Decentralization can help reduce deforestation when user ... - PNAS

Decentralization can help reduce deforestation when user groups engage with local government Glenn D. Wrighta, Krister P. Anderssonb,1, Clark C. Gibsonc, and Tom P. Evansd a Department of Social Sciences, University of Alaska, Southeast, Juneau, AK 99801; bInstitute of Behavioral Science, University of Colorado at Boulder, Boulder, CO 80309; cDepartment of Political Science, University of California at San Diego, La Jolla, CA 92093; and dDepartment of Geography, Ostrom Workshop, Indiana University, Bloomington, IN 47405

Policy makers around the world tout decentralization as an effective tool in the governance of natural resources. Despite the popularity of these reforms, there is limited scientific evidence on the environmental effects of decentralization, especially in tropical biomes. This study presents evidence on the institutional conditions under which decentralization is likely to be successful in sustaining forests. We draw on common-pool resource theory to argue that the environmental impact of decentralization hinges on the ability of reforms to engage local forest users in the governance of forests. Using matching techniques, we analyze longitudinal field observations on both social and biophysical characteristics in a large number of local government territories in Bolivia (a country with a decentralized forestry policy) and Peru (a country with a much more centralized forestry policy). We find that territories with a decentralized forest governance structure have more stable forest cover, but only when local forest user groups actively engage with the local government officials. We provide evidence in support of a possible causal process behind these results: When user groups engage with the decentralized units, it creates a more enabling environment for effective local governance of forests, including more local government-led forest governance activities, fora for the resolution of forest-related conflicts, intermunicipal cooperation in the forestry sector, and stronger technical capabilities of the local government staff. Bolivia

| Peru | decentralization | forests | governance

F

orests are complex systems that defy simplistic, one-size-fits-all governance approaches. Like other common-pool resources (CPRs), forests are susceptible to overuse and degradation because it is costly to exclude potential users and their use can degrade or even deplete the resource. To make governance even more challenging, forests take far longer to develop and recover than the sitting terms of parliamentarians or presidents. For a century, governments implemented top-down, centralized forest policy, considering it to be the superior approach to ensure effective protection and use. However, many scholars and policy makers now perceive such an approach failed to sustain both forests and the livelihoods of the groups that depend on them (1, 2). Starting in the 1980’s, many national governments and international donors responded to this new view by aggressively pursuing policies to decentralize the governance of forests, transferring many rights and responsibilities associated with forest governance from the central to subnational governments (3–5). Currently only a handful of developing countries have not decentralized forest governance (6–11). The core argument behind the decentralization reforms, which international organizations have used widely, is that local authorities have better information about local forests and users, and thus can develop better policy solutions (5, 12–14). Several experts, however, have started to question the effectiveness of decentralized governance of collective goods, such as forests, suggesting such reforms may result in worse outcomes or, at best, outcomes no better than under central government control (12, 15, 16). Few robust studies exist that test this proposition: Extant work employs either qualitative case studies with a small number of observations or tends to focus on the village-level effects of www.pnas.org/cgi/doi/10.1073/pnas.1610650114

the devolution of property rights to local user groups rather than on the decentralization reforms that target general-purpose, local government units, even though they are the most common targets of the decentralization policies (6, 17–20). We thus lack persuasive evidence for the effectiveness of these reforms in the very place that was the main target of the reforms: within the jurisdictions of local, general-purpose governments. The lack of relevant and robust evidence is particularly serious for the ongoing policy efforts to curb tropical deforestation, such as the international initiative on Reducing Emissions from Deforestation and Forest Degradation. Without credible studies, policy makers can know neither the effectiveness of current policy instruments nor how to alter them to increase their effectiveness in the future (21). Here, we draw on CPR theory (22–24) to develop an argument about the institutional conditions under which decentralization is likely to lead to improved forest governance outcomes. Specifically, we derive our argument from the work of Elinor Ostrom, who proposed eight design principles for sustaining CPRs (22). The achievement of most of these principles hinges directly on the degree to which local users are recognized and allowed to participate in forest governance activities, such as rule making, monitoring, and enforcement (Measurement of Community Engagement). This logic provides the foundation for our main proposition: When local user groups engage actively with local government officials, this engagement improves the conditions for effective CPR governance and makes it possible for decentralization to sustain forests. To test this argument, we constructed an original database measuring decentralization policy, local governance attributes, and forest Significance Decentralization is one of the most important innovations in environmental policy during the past 30 years. Despite the pervasiveness and large amounts of resources invested to implement these reforms, little is known about their environmental effects. Given worldwide interest in forest conservation, this lack of knowledge hampers efforts to improve the effectiveness of current policy initiatives. Using quasi-experimental methods, we find that the environmental effects of decentralization reforms depend on how the reforms affect the conditions for user groups to govern their forests. Our findings show that decentralization to general-purpose governments may be most effective in places where forest users take advantage of opportunities to engage with local politicians about forestry issues. Author contributions: K.P.A., C.C.G., and T.P.E. designed research; G.D.W. and K.P.A. performed research; G.D.W., K.P.A., and T.P.E. analyzed data; and G.D.W., K.P.A., C.C.G., and T.P.E. wrote the paper. The authors declare no conflict of interest. This article is a PNAS Direct Submission. 1

To whom correspondence should be addressed. Email: [email protected].

This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10. 1073/pnas.1610650114/-/DCSupplemental.

PNAS Early Edition | 1 of 6

SUSTAINABILITY SCIENCE

Edited by B. L. Turner, Arizona State University, Tempe, AZ, and approved November 14, 2016 (received for review June 29, 2016)

Fig. 1. Forest cover differences for matched Peruvian and Bolivian samples. Under decentralization, rates of deforestation are significantly lower (less negative). These results are shown in table form in Table S5.

cover change in 200 municipal territories in Bolivia and Peru. While sharing a number of biophysical, socioeconomic, and cultural factors, Bolivia’s central government passed a forestry decentralization reform in 1996 that gave local governments (municipios) substantial rights, responsibilities, and resources to manage some of their forested areas (25, 26). Over the same time period, Peru kept most powers over forests under the purview of the central government (27, 28). We use matching techniques to compare outcomes in the local government territories in the decentralized setting with outcomes in similar territories in the centralized setting. With these matched observations, we then use regression techniques to evaluate the environmental impact of decentralization and the conditions under which such reforms can help stabilize forest cover. Results Our results show that the decentralization of forest governance to general-purpose governments is associated with lower overall rates of deforestation. This relationship disappears, however, in cases where such governments fail to build relationships with local groups who use the forest: Community engagement appears to be a necessary factor for the successful decentralization of forest governance. The plots in Fig. 1 show differences between forest cover in carefully matched decentralized and centralized territories (details are provided in Materials and Methods). In terms of rates of forest cover change, decentralized territories have significantly more stable forest cover (P < 0.05). The average treatment effect associated with decentralization is 2.6% less forest lost per year. We then analyzed the effects of community engagement on deforestation across decentralized and centralized municipalities to see if the effect differed between these two groups. To do so, we generated an interaction term, the product of “decentralization” and “user-group engagement,” and included the interaction term, as well as both base terms, in a generalized estimation equation (GEE) regression model with the same control variables used for the matching analysis above. [Where an interaction term is included in a regression model, the significance of coefficients in the table is not substantively meaningful; therefore, as suggested by methodologists (16), we interpret the results by examining the confidence intervals in the graph of the marginal effect of 2 of 6 | www.pnas.org/cgi/doi/10.1073/pnas.1610650114

decentralization on change in forest cover, given different levels of user engagement (Fig. 2).] Fig. 2 shows a graph of the marginal effects of a change from a centralized to decentralized regime, conditional on the level of user-group engagement with the local government. The results provide support for this study’s central hypothesis: Where community engagement is low (i.e., where forest user groups rarely meet with local government officials to express opinions regarding forestry), there is no significant effect of decentralization on forest cover change. With greater community engagement, however, decentralization has a positive effect on forest cover change, leading to significantly lower rates of deforestation (P < 0.05). Our results also indicate that the Peruvian government’s decision to exclude forest governance rights and responsibilities from the municipal government mandate may have backfired. The regression analysis (Table S1) finds that community engagement in Peru had a negative effect on forest cover change. We attribute this result to the fact that Peruvian municipalities have no official mandate to work on forestry issues, although they do have a mandate and some public resources to facilitate agricultural development (29). Citizen engagement under such circumstances may not contribute to more and better interventions to protect forests or to promote forestry (Background for Comparison of Forestry Policy in Bolivia and Peru). On the contrary, it may result in higher deforestation rates because agricultural land use often competes directly with forestry and forest conservation activities. Discussion What explains these results? Why is the environmental impact of decentralization contingent on user-group engagement? We propose that user-group engagement with the local government in a decentralized setting is necessary for creating an enabling policy environment for the governance of CPRs, such as forests. When the local policy environment is favorable for CPR governance, deforestation rates are lower. To test this idea further, we apply Ostrom’s thesis about CPR governance (22, 23) to the study of decentralization and examine empirically the extent to which Ostrom’s “design principles,” a set of institutional conditions that she argues help to sustain CPRs, are present in our sample of municipal territories. (We use our field observations from 2008 for this part of the analysis.) The main idea here is that the fulfillment of these design principles is more likely when user groups are

Fig. 2. Effects of decentralization, based on the GEE regression models with matched units. The difference between centralized and decentralized municipalities is not significant where engagement is weak, but the effect of decentralization is strong and significant where community engagement is stronger. Dashed lines represent 95% confidence intervals.

Wright et al.

Fig. 3. Percentage of municipalities that meet four of Ostrom’s design principles. In these statistical tests, we compare four groups of municipalities in our sample: 26 decentralized municipalities with high community engagement, 74 decentralized municipalities with low community engagement, 43 centralized municipalities with high community engagement, and 57 centralized municipalities with low community engagement. *P < 0.1; **P < 0.05; ***P < 0.01.

Wright et al.

programs in the four types of local administrations. We find that such programs are more likely to exist in decentralized units where users are activity engaged (P < 0.01). According to Ostrom (22), the success of CPR governance depends on matching institutional arrangements with the local context. We propose that in order for a local governance system to match solutions to the specific local circumstances, the system needs to have technically competent personnel in the local government administration. Consequently, the units responsible for the creation and enforcement of rules about forest use need to have some technical knowledge about forestry. As a proxy for this condition, we calculate the number of local government employees with formal training in forestry or agricultural sciences. The results in Fig. 3 show that the decentralized units with high user engagement have a higher proportion of employees with technical training (P < 0.05). Finally, Ostrom’s eighth design principle states that effective local governance of large-scale CPRs will benefit from a nested governance system, in which local user groups and their institutional arrangements are nested within governance units that operate at broader spatial scales. Our proxy indicator for this principle is the existence of formal agreements between local governments to cooperate on forest governance activities. Our comparison shows that such cooperation exists at a higher rate in the decentralized units with high user engagement (P < 0.1). The results of these comparisons of proxy indicators suggest that the decentralized territories where users are more actively engaged experience better conditions for effective local forest governance compared with territories where users are not as engaged. Taken together, these results suggest that a possible process through which decentralized systems can maintain more stable forests is by organizing their work in ways that make involvement with forest users both possible and meaningful. Given the inherent uncertainties associated in all comparative analyses, we consider the possibility of unobserved differences that may explain the variation in deforestation rates between the local territories of the two countries. As plausible alternative explanations for our results, we consider three differences between Bolivia and Peru: (i) central government policies, (ii) political history, and (iii) market opportunities. Detailed tests and discussions of these three possible explanations can be found in Alternative Explanations to the Observed Results. Our conclusion from our analysis of these alternative explanations, however, is that our comparative analysis provides a stronger explanation of the observed patterns. All three alternative explanations imply that there might be unobserved differences in government policies, political history, and/or market incentives that would make deforestation more likely in Peruvian territories regardless of any decentralized policy. According to this logic, one would expect to see one or more of these contextual factors generating decisions and actions in Peru that lead to high anthropogenic pressure on forests and a resultant increase in deforestation rates in the aggregate (or at least higher than in Bolivia). One would also expect that such differences could be identified by examining deforestation rates (before the time of the Bolivian reform) in each of Peru’s local territories in our study. Such expectations, however, are not consistent with empirical data at either the national or local level. First, data on aggregate deforestation rates show the opposite: Peru has experienced lower overall deforestation rates at the national level than Bolivia during this period (31). Second, because our comparative analysis between local territories in Bolivia and Peru controls for historical deforestation rates, along with a number of other proximate drivers of deforestation in each territory, we can be more confident that the results of our analysis are not driven primarily by these differences. Reverse causality may also threaten our explanation (i.e., that areas with better forest condition or lower deforestation rates might somehow be more likely to have been decentralized in PNAS Early Edition | 3 of 6

SUSTAINABILITY SCIENCE

more actively engaged with the local government officials in decentralized entities (a more complete analysis is provided in Table S2). Of the eight principles in Ostrom’s theory, we have field data on four: (i) existence of forums for conflict management, (ii) monitoring and enforcement activities by individuals who are accountable to users, (iii) the ability of matching solutions to local conditions, and (iv) the institutional nestedness of forest governance arrangements. We use our field observations to test whether these enabling conditions vary among the municipal territories in our sample. We compare indicators for these four institutional conditions between four groups of municipal territories in Bolivia and Peru: (i) decentralized territories with high degrees of user-group engagement, (ii) decentralized territories with low user-group engagement, (iii) centralized territories with high user-group engagement, and (iv) centralized territories with low user-group engagement. Our theoretical expectation is that favorable institutional conditions exist to a greater extent in decentralized territories where user groups actively engage with local government officials. The results in Fig. 3 support this idea, showing that decentralized territories with a high degree of user engagement report consistently higher scores on our four proxy measures of Ostrom’s design principles (22). [The reported results in Fig. 3 are based on Pearson χ2 test statistics and associated P values for cross-tabulations that examine the degree of association between decentralized units (yes/no) and user engagement (high/low)]. According to Ostrom’s design principles, systems that enjoy easy access to fora for conflict resolution are more likely to govern their shared resources sustainably (22). As an indicator for this condition, we calculate the proportion of local governments that report having intervened in conflicts in the forestry sector where such conflicts exist. As shown in Fig. 3, there is more frequent intervention in conflicts in decentralized territories in which local user groups are actively engaged with local governments than in the rest of the sample (P < 0.01). A second design principle states that successful local governance of CPRs is more likely when the individuals responsible for monitoring and enforcement are accountable to the users (22). One of the mandated responsibilities of democratically elected local governments in decentralized Bolivia is the monitoring and enforcement of rule compliance in the forestry sector, but the extent to which local governments perform these duties depends, in part, on how committed the local politicians are to forest governance (30). Here, we examine the existence of monitoring

Bolivia). Such reverse causality is not the case, however; every local government in the country received the same rights and responsibilities over the forests simultaneously (26). A subtler endogeneity concern is that user groups would be more likely to engage with local governments in areas with more abundant and stable forest resources, but such an explanation is not supported by theory. A core finding by researchers examining local environmental governance is that forest user groups are more likely to engage in a resource’s management when it is salient, scarce, and perceived to be threatened, not when it is abundant and in good condition (30, 32). All comparative analyses warrant caution when interpreting the results. Although it is impossible to control for all contextual differences between territories, our design uses the careful matching of similar areas and longitudinal data, which increase our confidence in the inferences we draw from the analyses. Conclusion Our results show that decentralization is not a panacea. Decentralization does not automatically lead to more stable forests because outcomes likely depend on how local politicians choose to interact with other members of the local governance system. Our findings suggest that the interactions between local forest users and local politicians are particularly important because this relationship can strengthen the incentives for politicians to take action in the forestry sector and can help to make such action more effective. When local politicians perceive political incentives to take policy action in the forestry sector to support and monitor local people’s interactions with the local forests, decentralization stands a better chance to succeed in stabilizing forest cover. Forest user engagement with local government officials is also important because it allows these parties to gather useful information about how local problems and issues may be addressed, and this information exchange has implications for downward accountability. Consistent with the findings from the literature on democratic decentralization (5, 9, 33) with more frequent communication, local politicians can gather information about community needs and preferences, making it feasible to respond to local needs and, in this way, strengthen the support of their constituents and their chances for reelection. Strong user-group engagement also allows community members (voting constituents) to gauge the performance of local politicians, making it possible for community members to reward effective politicians with reelection and to punish ineffective or corrupt leaders by voting against them (21, 34). Even when local government territories experience more stable forest cover, however, it does not necessarily mean that people’s livelihoods are improved or that some form of distributive justice is served. It is entirely possible that the local user groups that engage with the local government administration are in relatively privileged positions and push for a more active forest governance program to strengthen their own narrow, selfinterested objectives in the forestry sector. Such processes of elite capture, which several studies report to be a common byproduct of decentralization reforms (35, 36), cannot be ruled out on the basis of our results. In sum, our findings show that decentralized regimes can, under certain conditions, perform better than centralized regimes. Achieving such improvements involves making sure that forest users have ample opportunities to participate in the decentralized governance process. Previous studies show an important role for external organizations in supporting such participation (37, 38). For example, inclusive governance is more likely when central governments require local governments to conduct participatory planning and budgeting activities and mandate the establishment of local committees to oversee local government spending (28, 39), as well as when nongovernmental organizations support user 4 of 6 | www.pnas.org/cgi/doi/10.1073/pnas.1610650114

groups to strengthen their organizational capabilities (40, 41). Such interventions, in combination with forestry decentralization, can improve the governance of the world’s forests. Materials and Methods There are four major data sources for this study: (i) surveys of local governance actors (2001 and 2008), (ii) census/archive data (2000 and 2007), (iii) satellite images (1993, 2000, and 2008), and (iv) digital elevation models of Peru and Bolivia. In each of 200 selected municipalities, we interviewed the elected mayor in two waves: 2001/2002 and again in 2008. In addition, we interviewed municipal forestry officials and community leaders to triangulate responses. Survey enumerators completed a survey instrument (258 questions) with municipal officials, which was designed to elicit information regarding the interviewee’s policy priorities, staff, relationship with central and nongovernmental agencies, and relationship with citizens. Biophysical data were generated from two sources: (i) digital elevation models to characterize steepness of terrain, and (ii) data on forest cover that were generated using remote-sensing analysis (Landsat Thematic Mapper satellite imagery and aerial photography). We used digital elevation models (30-m Shuttle Radar Topography Mission) to generate measures of surface slope to identify the percentage of land in each municipality above a 12% grade, that is, the slope above which commercial, large-scale agricultural production is not feasible. We also performed remote-sensing analysis of satellite images acquired to estimate forest cover change for our sample of 100 local government territories in Bolivia and for 35 Peruvian municipalities in the period. The methods used to calculate the dependent variable, forest cover change, are described in Measurement of Forest Cover Change. Decentralization is perhaps our most important independent variable of interest. At the time our data were collected, Bolivia had experienced a country-wide process of decentralization of forest governance. In 1994, the Congress of Bolivia passed the Ley Participación Popular, the “Popular Participation Law,” essentially a package of decentralization reforms that granted substantial authority and 20% of national tax revenues to municipal governments. The enactment of this law was followed by the 1996 Ley Forestal 1700, which decentralized substantial control over forests to local governments. The Ley Forestal 1700 lengthened the tenure of leases to forestry firms for timber exploitation, made these leases renewable, and improved the security of tenure for the forest-dependent citizens by creating new jurisdictions for the communal management of local forest resources (26, 42). Most importantly, it granted municipalities the power to monitor forestry operations and enforce forestry rules and regulations related to forest clearings within their territory (42). Unlike Bolivia, forestry decentralization had not yet touched the forestry sector in Peru at the time of our last survey wave in 2008. Although the Peruvian national government began to devolve power to local government (both regional and municipal governments) in the early 2000s, forest governance remained in the hands of national government agencies (28, 43). Forestry decentralization did eventually affect Peru (decentralization reforms were implemented shortly after our second, and final, wave of surveys was gathered), but even when these reforms took place, forestry governance was not devolved to municipalities, instead being granted to regional governments (roughly equivalent to states in the United States or departments in Bolivia) (29). In practice, the absence of decentralization reforms in Peru does not mean that local governments were never engaged in forest governance activities. Peruvian municipal governments do have a well-institutionalized system for citizen input in municipal politics (including nationwide municipal participatory budgeting processes and extensive advisory roundtables, in addition to a thick network of civil society organizations), but these institutions are rarely involved in systematic forest governance activities. Since the early 2000s, the Peruvian national government did, however, begin handing over more responsibilities for public services related to agricultural land uses (43). We present two independent variables of interest: decentralization reforms and degree of community engagement on change in forest cover (deforestation) over time. Decentralization is a dummy variable that identifies whether the municipality was located in a formally decentralized regime at the time when the survey data were collected for that municipality. Because Peru did not experience a decentralization reform during this period, this variable is coded 0 (meaning centralized) for both survey waves in Peru (2001 and 2008). For Bolivia, we coded decentralization as 0 (centralized) during the first wave (2001) and as 1 (decentralized) during the second wave (2008) of data collection. Decentralization was coded in this way because we believe, supported by policy literature in Bolivia, that the 1996 forestry decentralization reforms experienced a significant policy lag: a significant amount of delay in the efforts to implement the new regime, integrating municipal governments functionally

Wright et al.

(before the implementation of the Bolivian reform). Despite carefully developed comparisons, using multivariate matching techniques, there is still a great deal of uncertainty associated with the inferences drawn from comparative analyses, especially when such comparisons cross national borders. (Spatial autocorrelation is not a major concern for this analysis because in our matched sample of local government territories, municipalities are generally located at great distances from one another.) Because of the uncertainties associated with all comparative work, it is important to consider alternative explanations of the observed patterns and how unobservable differences might have affected the results. Three such alternative explanations are discussed in Alternative Explanations to the Observed Results. The Mahalanobis matching method matches observations (in this case, several treatment cases for each control) according to the “Mahalanobis distance” between them (49). The Mahalanobis distance is the distance between observations in a multidimensional space, in which each dimension is a control variable (a variable upon which the matching is to be based). These control variables include annual rate of deforestation (lagged), the proportion of municipal area with a slope over 12% (the percentage above which most mechanized agriculture is impossible), road density [kilometers per square kilometer, natural logarithm (ln)], population (ln), municipal budget size ($US million, ln) and municipal area (hectares, ln). By using this technique, it is possible to generate a set of matched cases in which treatment and control cases are not significantly different on observables, except for the treatment. In essence then, the technique, like other matching techniques, generates a “treatment” group and a “control” group that are statistically not significantly different on important observable control variables (49, 50, 54, 55). We also use propensity scores to improve the balance of our matched samples, such that control (centralized) cases are more comparable to treatment (decentralized) cases, as suggested by statistical methodologists (50, 51, 56). We generate propensity scores using several of the control variables listed above. These propensity scores are then used as a matching variable in our Mahalanobis matching models, in addition to other control variables. After generating a matched sample based on control variables and propensity scores, we used two-sample t tests to confirm that our matched samples did not significantly differ in terms of the mean values of the centralized (control) and decentralized (treatment) variables. To generate apples-to-apples comparisons, we eliminated poorly matching observations from the sample. In the end, we were able to generate a strong sample of cases with no significant differences in terms of the control variables in our model. Multivariate matching techniques enjoy a number of advantages over regression techniques, the standard approach in the social sciences. First, statistical tests using matching do not assume a linear, additive effect. Second, because we use statistical tests to ensure a balanced sample, extreme values of control variables cannot drive spurious results (54, 55). At the same time, matching is not useful when examining the interactive effects of multiple independent variables on a single dependent variable. Therefore, we use regression techniques to test hypotheses involving interactions between community engagement and decentralization. In these models, we also control for the biophysical variables listed above. In postestimation tests, we examined regression models with both matched and unmatched samples and found that regression models produced different results, suggesting that this standard approach is problematic because it tends to compare incomparable cases. Our approach, using regression models after preprocessing data with matching models, addresses this problem (56). The regression technique we use here, GEE regression, is used to deal with potential temporal autocorrelation in panel data (57–60).

1. Chomitz KM (2007) At Loggerheads? Agricultural Expansion, Poverty Reduction, and Environment in the Tropical Forests (World Bank Publications, Washington, DC). 2. Larson AM, Petkova E (2011) An introduction to forest governance, people and REDD+ in Latin America: Obstacles and opportunities. Forests 2:86–111. 3. Andersson K, Gibson CC, Lehoucq F (2006) Municipal politics and forest governance: Comparative analysis of decentralization in Bolivia and Guatemala. World Dev 34:576–595. 4. Phelps J, Webb EL, Agrawal A (2010) Does REDD+ threaten to recentralize forest governance? Science 328:312–313. 5. Agrawal A, Ribot J (1999) Accountability in decentralization: A framework with South Asian and West African cases. J Dev Areas 33:473–502. 6. Faguet J-P (2014) Decentralization and governance. World Dev 53:2–13.

7. Adam YO, Eltayeb AM (2016) Forest policy and poverty alleviation: A review. Forest Policy and Economics 73:300–307. 8. Smyle J, Collins S, Biason C (2016) Rethinking Forest Regulations (Rights and Resources Institute, Washington, DC). 9. Ribot JC (2003) Democratic decentralisation of natural resources: Institutional choice and discretionary power transfers in Sub-Saharan Africa. Public Adm Dev 23:53–65. 10. Miteva DA, Pattanayak SK, Ferraro PJ (2012) Evaluation of biodiversity policy instruments: What works and what doesn’t? Oxford Review of Economic Policy 28: 69–92. 11. Samii C, Paler L, Chavis L, Kulkarni P, Lisiecki M (2014) Effects of decentralized forest management (DFM) on deforestation and poverty in low and middle income countries: A systematic review. Campbell Systematic Reviews 10(10).

Wright et al.

ACKNOWLEDGMENTS. We thank Nathan Cook, Sean Sweeney, Aseem Prakash, Patricio Valdivieso, Donato Rojas, Ashwin Ravikumar, Tara Grillos, and two anonymous reviewers for substantive comments made on earlier versions of the paper. Patty Zielinksi of the Ostrom Workshop at Indiana University provided valuable editing help. This work was supported by National Science Foundation Grants DEB-1114984 and SES-0648447.

PNAS Early Edition | 5 of 6

SUSTAINABILITY SCIENCE

into the new governance regime (44, 45). Our explanation is that there would be a policy lag that delayed the impact of the newly implemented governance regime on rates of deforestation. Over time, as the new decentralized governance regime took hold, we would expect a cleaner signal of a relationship between decentralization and deforestation to emerge. To ensure our results are robust to an alternative coding of decentralization, we ran all of the models presented here in which Bolivia is coded as 1 (postdecentralization) in both survey waves (2001 and 2008). This alternative coding represents a test of an outcome where there is a near-immediate impact of the new decentralization regime on a reduction in deforestation (shorter policy lag). Although this alternative coding changed the balance of our matching sample significantly, the direction and significance of our results did not vary when using this alternative coding: Decentralization still reduces forest cover loss significantly (P < 0.05). Community engagement is a variable that denotes the degree to which a local government is connected through frequent interactions about forestry with community-based organizations. This variable is drawn from one of our survey questions, which asks respondents how often community-based organizations expressed opinions regarding forestry to municipal government officials on a range from 1 (“never”) to 5 (“very frequently”), which is a variable that has been shown to predict the extent to which democratically elected local governments involve citizens in both policy decisions and implementation (38). We averaged the responses from surveys with mayors, local forestry officials, and community-based organization leaders in each municipality to generate an overall measure of the degree of community engagement on forestry issues within a municipality. Our empirical tests use two multivariate techniques: (i) Mahalanobis matching with propensity scores and (ii) GEE regression using Mahalanobis matching with propensity scores as a preprocessing technique to eliminate noncomparable observations. Descriptive statistics for all of our values are included in Table S3, and a code book for these variables is included in Table S4. For the datasets used in this paper, see Datasets S1 and S2. The cases of Bolivia and Peru provide an opportunity to use comparative analysis to study the effects of decentralization and community engagement on forest outcomes. Even though decentralization reforms in forest policy have been applied to municipalities in Bolivia and not in Peru, a simple comparison between Bolivian and Peruvian municipalities in terms of land cover change and other forestry-related outcomes (the so-called “difference in difference approach”) is not appropriate in this case because we are likely to confuse differences between Peru and Bolivia with the effects of decentralization (46–51). Instead, we draw on recent studies in program evaluation to create a quasi-experimental research design that enables an approximation to a counterfactual analysis (52, 53). Through this research design, we compare what happened after the reform in Bolivia with a counterfactual scenario of what is likely to have happened in the absence of the decentralization reform. Because such a scenario does not exist in Bolivia, because all local governments were given the same rights and responsibilities through the reform, we use Peruvian local government territories as surrogates for the unobservable counterfactual scenario in Bolivian territories. Using Peru as a surrogate counterfactual scenario for Bolivia constitutes a hard “test case” because at the national aggregate level, Peru experienced slightly lower deforestation rates than Bolivia during the 2000–2010 period (31). This finding means that the comparative analysis of local governance outcomes is biased against Bolivian territories, and that Bolivian local governments face an uphill battle to exhibit lower deforestation rates than their Peruvian counterparts. Following this design, we use multivariate matching techniques to ensure that the Peruvian territories that represent the surrogate for a Bolivian counterfactual scenario are as similar as possible to the Bolivian territories when it comes to several contextual variables, such as population densities, topography, road densities, forest cover, and historical deforestation rates

12. Treisman D (2007) The Architecture of Government: Rethinking Political Decentralization (Cambridge Univ Press, Cambridge, UK). 13. United Nations (1992) Report of the United Nations Conference on Environment and Development, Rio de Janeiro (United Nations, New York). 14. United Nations Food and Agriculture Organization (2001) The State of the World’s Forests 2001 (Food & Agriculture Organization of the United Nations, Rome). 15. Keefer P (2007) Clientelism, credibility, and the policy choices of young democracies. Am J Pol Sci 51:804–821. 16. Tacconi L (2007) Decentralization, forests and livelihoods: Theory and narrative. Glob Environ Change 17(3):338–348. 17. Porter-Bolland L, et al. (2012) Community managed forests and forest protected areas: An assessment of their conservation effectiveness across the tropics. For Ecol Manage 268:6–17. 18. Baland J-M, Bardhan P, Das S, Mookherjee D (2010) Forests to the people: Decentralization and forest degradation in the Indian Himalayas. World Dev 38:1642–1656. 19. Somanathan E, Prabhakar R, Singh Mehta B (2009) Decentralization for cost-effective conservation. Proc Natl Acad Sci USA 106:4143–4147. 20. Andrews M, Shah A (2005) Citizen-Centered Governance: A New Approach to Public Sector Reform. Public Expenditure Analysis (World Bank, Washington, DC). 21. Andersson K, Gibson CC (2006) Decentralized governance and environmental change. J Policy Anal Manage 26:99–123. 22. Ostrom E (1990) Governing the Commons (Cambridge Univ Press, New York). 23. Poteete AR, Janssen MA, Ostrom E (2010) Working Together (Princeton Univ Press, Princeton, NJ). 24. Dietz T, Ostrom E, Stern PC (2003) The struggle to govern the commons. Science 302: 1907–1912. 25. Andersson K, Ostrom E (2008) Analyzing decentralized resource regimes from a polycentric perspective. Policy Sci 41:71–93. 26. Pacheco P (2006) Descentralización Forestal en Bolivia: Implicaciones en El Gobierno De Los Recursos Forestales Y El Bienestar De Los Grupos Marginados (Plural Editores, La Paz, Bolivia, Spanish). 27. Jaramillo M, Alcázar L (2013) Does Participatory Budgeting Have an Effect on the Quality of Public Services? (Inter-American Development Bank, Washington, DC). 28. Jaramillo M, Wright GD (2015) Participatory democracy and effective policy: Is there a link? Evidence from rural Peru. World Dev 66:280–292. 29. Kowler LF, et al. (2016) Analyzing Multilevel Governance in Peru (Center for International Forestry Research, Bogor Barat, Indonesia). 30. Gibson CC, Dodds D, Turner P (2007) Explaining community-level forest outcomes: Salience, scarcity and rules in Eastern Guatemala. Conserv ation and Society 5(3): 361–381. 31. United Nations Food and Agriculture Organization (2011) Global Forest Resources Assessment 2010 (Food & Agriculture Organization of the United Nations, Rome). 32. Gibson CC, McKean MA, Ostrom E (2000) People and Forests: Communities, Institutions, and Governance (MIT Press, Cambridge, MA). 33. Agrawal A, Gupta K (2005) Decentralization and participation: The governance of common pool resources in Nepal’s Terai. World Dev 33:1101–1114. 34. Smulovitz C, Peruzzotti E (2000) Societal accountability in Latin America. Journal of Democracy 11:147–158. 35. Persha L, Andersson K (2014) Elite capture risk and mitigation in decentralized forest governance regimes. Glob Environ Change 24:265–276. 36. Nygren A (2005) Community-based forest management within the context of institutional decentralization in Honduras. World Dev 33:639–655. 37. Wampler B, Avritzer L (2005) The spread of participatory democracy in Brazil: From radical democracy to participatory good governance. Journal of Latin American Urban Studies 7:37–52. 38. Andersson K, van Laerhoven F (2007) From local strongman to facilitator: Institutional incentives for participatory municipal governance in Latin America. Comp Polit Stud 40:1085–1111. 39. Roberts N (2004) Public deliberation in an age of direct citizen participation. American Review of Public Administration 34:315–353. 40. Buckland J (1998) Social capital and sustainability of NGO intermediated development projects in Bangladesh. Community Dev J 33:236–248.

6 of 6 | www.pnas.org/cgi/doi/10.1073/pnas.1610650114

41. Barnes C, van Laerhoven F (2015) Making it last? Analysing the role of NGO interventions in the development of institutions for durable collective action in Indian community forestry. Environ Sci Policy 53:192–205. 42. Andersson K (2003) What motivates municipal governments? Uncovering the institutional incentives for municipal governance of forest resources in Bolivia. J Environ Dev 12:5–27. 43. Jaramillo M (2009) The pre-decentralization baseline case. Local Governments and Rural Development, eds Andersson K, Gordillo De Anda G, van Laerhoven F (Univ of Arizona Press, Tucson, AZ), pp 113–138. 44. Pacheco P (2005) Decentralization of forest management in Bolivia: Who benefits and why. The Politics of Decentralization: Forests, People, and Power, eds Pierce Colfer C, Capistrano D (Earthscan, London), pp 166–183. 45. Andersson K, Benavides J-P, León R (2014) Institutional diversity and local forest governance. Environ Sci Policy 36:61–72. 46. Fisher RA (1966) The Design of Experiments (Hafner, New York), 8th Ed. 47. Splawa-Neyman J (1990) On the application of probability theory to agricultural experiments. Essay on principles. Section 9. Stat Sci 5:465–472. 48. Rubin DB (1990) Comment: Neyman (1923) and causal inference in experiments and observational studies. Stat Sci 5:472–480. 49. Rubin DB (1980) Bias reduction using Mahalanobis-metric matching. Biometrics 36: 293–298. 50. Sekhon JS (2009) Opiates for the matches: Matching methods for causal inference. Annu Rev Polit Sci 12:487–508. 51. Smith HL (1997) Matching with multiple controls to estimate treatment effects in observational studies. Sociol Methodol 27:325–353. 52. McConnachie MM, et al. (2016) Using counterfactuals to evaluate the cost-effectiveness of controlling biological invasions. Ecol Appl 26(2):475–483. 53. Ferraro PJ (2009) Counterfactual thinking and impact evaluation in environmental policy. New Directions for Evaluation 2009(122):75–84. 54. Brady HE, McNulty JE (2011) Turning out to vote: The costs of finding and getting to the polling place. Am Polit Sci Rev 105:115–134. 55. Heinrich CJ, Lopez Y (2009) Does community participation produce dividends in social investment fund projects? World Dev 37(9):1–15. 56. Ho DE, Imai K, King G, Stuart EA (2007) Matching as nonparametric preprocessing for reducing model dependence in parametric causal inference. Polit Anal 15:199–236. 57. Frees EW (2004) Longitudinal and Panel Data: Analysis and Applications in the Social Sciences (Cambridge Univ Press, New York). 58. Rabe-Hesketh S, Skrondal A (2008) Multilevel and Longitudinal Modeling Using Stata (Stata Press, College Station, TX). 59. Wu CT, Gumpertz ML, Boos DD (2012) Comparison of GEE, MINQUE, ML, and REML estimating equations for normally distributed data. Am Stat 55:125–130. 60. Zeger SL, Liang KY (1986) Longitudinal data analysis for discrete and continuous outcomes. Biometrics 42(1):121–130. 61. Rojas D (2003) Tasa de Deforestación de Bolivia, 1993-2000 (BOLFOR, Superintendencia Forestal, La Paz, Bolivia), Spanish. 62. Hansen MC, et al. (2013) High-resolution global maps of 21st-century forest cover change. Science 342:850–853. 63. Andersson K (2013) Local governance of forests and the role of external organizations: Some ties matter more than others. World Dev 43:226–237. 64. Sekhon J (2007) Alternative Balance Metrics for Bias Reduction in Matching Methods for Causal Inference (Survey Research Center, Berkeley, CA). 65. Zeger SL, Liang K-Y, Albert PS (1988) Models for longitudinal data: A generalized estimating equation approach. Biometrics 44:1049–1060. 66. Duncan TE, Duncan SC, Hops H, Stoolmiller M (1995) An analysis of the relationship between parent and adolescent marijuana use via generalized estimating equation methodology. Multivariate Behav Res 30:317–339. 67. Horton NJ, Lipsitz SR (1999) Review of software to fit generalized estimating equation regression models. Am Stat 53:160–169. 68. Bedoya E, Klein L (1996) Forty years of political ecology in the Peruvian upper forest: The case of Upper Huallaga. Tropical Deforestation: The Human Dimension, eds Sponsel LE, Headland TN (Columbia Univ Press, New York), pp 165–196. 69. Kaimowitz D, Smith J (2001) Soybean Technology and the Loss of Natural Vegetation in Brazil and Bolivia (Agricultural Technologies and Tropical Deforestation, Wallingford, Oxon, UK). 70. De Sy V, et al. (2015) Land use patterns and related carbon losses following deforestation in South America. Environ Res Lett 10(12):1–15.

Wright et al.

Supporting Information Wright et al. 10.1073/pnas.1610650114 Here, we provide more in-depth discussion of several methodological aspects of the paper. The purpose of this discussion is to provide sufficient detail on the methodological choices made in the study so that other researchers are able to replicate the study should they wish to do so. The supplement consists of six sections: (1) background information on forestry policy in Bolivia and Peru, (2) measurement of forest cover change, (3) measurement of usergroup engagement, (4) methodological choices about matching and regression analysis, (5) alternative explanations to the observed results, and (6) description of data files available for replication. Background for Comparison of Forestry Policy in Bolivia and Peru Bolivia and Peru are neighbors and have high-mountain and lowland Amazon jungle geographies. The two countries both have large indigenous populations, including Quechua- and Aymaraspeaking highland indigenous peoples and multiple lowland indigenous groups. Both are middle-income countries with high levels of inequality and a great deal of rural poverty. Both countries share a history of Spanish colonial domination and unstable, authoritarian regimes in the postindependence period. More recently, both countries experienced similar waves of neoclassical or “neoliberal” economic liberalization. Additionally, both countries have recently emerged from periods of authoritarian rule and now host vibrant, contentious, although often corrupt, democratic national governments. The democratization processes strengthened the customary land rights of indigenous populations in both countries. It also introduced democratic election procedures for political leadership positions within local-level, multipurpose governments (municipios in Bolivia and distritos in Peru) (42). When it comes to forest governance, however, Bolivia and Peru differ significantly on their degrees of decentralization. This difference makes it possible to compare municipalities on the Peruvian side of the border with similar municipalities on the Bolivian side to examine the effects of decentralization. In addition, longitudinal data make it possible to assess the effects of the reform by comparing deforestation outcomes in Bolivian municipal territories before and after the decentralization reform. Both types of analyses, the comparison between Peru and Bolivia and the before and after comparison within Bolivia, produced consistent results: Decentralization only seems to work when user groups are actively seeking to engage the local government officials. Just because local governments in Peru lack an official mandate in the forestry sector does not mean that they are completely inactive in this sector. Local governments in Peru (distritos) represent a wellestablished public organization. Ever since 1963, Peruvians have held democratic elections to select their municipal government representatives. For most Peruvians, these units are the primary “go-to” places for voicing concerns or asking for help, regardless of the issue at hand. Hence, even in a centralized forestry regime, many Peruvians who are concerned with forestry issues still approach their local government officials. Although these local officials do not have a mandate to intervene in the forestry sector, and they do not receive any resources from the central government to do so, they do have an official mandate and some public resources to promote agricultural development in their territory. As a result, citizen engagement under such circumstances may not result in more and better interventions to protect forests or to promote forestry. In fact, such engagement may result in higher deforestation rates. Measurement of Forest Cover Change We worked with remote sensing analysts in Bolivia and Peru to calculate forest cover change for two periods: 1993–2000 and Wright et al. www.pnas.org/cgi/content/short/1610650114

2000–2008. We developed a processing protocol based on the methods used in the Government of Bolivia’s effort to monitor forest cover change (61). The same protocol was used to process the Landsat Thematic Mapper (TM) imagery for Peru, using images that were within 1 month of the dates used for Bolivia. The technical teams used ERDAS Imagine software to carry out supervised classification of the Landsat layers. These layers were classified into two land cover classes: forest and nonforest. The teams used spectral signatures to distinguish between forest and nonforest areas using archives of high-resolution aerial photography. In addition, they used auxiliary data from their respective government agencies responsible for deforestation monitoring, including high-resolution video footage taken from flights over large swaths of the areas of the two countries that had experienced a great deal of land cover change. A database of ground truth data was generated from the aerial photography and georeferenced video data to calibrate and validate the supervised classification. Forest cover change was detected by overlaying the classified polygons from 1993, 2000, and 2008; identifying areas that had changed from one land cover category to another; and calculating the spatial extent of both deforestation and forest regrowth. When it comes to the minimal mapping unit, when classifying the forest polygons, the teams filtered polygons smaller than four pixels (directly on the TM scenes with 30-m pixels). There were, however, exceptions to this rule. Because deforestation events tend to be events clearing large spatial areas, individual pixels identified as changing from forest to nonforest in the middle of remote and largely forested areas are likely to be artifacts and errors associated with the classification process. To remove these single-pixel anomalies, a low-pass filter was used. For a random sample of 1,000 data point units distributed throughout the mapped territory in our study, we found that the overall classification accuracy of each generated land cover dataset exceeded 89%. For each municipal territory, we calculated the compound annual rate (CAR) of forest cover change during the three periods of interest, using the following formula: 

Ending  Value CAR  of   forest  cover  change = Beginning  Value



1 #of   years


− 1.

Finally, to make sure our land cover data are consistent with estimates from existing secondary datasets on forest cover change, we (i) compare our estimates of CAR of forest cover change with those estimates from Hansen et al. (62) and (ii) conduct a sensitivity analysis using dependent variables calculated from data generated by Hansen et al. (62). Comparison of Forest Cover Change Data. To examine the degree of similarity between the two datasets, we make a pixel-by-pixel comparison of the forest maps for the year 2000. The data from Hansen et al. (62) define forests as woody vegetation of 5-m minimum height and with at least 10% canopy cover, whereas our own data employ a less restrictive definition: an area of at least 1 ha with woody vegetation of at least 2- to 5-m height in its mature state and with at least 10% canopy cover (61). Because the two datasets use fundamentally different definitions of forest, we expect the data from Hansen et al. (62) to generate smaller estimates of forest cover by forest. These smaller estimates are indeed what we find: In areas outside of tropical forest areas in the lowlands, the data from Hansen et al. (62) consistently classify as nonforest land areas that we classify as highland 1 of 5

submesophytic, xerophytic, or subhigrophytic forest. We find the average difference between the estimated forest cover for the municipalities (n = 48) with tropical forest areas to be 1.9%, but for municipalities higher up along the eastern slopes of the Andes, in areas with various types of highland forests, the average difference is much higher: 68% (n = 31). For example, in nine municipal territories outside of the tropical biome, the data from Hansen et al. (62) estimated the forest cover to be less than 2%, whereas our maps found the forest cover to be 60–80%. In addition to the difference in forest definitions, these discrepancies may be due to the more extensive use of high-resolution imagery, as well as overflights, to aid in the classification for the processing of our own data. Because our sample includes municipal territories that are in a variety of biomes, not just tropical, we prefer using our own data on forest cover change because these data provide more accurate estimates of actual forest cover change across all biomes and for a much larger number of municipal territories. Sensitivity Analysis. To examine the degree to which our results are sensitive to our choice of forest cover measures, we rerun all empirical tests with the dependent variables recalculated based on the data from Hansen et al. (62). Because the earliest data point in that dataset is the year 2000, we use slightly different cutoff dates for the two periods in the reanalysis (2000–2002 and 2002–2012, respectively). Even though using this alternative dataset greatly reduces the number of observations in our analysis (15% fewer municipal territories are covered by these data), the results are consistent: The average treatment effect of decentralization is 6.0% and statistically significant (P < 0.05). In the matched regression analysis, all GEE regression coefficients maintain the same signs compared with the analysis using our own forest cover data as the dependent variable, but some of these coefficients lose statistical significance (P > 0.05), including the interaction term. The likely reason for this loss of precision is the much smaller sample when the dependent variables are calculated from the Hansen et al. (62) data.

Measurement of Community Engagement We measure the degree of engagement between representatives from community-based organizations and the officials from the local government administration. The ordinal measure is described in Materials and Methods. Here, we discuss the theoretical importance of this variable and why it works as a proxy measure for a series of institutional conditions that are considered to support decentralized governance of CPRs. User-group engagement is a good proxy measure for a host of institutional conditions that the literature on decentralized resource governance describes as conducive to effective governance of CPRs (22, 23, 63). More precisely, to achieve the institutional conditions described by most, if not all, of Ostrom’s design principles (22) in the context of decentralized forest governance, active user-group engagement with local government decision makers is necessary. In the absence of repeated interactions between local users and local government authorities, who are the targets of the decentralization reforms, it seems improbable to observe most, if not all, of Ostrom’s eight principles (23). In Table S2, we analyze how the achievement of each of the eight principles depends, to various degrees, on active user-group engagement in the local government decision making. Methodological Choices About Matching and Regression Analysis Ideally, to test the effects of decentralization on forest-related outcomes such as deforestation, we would use a randomized, controlled experimental approach in which decentralization reforms would be applied to randomly selected jurisdictions such as municipalities, whereas other jurisdictions would not receive the decentralization “treatment.” If decentralization were applied randomly to muWright et al. www.pnas.org/cgi/content/short/1610650114

nicipalities in Bolivia and/or Peru, for example, it would be possible to examine the effects of decentralization by comparing the average changes in forest cover in decentralized municipalities with changes in forest cover in cases that have not been “treated“ with decentralization. Such an approach is not practical because decentralization reforms were applied uniformly to all municipal territories throughout Bolivia and not at all in Peru. Instead, we attempt to use multivariate matching techniques to approximate randomization. Specifically, we use a matched sample in which municipalities in a decentralized setting are matched with nondecentralized municipalities that share a large number of key characteristics. We use Mahalanobis matching in this study. This approach matches observations (in this case, several treatment cases for each control) according to the “Mahalanobis distance” between them. The Mahalanobis distance is the distance between observations in a multidimensional space, in which each dimension is a control variable (a variable upon which the matching is to be based). By using this technique, it is possible to generate a set of matched cases in which treatment and control cases are not significantly different on observables, except for the treatment. In essence, then, the technique, like other matching techniques, generates a treatment group and a “control” group that are statistically not significantly different on important observable control variables (48–50). Quantitative methodologists suggest that the use of a propensity score as a matching criterion is helpful in improving the balance of matched samples, such that control (centralized) cases are more comparable to treatment (decentralized) cases (48–50, 64). We generate propensity scores [effectively, the likelihood that a municipality with the observed characteristics of a given sample municipality will appear in the treatment (decentralized) group] using several key biophysical variables, including annual rate of deforestation (lagged), the proportion of municipal area with a slope over 12% (the percentage above which most mechanized agriculture is impossible), road density (kilometers per square kilometer, ln), population (ln), municipal budget size ($US million, ln) and municipal area (hectare, ln). These propensity scores are then used as a matching variable in our Mahalanobis matching models, in addition to other control variables. After generating a matched sample based on control variables and propensity scores, we used two-sample t tests to confirm that our matched samples did not significantly differ in terms of the mean values of the centralized (control) and decentralized (treatment) variables. Where statistically significant differences between samples occurred, these differences indicated that our matched samples were not good comparisons; that is, we were, to some extent, comparing apples and oranges. Therefore, to generate “apples-to-apples” comparisons, we eliminated poorly matching observations from the sample by eliminating matches with significantly differing propensity scores. In the end, we were able to generate a strong sample of cases with no significant differences in terms of the control variables in our model. In short, our matched sample appears to be a good apples-to-apples comparison. Matching enjoys certain advantages over conventional regression techniques, but it is not useful when examining the interactive effects of multiple inputs on a single outcome. To deal with this shortcoming, as noted above, we use regression models to test hypotheses involving interactions between community engagement and decentralization. Because the data are cross-sectional, time-series data, we use a population-averaged GEE timeseries approach. GEE models are extensions of generalized linear models like Poisson and logit regression, but which allow analysts to compensate for serial autocorrelation by assuming a “working” within-unit correlation matrix and adjusting errors accordingly (65–67). All of the models here were tested in regressions that assume a range of different within-unit correlation matrices, with no substantive differences in our results. 2 of 5

The presence of many poorly matched cases in our matching analysis would suggest that regression results include inappropriate comparisons of cases that are different: an “apples-to-oranges” comparison. In our tests, we examined regression models with both matched samples and unmatched samples, and found that these tests produced different results, suggesting that regression models are inaccurate in the case of decentralization because they compare incomparable cases. The approach of using regression models after preprocessing data with matching models addresses this problem and allows us to enjoy some of the advantages of regression-based techniques while also comparing similar cases through matching preprocessing (56). Besides the results we report here, we also conducted a number of robustness checks on our regression results. These checks included the following: (i) retesting our models after removing observations with high deviance residuals and high leverage cases; (ii) including a fuller set of controls, including a range of mayoral characteristics (e.g., gender, ethnicity, years in office), indigenous population, human development index (HDI) and HDI squared terms, total municipal size (square kilometers, nl), total forest size (square kilometers, nl), and total number of municipal employees; and (iii) including and excluding these variables in a series of sensitivity tests. Using these tests, we were never able to change the direction or significance of our independent variables of interest. The models shown here use “percentage of municipality covered in forest” as a control for the relative availability of forest resources. However, we also tested all of the models shown here with a measure of absolute forest size (forest cover in square kilometers, ln). Both control variables produce the same substantive results. Table S3 shows descriptive statistics for the variables used in the statistical models presented here, Table S5 shows our matching results, and Table S1 shows our GEE regression results. The substantive meaning of these matching and regression tables is explicated in the main text. Alternative Explanations to the Observed Results We explore three plausible alternative explanations to our main results: (i) differences in central government policies; (ii) differences in political history, and (iii) differences in market opportunities. Differences in Central Government Policies Beyond Decentralization.

It seems reasonable to assume that Peruvian agricultural policies and systems for property rights, or other national policies, may be such that the incentives for deforestation are stronger there than in Bolivia and that differences in these policies, not decentralization, explain our results. If this logic holds, then one would expect to see such a policy environment producing high anthropogenic pressure on forests and eventually more deforestation in the aggregate. One would also expect that such policy differences would manifest themselves through differences in the historical deforestation rate before the year 2000 in each of the local territories in our study. These expectations are not consistent with empirical data at either the national and local level. First, data on aggregate deforestation rates during this period show the exact opposite: Peru has lower deforestation rates than Bolivia (31). Second, because our comparative analysis between local territories in Bolivia and Peru controls for historical deforestation rates in each territory, we can be more confident that the results are not primarily driven by such policy differences. Differences in Political Histories. Although the two countries share several important political features at a general level, including

Wright et al. www.pnas.org/cgi/content/short/1610650114

similar structured public administrations of unitary governments, similar indicators of government corruption according to international transparency, and similar turnover rates of elected officials, differences do exist when it comes to historical political events in the two countries. One of these potentially transformative events is the insurgence of the Peruvian guerilla Sendero Luminoso in the decades before our study. This insurgence has been documented to have affected land use decisions in some parts of rural Peru (68). As an example, the Sendero Luminoso guerilla movement in Peru is likely to have stymied investments in large land use conversion in the Peruvian Amazon because rural out-migration increased and agricultural productivity slowed down as a result of the insurgence, which is believed to have lowered deforestation rates overall during the 1980s and 1990s. A possible consequence of Peru’s dark political history is that rural people in Peru are less likely to get engaged in local public affairs. If true, then one would expect that local citizens in centralized Peruvian territories engaged less with local governments overall than citizens in decentralized territories in Bolivia. However, our data suggest the opposite; that citizens in Peru (both in the 2002 and 2008 survey waves) report higher overall frequency of communication with local government officials than their counterparts in Bolivia, and this difference is statistically significant (P < 0.05). The effect of this communication is what is different in Peru and Bolivia, which means that it is decentralization in combination with frequent communication about forestry that will help reduce deforestation, not communication or decentralization by itself. Market Incentives May Be Different in the Two Countries. It is possible that local people in Peru perceive stronger incentives to convert forests to other land uses. It is also at least theoretically possible that Peruvians have more advanced capabilities for removing forests (although there is no empirical evidence to suggest that this theory is true). For the sake of argument, it may be that the combination of these market opportunities and superior capabilities in Peru skew the analysis so that it appears that decentralized units in Bolivia, where users happen to be more engaged, have lower deforestation, but that this pattern would primarily be a result of a skewed comparison. This alternative explanation is not supported by the data, however, because aggregate deforestation data show that Peru has a lower deforestation rate than Bolivia, at least since 1990. Moreover, the Bolivian government has been more aggressive than their Peruvian counterpart when it comes to pursuing export-oriented policies in the agricultural sector in the 1990–2005 period. During this time, the Bolivian government introduced the removal of price controls, a currency devaluation, fiscal incentives for exporters, low taxes, road construction, new trade deals to promote export of agricultural crops, and government land grants (69). This difference is supported by the aggregate deforestation data (31, 70), and suggests that any sample of local territories in Peru is likely to experience lower deforestation rates than Bolivian territories. In that sense, local territories in Peru constitute a tough baseline against which decentralized territories in Bolivia are being compared.

Description of Data Files for Replication We make all of the data used in the analysis available. To facilitate replication, Table S4 serves as a codebook for the data file. The “do-file” includes all of the commands used in the analyses in the study and may be run on Stata 12 or higher [requires previous download of customized commands (e.g., “psmatch2,” “outreg”)].

3 of 5

Table S1. GEE regression results Variable

GEE regression

Decentralization Community engagement Community engagement * decentralization Slope Road density (ln) Population (ln) Municipal size, ha (ln) Budget $US million (ln) Forest cover, lagged (ln) Constant Observations No. of municipalities

0.012 (0.373) −0.011 (0.000)*** 0.003 (0.521) 0.000 (0.817) −0.003 (0.241) 0.004 (0.063)* 0.005 (0.032)** 0.010 (0.000)*** 0.000 (0.984) −0.099 (0.001)*** 101 79

Effect of community engagement with local officials across centralized and decentralized regimes. Note that methodologists suggest that regression tables like these tables are not helpful in the case of interaction models (16, 47, 48). P values are shown in parentheses. *Significant at P < 0.1; **significant at P < 0.05; ***significant at P < 0.01.

Table S2. Role of user-group engagement with local government officials for the achievement of Ostrom’s eight design principles Design principle (i) Clearly defined physical and social boundaries

(ii) Rules regarding the appropriation and provision of common resources that are adapted to local conditions (iii) Collective-choice arrangements that allow most resource appropriators to participate in the decision-making process

(iv) Effective monitoring by monitors who are part of or accountable to the appropriators (v) Scale of graduated sanctions for resource appropriators who violate community rules (vi) Mechanisms of conflict resolution that are cheap and of easy access (vii) Self-determination of the community recognized by higher-level authorities

(viii) In the case of larger CPRs, organization in the form of multiple layers of nested enterprises, with small local CPRs at the base level

How user-group engagement matters Without engagement from users, it is impossible for local government decision makers to know the boundaries of local de facto property rights Resource users know local conditions well and will often have a good appreciation of which types of rules make sense in a particular context The participation of resource appropriators in governance decisions cannot happen without active user engagement, and the actual creation of such policy is more likely if users pressure the local government to allow for user engagement (8) Effective rule monitoring relies on active engagement by users, both in monitoring activities as well as in holding monitors accountable (17) Unless users and local government officials interact regularly, community rules will not be known to the government authorities, and will not be enforced User-group engagement will help communicate to governmental authorities what constitutes cheap and easy access for them Self-determination could plausibly happen in the absence of user engagement, but when such rights are not recognized by the highest levels of authority, active community engagement with local government officials can increase the likelihood of achieving such recognition (8) Nested governance requires both communication and coordination among multiple governance actors, especially between local user groups and the local governments

Authors’ elaboration based on Ostrom’s Design Principles (22).

Wright et al. www.pnas.org/cgi/content/short/1610650114

4 of 5

Table S3. Descriptive statistics Variable Decentralization Community engagement Community engagement * decentralization Central government engagement Annual forest change, % Forest cover, % (lagged) Slope, % (>12%) Road density (ln) Municipal size, ha (ln) Population (ln) Budget, $US million (ln)

Observations

Mean

SD

Minimum

Maximum

472 422 422 375 256 268 465 249 465 405 423

0.21 2.99 0.65 2.47 −0.02 17.83 46.32 −2.47 11.08 8.27 −0.93

0.41 1.18 1.29 1.51 0.03 27.57 31.8 1.7 1.62 1.48 1.67

0 1 0 1 −0.14 0 0 −9.21 7.18 3.67 −8.19

1 5 5 5 0.13 100 94.8 0.47 15.79 12.63 5.22

Table S4. Code book for data files Variable name

Description

country_geoid Wave ann_pct_chg corg_opin_for

Unique identifier 1 = 2002 survey, 2 = 2008 survey CAR of forest cover change for each municipal territory Frequency of in-person meetings concerning forestry issues between community-based organizations and the municipal officials, on a Likert scale: 1 (never) to 5 (very frequent) 0 = Peru in 2002 and 2008, Bolivia in 2002 1 = Bolivia in 2008 Forest cover (%) for the municipal territory in the previous period Length of road in the municipal territory (km) per square kilometers of municipal territory (logged) Percentage of municipal area that has a slope greater than 12% Total land area of the municipal territory (ha, logged) Total population living in the municipal territory (logged) Total local government budget in $US million (logged)

decentralization for_cover_lag1 road_density_logged pa_sl12 munic_ha_ln pop_ln ln_budget_mill

Each variable used in the analysis in this paper, included in the data files, is described in this table.

Table S5. Effects of decentralization, matching results Dependent variable

Treatment cases

Control cases

Total cases

ATE

ATU

ATT

43

59

102

0.03

0.03

0.02 (2.47)*

Annual forest change, %

T-statistic is shown in parentheses. ATE, average treatment effect; ATT, average treatment on the treated; ATU, average treatment on the untreated. *Significant at the P < 0.05 level.

Other Supporting Information Files Dataset S1 (XLS) Dataset S2 (TXT)

Wright et al. www.pnas.org/cgi/content/short/1610650114

5 of 5