Farm Acreage Shocks and Food Prices: An SVAR Approach to. Understanding the Impacts of Biofuels. 1 Introduction

Transcription

1 Farm Acreage Shocks and Food Prices: An SVAR Approach to Understanding the Impacts of Biofuels Catherine Almirall, Maximilian Auhammer and Peter Berck Abstract From 2005 to 2006, s jumped 47%, and from 2006 to 2007 they rose another 28%. To what extent has biofuels production contributed to higher food s? Can we anticipate how food s will react to further increasing biofuels production, and how does this depend on the feedstock used for the biofuel? This paper analyzes how food s have responded to exogenous shocks in acreage supply over the last 50 years. Using a structural vector auto-regression framework, we examine both shocks to a crop's own acreage and shocks to total cropland. This allows us to estimate both the expected eect of dedicating existing cropland to biofuels feedstock production as well as the eect of dedicating non-crop lands. We nd that a negative shock in own acreage leads to a spike in crop s for both soybeans and. For a reduction in area of one million acres, we estimate a increase of $0.04 to $0.07 per bushel. A back-of-the-envelope calculation shows that the increase in ethanol production from 2006 to 2007 explains approximately 30 to 50% of the increase. We nd an increase in soy when soy acres are removed, and we nd a decrease in and soy s when non-crop farmland is removed. The results are robust to dierent specications. These results will hopefully inform the impact assessment literature for biofuels and allow us a better understanding of what could happen to food s if the US continues to increase biofuels production. 1 Introduction The last ve years have seen tremendous expansion in biofuels production, particularly in ethanol in the United States. At the same time, commodity s (like ) have experienced incredible spikes. From 2005 to 2006, s jumped 47%, and from 2006 to 2007 they rose another 28%. Not surprisingly, the Department of Agricultural and Resource Economics, University of California, Berkeley. Send correspondence to Catherine Almirall, Department of Agricultural and Resource Economics, 207 Giannini Hall, University of California, Berkeley, CA , USA; Phone: ; Fax: ; calmirall@berkeley.edu. Many thanks to Michael Jansson and Yuriy Gorodnichenko for helpful suggestions. Financial support provided by the Energy Biosciences Institute. All views are those of the authors alone and do not represent those of the EBI. Working paper; do not cite without permission of authors. 1

2 food versus fuel debate has economists and policy makers asking how much of the increase in s was a result of ethanol production. In 2007, the BBC wrote that [i]t is one of the most hotly debated environmental topics of the year - whether the drive to produce alternative so-called green fuels will take food from the mouths of the hungry. To what extent has biofuels production contributed to higher food s? Can we anticipate how food s will react to further increasing biofuels production, and how does this depend on the feedstock used for the biofuel? This paper analyzes how food s have responded to exogenous shocks in acreage supply over the last 50 years. We examine both shocks to a crop's own acreage and shocks to total cropland. This allows us to estimate both the expected eect of dedicating existing cropland to biofuels feedstock production as well as the eect of dedicating marginal lands. While biofuels in the US have so far been generally restricted to ethanol, a number of other technologies are being considered. The feedstocks for these second- and third-generation biofuels could be grown on land currently dedicated to or soybeans, or on marginal lands. One of the more promising biofuels appears to be ethanol made from a feedstock such as miscanthus or some combination of prairie grasses. Also being considered are cellulosic ethanols, including ethanol produced from stover. As mentioned above, economists, policy makers, and the general public have been especially concerned about the eect of ethanol on commodity and food s. Some research has emerged to analyze this question. Sexton et al (2007) estimate that in 2007, with 18% of US going to ethanol production, ethanol raised s at least 18 percent and perhaps as much as 39 percent, depending on elasticity assumptions. The FAPRI model used by Searchinger et al (2008) estimates that if 100 million new bushels of were to be exported (approximately 1% of US production), s would increase by approximately 2%. Ugarte et al (2000) estimate that if 10% of US crop acreage were devoted to biofuels production, food s would rise by approximately 8 to 14%. Rajagopal et al (2007) calculate that s will rise 21% when 15% of production is used for biofuels. However while some models, such as those mentioned above, have emerged to investigate food/fuel linkages, they have not been general enough to predict the eects of second- and third-generation biofuels production on food s. These biofuels use non-food crops (such as miscanthus) and have the potential to be grown on marginal lands. While it is generally acknowledged that ethanol policies can lead to spikes in food s, two questions remain unanswered. First, no consensus has been reached about the magnitude of the eect of ethanol policies, although the papers cited above generally report an elasticity of approximately one to two for the impact of ethanol on s. Second, little research has been done on the eects of non-food crops for biofuels production. This is an open question that should not be ignored: prairie grasses, if grown on croplands previously used for food grains, would be expected to 2

3 raise food s. On the other hand, if feedstocks are grown on marginal lands not previously used for crop production, crop s would not be expected to rise. The goal of this paper is to analyze how food s have responded to exogenous shocks in acreage supply over the last 50 years. We examine both shocks to a crop's own acreage and shocks to total cropland, estimating the eect of dedicating existing cropland to biofuels feedstock production as well as the eect of dedicating non-crop lands. The dynamic nature of agricultural production makes this question ideally suited for a structural vector auto-regression (SVAR) of the sort routinely used in the macroeconomic literature. This framework allows us to leverage the timing of planting decisions versus harvest outcomes with a classic time-series methodology. We estimate a system of equations to explain the relationship between total cropland, and soybean acreage, and and soybean spot and s. In particular, we use a factor-augmented structural vector auto-regression to allow for exogenous shocks to the entire system, including supply-side shocks such as a spike in farm input s or a demand-side shock such as increased foreign demand. Time-series models such as SVARs are widely used in macroeconomic applications where variables are jointly determined, adjustment to long-run equilibrium is not instantaneous (which implies the importance of including lags in the model), and the underlying data generating process follows a specic timing mechanism. The above are all characteristics of an agricultural supply model, in which and acreage are jointly determined and in which the eects of a shock can last several periods. We show that a structural VAR can be used to leverage the sequencing inherent in agricultural supply. The resulting model is able to capture dynamics that may be missed with other models. Also, the system's dynamic nature allows us to estimate forecast error variance decompositions (FEVDs), which explain the percentage of variance that comes from specic shocks, and impulse response functions (IRFs), which trace out the eect of exogenous shocks across time. Previous work on dynamic agricultural systems beyond the Nerlovian model is relatively limited. Mushtaq and Dawson (2002) use a recursive vector auto-regression approach to investigate acreage response of various crops in Pakistan. They nd that this approach is more appropriate than a Nerlovian partial-adjustment model, particularly in explaining adjustments to long-run equilibrium. However since their interest is in acreage response, they do not report the impact of shocks to acreage on s. We nd that a negative shock in acreage leads to a spike in crop s for both soybeans and. For a reduction in area of one million acres, we estimate a increase of $0.04 to $0.07 per bushel. To put this in perspective, consider that from 2006 to 2007, acreage dedicated to ethanol increased by about 6 million acres in the United States. At the same time, s rose $0.80 per bushel. Thus our model nds that approximately 30 to 50% of the increase was due to new ethanol production. 3

4 For a negative shock in soybean acreage, we nd a increase of $0.08 to $0.23 per bushel. This much larger magnitude is partly explained by the fact that US soybean acreage has a much larger share of world production than does US acreage. For the scenario in which one million acres is moved from non-crop farmland (including pasture and idle lands) to crop production (but holding or soybean acreage constant), we nd a decrease of $0.05 to $0.10 per bushel. The intuition for this result is that, while (or soybean) acreage is held constant, the production of substitute crops increases. Accordingly, demand for the crop falls. Historically, this has corresponded to increased production of other grains, but the intuition should also hold for the increased production of substitute biofuels feedstocks. Thus we nd robust results that removing acreage from food production, in order to grow biofuels feedstocks, increases food crop s. On the other hand, switching non-crop farmland to a substitute crop can lower food crop s. While economists have generally expressed concern about the impact of biofuels production on food s, empirical estimation using time-series methods has not been carried out. Importantly, we distinguish between crop and non-crop lands, adding a layer of complexity to the acreage/ relationship. These results will hopefully inform the impact assessment literature for biofuels and allow us a better understanding of what could happen to food s if the US continues to increase biofuels production. 2 Econometric Framework 2.1 Structural Vector Auto-Regression: Scenario 1 We apply a structural vector auto-regression 1 to analyze what we call scenario 1, in which food acreage is removed and dedicated to a biofuels feedstock. 2 In this framework, we use a system of equations to explain the relationship between and soybean acreage, total cropland, and and soybean spot and. We impose identication restrictions that take advantage of the timing of planting decisions in the United States. That is, agricultural producers set their acreage at planting time according to their expectation of harvest-time s. In the classic Nerlovian framework, this implies that current acreage is a function of past acreage, past s and s, and supply-side variables such as input s, which are the only variables observable to farmers at planting time. Price at harvest is then a function of production (acreage times yield) and a number of demand-side market forces. Accordingly, the vector of 1 In time series literature, structural vector auto-regressions refer to vector auto-regressions that allow for causal interpretations. This use of the word structural is dierent from that in the general econometrics literature; the system of equations need not explicitly model an optimization problem. 2 Note that what matters for this model is not which feedstock is grown (eg, versus switchgrass) but where that feedstock is grown (land previously dedicated to food production versus non-crop farmland). 4

5 variables of interest is y t t soy t supply variables t yield t acreage t total farmland t demand variables t harvest t soy harvest t A generic vector auto-regression with exogenous variables x t has the following structure Ay t = A 1 y t 1 + A 2 y t A k y t k + Cx t + Bε t (1) where ε t N(0, I K ) and E(ε s ε t = 0), s t. We then impose restrictions according to our identifying assumptions. We can re-write the above equation as follows y t = A 1 A 1 y t 1 + A 1 A 2 y t A 1 A k y t k + A 1 Cx t + u t (2) where u t = A 1 Bε t, implying that u t follows a white noise process. Our identifying assumptions are A = a 31 a a 41 a 42 a a 51 a 52 a a 61 a 62 a a 71 a 72 a 73 a 74 a 75 a a 81 a 82 a 83 a 84 a 85 a 86 a a 91 a 92 a 93 a 94 a 95 a 96 a

6 B = b b b b b b b b b 99 Note that there are no restrictions on the lags (A 1 through A k ) or on the coecients on exogenous coecients. The restrictions on A come from the timing of the agricultural production process in the United States. That is, and soybean (which are observed in March for harvest-time delivery) and supply-side variables (such as input costs and loan rates) are generated before acreage decisions or fall s have been observed; accordingly, they are functions of only lagged and exogenous variables. Total farmland, acreage, and yields are determined after s have been observed but before fall s are known. Thus they are functions of s, supply-side variables, and lagged and exogenous variables. This is similar to the acreage function in a Nerlovian (partial-adjustment) framework. Demand-side variables (such as foreign production, aecting demand for US exports) that are determined in the summer are functions of s, supply-side variables, US acreage and yields, and lagged and exogenous variables. Finally, harvest time s are a function of that year's s, supply-side variables, acreage and yield decisions, and demand-side variables. The structure of B imposes orthogonality of contemporary structural shocks. The exogenous variables x t are constants and time trends. We focus on the case of one lag (k = 1) but examine the robustness of our estimates to the inclusion of additional lags. The system is estimated via maximum likelihood. 2.2 Diusion Indices As mentioned above, the orthogonality conditions on the matrix B require that there be no omitted variables. Accordingly we control for spring-time supply-side variables (such as input s and agricultural loan rates) and summer-time demand-side variables (such as US income and foreign production, which aects demand for US exports). However the curse of dimensionality prevents us from including all of these variables in the 6

7 system; we would quickly run out of degrees of freedom. Thus we include factors (also known as diusion indices or principal components) to control for these variables while avoiding the curse of dimensionality inherent in large VAR models. Stock and Watson (2002) show that a large number of time series variables can be summarized with a few indices using principal components analysis. The end result is a linear combination of the original time series, with the linear coecients chosen to incorporate as much of the variation in the original series as possible. This nonparametric approach begins with the objective function ( ˆF, ˆΛ) = argmin[(nt ) 1 i (x it λ i F t ) 2 ] (3) t where (F ) are the factors and (Λ) the factor loadings. This is solved by setting ˆΛ equal to the eigenvectors of X X corresponding to the largest eigenvalues. ˆF is then found by setting ˆF = X ˆΛ/N. This approach is applied separately to two sets of time series variables, one consisting of variables aecting crop supply and one of variables aecting crop demand. For each variable used in the indices we test for a unit root, take the rst dierence of the natural log, and standardize to mean zero and unit variance. Then we estimate a supply-side diusion and a demand-side diusion, which are both incorporated into the structural vector auto-regression. Thus we are able to control for international and domestic macroeconomic disturbances. 2.3 Robustness Checks One potential concern with the above factor-augmented SVAR is its large size. Generally, smaller systems perform better in this framework than do larger systems. This concern obviously needs to be balanced with the potential of omitted variables bias. To address the concern, we also estimate a sparser model as a robustness check. This model contains only and soybean s, acreage, and and soybean farmgate s. Accordingly, the identifying restrictions on matrices A and B are: A = a 31 a a 41 a 42 a a 51 a 52 a

8 B = b b b b b 55 Additional robustness checks include using a log/log specication, allowing additional lags, varying the time period studied, and varying the factor indices for supply- and demand-side variables. 2.4 SVAR Framework for Scenario 2 In scenario 2, we consider the eect of growing a biofuels feedstock on acreage not previously dedicated to food-crop production. Thus we hold own ( or soybean) acreage constant, while increasing total cropland (by decreasing non-crop farmland). Accordingly, the vector of variables of interest is y t t soy t supply variables t yield t acreage t total cropland t demand variables t harvest t soy harvest t The identifying restrictions on A and B are the same as for scenario 1. We again consider various robustness checks, including a sparser specication, additional lags, a log/log functional form, dierent supply- and demand-side es, and a varying time frame. 2.5 Forecast Error Variance Decomposition and Impulse Response Functions The dynamics of the system imply that interpretation of either the reduced form or structural coecients is dicult. Two tools for analyzing the coecients are forecast error variance decompositions (FEVDs) and impulse response functions (IRFs). Forecast error variance decomposition tells us the percentage of the 8

9 forecasting error for a variable due to a specic shock at a given horizon. We can write the h-step forecast made in time t as ŷ t (h) = A 1 A 1 y t 1 + A 1 Cx t (4) Then the h-step forecast error is, as shown by Lutkepohl (1993) h 1 y t+h ŷ t (h) = Θ i u t+h i (5) with Θ i = (A 1 A 1 ) i. Since the reduced form residuals u t are contemporaneously correlated, their individual contributions to the forecasting error cannot be determined. Accordingly we consider the structural FEVD by substituting u t = A 1 Bε t : i=0 h 1 h 1 y t+h ŷ t (h) = Θ i A 1 Bε t+h i = Ψ i ε t+h i (6) i=0 i=0 with Ψ i = Θ i A 1 B = (A 1 A 1 ) i A 1 B. Let ψ mn,i denote the mn-th element of Ψ i, so that the h-step forecast error of the j-th component of y t is h 1 K y j,t+h ŷ j,t (h) = (ψ j1,i ε 1,t+h i ψ jk,i ε K,t+h i ) = (ψ jk,0 ε k,t+h ψ jk,h 1 ε k,t+1 ) (7) i=0 k=1 and the mean-squared error is E(y j,t+h ŷ j,t (h)) 2 uncorrelated and have unit variance. = K k=1 (ψ2 jk, ψ2 jk,h 1 ) since the ε k,t are We can then dene ω jk,h = ψ2 jk, ψ2 jk,h 1 MSE[ŷ j,t(h)] Thus the FEVD at horizon h (for instance, h = 2) estimates the percentage of the total forecast error that comes from each orthogonalized structural shock. The dynamic nature of the above system also allows us to estimate impulse response functions (IRFs), which trace out the eect of exogenous shocks on realizations of the random variables across time. We rewrite the above VAR with its moving average representation: y t = µ + Θ i u t i (8) i=0 with Θ i = (A 1 A 1 ) i. A simple impulse response function plots Θ i across time, but it has no causal interpretation since the reduced form residuals u t are contemporaneously correlated. The structural residuals 9

10 ε t are not contemporaneously correlated, so a causal interpretation is appropriate. Substituting u t = A 1 Bε t into the above equation, we have y t = µ + Θ i A 1 Bε t i = µ + Ψ i ε t i (9) i=0 i=0 with Ψ i = Θ i A 1 B = (A 1 A 1 ) i A 1 B. Thus the structural impulse response function traces each element of Ψ i for each time period following a shock in period t. Standard errors are bootstrapped, as it has generally found that the delta method is inadequate for this sort of highly non-linear system. 3 Data Data is obtained for US production of and soybeans from 1956 to Data on farmland, cropland and planted acreage, measured in thousand acres, is obtained from the National Agricultural Statistics Service (NASS) at the USDA. Crop s paid to farmers, dollars per bushel, are also obtained from NASS. These are then deated by the GDP deator, obtained from the Bureau of Economic Analysis (BEA), into 2007 dollars per bushel. For the supply-side diusion, eleven variables are used. As these aect farmer planting decisions, they reect only information available up until planting time (for and soybeans, rst quarter data). Oil s are given by West Texas Intermediate Oil Prices at the end of the rst quarter, available from Global Financial Data (GFD). The national average loan rate for, soybeans, and wheat is available from the Commodity Research Bureau (CRB) from 1956 to 2003, and from the Economic Research Service (ERS) at the USDA for 2004 to Input s, including automobiles, two USDA-calculated indices of producer s paid, building materials, and farm wages are available from NASS. All s are deated by the BEA's GDP deator. For the demand-side diusion, data should reect information available up until (and including) harvest time. For and soybeans, this implies fourth quarter data. Oil s are given by West Texas Intermediate Oil Prices, available from Global Financial Data (GFD). Third-quarter US Gross National Product is obtained from GFD. Oil s and GNP are deated by the US GDP deator, also from GFD. Corn and soybean production from Southern Hemisphere countries (Argentina and Brazil), which harvest during the US summer months, are available from the Food and Agriculture Organization of the United Nations. The ideal diusion would also incorporate GDP from importing countries, but this is not reliably available on a quarterly basis. Since the timing of innovations is important for the ordering of the VAR, incorporating GDP data updated yearly would be inappropriate. The rst supply accounts for 10

11 30% of the variation in all the series and loads primarily onto the two USDA-computed input s series and oil s. The second supply accounts for 20% of the variation and loads primarily onto loan s. The rst demand accounts for 24% of the variation in all the series and loads onto all six series. The second demand accounts for 23% of the variation in all the series and loads primarily onto soy and production in Argentina. All variables are examined for evidence of unit roots. We consider augmented Dickey-Fuller unit root tests and Phillips-Perron tests, with and without trends, for all variables. Results are presented in table 1. A unit root is rejected at the 5% level for acreage, and soy yields, and the supply and demand diusion indices. We also perform a Kwiatkowski-Phillips-Schmidt-Shin (KPSS) test for each variable, with and without trends. Results are presented in table 1. Trend stationarity is rejected at the 5% level for and soy spot and, soy acreage, farmland, non-crop farmland, and soy yields. [Table 1: Unit Root Tests] 4 Results Estimation of the one-lag nine-equation SVARs for gives the following coecients (table 2.1): [Tables 2.1 and 2.2: SVAR coecients] Thus in scenario 1, the spring depends positively (and signicantly) on lagged acreage and lagged and soy spot s, as expected. The fall spot depends positively and signicantly on the contemporaneous supply and farmland and lagged yields, farmland, and and soy spot s. It depends negatively on the contemporaneous, contemporaneous yields and acreage, and lagged. Note, however, that these coecients are dicult to interpret because a ceterus paribus assumption does not hold in a vector auto-regression. For instance, s aect yields and acreage, which in turn aect the spot. Therefore interpreting the coecient on the in the acreage equation is not straightforward. Coecients for scenario 2 are given in table 2.2. Estimation of the one-lag nine-equation SVARs for soy gives the following coecients (table 2.3): [Tables 2.3 and 2.4: SVAR Coecients] For scenario 1, the spring soy depends positively and signicantly on lagged soy acreage and soy spot. The fall soy spot depends positively on the contemporaneous supply and lagged soy yields and spot s. It depends negatively on soy yields and soy acreage. Interpretation of many of the coecients is again dicult, with the ceterus paribus assumption not appropriate. Coecients for scenario 2 are given in table

12 As described in the modeling section, more intuitive ways to analyze the estimated model are to consider the estimated Forecast Error Variance Decompositions (FEVDs) and Impulse Response Functions (IRFs). The one-lag nine-equation SVAR gives the following structural FEVD for scenarios 1 and 2 with acreage (tables 3.1 and 3.2): [Tables 3.1 and 3.2: FEVD] Not surprisingly, the greatest contributor to forecast error in the is the spot and s in both scenarios. Similarly, the largest contributor to forecast error in the spot is the spot itself. Exogenous shocks to acreage contribute to 5% of the forecast error of the spot (scenario 1) and exogenous shocks to non-crop farmland contribute to 4% of the forecast error (scenario 2). The structural FEVDs for the soy scenarios are as follows: [Tables 3.3 and 3.4: FEVD] Soy spot and s area gain large contributors to each others' forecast errors. The supply is also a large contributor to the forecast error for both soy spot and s. Exogenous shocks to soy acreage contribute to 7% of the error in the soy spot (scenario 1) and shocks to non-crop farmland contribute to 10% of the forecast error. Thus for both crops and both scenarios, shocks to acreage contribute to only a fraction of the two-step forecast error in spot s. This is not surprising, since spot and s are expected to heavily impact forecast error. Figure 1 shows some of the structural impulse response functions estimated from the one-lag nine-equation model for scenarios 1 and 2. In particular, we show the IRF graphs for the eect of negative (and soybean) acreage shocks to own and the eect of negative non-crop farmland shocks to own and soybean s. 3 For every 1 million acres of production removed, increases in the rst period by $0.04 per bushel. The eect lasts one additional period, and then falls back to zero. For every 1 milion acres of soy production removed, soy increases in the rst period by $0.23 per bushel. The eect is more persistent than it was for, although it does fall back towards zero. We expect the eect to be higher for soybeans than for, as US production of soybeans has historically had a much larger share of world production than has US. For a negative one million acre shock to non-crop farmland (holding acreage constant), falls. The initial eect is $0.02 per bushel, peaking at $0.05 per bushel. Eventually the eect returns to zero. Holding soybean acreage constant, the negative shock to non-crop farmland is $0.08 per bushel. What appears to be happening in the models for both crops is that, holding own acreage constant, acreage of other crops is rising. Since grains are largely substitutable, this takes pressure o of the demand for the crop. 3 Each SVAR has nine equations and therefore 81 estimated IRFs. These results are available upon request. 12

13 Since own acreage was held constant, supply doesn't change, and the fall in demand lowers s. This intuition could correspond historically to increases in, for instance, other food crops like wheat. There is no reason to expect the story not to hold for other biofuels feedstocks, like miscanthus or switchgrass. 5 Robustness Checks As described in the modeling section, a number of robustness checks are considered. For instance, a far sparser SVAR (with only ve equations) is considered for both crops and both scenarios. Results are quite similar (gure 2). A negative one million acre shock to production raises s by $0.07. For soybeans, the initial increase is $0.23 per bushel for a one million soy acre shock. Decreasing non-crop farmland (holding own acreage constant) lowers s by $0.05 and soy s by $0.10. The magnitude of the eects is again similar for a log/log specication. [Figure 2: Robustness Checks] The SVAR is also estimated with additional lags allowed in the system (gure 2). The BIC-selected model is one lag for both scenarios and both crops. For scenario 1 with, the estimated IRF is quite similar for two lags. With three lags the initial eect is similar but the eect in later periods is unstable (and implausible). For scenario 1 with soy, the general eect is similar but (implausible) oscillations appear. For scenario 2, the eect on s is not robust to the inclusion of additional lags. The eect on soy s is robust to two lags but not three. Next the SVAR is estimated with dierent diusion es (gure 2). The above results were for the inclusion of the primary diusion es, which loaded primarily onto input s (supply ) and fairly evenly across US GNP and southern hemisphere agricultural production. The model is also estimated with the secondary diusion es, which loaded primarily onto loan s (supply ) and Argentine and soy production. The results are again similar to those in the main specication (gure 2). Finally, the SVAR is estimated with a varying time period. The same SVAR is estimated eleven times, with 39 years included in each estimation (i.e., 1958 to 1997, 1959 to 1998, etc.). As can be seen from the estimated IRFs (gure 3), the results are quite robust to varying the time frame: [Figure 3: Rolling Time Frame] 6 Conclusion: summary, results, contributions, caveats, future work Using a dynamic system of simultaneous equations, we explore the impacts on crop s of changes in land use. We develop a structural vector autoregression, allowing us to analyze impulse response functions and 13

14 forecast error variance decompositions. These econometric tools, common to macroeconomic applications, provide elegant descriptions of the dynamics of the agricultural production process. We nd signicant and sustained increases in and soybean s when crop acreage is removed. For a reduction in area of one million acres, we estimate a increase of $0.04 to $0.07 per bushel. To put this in perspective, consider that from 2006 to 2007, acreage dedicated to ethanol increased by about 6 million acres in the United States. At the same time, s rse $0.80 per bushel. Thus our model nds that approximately 30 to 50% of the increase was due to new ethanol production. For a negative shock in soybean acreage, we nd a increase of $0.08 to $0.23 per bushel. This much larger magnitude is partly explained by the fact that US soybean acreage is a much larger share of world production than is US acreage. We also nd signicant and sustained decreases in crop s when own acreage is held constant and total cropland is increased. A 1 million acre increase in US cropland leads to an approximately $0.05 to $0.10 decrease in and soybean s. What sets our model apart from most is that the results extend to the production of biofuels besides ethanol. Any biofuel feedstock that is grown on land previously dedicated to will increase s; this is crucial as the US considers the production of second- and third-generation biofuels. A number of caveats should be mentioned. First, the magnitudes we see depend on the US share of world production. If this were to change substantially, we might expect a dierent multiplier. Second, the pathways for the observed responses are only hypothesized. For a scenario in which crop acreage is removed, it is intuitive that the crop's will rise. Supply is constrained by the removal of acreage, a crucial input in the production process, and demand for the food crop has not changed. The intuition for increasing production on non-crop farmland is also fairly intuitive: the crop's supply has not changed (acreage for that crop is held constant) and demand will fall (as a substitute crop is grown). However, a structural model is needed to conrm these ndings. Our ndings open up a number of possible extensions and future research projects. Scientists and policymakers have expressed hope that new biofuels feedstocks will be grown on land that does not compete with food crops, thus avoiding the eects of biofuels on food s. Our evidence is suggestive that, if these additional crop lands released pressure from the market, s could indeed decline. Verifying this, by analyzing the causal pathways at work, will be crucial as biofuels policies move forward. References [1] BBC Will biofuel leave the poor hungry? [2] Bureau of Economic Analysis. 14

15 [3] Commodity Research Bureau Commodity Yearbook. [4] Economic Research Service. [5] Food and Agricultural Organization of the United Nations. [6] Food and Agricultural Policy Research Institute Documentation of the FAPRI Modeling System. FAPRI-UMC Report # [7] Global Financial Data. [8] Hamilton, J. D Time Series Analysis. Princeton, NJ.: Princeton University Press. [9] Lutkepohl, H Introduction to Multiple Time Series Analysis. Lutkepohl, Helmut Berlin: Springer-Verlag. [10] Mushtaq, K. and Dawson, P. J Acreage Response in Pakistan: A Co-Integration Approach. Agricultural Economics 27: [11] National Agricultural Statistics Service. [12] Rajagopal, D., Sexton, S. E., Roland-Holst, D., and Zilberman, D Challenge of Biofuel: Filling the Tank Without Emptying the Stomach? Environmental Research Letters 2:1-9. [13] Searchinger et al Use of U.S. Croplands for Biofuels Increases Greenhouse Gases Through Emissions from Land-Use Change. Science 319: [14] Sexton, S., Rajagopal, D., Zilberman, D., and Hochman, G Food Versus Fuel: How Biofuels Make Food More Costly and Gasoline Cheaper. Agricultural and Resource Economics Update. Sep/Oct [15] Sims, C. A Macroeconomics and Reality. Econometrica 48(1): [16] Stock, J. H. and Watson, M. W Forecasting Using Principal Components from a Large Number of Predictors. Journal of the American Statistical Association 97(460): [17] Ugarte, D.G., et al., The Economic Impacts of Bioenergy Crop Production on US Agriculture. Report prepared for the US Department of Energy and the US Department of Agriculture. 15

16 Table 1: Unit Root Tests Augmented Dickey-Fuller Test KPSS Test with with Time Trend Time Trend test statistic p-value test statistic Corn Soy Corn Soy Corn acreage Soy acreage Farmland Noncrop farmland Corn yield Soy yield Supply Diffusion Index Supply Diffusion Index Demand Diffusion Index Demand Diffusion Index The null hypothesis of the ADF test is that the variable contains a unit root. The null hypothesis of the KPSS test is that the variable is trend stationary. A Phillips-Perron test with time trend gives similar results, as do ADF and PP tests without time trends. Exceptions are yields, where the null is (not surprisingly) not rejected in the tests without trends. The 5% critical value for the null hypothesis in the KPSS test with time trend is

17 17 Table 2.1: Estimated SVAR Coefficients Scenario 1, soy supply contemporaneous yields farmland acr. demand harvest soy harvest soy supply one period lag soy supply yields farmland acreage demand harvest soy harvest yields farmland acr. demand harvest soy harvest linear trend Corn and soy and harvest in 2007 dollars per bushel. Farmland and acreage in million acres. Asymptotic standard errors in parentheses. Intercept not reported

18 18 Table 2.2: Estimated SVAR Coefficients Scenario 2, soy supply contemporaneous yields acr. noncrop farmland demand harvest soy harvest soy supply one period lag soy supply yields acreage noncrop farmland demand harvest soy harvest Corn and soy and harvest in 2007 dollars per bushel. Non-crop farmland and acreage in million acres. Asymptotic standard errors in parentheses. Intercept not reported yields acr. noncrop farmland demand harvest soy harvest linear trend

19 19 Table 2.3: Estimated SVAR Coefficients Scenario 1, soy soy supply contemporaneous soy yields farmland soy acr. demand harvest soy harvest soy supply one period lag soy supply soy yields farmland soy acreage demand harvest soy harvest soy yields farmland soy acr. demand harvest soy harvest linear trend Corn and soy and harvest in 2007 dollars per bushel. Farmland and soy acreage in million acres. Asymptotic standard errors in parentheses. Intercept not reported

20 20 Table 2.4: Estimated SVAR Coefficients Scenario 2, soy soy supply contemporaneous soy yields soy acr. noncrop farmland demand harvest soy harvest soy supply one period lag soy supply soy yields soy acreage noncrop farmland demand harvest soy harvest Corn and soy and harvest in 2007 dollars per bushel. Non-crop farmland and soy acreage in million acres. Asymptotic standard errors in parentheses. Intercept not reported soy yields soy acr. noncrop farmland demand harvest soy harvest linear trend

21 response variable response variable Table 3.1: Forecast Error Variance Decompositions (Two-Step Ahead) Scenario 1, impulse variable soy supply yields farmland acr. demand harvest soy harvest soy supply yields farmland acreage demand harvest soy harvest Table 3.2: Forecast Error Variance Decompositions (Two-Step Ahead) Scenario 2, impulse variable noncrop farmland demand harvest soy harvest soy fut- supply ures yields acr soy supply yields acreage non-crop farmland demand harvest soy harvest

22 response variable response variable Table 3.3: Forecast Error Variance Decompositions (Two-Step Ahead) Scenario 1, soy impulse variable demand harvest soy harvest soy fut- supply soy farmures yields land soy acr soy supply soy yields farmland soy acreage demand harvest soy harvest Table 3.4: Forecast Error Variance Decompositions (Two-Step Ahead) Scenario 2, soy impulse variable noncrop farmland demand harvest soy harvest soy fut- supply soy ures yields soy acr soy supply yields acreage non-crop farmland demand harvest soy harvest

23 change in soybean, dollars per bushel soybean change, dollars per bushel 23 change in, dollars per bushel change in, dollars per bushel Figure 1: Impulse Response Functions Scenario 1: negative shock (1 million acres) to acreage Scenario 2: negative shock (1 million acres) to non-crop farmland year year Scenario 1: negative shock (1 million acres) to soybean acreage 0.5 Scenario 2: negative shock (1 million acres) to non-crop farmland year -0.2 year

24 change in soy, $ per bushel change in soy, $ per bushel 24 change in, $ per bushel change in, $ per bushel Figure 2: Robustness Checks 0.08 Scenario 1, Corn Sparse 0.08 Scenario 2, Corn Sparse 0.06 Sparse, Two Lags 0.06 Sparse, Two Lags Sparse, Three Lags Main Main, Two Lags Main, Three Lags Sparse, log/log Main, Secondary DIs Sparse, Three Lags Main Main, Two Lags Main, Three Lags Sparse, log/log Main, Secondary DIs 1 Scenario 1, Soy Sparse 0.15 Scenario 2, Soy Sparse Sparse, Two Lags Sparse, Three Lags Main Main, Two Lags Main, Three Lags Sparse, log/log Main, Secondary DIs Sparse, Two Lags Sparse, Three Lags Main Main, Two Lags Main, Three Lags Sparse, log/log Main, Secondary DIs

25 change in soy change in soy 25 change in change in Figure 3: Rolling Time Frame The estimated IRF for the entire sample is shown with a dashed, red line. The rolling window estimation results are shown in gray. Scenario 1, Scenario 2, Scenario 1, soy Scenario 2, soy