.02 -.08 0 -.06 .02 -.04 .04 -.02 0 .06 Figure A1 -5 0 5 t 10 20 -5 X/GDP ADL(2,2) 0 5 t M/GDP ADL(1,1) 10 15 20 M/GDP ADL(2,2) -.1 -.08 -.06 -.04 -.02 0 X/GDP ADL(1,1) 15 -5 0 5 NRB/GDP ADL(1,1) t 10 15 20 NRB/GDP ADL(2,2) Note: the figure illustrates the impulse response function for the ADL(1,1)-models of column (1)-(3) in table 4 together with the impulse response functions given the OLS-estimates of the following ADL(2,2)-model: Yit/GDPit = ρ0Yit-1/GDPit-1 + ρ1Yit1/GDPit-1 + a0RESit/GDPit + a1RESit-1/GDPit-1 + a2RESit-2/GDPit-2+ b0AIDit/GDPit +b1AIDit-1/GDPit-1 + b2AIDit-2/GDPit-2+ country controls + time controls+ country/ time controls + uit. Table A1: Cross sections, RS sample Estimation method: Dep. Variable RES/GDP AID/GDP FSA/GDP FMA/GDP Area Island Constant Observations R-sq OLS (4) X/GDP -0.298** (0.143) -1.257*** (0.380) 0.274* (0.152) -0.154 (0.136) -0.037*** (0.010) 0.023 (0.044) 0.730*** (0.140) 69 0.42 1960-2000 OLS (5) M/GDP 0.326*** (0.119) -0.430 (0.374) -0.325* (0.178) 0.418** (0.158) -0.039*** (0.010) 0.029 (0.043) 0.777*** (0.139) 69 0.38 OLS (6) NRB/GDP -0.635*** (0.073) -0.900*** (0.149) 0.605*** (0.050) -0.565*** (0.048) 0.005* (0.003) 0.006 (0.012) -0.089** (0.041) 69 0.83 OLS (4) X/GDP -0.394*** (0.112) -1.136*** (0.352) 0.561 (0.584) -0.367 (0.531) -0.034*** (0.011) 0.022 (0.045) 0.688*** (0.153) 71 0.45 1980-2000 OLS (5) M/GDP 0.251** (0.111) -0.173 (0.379) -0.028 (0.719) 0.178 (0.649) -0.034*** (0.012) 0.045 (0.048) 0.699*** (0.163) 71 0.34 OLS (6) NRB/GDP -0.660*** (0.054) -1.002*** (0.142) 0.589*** (0.175) -0.522*** (0.166) 0.005 (0.003) -0.004 (0.013) -0.074* (0.037) 71 0.84 27 Note: Robust standard errors in parentheses. *, ** and *** denote significance at 10%, 5% and 1%. All variables are averages across 1980-2000. RS sample refers to the sample of Rajan and Subramanian (2008) for the corresponding time period. Table A2: Panel estimation using both within and between country variation OLS estimations OLS OLS OLS (1) (2) (3) Dep. variable: X/GDP M/GDP NRB/GDP RES/GDP -0.501*** 0.115*** -0.615*** (0.022) (0.025) (0.018) AID/GDP -0.643*** 0.064 -0.708*** (0.061) (0.068) (0.038) FSA/GDP 0.010 -0.015 0.025*** (0.010) (0.010) (0.004) FMA/GDP 0.102*** 0.215*** -0.113*** (0.029) (0.024) (0.017) Area -0.044*** -0.045*** 0.001 (0.002) (0.002) (0.001) Island -0.032*** -0.032*** 0.001 (0.006) (0.007) (0.003) Observations 2579 2579 2579 Countries 126 126 126 R-sq 0.45 0.44 0.60 Note: Robust standard errors in parentheses. *, ** and *** denote significance at 10%, 5% and 1%. Time period: 1975-2007. Year fixed effects are included. Area is included in the log form. R-squared refers to within R-squared. Est. method: 28 Table A3: Dynamic panel estimates with GDP HP-trended as recommended for annual data (lambda=6.25) Observations Countries R-sq (1) FE X/GDPhp 0.814*** (0.032) -0.441*** (0.076) 0.351*** (0.079) 0.045 (0.038) 0.017 (0.039) -0.004 (0.004) 0.015 (0.018) 2376 126 0.78 (2) FE M/GDPhp 0.785*** (0.023) -0.052 (0.051) 0.099*** (0.035) 0.246*** (0.055) -0.023 (0.063) -0.017*** (0.004) 0.073*** (0.017) 2376 126 0.72 (3) FE NRB/GDPhp 0.685*** (0.024) -0.392*** (0.107) 0.195** (0.087) -0.211*** (0.062) 0.005 (0.056) 0.014*** (0.005) -0.063*** (0.023) 2376 126 0.58 (4) IV X/GDPhp 0.815*** (0.032) -0.444*** (0.091) 0.351*** (0.088) 0.141 (0.288) -0.145 (0.270) -0.005 (0.005) 0.017 (0.018) 2376 126 0.78 (5) IV M/GDPhp 0.813*** (0.032) -0.060 (0.037) 0.093** (0.036) 0.935* (0.565) -0.847 (0.547) -0.018* (0.010) 0.077*** (0.024) 2376 126 0.67 (6) IV NRB/GDPhp 0.712*** (0.032) -0.386*** (0.098) 0.211** (0.085) -0.745 (0.456) 0.619 (0.440) 0.014 (0.011) -0.066*** (0.025) 2376 126 0.53 Long run coeff.: Res Long run P-val: Res Long run coeff.: Aid Long run P-val: Aid -0.48*** 0.00 0.33* 0.08 0.21 0.19 1.04*** 0.00 -0.63*** 0.00 -0.65*** 0.00 -0.50** 0.04 -0.02 0.97 0.18 0.23 0.47 0.49 -0.61*** 0.00 -0.44 0.31 Lagged dependent var. RES/GDPhp L.RES/GDPhp AID/GDPhp L.AID/GDPhp FSA/GDPhp FMA/GDPhp Panel B: First stage estimation. Dependent variable: AID/GDP MS x Grant Sum/GDPhp 1.210*** 1.221*** (0.325) (0.326) L.MS x Grant Sum/GDPhp 0.210 0.198 (0.270) (0.270) R-sq 0.25 0.26 F instr 0 lag 29.38 29.94 Part. R-sq instr 0 lag 0.20 0.20 1.219*** (0.328) 0.171 (0.270) 0.26 28.28 0.19 Panel C: First stage estimation. Dependent variable: L.AID/GDP MS x Grant Sum/GDPhp 0.510 0.528* 0.523* (0.313) (0.305) (0.303) L.MS x Grant Sum/GDPhp 0.970*** 0.947*** 0.904*** (0.359) (0.351) (0.347) R-sq 0.23 0.26 0.26 F instr 1 lag 28.92 30.71 29.05 Part. R-sq instr 1 lag 0.18 0.19 0.18 Underid. test p 0.08 0.09 0.09 Note: Robust standard errors in parentheses. *, ** and *** denote significance at 10%, 5% and 1%. The sample is an unbalanced panel over 1975-2007.See the text for the construction of the instrument MS x Grant Sum/GDP used to instrument AID/GDP in column (4)-(6). All controls in the second stage are included in the first stage; not shown to save space. The long run coefficient for, say, RES is calculated as: (a0+a1)/(1-ρ). The p-values presented below the long-run coefficients are calculated with the non-linear test procedure “testnl” in Stata, and indicate the level of significance on which we can reject that the long run-coefficient is zero. Country and year fixed effects are included. R-squared refers to within R-squared. GDPhp refers to a standard HP-trend of GDP (lambda=6.25). 29 Table A4: Dynamic panel estimates with GDP HP-trended to be very robust to shortrun fluctuations (a relatively rigid trend, r=rigid, lambda=1600) Panel A: FE and IV estimations (2) (3) (4) FE FE IV M/GDPhpr NRB/GDPhpr X/GDPhpr 0.749*** 0.708*** 0.772*** (0.065) (0.023) (0.049) -0.026 -0.387*** -0.416*** (0.032) (0.120) (0.105) 0.166*** 0.210* 0.389*** (0.042) (0.108) (0.106) 0.214*** -0.263*** 0.578 (0.073) (0.057) (1.073) -0.066 -0.009 -1.007 (0.105) (0.049) (1.089) -0.015*** 0.020*** 0.002 (0.005) (0.004) (0.005) 0.120*** -0.094*** 0.039* (0.023) (0.020) (0.021) 2376 2376 2376 126 126 126 0.79 0.65 0.73 (5) IV M/GDPhpr 0.837*** (0.059) -0.022 (0.039) 0.115** (0.050) 2.307 (1.792) -2.533 (1.716) -0.016 (0.011) 0.120*** (0.034) 2376 126 0.56 (6) IV NRB/GDPhpr 0.761*** (0.047) -0.393*** (0.098) 0.255*** (0.099) -1.552 (0.984) 1.351 (0.939) 0.019** (0.009) -0.094*** (0.024) 2376 126 0.47 0.57*** 0.01 -1.39 0.50 -0.58*** 0.00 -0.84 0.33 Panel B: First stage estimation. Dependent variable: AID/GDP MS x Grant Sum/GDPhpr 0.963*** 0.988*** (0.203) (0.196) L.MS x Grant Sum/GDPhpr 0.247 0.210 (0.232) (0.218) R-sq 0.17 0.19 F instr 0 lag 45.02 46.91 Part. R-sq instr 0 lag 0.13 0.13 0.958*** (0.207) 0.209 (0.239) 0.19 43.54 0.12 Estimation method: Lagged dependent var. RES/GDPhpr L.RES/GDPhpr AID/GDPhpr L.AID/GDPhpr FSA/GDPhpr FMA/GDPhpr Observations Countries R-sq Long run coeff.: Res Long run P-val: Res Long run coeff.: Aid Long run P-val: Aid (1) FE X/GDPhpr 0.745*** (0.074) -0.414*** (0.104) 0.394*** (0.097) -0.046 (0.054) -0.054 (0.122) 0.004 (0.005) 0.030 (0.020) 2376 126 0.77 -0.08 0.44 -0.39 0.42 0.56*** 0.00 0.59 0.35 -0.61*** 0.00 -0.93*** 0.00 -0.12 0.56 -1.89*** 0.01 Panel C: First stage estimation. Dependent variable: L.AID/GDP MS x Grant Sum/GDPhpr 0.565** 0.603** 0.546** (0.269) (0.254) (0.276) L.MS x Grant Sum/GDPhpr 0.746*** 0.683*** 0.693** (0.286) (0.258) (0.304) R-sq 0.18 0.22 0.21 F instr 1 lag 33.78 37.73 30.71 Part. R-sq instr 1 lag 0.12 0.13 0.12 Underid. test p 0.07 0.08 0.08 Note: Robust standard errors in parentheses. *, ** and *** denote significance at 10%, 5% and 1%. The sample is an unbalanced panel over 1975-2007.See the text for the construction of the instrument MS x Grant Sum/GDP used to instrument AID/GDP in column (4)-(6). All controls in the second stage are included in the first stage; not shown to save space. The long run coefficient for, say, RES is calculated as: (a0+a1)/(1-ρ). The p-values presented below the long-run coefficients are calculated with the non-linear test procedure “testnl” in Stata, and indicate the level of significance on which we can reject that the long run-coefficient is zero. Country and year fixed effects are included. R-squared refers to within R-squared. GDPhpr referes to a relatively rigid HP-trend of GDP (lambda=1600). 30 Table A5: IV for RES Res/GDP L.Res/GDP Aid/GDP L.Aid/GDP FSA/GDP FMA/GDP L.X/GDP L.M/GDP Panel A: IV estimations (1) (2) IV IV X/GDP M/GDP -0.281* -0.442** (0.156) (0.178) 0.028 0.376** (0.187) (0.190) 1.311*** 1.776*** (0.390) (0.522) -1.310*** -1.819*** (0.467) (0.582) -0.000 -0.014 (0.007) (0.011) -0.001 0.062** (0.027) (0.028) 0.801*** (0.037) 0.838*** (0.036) L.NRB/GDP Observations Countries R-sq Long run coeff.: Res Long run P-val: Res Long run coeff.: Aid Long run P-val: Aid 2367 126 0.65 -1.27*** 0.01 0.00 1.00 2367 126 0.48 -0.41 0.58 -0.27 0.87 (3) IV NRB/GDP 0.192 (0.157) -0.407* (0.211) -0.420* (0.234) 0.403* (0.238) 0.015** (0.007) -0.068*** (0.025) 0.644*** (0.047) 2367 126 0.44 -0.60** 0.06 -0.05 0.90 Panel B: First stage estimation. Dependent variable: RES/GDP P GROSSRESM -0.024*** -0.025*** -0.025*** (0.008) (0.007) (0.007) P GROSSRESX 0.047*** 0.045*** 0.047*** (0.007) (0.007) (0.007) R-sq 0.18 0.14 0.20 F instr 25.92 25.63 29.44 Part. R-sq instr 0.06 0.06 0.06 Panel C: First stage estimation. Dependent variable: L.RES/GDP L.P GROSSRESM -0.025*** -0.026*** -0.027*** (0.008) (0.008) (0.007) L.P GROSSRESX 0.030*** 0.034*** 0.029*** (0.009) (0.009) (0.008) R-sq 0.19 0.12 0.22 F instr 14.33 14.85 18.55 Part. R-sq instr 0.05 0.05 0.05 Panel D: First stage estimation. Dependent variable: AID/GDP MS x Grant Sum/GDP 1.283*** 1.301*** 1.281*** (0.176) (0.175) (0.174) R-sq 0.31 0.32 0.32 F instr 22.72 23.24 23.06 Part. R-sq instr 0.06 0.06 0.06 Panel E: First stage estimation. Dependent variable: L.AID/GDP L.MS x Grant Sum/GDP 1.341*** 1.273*** 1.317*** (0.315) (0.307) (0.309) R-sq 0.28 0.31 0.30 F instr 13.40 13.68 14.53 Part. R-sq instr 0.06 0.06 0.06 Underid. test p 0.00 0.00 0.00 Overid. test p 0.33 0.67 0.58 31 Note: Instruments for net resource exports, RES, are price country-year specific price index for imports and exports of (ca 50) different commodities (“mean weight RESX+RESM”). The prices used in constructing the indexes are world market prices, while the weights are country-specific based on disaggregated import and export data for each country. To avoid endogeneity in the weights, we use the mean weights over the entire sample (1975-2006).