EBB and f(al)low
On 2 September, the European Biodiesel Board (EBB) presented in Brussels results of a new study of indirect land use change, performed by a group of North American modelers using the Global Trade Analysis Project (GTAP) economic model. The modelers are independent of EBB, but for brevity I’ll refer to them as ‘the EBB modelers’ and the study as ‘the EBB report’ henceforth. The headline result of the study was a relatively low prediction for iLUC (indirect land use change) emissions, the emissions caused when agriculture expands to meet increased biofuel demand. EBB have presented this result as a major challenge to a proposal shortly to be voted on by the European Parliament that ‘iLUC factors’ based on an earlier study by the International Food Policy Research Institute (IFPRI) (Laborde (2011), should be applied in European legislation (as they already are in California’s LCFS, for instance).
There are two main themes in the EBB press release accompanying the study. The first is to argue that this study shows that iLUC emissions are lower than previously thought, “Biodiesel could account for as little as 2,33gCO2eq/MJ, compared to current 55gCO2eq/MJ allocated in a Commission proposal”1. The second contention is that by producing different values, the study shows that iLUC is too uncertain to regulate, “The divergence of results due to a slight change in assumptions, once again, opens the floor to question the validity of ILUC science for policy making.” However, the Laborde (2011) work remains in my opinion the best available EU scientific evidence on iLUC (as I argued in GCB Bioenergy last year), and the usefulness of the new modeling is undermined by some very strong assumptions and technical modeling issues.
Why are the numbers so small?
There are several aspects of this new modeling that make it unsurprising that the results come out differently than the Laborde (2011) modeling. As with any economic modeling exercise, this is a complicated and detailed modeling effort, and there are many interesting points of comparison that could be considered, but there are five things which I think are crucial to understanding this result:
- The GTAP biofuel model developed by Purdue has been tuned to represent the U.S. market. Setting up new database categories (such as disaggregating rapeseed, palm and soy vegetable oils) and calibrating to the European market is not a trivial task, and this modeling team have not had the same time and capacity to develop their model version as IFPRI has with MIRAGE. It would be surprising if the modelers adapting GTAP were able to achieve comparable model and data quality ‘first time’.
- From the results, it looks like the new model fails to capture adequate connectivity in vegetable oil markets. We’ve argued before that vegetable oils are fungible at the macro level, and that increased European vegetable oil demand of any sort will drive increased palm oil production, and many experts see palm oil as the world’s ‘marginal’ oil (the one most responsive to additional demand). At the moment, the new model seems to have this backwards—palm oil demand is a strong driver of deforestation in the west (according to Table 17 forty percent of the iLUC attributed to palm oil biodiesel comes from deforestation in the U.S., EU and Canada) but other oils only very weakly drive palm demand.
- GTAP is not able to model the conversion of unmanaged land, such as rainforest or natural grassland, whereas MIRAGE does. This is especially problematic in areas like Southeast Asia where we know that a great deal of new land comes from unmanaged peat forests.
- Overruling the economics of the model, it is assumed that even in managed forest areas there can be no deforestation in the U.S., EU or Canada, regardless of economic signals.
- Most importantly, fallow land has been modeled in a way that biases the model towards very low iLUC. European fallow land availability has been modeled as a yield effect, rather than as a land category, which is difficult to justify. Rather than determining that fallow land will be used to grow biofuel feedstock through the economics in the model, this result has been assumed and imposed exogenously. There is no data to support the specific parameter used, the modelers admit that the way it has been implemented is far from ideal, and the assumption that there is zero carbon loss when fallow areas are returned to agricultural production is generally not correct.
Even ignoring the first three points, if one were to remove only the new treatment of fallow land and rule against deforestation in EU/U.S./Canada, the iLUC for rapeseed would have been 20 gCO2e/MJ according to the report, a result which is rather closer to the normal range for biodiesel results from iLUC studies.
What does this mean for certainty?
While the European Biodiesel Board has presented this modeling effort as new and improved, they have also said in public statements that they can’t say for sure whether it is better or worse than MIRAGE. This, the argument goes, serves to emphasize the uncertainty in the field of iLUC modeling, and that it’s not possible to regulate using iLUC factors based on models. The ICCT have always recognized that there is uncertainty in iLUC results (for instance we apply a Monte Carlo analysis to quantifying it in this paper from last year), but we have argued that there is still an adequate body of evidence to use iLUC numbers in regulatory treatments. This is not the only new study on iLUC presented recently. Results from a JRC assessment of iLUC based on historical data, presented in the European Parliament earlier in the year, put iLUC numbers similar to or higher than the Laborde (2011) results. A new study by the JRC with AgLink estimates significantly larger areas of land use change than Laborde does. Modeling iLUC is a complicated process, and as the EBB note changing important parameters can affect the results significantly (we shouldn’t expect anything else)—however, simply producing a new study doesn’t render previous work irrelevant, especially while there are still important questions to be answered about this modified GTAP framework. There are a variety of reported results in the literature, and while this one gives a lower value than MIRAGE, several give larger numbers. For all this, Laborde (2011) is still the most convincing study available in terms of methodology and coverage, and the use of iLUC factors in European legislation, based on the best available evidence for the expected magnitude of iLUC, would still be a reasonable policy response.
More on the modeling
The five features I just listed are at the core of explaining the very low results from this modeling exercise, but (for those who are interested) I’ll run through a bit more detail about how this study differs from the MIRAGE work by IFPRI.
GTAP can be thought of as a family of models with a common core, tailored to study different issues. The version of GTAP used by the authors of the EBB study is based on a line of model development led by Purdue University. It has branched from the version of the model used to model iLUC for California’s Low Carbon Fuel Standard, but has had several changes made in the intervening period. The EBB study’s authors have then modified the model further still to enable it to investigate issues relevant to the EU (the Purdue model is tailored to U.S. markets). Indeed, when Purdue ran the LCFS version of the model2 for a 2010 study by the EU’s Joint Research Centre (JRC), it came out with relatively high iLUC values—about 60 gCO2e/MJ for a European biodiesel mix, and over 150 gCO2e/MJ for European wheat ethanol.
There have been several changes made by Purdue since 2010. Here, I want to talk about three points in particular—one change introduced by Purdue in 2010, and two made specifically for the EBB study. The innovation introduced by Purdue affects the yield of new land brought into production, the changes made for EBB relate to the way the model deals with vegetable oils and biodiesel, and the attempt to represent fallow land.
There are several innovations introduced by Purdue since they modeled EU biodiesel for the JRC that tend to reduce the iLUC numbers. The EBB paper highlights the treatment of marginal yield in particular—the model is more optimistic about yields on new land than it was in the past. This comes from a complicated assessment of biological potential based on the ‘Terrestrial Ecosystem Model’. I won’t go into the details here, but suffice to say that while this approach is certainly more sophisticated than what is done in MIRAGE, it is unclear to me at least whether the results it delivers are really more accurate (this is something that we discussed extensively at the Advisory Panel to the California Low Carbon Fuel Standard without a clear conclusion). It’d be great to see data showing that the TEM results are really a good approximation to marginal yields, but other experts (such as JRC, 2010) have argued that in fact marginal yields are lower rather than higher than MIRAGE predicts.
The version of GTAP obtained from Purdue does not distinguish between different vegetable oils—in particular, it does not disaggregate soy oil from rapeseed oil from palm oil. The EBB authors therefore had to amend the model and database to allow them to model soy, rapeseed and palm biodiesel separately. The way this is done is important to the results—depending on the way that the different oil markets are parameterized, we will see land use change in different areas. For instance, I have previously argued that there is good reason to believe that rapeseed demand for biodiesel drives palm oil demand. If the model reflects such as assumption, we might expect rapeseed to cause deforestation in Indonesia. If, on the other hand rapeseed is not linked to palm, the land use change might all happen in Europe. The EBB report does not include the same detail about where land use changes happen as is reported by Laborde (2011), but we can use some proxy data to get a sense of what is going on. EBB have followed IFPRI in assuming that palm oil expansion is linked to peat loss, and thus peat emissions. From Table 15 in the EBB report, we can infer that about 40% of the iLUC emissions associated with palm oil biodiesel are from peat loss—i.e. that if palm expansion is significant in a scenario, we expect to see significant peat emissions. In the rapeseed biodiesel scenario, however, peat emissions constitute only perhaps around 3% of the iLUC, compared to nearly 30% in Laborde (2010). The story is similar for soy biodiesel (about 10% of emissions coming from peat, compared to 30% in Laborde). Clearly, in this version of GTAP, the links between soy or rapeseed oil and palm oil are relatively weak, and therefore it seems likely that this GTAP modeling does not provide a good description of the global vegetable oil market.
The second modeling innovation for the EBB that I want to talk about is the way the authors have tried to capture the possibility of using fallow land. They argue that there are extensive areas of fallow land available in Europe, and that these areas will be used for biofuel in preference to, for instance, converting forest. The modeling problem is that GTAP does not explicitly represent these fallow areas, and so the authors have to find a ‘fix’ to represent the effect they expect to see. Specifically, the ‘elasticity of yield to price’ has been raised by a factor of four, the argument being that because bringing fallow land back into production requires no ‘new’ land, it is in some sense comparable to increasing yields. When the elasticity of yield to price is increased, it tells the model to assume that more biofuel feedstock will come from yield and less from new land, so doing this is bound to reduce iLUC emissions.
There are several problems with doing things this way though. Firstly, the case that fallow land will be preferentially used is not really adequately made. The paper presents a limited correlation between expanding rapeseed area and reducing fallow, but this isn’t sufficient to justify the very strong assumption this paper makes in favor of fallow land. Biofuel industry representatives have repeatedly made this argument that there is abandoned or fallow land in Europe that they believe will be used before other land is converted, but the case has never been made convincingly. To really justify this argument, you would need to show:
- that there is indeed fallow land available;
- that it will be economically preferable to bring fallow land back into production rather than displacing other crops or increasing imports;
- that the fallow land will only be used because of biofuel demand, and would not otherwise be brought back into production.
While the first point is somewhat valid, it is irrelevant to state that an area of land exists unless there is compelling reason to believe that it will be used in preference to other areas to meet biofuel demand specifically. Because this new report assumes that fallow land will be used, rather than actually predicting it from fundamental economics, it doesn’t really move this debate forward.
Aside from the fact that the use of fallow land is effectively input into the model rather than output from it, the way that yield has been used to implement it is problematic. The authors would have preferred to introduce a fallow land category to the GTAP database, but as they were not able to do this they have ratcheted up the yield response as an alternative. Introducing fallow land in this way will introduce various behaviors into the model that are not really descriptive of the availability of an additional land area. The chosen yield elasticity, 1.0, is four times the value used in GTAP (which has been already criticised by some experts as being potentially rather too high in the first place). The report makes no justification for picking this level rather than any other, and between the lack of compelling data to support the basic hypothesis that most biofuel will be grown on fallow land, the use of a yield parameter to represent a land type, and the lack of empirical basis for the parameter choice, it’s hard to see this treatment as preferable to the treatment of fallow land in MIRAGE (where it is included in the database but largely already brought into production in the 2020 baseline, before extra biofuel demand is added).
A note on transparency
In parallel to presenting this new work, EBB has accused the MIRAGE model and Laborde (2011) of being intransparent. One difference between the MIRAGE and GTAP models is in the way that the models are ‘maintained’. The version of MIRAGE used by David Laborde for the European Commission modeling is curated by the International Food Policy Research Institute (IFPRI). It is coded by IFPRI, and cannot be run by third parties without IFPRI’s assistance. GTAP, on the other hand, is an open source model (or, more accurately, a family of open source models) meaning that the source code is available on the internet for anyone to read. The ICCT is certainly not against open source modeling, but there is much more to transparency than whether a model is open sourced or not. Especially in the context of public policy making, there is a real limit to the value to most stakeholders of being able to access source code. I have indeed downloaded and poked around in the GTAP source code3, and while this is a valuable opportunity to experts with an adequate background, I doubt anyone would disagree that it’s not normally the best way to learn how a model actually works—what’s much more important is good documentation. While the code of MIRAGE is not available, the biofuel model is documented thoroughly in IFPRI’s 2010 and 2011 reports for the Commission, and (similarly to GTAP) in the back catalogue of publications through which the model was developed. In addition to content from IFPRI, the ICCT has written an introductory briefing on the model and its results, and It has been reviewed in publically available papers at least three times, by us, Kiel Institute and (S&T)2. The Commission has presented the results at public workshops, and David Laborde and other experts with a degree of understanding of the model have presented results and answered questions at conferences and workshops on numerous occasions. Aside from anything else, Laborde (2011) actually presents a larger set of results and a broader discussion of those results and data than is presented in the EBB paper—that is meant not as a criticism of the presentation of the EBB work, but to again underline that there is much more to transparency than simply having source code available.
There have been nearly two years since the Laborde (2011) report was published during which stakeholders have been able to contest, discuss and examine it as the Parliament and Council move to define their respective positions on the Commission’s iLUC proposal. I would argue that it has stood up well to that scrutiny. This new study from EBB is entirely appropriate as a contribution to that discussion, but being published only a week before Parliament votes, there is no time for it to be put through a comparable process of scrutiny—although, as outlined here, there are several questions that are immediately apparent. There is always space for improved public engagement on difficult technical issues such as iLUC, but it seems to me to be a little unfair of EBB to criticize IFPRI in the way it has for a lack of transparency, when the EBB itself has on occasion neglected to take advantage of the opportunities it has had to have any questions constructively answered, on several instances preferring instead to attack the model, the modelers and indeed the whole premise of iLUC modeling.
1. U.S. and UK readers should note that in continental Europe standard usage has commas to delineate the decimal point, so 2,33 is the continental equivalent of 2.33, and the EBB is not asserting iLUC emissions of two thousand three hundred and thirty gCO2e/MJ.
2. To be precise, while the model run for the JRC was much closer to the LCFS version of GTAP, there were still several differences. JRC (2010) note, “Some elements of the model and the data base used in the analysis reported here differ from the data and model used for the California Air Resources Board. With respect to the data base, the differences in particular include a) new structure of oilseeds biodiesel production and production of oilseeds meal–by-product of vegetable oil19, and b) improved representation of the EU wheat ethanol sector. More specifically, in the data used in this project, oilseeds meal is a by-product of vegetable oil, not a by-product of oilseed biodiesel.”
3. A representative example of a chunk of GTAP code runs as follows:
!SSA Yield Stuff!
ZERODIVIDE (ZERO_BY_ZERO) DEFAULT 0 ;
ZERODIVIDE DEFAULT 1 ;
THETAi(j,r) #Endowment's cost share in value-added#;
THETAi(j,r) = LANDCOST(j,r)/TOTALC(j,r);
YDONOFF(j,r) #Turn off the yield calibrated ESUBVA's by setting to zero#;
Read YDONOFF from file GTAPPARM header "YD01";
YDREGSCALE(r) #Scale up or down the target yield elast for a region#;
Read YDREGSCALE from file GTAPPARM header "YDRS";
YDE_Target #Scalar yield elasticity target, read from PARM for SSA#;
Read YDE_Target from file GTAPPARM header "YDEL";
Coefficient (parameter) (all,j,ALL_INDS)(all,r,REG)
# elasticity of substitution in value-added-energy subproduction #;
ESUBVA1 from file GTAPPARM header "ESBV";
Coefficient (parameter) (all,j,ALL_INDS)(all,r,REG)
# elasticity of substitution in value-added-energy subproduction #;
ESUBVA(j,r) = if(THETAi(j,r)*YDONOFF(j,r) le 0, ESUBVA1(j,r))
+ if(THETAi(j,r)*YDONOFF(j,r) gt 0,
!SSA Yield Stuff!