Data.

We assembled a country, regional, and road-specific dataset that combines information about deforestation, land use, road network (including road length and surface type—paved, gravel, or earth), and other possible determinants of deforestation, potential environmental variables (e.g., ecoregion and biomass), financial and maintenance costs of proposed road projects, number of schools and hospitals within 20 km of each proposed road, and other social and cultural variables. In total, the analysis considered 30 variables. For each, we used the most recent high-quality data source available (see SI Appendix, Table S4, for a description of all data used).

Road projects were identified for inclusion based on two steps. First, we reviewed official documentation and consulted with road agency staff to identify as many road projects as possible that were prioritized for implementation by national governments. Identified projects included both road improvements and new construction. Second, we eliminated from this group both projects outside the Amazon Basin as delimited by the Amazon Geo-Referenced Socio-Environmental Information Network (RAISG) and those for which basic information (e.g., road surface type or specific route) was insufficient to carry out further analysis. The selected projects cover 12,263 km (7,620 miles), with a total proposed investment of US$ 27 billion.

Method.

We used a multicriteria approach that integrates a diverse set of issues, including standard road costs and benefits, into a single index. In this section, we present a detailed description of the assumptions and methods used to calculate the environmental, social, and economic indicators, as well as the efficiency measure.

Central Assumptions and Parameters.

We used two general parameters for analysis throughout the paper: 1) an evaluation period of 20 y, including the construction period, and 2) a discount rate of 7%. The first was chosen based on a standard road project life cycle (29). The second corresponds to the long-term interest rate in Brazil, the country with the largest set of road projects in our sample. In Table 1, we describe additional assumptions and discuss any potential bias they might cause. Positive bias implies that the assumption may overestimate the value of the final efficiency indicator calculated as the ratio between the economic impact and the socioenvironmental damage (i.e., a positive bias means less socioenvironmental risk and/or more economic return), while negative bias implies the opposite. These assumptions are also included in the text.

Deforestation Prediction.

To predict deforestation around each project, we used Dinamica EGO (33), which simulates future land cover change based on a probabilistic model of past deforestation, as explained by biophysical and socioeconomic variables (34). The approach has four main steps: 1) calculate transition matrices, 2) calculate a transition probability map, 3) set up and simulate the model with and without the road project, and 4) validate the simulation.

We calculated transition matrices based on observed deforestation (i.e., loss of native vegetation) from 2011 to 2015, considering three land cover classes: deforested, forested, and nonforest (35). It is possible that road-induced deforestation is overestimated in our study because the data source includes deforestation not caused by humans, as well as conversions from primary forests to, for example, plantations, which are considered forests by some countries. In the absence of better data, however, Hansen et al. (36, 37) have been accepted and used as a proxy for deforestation in the literature.

Because the importance of each potential driver of change varies by location, we calculated individual transition matrices for the specific area in which each potential project would occur. In some particularly remote places, however, there are no existing roads from which transition probabilities could be estimated. In these cases, we derived the transition probabilities from nearby regions that already have roads. This approach is justified by the influence of broader patterns on the local scale (38).

For step 2, Dinamica EGO uses the weight of evidence approach to calculate the influence of a set of spatial features on the probability of each pixel transitioning to another class. We use distance to/presence of/value of the following features: roads, dams, river transport routes, mining, protected areas, indigenous lands, military areas, cities, and elevation. Since the weight of evidence approach only applies to categorical data, all continuous variables such as elevation and distance were transformed into categorical variables. More specifically, the continuous variables were transformed into bins where the number of intervals and their buffer size were selected to best preserve the original data structure. Once the bins were calculated, the effect of each variable on the spatial probability of a transition (i.e., from native vegetation to deforestation) was also calculated. Based on these effects (or the weights of evidence), we were able to produce a transition probability map (39).

Additionally, to use the weights of evidence approach, all variables must be spatially independent. We estimated the correlation between the independent variables using Pearson's correlation test to measure the pairwise correlation. Where the correlation coefficient was more than 0.7, we excluded the one that made the performance of the model worse (40, 41).

We used site-specific conditional transition probabilities to simulate future changes in land cover considering two scenarios: first, with no change to the existing road network, and second, adding the proposed roads or changing road surface type from unpaved to paved, as appropriate. The difference in deforestation between scenarios is the additional deforestation that would be caused if a given road project was implemented. This relationship can be considered causal as long as 1) deforestation does not drive the decision to implement roads (i.e., no inverse causality) and 2) all channels through which roads drive deforestation are controlled for in the model (i.e., no omitted variable bias).

To validate model predictions, we simulated deforestation in 2016 and compare the results to observed values from the same year. Omission errors, in which the model fails to correctly predict deforestation, averaged 2.82%. The average commission rate, in which the model predicts deforestation that did not occur, is 0.56%. These results suggest that our deforestation model is reasonable but that predictions are underestimated for some road projects. Accuracy of predictions as tested using a decay window size method calculated automatically in Dinamica EGO was also satisfactory (SI Appendix, Table S5).

Environmental Impact.

To estimate environmental damage, we overlaid projected additional deforestation caused by the proposed roads on map layers accounting for five elements of ecological importance: species diversity, ecosystem coverage, surface water, carbon storage, and national protected areas. We then calculated indices of impact for each, accounting for both scope and significance of damage.

Potential impact on species diversity was calculated using range maps of endangered amphibians, reptiles, birds, and mammals. For each cell that would be cleared due to the road project, we calculated the increased risk of extinction as the sum of each species multiplied by a weight assigned to its vulnerability status SI Appendix, Table S14 (b) according to the International Union for Conservation of Nature (IUCN) Red List (42), i.e.,∑jsj⋅vulnerabilityj,

where sj is species j density (measured as the species count for each cell) and vulnerabilityj is the status of species j in the IUCN Red List, with weights corresponding to extinction risk (SI Appendix, Table S14; ref. 43). For each road, the total risk is given by summing the species diversity risk calculated in each cell that would be cleared.

To calculate the biome impact, we used ecoregions as reference (44). For each road, we calculated the risk to each biome as the area of that biome that would be deforested multiplied by the inverse of its total coverage in the study area, i.e.,∑ecoregion1Areaecoregion⋅(Area of the ecoregion that would be cleared).

We then calculated the total biome-level damage score for by each road as the sum of the risk to each biome impacted.

We calculated the risk of interfering with hydrological processes using global surface water data (45). Risk was quantified as the sum of deforested areas with presence of surface water for at least 25% of the period of observation (1984 to 2015). This approach may underestimate risk in two ways. First, surface water obscured by standing vegetation cover can go undetected by remote sensing, such that heavily forested river segments may not be identified. Second, seasonally flooded areas will typically be excluded by the 25% presence threshold.

Impact on the global climate was represented by carbon emissions, calculated as the tons of carbon present in deforested cells (46). Finally, we calculated the area that would be cleared inside national protected areas, as a broader indicator of importance, independent of physical and biological features. Regional protected areas, whose management and management objectives vary dramatically within and between countries, were not included.

To combine types of environmental impact, we first normalized each of the five environmental risk variables as follows:normalized riski=(riski−mini)(maxi−mini),

where mini and maxi are the corresponding statistics of the variable i. This normalization guarantees that the values are in the range [0,1], following one of several standard practices in statistics for normalization where the goal is to compare measurements that have different units. We then calculate a combined environmental damage score for each road by averaging the scores for each of the five normalized variables, assigning equal weights to each.

SI Appendix, Table S6, shows the final environmental damage score for all road projects, as well as the individual risk scores for each of the five variables.

Social Impact.

To calculate social impact, we identified both positive and negative effects. We used spatial data to calculate three indicators of benefit related to improved access to health care and education: 1) the number of schools and health centers inside the 20-km buffer around each road, 2) the average distance between these and the proposed road, and 3) the total population of the municipalities through which the road would pass. The final nonmonetary social benefit measure is a linear combination of these indicators.

To calculate social costs, we used both spatial and survey data. We used the former to calculate the length of each proposed road that would pass inside the territory of indigenous people in voluntary isolation (i.e., indigenous people who do not maintain or have contact with other peoples). Specifically, we first identified which proposed roads would go through the territories of indigenous people in voluntary isolation by overlapping two relevant spatial layers. Second, we calculated the length of road (in km) that would pass through the indigenous territories. Third, using the same methodology as for the other indicators, we normalized the resulting lengths to combine this indicator with those for the additional social variables.

We also used data from a questionnaire given to experts in each country (n = 33) on two issues: 1) the degree of rejection by people directly affected by each road project and 2) violation of any legal norms. The former was defined according to five categories from “very low” to “very high” levels of rejection. To each category, we assigned a weight from 1 to 5 to capture the different levels of rejection: 1 for the lowest and 5 for the highest rejection. The answers were then combined linearly through a weighted average. The second question was posed as a simple yes or no choice. In this case, we assigned the value of 1 to roads with no legal violation and 5 otherwise. As before, the answers were linearly combined.

To create the unified social impact score, we follow the same normalization procedure used for environmental damage, allowing benefits to reduce the score and negative impacts to increase it. SI Appendix, Table S7, shows the final social impact scores for all road projects.

Economic Return.

There are different methods to calculate the economic return from road investments. Approaches following the trade literature (47) estimate the welfare impact of a reduction in transportation costs in terms of reduced prices and increased economic activity. Other approaches parameterize explanatory regression models using cross-sectional data on gross domestic product with road density as an explanatory variable (SI Appendix, Table S14 (b–c)). A common approach among development banks is to calculate the increase in consumer surplus resulting from reducing transportation costs, based on the elasticity of demand for transport and other factors. For example, the Fourth Highway Development Model (HDM-4) and Roads Economic Decision (RED) model, both developed by the World Bank, are broadly used for this purpose (SI Appendix, Table S14 (d); ref. 48).

In this study, we used the third approach. In particular, we applied the RED model, which estimates the NPV of building or improving rural roads with low traffic, as is the case for the road projects in this study (49). In the context of multiple theoretically valid approaches, this choice was made to facilitate use of the results by decision-makers, for whom the consumer surplus approach in general and HDM-4 and RED, in particular, are most familiar. The government of Brazil, for instance, directly uses these approaches.

The RED model evaluates one road at a time, comparing project implementation to a scenario in which the project is not implemented. For the road improvement projects, our without-project scenario was the road in its current state. For the new road projects, because current demand for transit could not readily be observed, we estimated transit in the without-project scenario based on a hypothetical road in the same location but in the worst possible condition.

Gross benefits are calculated based on reduced vehicle operating costs and travel time. Key inputs include the International Roughness Index (IRI) scores (based on road condition), which determines vehicle operating costs and travel speed, and AADT, which is used directly to calculate consumer surplus. To overcome the lack of data for the projects evaluated, we used econometric models based on existing observations to estimate road conditions and normal traffic in the with-project scenario. This estimation was done in two steps. First, we regressed road and local characteristics on observed AADT for all roads in the Amazon with such data. Second, we used the fitted model to predict traffic for all road projects in our analysis. Generated traffic (existing users driving more frequently or driving farther, as well as new users) was computed internally by the RED model based on a defined price elasticity of demand for transport. We used an elasticity of 1 for cars and buses and 0.6 for trucks (50).

Our model specification for AADT and IRI waslogY=β0+γX1+δX2+error,

where Y is road condition or AADT, X1 is a matrix containing variables related to road characteristics (length, surface type, and classification), and X2 is a matrix containing local characteristics (population density, land cover, elevation, and gross domestic product in 2017 US$). Descriptive statistics are given in SI Appendix, Table S8. We used ordinary least squares regression for the traffic model and ordered logistic regression for the roughness model to account for a categorical dependent variable (very poor, poor, fair, good, and very good) in the latter case.

SI Appendix, Table S9, shows results for the traffic model. SI Appendix, Tables S10 and S11, show results for the road condition model. None of the roads for which we had observed roughness data were classified as in “very good” condition; as a result, estimated conditions also exclude this category. We transformed road condition predictions into IRI scores using the standard values in the RED model (SI Appendix, Table S12).

For the purpose here, the models seek only to estimating missing traffic and predict road conditions. There is no need to understand the precise causal relationship between coefficients on the regressors and the dependent variables (51). We note that our sample consists of 90 observations which could potentially lead to imprecise estimates. However, observed traffic data in the Amazon region are extremely scarce. The models for AADT and IRI could be improved if more data on traffic were available.

Road costs include initial construction and ongoing maintenance. Official data on initial investment were available for 65% of the roads analyzed. For the remaining projects, we used the average per km investment from our sample in the relevant country. Annual maintenance costs were estimated using standards reported in the World Bank’s Roads Cost Knowledge System (ROCKS) database (52). These maintenance costs are the average among the 67 projects for which information was available in the ROCKS database (SI Appendix, Table S13).

The net economic benefit is calculated by RED asNet economic benefit=−initial investment+∑benefitt−maintenance costt(1+d)t,

where t is the year and d is the discount rate. We used a 20-y horizon and a discount rate equal to 6.87% per year, which is the long-term interest rate in Brazil, the country with the largest set of road projects in our sample.

Finally, we normalized the NPV following the same procedure used for the environmental and social damage scores, with the goal of obtaining a measure of economic benefits that is comparable to these other measures.

Efficiency.

For road projects with NPV > 0, we first calculated a socioenvironmental damage score as the linear combination of environmental and social scores with equal weight assigned to each. Second, we calculated the efficiency with which a given project delivers economic benefit by dividing NPV by the socioenvironmental damage:efficiency=Net economic benefitSocioenvironmental damage score.

SI Appendix, Figs. S4 and S5, rank the NPV > 0 road projects from most to least efficient, considering all countries together and then individually. In each panel, efficiency is scaled from 0 to 1. We do not rank the efficiency of projects with NPV < 0, as there is no rational basis for spending scarce public resources to generate not only socioenvironmental harm but also economic loss. We acknowledge that national and local governments may continue to prioritize and implement road projects that lead to apparent lose–lose situations. Understanding why these proposed roads remain on the agenda is an important area for study. In the meantime, we hope the framework and improved information presented here can support movement toward better infrastructure decisions for people and the environment.

Data Availability.

All data used in this study are publicly available from the sources mentioned in the text. The combination of these different datasets, used to create our database, is presented in SI Appendix, Tables S1–S3, S6 and S7, and S11–S14.