One cannot overstate the policy importance of obtaining reasonably accurate estimates of the effects of road improvements on both economic activity and land clearing in the Amazon. The Brazilian government has earmarked billions of dollars for road construction and paving through the region under the Avança Brasil plan and the future of the Amazon forest, and the well being of its many human inhabitants, will be significantly effected for better or for worse accordingly. A better understanding the relative trade-offs between economic growth and environmental damage will greatly enhance the ability of policy makers to design more sensible road systems that are consistent with societies’ values (which society's values is another paper).
That a very strong, robust correlation exists between the level of clearing and the level (i.e. extent) of roads cannot be denied. However, on the basis of this strong levels-correlation very strong conclusions have been drawn about the causal nature of this relationship. Nevertheless, a strong levels-correlation accompanied with a plausible sounding theory does not prove that causality goes from roads to land clearing. Pragmatically, even if there is a great deal of causality from roads to deforestation, if there is also reverse causality then estimates that do not take this into account could overstate the effect of exogenous road building on the environment.
Thus in this paper we have taken a rich data set with both time series and spatial variation at the municipio level for legal Amazonia and subjected it to battery of dynamically rigorous tests. We specify dynamic rather than static models, difference out levels fixed effects, and condition on past variables to minimize endogeneity. We further allow for municipio-specific heterogeneity in the relationship between transport costs and land clearing.
In the first instance we follow Andersen et al. (2002) and use an automated “random reduction” technique for dealing with a large number of candidate explanatory variables without possibility of bias or interference from the researchers. These regressions suggest that there is a great deal of heterogeneity in the response of land clearing to changes in transport costs. In particular, decreasing transport costs in cleared areas lowers deforestation, while road building (and paving) in less settled areas definitely increases deforestation. Given that these are rough estimates based on aggregate data we may justifiably emphasize that the pattern of land clearing is most important for policy purposes. In fact the policy implications of our analysis is quite clear: to minimize environmental damage policymakers should focus on improving transport costs where there is already an established economic center of activity and avoid uncleared land (whether this is an optimal strategy from an economic perspective is another study).
In order to investigate why our dynamic results differ so greatly from most of the existing (static) studies, we then subject the transport cost and land clearing variables to a series of strict out-of-sample forecasting exercises to ascertain whether past information of each variable is useful in predicting the evolution of the other variable. We find consistent evidence that the evolution of land clearing patterns has greater power explaining the future evolution of transport costs, suggesting that causality could run from land clearing to transport costs. If there in fact is reverse causality in this relationship, static analyses will be prone to endogeneity bias. We then directly address the possibility of endogeneity in the relationship by instrumenting for transport costs in a regression equation of land clearing. We find that once variation in transport costs has been rendered exogenous by instrumenting, they are not correlated with land clearing. In fact, just by introducing heterogeneity (i.e. interaction terms with extent of initial cleared land) into the regression, even the transport cost variables in the OLS contemporaneous regression lose strong statistical significance and start to show evidence of the same heterogeneous pattern detected more robustly in the dynamic estimation. Thus we find two likely reasons to explain why our results differ from the more conventional findings; first because our dynamic approach suffers less from endogeneity bias and, second, because we allow for heterogeneity in the relationship across municipios.
Taking these results at face value suggests several possible, but not necessarily mutually exclusive, explanations. The first scenario is a variation on the Kaimowitz and Angelsen hypothesis: paving roads (i.e. decreasing transport costs) in areas that have established settlements raises land values and encourages agricultural intensification, lowering the pressure on surrounding forest areas. Decreasing transport costs into relatively pristine areas, however, has the expected effect of increasing agricul