1. Land Use and Transportation Costs in the Brazilian Amazon Diana Weinhold, LSE Eustaquio Reis, IPEA LBA III Conferencia, Brasilia July 28, 2004 "Economists on the loose again, Oh, dear." -Susanna B. Hecht, Prof. of urban planning, UCLA

2. Many previous empirical studies have tended to show a positive correlation between the extent of roads network and the degree of land clearing: Chomitz and Gray (1995) in Belize Cropper, Griffiths and Mani (1996) in Thailand Nelson and Hellerstein (1995) in Mexico Liu, Iverson and Brown (1993) in the Philippines Pfaff (1998) in Brazil.

3. One characteristic that most previous studies share is that they are more or less static, examining the (usually) contemporaneous relationship between the levels of transport costs and the levels (i.e. extent) of cleared land (or probability of clearing). For example, Laurance et al. (Science 2001) overlays the 1995 road network on a 1992 map of the Amazon and calculates the extent of clearing on each side of the road. Thus they causally attribute all of the cleared land to the road in question.

4. NASA: “The pattern of deforestation spreading along roads is obvious in the lower half of the image. “ SOURCE: NASA

5. Two more recent studies, by contrast, have estimated explicitly dynamic models of land use in the Brazilian Amazon. Andersen et. al (2002) Moreira (2003) Contemporaneously, Pfaff, Reis and Walker are utilizing an equally dynamic framework for analyses at a much greater level of disaggregation, with the preliminary results to be presented here at LBA.

6. This emerging body of literature suggests that the relationship between roads and clearing may be somewhat more complex than the conventional wisdom implies. In theory at least, lower transport costs could either raise, lower, or have no effect on deforestation depending on the assumptions. For example, Schneider (1994) and Chomitz and Gray (1995) both point out that (in theory at least) road intensification, as opposed to extensification, could possibly be a “win-win” strategy that both increases economic output while decreasing land clearing.

7. This paper: we revisit Andersen et. al. (2002) using the preferable transport cost data from Moreira (2003) Purpose of analysis is specifically to examine relationship between transport costs and land clearing Keep in mind any challenge to a well established result bears the burden not only of demonstrating robust conclusions, but it should also be an encompassing study in that in addition it should be able to explain why previous studies obtained the results that they did.

8. DATA: DESMAT data from IPEA Hundreds of variables of ecological, economic and agricultural conditions collected for the years 1975, 1980, 1985 and 1995 257 consistently defined areas in Brazilian Amazonia transport cost data from GIS analysis Basic Functional Form: following Andersen (2002) Growth_endog.  [time invariant controls], [lagged (1985-1995) endog. vars], [lagged exog. vars], [spatial vars] Where: time invariant controls include state and city dummies, area, natural vegetation, soil type, distance to state and federal capital, rivers, ave. temp and rainfall

9. Random Reduction results for Cleared Land Net total effect of level of TC (Sao Paolo) = +1.45

10. Our results suggest that for municipalities with a relatively low percentage of cleared area, decreasing transport costs will indeed increase deforestation rates. However, for municipalities with a certain extent of existing settlement, this relationship is weaker and might even be reversed. Theoretically our results are consistent with the idea that a fall in transport costs induces land use intensification and higher productivity in agriculture, which, in theory, can lead to less deforestation if the price elasticity of demand is low. There is some evidence that this might indeed be the case: Moreira and Reis (IPEA 2002) find that lower transport costs increase productivity Santana (FCAP 2002) finds that demand for beef in the Amazonian state of Pará is price inelastic.

11. Out of Sample Model Evaluation Exercises Based on Granger and Huang (1997) Models 1 and 2 are competing models of some variable Y Model 1 contains information on the past of Y and on the past of X, while model 2 contains information only on the past of Y If model 1 out-performs model 2 in out-of-sample forecasting, then we can say that the information from past X was useful; otherwise not and there is evidence of Granger non-causality. Out-of-sample forecasting ability is a much tougher ‘hurdle’ for a model to pass than simple in-sample fit

12. Out of Sample Model Evaluation Exercises 1985-1995 Note: Each regression also controls for State dummy variables, municipality size, distance to state and federal capital, and a CITY dummy for large urban municipalities.

13. Out of Sample Model Evaluation Exercises 1980-1985

14. Instrumental Variables Estimation: First Stage Regressions All first stage regressions also additionally control for: STATE1, STATE2, STATE3, STATE4, STATE5, STATE6,STATE7, STATE8, LAREA, LDIST1, LDIST2, and CITY.

15. Instrumental Variables Estimation Dependent variable: growth of cleared land ‘85-’95 Note: Additional control variables: State dummies, municipality area, distance to state and federal capital, CITY dummy.

16. Instrumental Variables Estimation Including interaction terms

17. Conclusions and Discussion Strong case for dynamic versus static analyses Preliminary evidence suggests that road intensification in settled areas could be a win-win strategy (?? Better data from Pfaff et. al. may shed more light on this) Either way there are strong policy implications for the road planning under Avanca Brasil Some indication that direction of causality may be more complex than conventional wisdom Needs more research Taken at face value, casts serious doubt on a number of influential studies (in Science, for example) that extrapolate from simple levels-correlations when making climate and environmental predictions about the effects of roads.