Land cover change effects from community forest management in Michoacán, Mexico

More than half of Mexico’s forests and about a third of the forests of the world are communally owned. Despite this, community forest management (CFM) is the least studied forest management policy, and existing studies have focused on the effects of CFM on deforestation. In this paper, we evaluate the effect of CFM on land cover more broadly, to understand how CFM affects a community’s land use decisions. Mexico’s forestry administration mandates that to legally sell timber, communities must adopt forest management plans designed by a certified forester. In this study, we use differential access to foresters to identify the effect of community management on land cover and deforestation. We find that over time communities that adopt management plans see relatively more primary forest, a limited expansion of the agricultural frontier, and a decrease in deforestation. The decrease in deforestation is economically significant since the economic benefits from the avoided CO2 emissions alone could far outweigh the costs of adopting the management plans.


Introduction
Forests act as a global reservoir of carbon, provide habitat to thousands of species, and are the source of livelihood for millions of the poorest people in the world. As such, their continuing loss has negative socioeconomic and environmental effects that extend beyond the boundaries of any country. Deforestation contributes to 11% of global greenhouse gas emissions, having the same effect on emissions as the global industrial sector (IPCC 2014). It also has a negative effect on habitat and biodiversity (Sánchez-Cordero et al 2009) and erodes the livelihood of poor communities in developing countries, which are more likely to be located near dense forest areas (Sunderlin et al 2008).
Most forests are common pool resources, in that they are non-excludable (users cannot be barred from consuming the resource) and rival (consumption by one person affects the ability of others to consume or the benefit others gain from consumption). To address their potential over-exploitation, forest users have developed systems of governance to regulate their use (Ostrom 1990, Ostrom andNagendra 2006). How these governance structures affect deforestation and the conversion of forestlands to other uses is not well understood. While conservation policies such as protected areas (PAs) or payments for environmental services (PES) have received substantial evaluation, less work empirically evaluates the effect of community forest management (CFM) on forest outcomes.
Common-pool resource governance structures are particularly pertinent for communally-owned forests, where the community can develop their own use rules (Hayes and Persha 2010). Communally owned forests, such as those legally owned by Indigenous people and communities, represent 28% of the world's forestlands (Gilmour 2016). In Mexico, over half (53.4%) of the total land in the country is under a communal-based ejido system (Morett-Sánchez et al 2017) where ejidos and indigenous communities 5 hold about 60% of the country's forestlands (Bray et al 2003, Torres-Rojo et al 2022. This makes Mexico the second largest holder of communal forests in the world, after Papua New Guinea (Barnes 2009). Additionally, these forests are threatened by overuse. According to Hansen et al's (2013) updated data, the average forest loss per year in Mexico between 2001 and 2012 (our sample period) was 186 500 ha yr −1 , making Mexico the country with the 15th highest level of deforestation in the world. The combination of significant communally owned forestlands and high deforestation make Mexico an important case study in exploring factors that promote conservation of common pool forest resources.
To decrease deforestation, stop the conversion of forestlands into other uses, and improve community management, the Mexican government introduced two key changes to the regulatory framework governing forestry in Mexico. In 1986, the government ended the practice of leasing land to private firms for timber harvest and allowed communities to harvest their own forests (Merino et al 1997). Then, in 1992 the government mandated that to legally harvest timber, communities need an integrated forest management plan drawn up by a certified forester. These management plans focus both on the commercial production of the forest (where and how much to harvest every year and what trees to fell) as well as on the ecological services that the forest provides (Antinori 2000). These regulatory changes allow us to evaluate how the participation of communities in legal timber markets, enabled through CFM plans, affect the land cover within the communities. This is particularly relevant in a context where illegal timber markets, whose estimated volumes of timber can exceed those of the legal market (Torres-Rojo 2021), create great pressure on the existing forests.
It is unclear whether forests are best conserved by allowing sustainable harvest or by 'sparing' the forest and making harvest illegal (Mohren 2019, Runting et al 2019, Blackman and Villalobos 2021. Participation of communities in legal timber markets, enabled by CFM, has theoretically ambiguous effects on land cover and deforestation. Communities may make an active decision to harvest their forest, through the adoption of CFM. This increases the value of the forest for that community and can lead to an increase in the timber harvested (potentially beyond a sustainable level), which in time would decrease the size of the forest and lead to the conversion of forestlands into agriculture and/or pasture. On the other hand, 5 Both 'ejidos' and 'comunidades indigenas' (indigenous communities) are a form of communal based land tenure system. Therefore, throughout the paper we will refer to these communities either by calling them ejidos or communities, and the term would refer to both land tenure systems, unless it is specified otherwise. the increase in forest value and decreased cost of collective action afforded by the management plan may lead the community to protect the forest from both internal and external threats and limit the harvest to sustainable levels (Alix-Garcia 2007). This would decrease deforestation in communities with CFM, increase the size of their forests, and in the cases where conversion of forestland happens, it would be limited to marginal (less valuable) forestlands. Alternatively, under the sparing scenario, the status quo of no active legal harvesting by the community is maintained and given the common pool resource nature of the forest, illegal logging from both within and outside the community can erode the forest and ultimately lead to the conversion of the forestlands into other uses.
The purpose of this paper is to empirically estimate the effect of a forest management plan on land cover and deforestation for communities in the state of Michoacán between 2001 and 2012. The importance of understanding the effect of CFM on land cover and not only on deforestation is that different land covers provide different ecosystem services, with lands with less human intervention generally having higher biodiversity and providing more ecosystem services (Haines-Young 2009, Martínez et al 2009. To obtain a causal estimate of these effects, we use exogenous variation in the supply of foresters that design management plans. By using the variation in access to foresters to predict adoption, we avoid the problem that communities that adopt management plans may be systematically different from those communities that do not. We find that communities that adopt CFM plans see a reduction in the total land dedicated to perennial crops and a preservation of land in primary forest cover. These communities also see a decrease in deforestation, with the effect increasing over time.
Our approach contributes to the existing literature in three ways. First, our focus on land cover change expands the scope of our study beyond the traditional focus on deforestation (Hajjar et al 2016) and to the best of our knowledge, is the first study to estimate causal effects of CFM on land cover change (Gautam et al 2002, Ellis and Porter-Bolland 2008, Niraula et al 2013, Bowler et al 2012. To do this, we use newly available remote sensing land cover data specifically developed for Michoacan by Mas et al (2017), which allows us to evaluate the effects of CFM at a broader scale than previous case studies (Hajjar and Oldekop 2018). Moreover, existing studies of CFM in Mexico (Alix-Garcia et al 2005, Blackman and Villalobos 2021, Ordonez 2021, as well as studies of other contexts, like Madagascar (Rasolofoson et al 2015) and Nepal (Oldekop et al 2019) focus primarily on deforestation and only in some cases on forest conditions or resulting agricultural uses. We believe this responds to recent calls to expand the scope of the quasi-experimental literature (Miteva et al 2012. Second, our study offers new evidence of the effects of an understudied forest management instrument through the use of quasi-experimental methods (Baylis et al 2016, Busch andFerretti-Gallon 2017). While globally CFM is one of the most common forest management practices, it is also one of the least studied of the most used forest management instruments, unlike PES schemes and PAs which have received much more attention by the literature. Several studies have evaluated different PES programs, including Mexico's, which was established in 2003 and gives annual payments to landowners in exchange for the preservation of their forestland. Alix-Garcia et al (2012) find that the program had moderate impacts in reducing deforestation. Costedoat et al (2015) find significant effects of Mexico's biodiversity PES on deforestation. Similarly, the effects of establishing PAs have been widely studied (Shah et al 2021). In Mexico specifically, Blackman et al (2015) find no significant difference between deforestation inside versus outside established PAs. In a study that integrates the available evidence from conservation policies in Mexico, Sims and Alix-Garcia (2017) find that both PAs and PES reduced expected forest loss and that PES had a small but significant effect on poverty reduction.
The current evidence base for CFM in Mexico, however, is scarce. The existing literature is heavily biased toward cases in South Asia, although there are almost eight times more communally owned forestlands in Latin America than in South Asia (Hajjar et al 2016). A recent study by Blackman and Villalobos (2021) is the first nationwide study of this policy in Mexico. They focus on the same harvesting permits we study here and on the main 16 states where forestry is prevalent, using propensity score matching and a difference-in-difference design to estimate the causal effect on deforestation, and find no overall effect. Previously, Alix-Garcia et al (2005) focused on the communities' decision to participate in CFM to estimate the effect of CFM on deforestation, finding that communities who decide to harvest their forest commercially have a higher rate of deforestation than communities that choose not to. However, their analysis is constrained by the use of cross-sectional data on a limited number of communities across Mexico. In this regard, our contribution to this literature is not only our focus on land cover change in addition to deforestation, but also our focus on the dynamic effects of CFM adoption on these outcomes.
Finally, we use two methodological innovations. The first is a novel identification strategy that addresses the selection bias into CFM and uses machine learning to predict the adoption of CFM. We justify our identification strategy from interview data collected from the foresters themselves. Additionally, our data (including data on foresters' market) allows us to perform a back of the envelope calculation of the cost-effectiveness of this policy, which is both a contribution to the existing literature as well as a useful tool for policy makers.

Data
We combine community-level forest management data with forest cover, weather, demographic and economic data to construct a 12 year panel dataset for all communities in the state of Michoacán. Michoacán provides an interesting case study, since it is home to multiple forest ecosystems, and faces substantial deforestation threats (Mas et al 2017). For those communities with CFM, we have detailed data on the management plans, which forester designed the plan, its duration, the area covered, and the amount of wood to be harvested (SEMARNAT-Department for the Environment and Natural Resources). SEMARNAT issues two kinds of permits, a long term (typically ten year) harvesting permit, that usually covers the whole forest, and a short-term permit that is issued in case of emergencies (such as fires and pest outbreaks). We focus on harvesting permits, which represent 99% of all the permits in our sample.
Information on land cover (including primary and secondary forests) come from detailed data for the years of 2004, 2007, from (Mas et al 2017. The land cover data are constructed from SPOT images, with a 10 m resolution, taken during the dry season of each year, and then classified using a visual classification method, together with a semiautomated method. The final maps are at a scale of 1:50 000. Additionally, we use data on deforestation from Hansen et al (2013), where they use Landsat 7 Enhanced Thematic Mapper Plus (ETM+) images to construct a worldwide dataset of forest cover and loss between 2001 and 2012. We chose this data because it provides information on annual tree loss, allowing us to estimate the effects of CFM on a year-to-year basis.
We also include yearly averages of temperature and rainfall, which we obtained by matching the communities to the nearest of 38 weather stations throughout Michoacán (the ones with data for these years). We complement these with satellite weather data, with the monthly rainfall data from the climate hazards group infrared precipitation with stations dataset (Funk et al 2015), and the monthly mean of the maximum temperature from (Funk et al 2019). The variables and data sources are in table 3 in the appendix.
For the demographic characteristics of the communities, we use data from the censuses conducted by INEGI (the Mexican statistical office) for 2000, 2005 and 2010. We define the boundaries and the area for each community, using data from the National Agrarian Registry (RAN, in Spanish). We also use price data on major forest and agricultural products. For wood, we have the average price for pine and oak, and the producer price index for sawmills and other wood products. For the pine and oak prices, we have data from the state of Chihuahua (33% of the national supply), Durango (20%), Michoacán (12.5%), and the average national price. These prices come from the annual outlook report by CONAFOR, the national committee in charge of forests in Mexico. For agricultural products, we have price data for the main crops in Michoacán.
Importantly, our sample of 1759 communities provides us with enough statistical power, so that at the 80% conventional power level, we can detect effect sizes of 0.012 SDs, which is an effect size that is much smaller than the 0.1 SDs from other studies on community monitoring of common pool resources (Ferraro and Agrawal 2021).

Empirical strategy
To estimate the effects of CFM on deforestation, we must control for the fact that communities chose to adopt CFM, and therefore communities that adopt a forest management plan are likely different than those that never adopt. While we control for time-invariant community characteristics with fixed effects, we might be concerned that there might be a time-varying unobservable that induces a community to adopt a CFM in a particular year that might also affect forest outcomes, which would bias our results. For example, we might be concerned that those communities who have different objectives for their forest might have different levels of demand for CFM, and have different land cover outcomes. To address this concern, we would like to have an external factor that affects the probability of a community adopting CFM without directly affecting their forest use. Thus, we focus on factors that affect the potential supply of forest management plans. Specifically, we focus on the foresters themselves, since they are responsible for designing the plans, and paying for the design of a plan represents a significant financial investment for the communities. From our data, we know that only a handful of foresters are responsible for designing most of the forest management plans in the state of Michoacán.
We find that the market is highly concentrated in the top performing foresters, and these foresters are scattered throughout the state, with clusters of communities dominated by only a few foresters (see figure 3 in the appendix). Only 16 foresters are responsible for 63% of all the management plans adopted between 1993 and 2013 (table 1). We define a top forester as a forester who has drawn at least eight of the management plans in our sample (or 2% of all the plans in our sample), which would place them in the top 20% of foresters. We believe that having access to a top performing forester lowers the cost of adoption for a community given their higher productivity, and so a community that is exposed to at least one of these foresters, is more likely to adopt a management plan.
We use the exposure to these different foresters as a source of exogenous variation in the probability of adoption of CFM. We construct spatial lags associated with each forester to capture the exposure of communities to these foresters and then use the lags as instruments for the adoption decision of each community (see the appendix for details of how we construct these variables). The idea is that if community A hires a top performing forester, it will be easier and/or cheaper for the neighboring community B to adopt a management plan in the following years. We are aware that the definition of who the top foresters are is an arbitrary one, and so we test the sensitivity of our results to different definitions of who we consider to be a top forester. These results are in section 'Sensitivity of results to changes in the definition of top foresters' in the appendix.
We estimate the effect that the predicted probability of an active CFM has on forest loss and land use, through a standard two stage least squares estimation, where we select the instruments and the controls, using the post-lasso double selection method (Belloni et al 2014(Belloni et al , 2016; see the appendix for more details). We control for both community and year fixed effects to strip out time-invariant unobservables and annual shocks that are common to communities across the state. Finally, note that the estimated effect of CFM on deforestation is the local average treatment effect (LATE). This measure captures the effect that the adoption of CFM has on land cover and deforestation for those communities who decide to adopt a CFM after being exposed to a highly productive forester.
The validity of the spatial lags for the different foresters as instruments for the adoption of CFM relies on two key assumptions. First, having a community served by a productive forester must have a significant effect on the probability of a neighboring community adopting CFM. We believe that this effect comes from two channels. The first channel is that these productive foresters are more skilled at 'selling' these management plans to other communities, with many of them mentioning in the interviews we did, that the promotion of their services is something they actively worked on. The second channel is related to the referrals that community leaders give to one another. A forester we interviewed said: 'Being a good forester is like being a good doctor, if you treat your patients well and heal them, they will come back to you whenever they have a need and they will recommend you to other patients' .
The second assumption is that the instruments have no effect on land cover and forest loss, other than through the effect they have on the probability of a community having an active CFM plan. In our setting, because we control for community-level fixed effects, we are specifically concerned whether a change in availability of foresters might be associated with a change in land cover. One might legitimately be concerned about whether this second assumption holds.
To understand the factors underlying the spatial distribution of foresters, we conducted in-depth interviews of seven foresters, one former member of the forestry police, and six community leaders. From those interviews, we found three common themes explaining why foresters end up working where they do. The first one is that foresters tend to stay (or return) to the places where they first started their careers, suggesting very little change in location throughout their career. A fundamental part of the relationship between foresters and communities is based on mutual trust, and this trust is built through time. A young forester will usually start working for a certified forester and will build their reputation with the local communities. This social capital will help them get their own clients once they become certified foresters, since referrals from other community leaders are an important factor in the forester hiring process. Second, foresters also tend to be in areas where they have close personal ties (such as their family or the family of their spouse). This factor will also affect where they will start their professional careers, and as such, where they will end up working later in their careers. Finally, we found that transportation costs represent a significant hurdle and so foresters will usually work with communities that are relatively close to where they live (within 2 h driving distance, on average; see table 4 in the appendix). Thus, we believe that the location of these foresters is exogenous to the deforestation and land cover change in any given year, in any given region.
Furthermore, for a given community, who their neighbors hired as a forester in previous years, is an exogenous decision, and it is reasonable to assume that it is uncorrelated with the changes in land cover (and forest loss) experienced by that community. Similarly, we believe that a given forester from a neighboring community would have no effect on another community's future changes in land cover. This implies that there is no leakage from the adopting community to its neighboring communities, other than through the adoption of CFM by neighboring communities. We cannot directly test for this type of violation of the exclusion restriction, but we can control for the effect of leakage by including spatial lags for forest loss.
It is possible that the most productive foresters would tend to work in regions where forestry is more prevalent, such that we would observe both a higher presence of these foresters and a higher rate of adoption of CFM, together with higher levels of forest loss. However, if the prevalence of forestry is determined by time invariant factors such as geography and resource endowment, these factors are controlled for by the community fixed effects. Given these conditions, we believe that our assumptions hold, and so our instrumental variables approach would allow us to get an unbiased estimate of the effect of CFM on deforestation.

Results
We start by looking at the characteristics of communities with and without CFM (table 2). We see that communities with CFM at any point between 2001 and 2012 are different to those with no CFM. The average forest loss at the beginning of the sample period is of 1.2 ha for all the communities, but it is much higher for communities with CFM. These communities also have more forest cover at the beginning of the period (in 2000). In terms of geographic characteristics, communities with CFM are closer to Morelia (the state capital), are at a higher elevation and thus have lower mean temperatures and higher rainfall. There is also a difference in terms of the potential crop yields, with CFM communities having a lower potential than non-CFM ones for maize, sorghum and sugarcane, and a higher potential for wheat. None of the demographic characteristics are significantly different.
We focus on two sets of results. The first set of results examines how CFM affects the total area of various land covers in 2004, 2007 and 2014. The second one estimates the effect on forest loss from having an active CFM, by using the yearly forest loss from the Hansen dataset. We present the first set of results in figure 1 (results in table 5 in the appendix), where the vertical axes show the hectares for each land cover and the values on the horizontal axes represent the years since a CFM plan was adopted. We find that as more time passes since the adoption of CFM, adopting communities have more areas in primary forest than communities with no CFM. This difference is statistically significant after five years, at which point communities with CFM have 250 ha of primary forest more than communities with no CFM. There is also a decrease in the area dedicated to perennial crops, which is especially relevant in Michoacán, where there is evidence of avocado orchards replacing forests (Mas et al 2017). Finally, there are no statistically significant effects on the amount of secondary forest or pasture.
The second set of results are in figure 2 (table 6 in the appendix has the full set of results), where the vertical axis shows annual deforestation (in hectares) and the values on the horizontal axis represent the years since a CFM was adopted. We find that in the first year following the adoption of a CFM plan, there is an increase in forest loss, which may make sense given that the CFM legalizes timber harvest, even though the estimated effect is not significant. But after the first year, the trend reverses and there is a decrease in deforestation up to five years after adoption. However, this decrease is only significant in the fifth year, at which point the decrease in deforestation is equal to 10.3 ha for the average community.

Discussion
The main assumption for the validity of our results is that the presence of highly productive foresters only affects land cover change and deforestation through the effect that it has on the probability of CFM adoption (the exclusion restriction). This is an untestable assumption and is equivalent to assuming that foresters are spatially distributed randomly after controlling for time-invariant community characteristics. As discussed in the Empirical strategy section, we provide qualitative data from in-depth interviews to foresters and members of the forestry community in Michoacan, showing that the deciding factors that influence foresters' location choice are exogenous to factors that affect land cover change and deforestation within communities. If this was not the case, we would expect to see that changes to land cover and deforestation prior to the arrival of these top foresters, are correlated with the presence of these foresters in a given region. We find no such evidence: we find that both communities and municipalities exposed to these top foresters do not seem to have different levels of deforestation, degradation, and area of undisturbed forests before 1993, when the current forestry regime began. Also, communities who adopt CFM plans designed by top foresters do not experience more or less forest loss up to five years before adoption, compared to communities with no CFM or communities that adopt CFM from  another forester (see section 'Evidence in support of the validity of the exclusion restriction' in the appendix for these results). However, we are aware that given that we cannot conclusively prove that foresters are randomly distributed spatially, there is still a risk that this key assumption does not hold, despite all our suggestive evidence, in which case our results would be biased.
With that caveat, we want to highlight that our results show that as more time passes since the adoption of CFM, there is a relative increase in primary forest cover compared to non-adopters, which reaches a maximum after five years. There is also a decrease in land with perennial crops and in pasture. The net effects after five years are that communities that adopt CFM have 126 more hectares of primary forest, 140 more hectares of secondary forest, 115 less hectares of perennial crops and 332 less hectares of pasture. Interpreted through the theoretical predictions from Alix-Garcia (2007), these results are consistent with a situation where the demand for total deforestation within the communities decreases to a point where there is no conversion of forestlands into agriculture/pasture. Additionally, the fact that there is no conversion of what could be considered as more marginal forestlands, i.e. secondary forest, is also consistent with this decrease in the total demand for deforestation.
In terms of the annual changes to tree cover loss, we find that there is a decrease in deforestation. After five years, the total avoided deforestation is equal to 13.9 ha. With 175 communities that adopt CFM during our sample period, the average avoided deforestation amounts to 2433 ha.
We can use these results to estimate the social benefits associated with the avoided CO 2 emissions from the avoided deforestation. Our back of the envelope estimates of the avoided CO 2 emissions, based on the mean carbon content of forest land in Mexico of 52 ton ha −1 (Saatchi et al 2011) yields an estimate of 464 218 tons of avoided CO 2 emissions. Using a price of CO 2 of $40 USD ton −1 (High-Level Commission on Carbon Prices, 2017), we then estimate that the value of the emissions saved from the avoided deforestation amount to a total of $18.65 million USD (see table 13). According to the information provided to us by the foresters, on average a forester would charge $7 USD per ha to draw a management plan, so that for the average plan in our sample (380 ha), the total cost would be $2660 USD. The total cost for all the plans in our sample would then be $465 500 USD, which means that the value of the externality from the avoided CO 2 emissions after five years, is 40× times higher than the cost of designing all the management plans. Note that we would expect the benefits from the policy to be even larger if we include the additional environmental services provided by the forests (e.g. erosion control and biodiversity), as well as the increased value of timber harvest. However, this analysis allows us to show the magnitude of the effects when accounting for only the CO 2 emissions from the avoided deforestation. Importantly, this cost-benefit analysis does not allow us to say anything about the net benefits to the communities but is focused on the social benefits associated with the positive externalities from the avoided deforestation in the form of the avoided CO 2 emissions.
Our results are in line with those from (Blackman and Villalobos 2021), where they find that for the same region where Michoacán is located, CFM decreases forest cover loss. We extend their results by providing a more comprehensive picture of land cover within the communities, showing both the dynamics of deforestation and that the lands with primary forests are larger and the lands in perennial crops and pasture are smaller than they would have been in the absence of CFM. Our results are also in line with the results from other studies, which have shown that there is an association between CFM and more primary (denser) forest and smaller areas in agriculture (Gautam et al 2002, Niraula et al 2013. We believe that it is important to understand how CFM affects the patterns of land cover, since the ecosystem services provided by different land covers will also be different. For example, land covers with less human intervention have higher biodiversity and provide more ecosystem services (Haines-Young 2009, Martínez et al 2009) and thus our results in terms of the preservation of the primary forests, provide evidence about the additional benefits associated with CFM. This would not have been possible if we only focused on the effects of CFM on deforestation.
Finally, we want to highlight two future areas of research. First, we believe that understanding the mechanisms through which CFM affects land cover change and deforestation could help in the design of policies that improve conservation outcomes. We know that the plans designed by different types of foresters are all very similar to each other (table 10 in the appendix), but we do not know how these plans are being implemented by the different communities. As such, we believe that future research should explore what these mechanisms are, together with a deeper exploration of how alternative types of management plans result in different outcomes. Second, we believe that our use of the foresters' effects on the adoption of CFM is one that should be studied further in other contexts and possibly used as a source of identification for the estimation of the effect of CFM on other outcomes. Understanding the supply of forest management plans and how it affects CFM and its outcomes has not been studied (to our knowledge) and is potentially an important component of the success or failure of the 'use' approach to forest conservation.

Data availability statement
The data cannot be made publicly available upon publication because they contain sensitive personal information. The data that support the findings of this study are available upon reasonable request from the authors.
(PAPIIT-IN301215). We wish to thank Ulises Sánchez, Fernando Hernández and Luis García for their help in putting together the data for the communities and the forest management plans. We are also thankful for the feedback we received from participants at the AERE Summer Conferences in 2019 and the FLARE annual meeting in 2018. Finally, we are especially indebted to all the members of the forestry community in Michoacán who took the time to speak to us.

Ethical statement
This study was granted a waiver of approval by one of the author's Institutional Review Board. No identifying information was collected from respondents and a type 2 exemption was granted. Consent was obtained by all study participants.

Methods-instrumental variables approach
To construct the instrumental variables we use in our analysis, we create a dummy variable that takes the value of 1 if a community has a forest management plan drawn up by any of these foresters (for ikt = 1, with k = 1, …, 16). We then estimate the spatial lags associated to these dummy variables. For this, we construct several spatial weight matrices (W ij ) using different distance cutoffs, which will define for each community who their neighbors are. The weight assigned to each neighbor is the inverse of the distance between the two neighboring communities, such that neighbors that are further apart will have a lower weight. The spatial weight matrix is defined as where d ij is the distance between community i and community j, and 'd' is the distance cutoff. We have seven spatial lags for each forester, where the inverse distance spatial weight matrices are determined using seven different distance cutoffs, starting at 35 km (the minimum distance at which every community has at least one neighbor) and up to 65 km, with 5 km increments. With 16 different foresters, this amounts to 112 spatial lags. In turn, we also lag (temporally) each spatial lag for up to eight years, which results in 784 possible instruments. We use several instruments, all of which come from the temporal and spatial lags of the foresters. When communities decide to adopt a forest management plan, they are then 'locked' in that state for the duration of the plan. The active CFM indicator variable shows if the community has an active CFM (in which case AC it = 1) and is going to be a function of the previous adoption decision, and of the duration of the plan. We instrument for the active CFM variable, by using the spatial lags (at different distance cutoffs) of the foresters, from t-1 to t-8, and then use this estimated probability in the second stage. The first stage equation is (2) where AC it−1 is a dummy variable that indicates if a community had an active CFM in the previous year and FOR kt−h is the spatial lag for forester k in the h previous periods. X it−1 is a vector of controls (population, wood and crop prices, weather).
The second stage equation is where F it is the land cover or yearly forest loss variable (in hectares) for each community, X it−1 is the same as above, and AC it−1 is the estimated probability of having an active CFM. Then, the effect of having an active CFM on forest loss is going to be captured by β 1 . The choice of the appropriate spatial and temporal lags to use as our instrumental variables is not a trivial one. The high dimensionality of the model, with a total of 784 possible instruments, together with our ignorance of the right functional form, leads us to take an agnostic approach to this variable selection problem. Following (Belloni et al 2014) and (Belloni et al 2016), we use the post-lasso double selection method. This method uses LASSO (Tibshirani 1996) to select variables from the estimation of a system of reduced form equations (one for the active CFM equation (2) and another for the forest loss equation (3)). For both equations, it selects the controls that minimize the sum of squared residuals, while penalizing the total size of the model, with a penalty factor on the sum of the absolute values of the coefficients. For the active CFM equation, it also selects the instruments that have the most predictive power for the probability that a community has a CFM. Thus, it will select the spatio-temporal lags from (2) that better predict the adoption of CFM, after controlling for the selected explanatory variables. The fact that we select the controls for both equations, and then use the union of the selected control to estimate the effect of CFM on deforestation, reduces the probability of regularization bias (Belloni et al 2016). Additionally, by choosing the instruments and the controls with the highest predictive power for both the probability of adoption, and the deforestation levels, we are minimizing our standard errors, and will have more precise estimates of the effects of CFM on deforestation.

Sensitivity of results to changes in the definition of top foresters
The definition we use of the top foresters is an arbitrary one, and so here we present evidence that our results are not sensitive to changes in the definition of who is considered a top forester. We start by defining it as the top 15% of foresters (foresters 1-13) and increasingly include foresters who have designed less plans (i.e. less productive foresters), up until forester 20, which would be the top 50% forester (see table 7).
In each case, we calculate the spatial lags associated with each of these foresters and then include them as possible instruments as we expand the definition of 'top' forester.
Our results are very stable across the different specifications, even when the spatial lags associated with the different foresters that are chosen as our instruments change as a result of allowing marginally less productive foresters to be included as potential instruments. It is important to note two things. First, it is expected that the results will change as we include more possible instruments, since the chosen instruments also change, and so the 'complier' population is also changing. Thus, the LATE that we are estimating will change as well. Second, given our agnostic approach to the instrument selection, the instruments that are chosen by the lasso double-selection process are the same when including all the top 20th or all the top 50th foresters. This highlights the robustness of our estimated results to the definition of who we consider to be a 'top' forester, as well as the importance of allowing for an instrument selection process that is based on the predictive power of the chosen instruments. Below we present the results (see tables 8 and 9).

Evidence in support of the validity of the exclusion restriction
Our exclusion restriction requires that the presence of this high-productivity foresters have no impact on land cover and forest loss, other than through the adoption of CFM. One possible violation of the exclusion restriction could be due to the differential management system implemented by these foresters. We find no evidence of this. In table 10, we can see that the characteristics of the plans drawn up by both types of foresters, are not statistically different. We find that the total area of the community, the total area under the management plan, the proportion of the land under the CFM, as well as the total volume of wood per hectare, are smaller for communities that adopt a CFM from a high-productivity forester. However, these differences are not statistically significant, for either of our two definitions of the most productive foresters.
Similarly, another concern is that the most productive foresters are targeting communities that are intrinsically different to other adopting communities. These differences could reflect differences in factors that lead communities to have different trends in deforestation, and as such, could also reflect possible violations of our exclusion restriction. We clearly cannot test for differences in unobserved characteristics, but we believe that by testing for differences in observed characteristics, we can gain information about the underlying differences between the communities. The results from the t-test of differences in the characteristics of the communities by the type of forester they hire, can be found in table 11. We find very few differences between communities, by forester type, and the ones that are statistically significant, are only so at the 10% significance level. We do a similar analysis for the potential yield of a group of crops and find that there are no differences either (see table 12).
Finally, the exclusion restriction implies that the highly productive foresters' choice of location is uncorrelated to the forest conditions of the communities that are exposed to these foresters. This implies that if top foresters' location is random relative to land cover and forest loss, then we would expect to see no changes in those outcomes prior to the arrival of these top foresters to a given community. Here, we present evidence that we believe supports the assumption of the top foresters' randomized location relative to land cover and forest loss.
First, we show that for communities that will be more exposed to top foresters, the forest conditions before 1993 are not different to those of communities not exposed to these foresters. We focus on the years before 1993, given that this is the year where the current forestry regime went into effect and CFM plans were allowed. To define exposure to top foresters, we use the 35 km spatial lags for all the top foresters, and then define the exposed communities as those for which the spatial lags associated with any of the top foresters is above the 75th percentile of the values for all the time period. The way we do this is that for each community we calculate the average of the 35 km spatial lags for each forester, then define those as treated, if the average exposure for any of these top foresters spatial lags, are above the 75th percentile of the average exposure for all the other communities. We also do this same exercise but instead of using the 75 percentile, I use the 50th percentile and the results are very similar in both cases. We use forest outcomes data from Vancutsem et al (2021), who use Landsat images to estimate deforestation and degradation in tropical moist forests, from 1990 to 2019. We then run a set of regressions of forest outcomes on the interaction between year dummies and a top forester exposure dummy, while controlling for quarterly rainfall and temperature per year (and lagged year). The regression coefficients Significance levels: * * * (1%), * * (5%) and * (10%). Significance levels: * * * (1%), * * (5%) and * (10%).    of the interaction between the top forester exposure dummy and the years is the conditional mean of the forest outcome for each year, for the municipalities/communities where the top foresters will be located in the future. We find that that the trends for both groups of communities move in a similar fashion, with none of the means being statistically significant ( figure 4). This reassures that our choice of instrument does not seem to be correlated with changes in forest outcomes prior to the implementation of CFM. We do a similar analysis at the municipality level. In this case, we define the base municipality from which each top forester operates from, as the municipality where they have done the most CFM plans.
We compare these to the municipality of residence reported by the foresters in our interviews, and then match this definition of the base municipality for the ones we have data on. We then classify these municipalities as the municipalities with the most exposure to the top foresters, and construct a panel with data at the municipality level on forest outcomes and weather. We run regressions where the outcome variables are the total area in hectares of undisturbed tropical moist forests, the total deforestation and degradation, and regress them on the interaction of year and the top foresters base municipality dummy, controlling for rainfall, temperature (both annually and quarterly) and municipality fixed effects, with the  standard errors clustered at the municipality level. We find that there seems to be no difference in the conditional means of the municipalities where these top foresters are based ( figure 5).
Lastly, we believe that if top foresters' location is uncorrelated with forest loss, then we would expect to see no changes in those outcomes prior to the arrival of these top foresters to a given community. To show that this is the case, we run an event study where the treatment is the adoption of a CFM plan designed by one of the top foresters, and the outcome is the forest loss per year inside of the community. We find that prior to the adoption of a CFM designed by a top forester, there is no difference in forest loss compared to other communities with no CFM or with a CFM designed by another forester (figure 6). We believe that if top foresters were targeting certain types of communities (for example, those that are less prone to cut down trees), then we would be able to observe that the behavior of those communities would be different in ways that would be reflected in their forest loss. The fact that we see no statistically significant difference in terms of forest loss provides evidence in support of the randomization assumption.
Values used for the cost-benefit analysis of avoided CO2 emissions