Introduction

Forest restoration is a powerful nature-based solution to mitigate climate change1,2, safeguard biodiversity3, and preserve ecosystem services4. Prominent global agreements, like the Bonn Challenge and the Sustainable Development Goals, call for large-scale forest restoration. To succeed, restoration must include private landowners, who own substantial parts of available restorable land globally5,6,7 and who face competing pressures to use land for agriculture, pasture, or peri-urban development. Many of these private land managers have highly land-dependent and often precarious economic livelihoods8. Even when restoration produces long-term profits or improves local water availability, landowners can face high up-front investments, labor costs, and knowledge barriers that hamper meaningful restoration activity. To reduce these barriers, restoration programs, such as those supported by the Atlantic Forest Restoration Pact (PACTO) in Brazil, offer economic and technical assistance to private landowners9,10. Although such initiatives are expanding rapidly worldwide11, there is little empirical evidence regarding their ability to generate additional forest restoration relative to a plausible comparison group of properties.

We evaluate the forest recovery impacts of PACTO —a large-scale restoration coalition targeting private land in the Brazilian Atlantic Forest (BAF). The BAF is an important setting in which to implement and evaluate large-scale forest restoration because it is sizeable, ecologically significant, and 75% privately owned12. Nearly 90% (roughly 1.1 million km2) of the BAF has been cleared or converted for agriculture, pasture, logging, mining, or urban centers over the last 500 years13. Reversing this trend through restoration is challenging since approximately one-third of South America’s population lives inside the original forest boundaries, and many economic livelihoods are dependent on agricultural and pastoral land use. Conserving and restoring the remaining fragments of the forest may help protect unique species in danger of extinction, in addition to providing important carbon, erosion, and watershed function services14,15,16,17,18.

PACTO is a collective agreement and regionalized knowledge-sharing network of over 250 stakeholders across the 17 states of the BAF, which encompasses nearly 30 degrees of latitude and a wide variety of forest types. The PACTO network includes government agencies, NGOs, companies, and research centers that implement forest restoration by directly assisting private landowners. The pact aims to democratize science-based restoration methods, apply a common monitoring protocol across the region, and assist implementers in securing funding. PACTO implementers, usually NGOs or government agencies, support forest restoration by providing landowners with free or subsidized materials, labor, training, and other activities. Depending on location and ecological conditions, they may support active restoration (i.e., direct seeding or planting young trees), passive restoration (i.e., allowing forests to regenerate naturally), and assisted natural regeneration (i.e., fencing off areas or removing non-native species)19,20.

PACTO operates in the context of Brazil’s federal and state initiatives to protect and enhance forests. The country’s key forest policy is the federal Forest Code, presently formalized in the Native Vegetation Protection Law (NVPL) of 2012. This law requires all landowners in the BAF to set aside 20% of their land as a Legal Reserve of native vegetation conservation, with additional requirements for ecologically-important areas (such as riparian zones) known as “Areas of Permanent Protection” (APPs)21. Since landowner compliance with these requirements is extremely low—there are 6.2 Mha (million hectares) of native vegetation deficits on private land in the BAF22—restoration needs are ubiquitous across the PACTO states. However, the explicit and implicit costs of restoration for private landowners in the BAF are high, particularly for active restoration10. PACTO’s approach recognizes that landowners often need material and knowledge support to restore forest effectively or at all. A core aim of the program is to help landowners comply with the law’s 20% native vegetation minimum19,20.

We use a quasi-experimental approach to estimate the impact of collaboration with PACTO on net forest recovery within private land in six states of the Atlantic Forest from 2000 to 2018 (Fig. 1). Our analysis compares restoration outcomes for land that received support from a PACTO implementer after 2009 to similar land units that did not. PACTO implementers began working with landowners on restoration in 2009, with participation growing over time. Because information on the year of first enrollment is not available, we consider all properties that received support to be treated from 2010 onward. We divide the landscape into 2 ha (hectare) units of analysis and define treatment according to having any land enrolled as well as by the proportion of the unit that overlaps with a PACTO-supported parcel.

Fig. 1: Map of parcels of land enrolled in PACTO-supported restoration.
figure 1

Our study area is the Southern coast of Brazil within the Atlantic Forest. The red areas on the maps are the parcels supported by the Brazilian Atlantic Forest Restoration Pact (PACTO) evaluated in this study. Parcels are located in the six states of Bahia (BA), Espírito Santo (ES), Paraná (PR), Rio de Janeiro (RJ), and São Paulo (SP). Black lines represent the borders of Brazilian states. The basemaps for both the main figure and the inset are provided by ESRI45,46.

Full size image

The estimation includes two steps. First, we use statistical matching to create a comparison group of land units that were similar to the PACTO group in key aspects driving likely program enrollment and forest change prior to 2009 but that did not participate in PACTO. The choice of matching covariates is informed by economic theory and the results of interviews with stakeholders (see Supplementary Discussion). We select potential matches for the treated grid cells from the pool of units within 0.5-5 km of each treated cell, using propensity score matching within municipality. This control group, constructed via matching, plausibly represents the unobservable counterfactual scenario of what would have occurred on PACTO-assisted land without PACTO program intervention.

Second, we compare outcomes for the PACTO-supported and non-PACTO-supported land over time using difference-in-differences regression analysis. The difference in net forest recovery between the groups after PACTO assistance began (2010–2018) versus before (2000-2009), conditional on included controls, represents the causal effect of PACTO support. Similar methods have been used to evaluate global forest protection efforts such as parks or payments for ecosystem services23,24,25,26, but causal evaluations of reforestation programs are rare27.

To understand the potential mechanisms and drivers of restoration and PACTO’s role in the process, we conducted semi-structured field interviews with stakeholders from several points on the restoration supply chain (See Fig. 2, Supplementary Discussion, Supplementary Note 1). These interviews helped to identify key barriers and facilitating conditions relevant to the BAF context, which inform our choice of matching variables. In addition, economic theory and our interviews elucidated important drivers of change and led to testable hypotheses about heterogeneity in program effectiveness. Specifically, we test for differential impacts of PACTO by distance to cities as a proxy for opportunity costs (the foregone profits from alternate uses of land), and across state entities, because the risk of legal sanction for violating minimum native vegetation requirements varies by state and was highlighted in the interviews as a key factor driving restoration uptake and maintenance activities.

In this work, we find evidence of restoration gains attributable to PACTO network support. Between 2010 and 2018, we estimate that net forest recovery increased by 10-20 percentage points on average in PACTO-supported land relative to the counterfactual. Consistent with the financial and legal compliance motives for restoration found in the interviews, we find that impacts were greatest in more remote areas–where opportunity costs are likely to be lower due to higher transport costs for agricultural products–and in states with higher enforcement activity, where the benefits of legal compliance are higher. These findings demonstrate both the promise of forest recovery programs on private land and the importance of the economic and institutional context in scaling-up restoration efforts.

Results

Drivers of restoration success

Figure 2 shows the organizational structure of the PACTO network. Structured interviews with stakeholders in each of these categories (Fig. 2, Supplementary Note 1, Supplementary Discussion) gave us an understanding of the restoration context and drivers of impact. According to these interviews, the decision to restore is primarily financial. This is consistent with a basic economic theory of restoration where a landowner’s choice to restore forest on a given parcel is driven by a comparison of their own private benefits and costs of restoration (see Supplementary Discussion). Indeed, while speaking about the incentives for restoration, one stakeholder noted: “the most painful part of the body is the pocket”. While the landowners themselves have the ultimate decision authority over their land, their decisions are strongly conditioned on the expected financial gains and losses of land-use options, the influence of government regulations and programs, contractual relationships with large companies, and support provided by restoration NGOs and the PACTO leadership council. We describe the interview results and the Forest Code in greater detail in the Supplementary Discussion.

Fig. 2: PACTO organizational structure and interviewees.
figure 2

This figure illustrates the organization of the Brazilian Atlantic Forest Restoration Pact (PACTO), partner activities, and the subjects of our field interviews. Panel (a) shows PACTO’s organizational structure and main partner activities. PACTO is led by a coordination council, but regional units oversee individual partner activities within each state. Most partners are restoration implementers, and many partners engage in multiple activities. Restoration implementers are PACTO partners that work directly with landowners to carry out restoration, with consultants sometimes serving as an intermediary. Panel (b) describes the eleven interviewees from field visits, with activities of partner interviewees referenced by color. Partner interviewees included two state-level government bodies (Gov 1 and Gov 2), two non-governmental organizations (NGO 1 and NGO 2), and one company (Comp 1). The two land consultant interviewees were employed by Gov 2 to recruit and manage restoration on private land, and the two groups of landowner interviewees were receiving restoration assistance from NGO 1.

Full size image

Overall, the benefits of forest restoration include the reduced risk of government sanction because of better compliance with the Forest Code, access to agricultural credit, additional income, and increased environmental quality or ecosystem service flows. Because of the risk of financial sanctions due to violations of the 20% forest coverage mandate in the Forest Code (according to at least one restoration consultant and a state prosecutor that were interviewed), the perceived benefits of forest restoration are related to the probability that the law is enforced. PACTO’s support may help landholders come into compliance with the law, which lowers the risk of fines. In addition, compliance with the Forest Code may also increase producer access to certain government-supported credit programs. For example, article 41 of the Native Vegetation Protection Law states that agricultural credit, agricultural insurance, tax credits, funding, and other benefits are available to landowners as compensation for compliance with forest restoration and conservation requirements of the law. This may be beneficial for some participants. In our interviews, some producers also noted a link between forests and water, both for consumption and for agricultural use, and others noted that over the long term, restoration projects that include agroforestry may help provide additional income from tree crops (see Supplementary Discussion). Reforestation projects may also improve soil stability and reduce erosion or sedimentation18,19.

On the cost side, interview participants articulated a wide range of factors mediating large-scale private land restoration, and detailed how the PACTO collaboration network works to reduce cost in order to make restoration more attractive. There are implicit and explicit costs for each hectare of land restored, which may differ considerably across locations. Because restoration directly precludes highly profitable land uses, like agriculture, and cattle ranching28, the main implicit cost is the foregone profits from alternate uses (the opportunity cost of restoration). In interviews, PACTO leadership and its implementation partners emphasized the strategy of targeting degraded or abandoned land which would have lower opportunity costs, facilitating engagement in restoration. The explicit costs of restoration vary depending on the method. PACTO implementing partners emphasized the high cost of active restoration, which re-creates forest on abandoned pasture or fallow land. This is the dominant form of restoration in the BAF and includes planting and caring for a variety of delicate tree species until they are able to survive, which can take 3–5 years. It requires both capital (seedlings, fencing) and labor (for weeding, planting and maintenance) costs. This cost can reach $8000 US dollars per hectare in the initial investment stage10. NGO implementers associated with PACTO state that they often covered most or all of the direct costs of restoration (Supplementary Discussion), providing a key form of support.

Interviewees also revealed the importance of networks in reducing informational and direct costs. PACTO supports information provision by establishing principles of restoration management and disseminates information to 16 “Regional Units.” Institutions affiliated with PACTO are involved in seedling production and scientific experiments in restoration that are aimed towards cost reduction through technological innovation and strategic use of inputs. This knowledge is valuable, because effective restoration requires complex technical expertise about ecological succession, soil, species, planting techniques, and market values.

Restoration may fail without sufficient attention to the match between local conditions and restoration methods, or the marketability of agroforestry products10. The decentralized nature of PACTO means that implementers have deep local knowledge in addition to strong connections with scientific institutions currently researching restoration methods. Interviewees emphasized that implementers use this knowledge to provide the most effective technical assistance possible, and also to educate landowners about longer-term benefits of restoration on their specific parcel of land.

Participation in PACTO induced additional restoration

The PACTO enrolled areas include 33,011 hectares across six states in the Brazilian Atlantic Forest (Fig. 1). This is comprised of 4697 unique parcels with an average parcel size of seven hectares. Much of this is non-forested farmland (Supplementary Fig. 1). A substantial portion of the parcels (32%) are in Bahia, with 18% of parcels in each Paraná and Espírito Santo, 14% in Rio de Janeiro, 11% in Minas Gerais and 6% in São Paulo.

PACTO began work in a period of high pressure on the Atlantic Forest. Figure 3 demonstrates that net forest recovery was declining in both the land that would receive support from PACTO and the matched control areas before 2009 (after a peak in 2003). However, trends across the enrolled and unenrolled groups are not statistically different prior to the intervention (Supplementary Table 6). After 2009, trends for the PACTO-supported land and the control areas diverge. The decline in net forest recovery reversed among PACTO-supported areas after 2009 (until 2018), while net forest recovery continued to decline among control areas.

Fig. 3: Annual net forest recovery.
figure 3

Average net forest recovery trends (in hectares) for all cells in post-matching treated (PACTO-supported; navy line; N=42,056) and untreated (not PACTO-supported; dashed red line; N=37,299) groups are shown before and after the beginning of the program in 2009 (dashed vertical line). In the pre-treatment period from 2000 to 2009, average net forest change is similar for treated and untreated groups. In the treatment period from 2010 to 2018, trends diverge with the treated group experiencing larger net forest recovery.

Full size image

After further controlling for potentially confounding differences between matched PACTO- supported and control areas using regression, we find that net forest recovery increased substantially relative to the counterfactual (Figs. 4 and 5). The difference-in-differences estimates with the matched sample show that PACTO increased net forest recovery by 10-20 percentage points on average across 2010–2018 relative to the previous period (Fig. 4 and Supplementary Table 1). The range of estimates corresponds to different definitions of treatment for the analysis units, which depend on the percent of overlap with PACTO areas. When treatment is measured as having any area under restoration in an analysis unit (a 2 ha grid cell), we find that on average PACTO converted 9.9 more percentage points of a cell’s initial non-forested area to restored forest than the counterfactual (Fig. 4a and Supplementary Table 1, column (3)). When treatment is measured by the percent of a cell treated, the measured impact is about 5 percentage points at 25% intensity, 10 percentage points at 50% intensity, 15 percentage points at 75%, and for fully covered parcels, approximately 20 percentage points (panel b of Fig. 4 and Supplementary Table 1, column (6)).

Fig. 4: Participation in the Brazilian Atlantic Forest Restoration Pact (PACTO) increases area restored relative to baseline restorable area.
figure 4

This figure shows the estimated marginal effects based on an interaction term between treatment and the post period indicator from Eq. (1). The outcome is net hectares of forest recovered divided by hectares of initial non-forested area. The hollow dot in a denotes the simple binary treatment effect from Supplementary Table 1, column (3), which is an ordinary least squares estimation with cell fixed effects, time, temperature and precipitation controls (N = 79,285; treated: 42,056; control: 37,229). The estimates indicated by solid dots in b are from the same table, column (6), also using the cell fixed effect specification of Eq. (1) (N = 79,285). The treatment variable in this case is continuous, and the dots indicate the treatment effect evaluated at 25, 50, 75, and 100% of the cell under restoration. The average treatment intensity for cells is 35%. Navy lines are 95% confidence intervals. Standard errors are clustered at the municipality level. Statistical significance is based on two-sided t-tests for which p-values are reported in Supplementary Table 1.

Full size image

Additionally, we estimate PACTO impacts for specific years using an event-study estimation (Fig. 5, see “Methods”). Here, the estimates trace the dynamic, annual treatment effects of PACTO on net restoration over time, evaluated at the mean proportion treated (35%). We find that effects are positive and statistically significant from 2011-2018. The estimates also show no statistical difference between treatment and control units in the pre- 2010 period, confirming that prior to treatment, both types of land were being managed in similar ways. It is important to keep in mind that this graph shows annual net forest recovery, so the total effect of PACTO over the 2010–2018 period is a cumulative function of the point estimates in Fig. 5.

Fig. 5: Event study estimation of participation in PACTO.
figure 5

Figure shows point estimates from a regression whose outcome is the inverse hyperbolic sign of net forest recovery measured at the annual level (N=79,285; treated: 42,056; control: 37,229). The estimation equation is given in the supplementary equation (1). Each dot shows the estimated coefficient on the interaction term between the year and the proportion treated, evaluated at the mean proportion treated (35%). The regression includes cell and year fixed effects, as well as annual aggregates of temperature and precipitation for a given grid cell. The omitted year is 2009, where the vertical gray dotted line represents the start of PACTO support. Navy lines represent 95% confidence intervals. Standard errors are clustered at the municipality level. Statistical significance is based on two-sided t-tests for which p-values are reported in Supplementary Table 16.

Full size image

These results are robust to several specification checks. First, our main estimation uses controls drawn from a pool within 5 km around the treated parcel, but excludes possible matches within 500 m of an enrolled grid cell. We use this strategy to identify similar parcels but to minimize possible contamination of estimates due to spillover effects. In particular, we may be concerned that a landowner who is restoring in one area might increase deforestation in another portion of their land, or that their collaboration with PACTO might result in learning or information that affects neighbors’ behavior. We examine alternate specifications where we include all possible cells in the pool of potential matches (Supplementary Table 13) and where we more restrictively exclude potential matches within 1000 m (Supplementary Table 14). We find no difference in the impact estimates across these three samples, indicating a lack of evidence for spillovers immediately adjacent to parcels where PACTO affiliates are working. All of these effects are stable under alternative fixed effect specifications and for alternative measures of forest recovery (Supplementary Tables 1, 10, 11, 12).

Using the marginal effect from the proportional treatment estimation (Supplementary Table 1, column 6), the average treatment intensity in our sample, and the amount of non- forested land in treated cells (66,851 ha), we calculate that the total restoration due to the PACTO intervention was approximately 4600 ha of forest. This level of restoration amounts to 21% of the total restorable land (21,800 ha) in our study sample of fully-treated PACTO parcels. For reasons that we discuss below, this is likely to be a lower bound on the causal effect. In addition, since our sample covers only a small part of the properties enrolled in PACTO, this is also a lower bound of the total impact of PACTO.

Impact heterogeneity consistent with key drivers of success

In light of the restoration dynamics and the key barriers identified in the interviews, we expect heterogeneous impacts of PACTO across space. One prediction is that restoration is more likely to be effective where opportunity costs are lower (see Supplementary Discussion). We test this hypothesis by adding an interaction term between the treatment effect and proximity to the nearest city (measured by 1 divided by 100 km to nearest city) in our baseline specification (Supplementary Fig. 3). This transformation of distance allows for impacts to vary depending on distance to city in a non-linear way, which may capture rapidly decreasing opportunity costs further from cities. Consistent with the hypothesis that restoration is less costly further from markets, we find that the effect of restoration is twice as high at the 95th percentile of distance (around 130 km) as it is at the 5th percentile (less than 1 km).

In addition, interviewees suggested that enforcement of the Forest Code varies by state and property size, creating heterogeneous compliance incentives across space. Landowners experiencing high enforcement will be more likely to restore. To test how the program impact varies by state, we estimate separate regressions with our matched subsample using the most restrictive set of controls (grid cell fixed effects and time-varying covariates). Figure 6 shows the impact by state using the binary treatment variable (coefficients are from Supplementary Table 2). Supplementary Table 3 provides the same estimation using the proportion treated variable. The largest treatment effects are in the states of Bahia and Paraná.

Fig. 6: Impacts and fines are highest in the states of Bahia and Paraná.
figure 6

Panel a shows the coefficient on the interaction between having any part of the analysis unit treated and the post period indicator for each state. The outcome is net hectares of forest recovered divided by hectares of initial non-forested area. All estimations include cell fixed effects and time-varying controls, as specified in Eq. (1). Markers denote the marginal effect of treatment and lines are 95% confidence intervals. Standard errors are clustered at the municipality level. Estimated values are reported in Supplementary Table 2, columns (2)–(7). Statistical significance is based on two-sided t-tests for which p-values are also reported in Supplementary Table 2. States are ordered by the magnitude of the point estimates. The number of observations used in estimation for each state are as follows: Bahia (N = 25,944; treated: 13,388; control: 12,556); Paraná (N = 13,780; treated: 7569; control: 6211); Espírito Santo (N = 15,316; treated: 7950; control: 7366); Rio de Janeiro (N = 10,616; treated: 5940; control: 4676); São Paulo (N = 4466; treated: 2342; control: 2124); Minas Gerais (N = 9163; treated: 4867; control: 4296). Panel b shows the value of forest-related fines per hectare of forest per rural individual for each state in the sample between 2000 and 201829. States referenced in b are shown in the same left-right order as in a.

Full size image

A possible explanation for the variation in the impact across states is that some states impose greater institutional pressure than others. To explore this hypothesis, we calculated the value of forest-related fines divided by the hectares of forest per rural population. This measure will be higher if forest per person tends to be smaller, capturing a mechanical relationship of more fines if there are more properties. It will also be higher if there is more enforcement, all else equal. Figure 6 provides evidence of differential enforcement of the Forest Code using data on forest-related fines from 2000 to 201829. Fines per forest ha per rural capita are highest in Bahia and Paraná, states where we observe higher restoration impacts. We see a similar trend across states for fines per capita or fines per forest ha (Supplementary Fig. 4). Fines alone do not represent the full institutional enforcement picture, as enforcement of the Forest Code is also driven by state governments through regional prosecutors of environmental law and state-level ministries of the environment. We also cannot rule out that these correlations are driven by opportunity costs, which certainly vary across states and may drive some of this pattern. However, these patterns are consistent with institutional context shaping the effectiveness of restoration.

Discussion

We have presented evidence that a large consortium of institutions implementing forest restoration projects in the Brazilian Atlantic Forest has been successful. Over a 10-year period, we estimate a causal increase of at least 4600 hectares of restored forest on previously deforested land. This analysis is important because there is very little evidence that forest restoration can be successful on private land and at a large scale, and even less that this can occur in low and middle income countries27. The vast majority of evidence on restoration effectiveness involves experimental plots, small projects, and comparisons of before and after measures of forest cover that may be confounded with general time trends27. Studies in Chile and China provide some related insights using quasi-experimental methods. In Chile, a quasi-experimental study found that forest area on private land can be increased via subsidies to landowners, although this expansion came at the expense of native forests30. Causal estimates of the impact of China’s Sloping Lands Conversion Program for reforestation have demonstrated that PES-type payments can induce forest planting31, but the program has been criticized for low diversity of the species chosen to be planted32. Our study contributes by estimating restoration impacts on already cleared land and providing an evaluation of decentralized reforestation efforts on private land in a biodiversity hotspot.

Our method of evaluation has important limitations, but these are likely to lead to underestimates of the true program impacts. First, there is the issue of detecting restoration using satellite data. As we do not know the exact enrollment date of each property, we assume that interventions begin at the earliest possible date any landowners began receiving support. Many properties may have enrolled later than that, and recent work suggests that it takes 3–5 years, depending upon growing conditions, for restored forest to be detectable via 30 m satellite imagery, such as that used for this study33. Locations that begin restoration in 2017 would only have one year of detectable restoration in our dataset. In general, restoration in later-enrolled units may not appear in the impacts we measure, potentially leading to underestimates. Second, our dataset represents areas of restoration that were compiled from thirteen PACTO affiliates as part of a pilot project intended to map out all land in restoration under the consortium. These affiliates included ten environmental NGOs and three companies. Because not all of the land has been mapped yet, the proportion of the total PACTO effort that is represented by the data is unknown. Some control parcels may in reality have had PACTO support, which would reduce our estimated impact compared to the true treatment effect. In addition, we are not able to estimate the full impact in terms of hectares restored because we do not have the full set of supported parcels.

Several aspects of restoration deserve further study. Our estimates show significant variation in restoration effectiveness across states, and the relationship with forest-related fines is consistent with information from our interviews indicating that institutional support is a key factor driving restoration decisions. Further, we observe greater restoration impacts farther from urban areas, which implies that lower opportunity cost land has greater restoration outcomes. However, although these findings are consistent with expectations, better information on land value34 and landowner restoration decisions35 would enable deeper exploration of the interaction between opportunity costs and restoration success and enable more cost-effective targeting of programs like PACTO. Further, we are unable to explore whether or not there could be a reason for differences in restoration efficacy due to differences in the direct costs of restoration across states and over time or by implementer organizational capacity. Changes in cost could occur because some soil types and locations facilitate restoration, because labor costs differ, or because particular organizations working with PACTO are more efficient in their enrollment and implementation processes. Understanding the dynamic relationships between cost and effectiveness is also a key area for future research. Another dimension ripe for exploration is variation in treatment across landholding size. In field interviews, we learned that large landholdings (such as eucalyptus-producing companies) are subjected to heightened institutional monitoring and enforcement, while small landholdings often receive more leniency and less monitoring (for further discussion see Supplementary Discussion). The dynamics of enforcement could be explored with a more detailed dataset. Finally, drivers of restoration and enforcement threats may change over time, meriting additional study of the longer-term impacts of programs like PACTO.

Overall, our current findings regarding the importance of institutional support and the interaction with opportunity costs for producers provide guidance for practitioners of restoration and for governments hoping to support natural climate solutions. PACTO and other restoration programs may produce more restoration at lower cost in an environment with strong complementary institutions and spatial targeting to lower opportunity cost areas. The PACTO initiative provides an example that demonstrates the possibilities for decentralized forest restoration both in Brazil and globally, and highlights important questions that remain to be answered.

Methods

Unit of analysis and data sources

For the forest recovery analysis, the landscape is divided into 2-hectare grid cells, creating a pool of uniformly-sized units. We choose 2-hectare units following the guidance of Avelino et al.36. Treatment is defined by overlap between PACTO parcels and these grid cell units (Fig. 7). Parcel data on PACTO-supported land was provided by the consortium as a geospatial shapefile and included size, location, and shape of each parcel. Data on the specific year of enrollment was not available; we therefore classify all enrolled parcels as treated after the first year of PACTO support in the region (2010 and after). The parcels include data from thirteen implementing organizations, including three companies and ten environmental NGOs.

Fig. 7: Treated and untreated grid cells.
figure 7

PACTO parcels are shown in yellow, and part of the 5 km untreated buffer area in pink. Units of analysis are defined by a grid of 2 ha cells. Units that have any overlap with a PACTO parcel are considered treated for the purposes of matching and the binary treatment analysis. The percent covered by a PACTO parcel is also calculated and used for the continuous treatment analysis. Parcel shapes vary, with some more centrally-condensed as in (a) and others more snakelike as in (b). The latter leads to a larger proportion of partially-treated units.

Full size image

Since parcels have different shapes, they may pass through only a portion of the square cell that is the unit of analysis, creating partially-treated cells (Fig. 7). The gridding procedure produces 42,056 treated units that have an average treatment intensity of 35%, defined by the extent of area overlap between a parcel and a cell. We use both a binary treatment variable and a continuous treatment (proportion of grid cell enrolled) as alternate measures of PACTO support in our empirical strategy.

We source land cover and forest change variables from the MapBiomas Collection 6, which provides these at a 30m spatial resolution33. The land cover data give snapshots of land use proportions across parcels in 2000 and 2010. High-level land cover categories include Forest, Farming, Water, Non-Forest Natural Formation, and Non-Vegetated Area. The collection separately classifies natural forest and forest plantations. Forest plantations are allocated to the farming category. Change data include annual forest gain and loss. Specifically, forest losses are suppression of primary forest vegetation and suppression of secondary forest vegetation, and forest gain is recovery of secondary forest vegetation. To ensure forest change events do not reflect transition noise in remote sensing data, MapBiomas applies a per pixel classification trajectory analysis that explicitly validates forest regrowth and deforestation events. The analysis utilizes a 5-year moving leads and lags window to verify change persistence. For example, a pixel classified as forest regrowth in 2015 must have an anthropic land use classification in 2013 and 2014, followed by three years of native vegetation from 2015 to 2017, among other persistence criteria.

The restoration outcome of interest is net forest recovery, which we define using the change data as the sum across years of annual forest gain minus the sum across years of annual forest loss. In our main regression specification, we divide net forest recovery by the baseline non-forested area (as measured in the land cover data at the start of each period) in order to create a proportional measure of restoration. In addition, we consider regressions using the net forest recovery measure in hectares in robustness checks (Supplementary Tables 10, 11).

We downloaded the remote sensing data in early 2022, which meant that forest recovery variables were only available through 2018. We extract forest change within units from 2000 to 2018 in hectares of gain and loss. This is then aggregated to two nine-year periods for the main analysis, a pre-treatment period from 2000 to 2009, and a treatment period from 2010 to 2018. The data are also aggregated by year for the event study. As an additional robustness check, we verify that our main results hold with data from MapBiomas Collection 9, published in August 2024. This most recent collection uses a substantially updated classification scheme, which measures much less forest gain than previous Collections. However, the results remain similar in magnitude even when extended through 2021. The matched raw forest measures over time can be found in Supplementary Fig. 7. Compared to Collection 6, Collection 9 appears to measure less restoration generally. However, the results for the binary and continuous specification are in Supplementary Table 15, which depend on the difference between treatment and control groups, and are nearly equivalent to the main results.

We also collect data on time-invariant covariates that may influence land-use decisions for use in matching and the difference-in-differences regressions. The matching covariates include slope, elevation, land cover type (including the presence of water), bioclimate, geology, initial forest coverage, distance to nearest city, population of nearest city, protected areas, municipality, and state. Slope, elevation, land surface type, and bioclimate data come from the United States Geological Survey database (www.USGS.gov). Protected area boundaries are from the World Database of Protected Areas (WDPA), downloadable at Protected Planet (www.protectedplanet.net/country/BRA), and city data come from GISMAP. We verify that any deviations between WDPA boundaries and the boundaries from the Brazilian System of Protected Areas (SNUC) do not materially affect our methods or results. State and municipal boundaries were downloaded from the Instituto Brasiliero de Geografia e Estatística.

In addition, we compile time-varying data on temperature and precipitation, which are included in the regressions. These are extracted annually at the municipality level from the CHIRPS dataset37. Temperature and precipitation affect both the measurement of the land use outcomes and also the potential success of restoration. We use these variables in our estimation as covariates, but not in the matching analysis. Finally, although we use the data only to establish state-level patterns, we do report summary statistics from annual environmental fines data29. Reported fines are those issued by IBAMA – the Instituto Brasileiro do Meio Ambiente e dos Recursos Naturais Renováveis, which is the federal agency in Brazil responsible for enforcing environmental law – for flora-related violations affecting native vegetation. Geoprocessing was performed with Python and tools in the ArcGIS Pro v3 console. Further data processing and all statistical analyses were performed with Stata 16.

Pre-estimation matching

Because PACTO implementers target land they identify as ideal for restoration (e.g. degraded agricultural land; abandoned pasture; land owned by individuals or companies thought likely to engage in restoration maintenance), treatment is nonrandom. To account for selection into restoration, we use propensity score matching to construct a counterfactual group that is, on average, similar to the treatment group with respect to observable characteristics in the pre-treatment period. We select matches from a buffer of 0.5-5 km around all treated parcels in order to establish a potential pool of controls from the set of 2 ha units.

Matches are chosen using one-to-one probit propensity score matching with replacement. Propensity scores from probit regressions estimate the probability of treatment for each cell. Matching variables used for probit regression include the following observable characteristics: coverage area in 2000 (for forest, non-forest natural formations, farming, non-vegetated area, and water), forest change area between 2000 and 2010 (for primary vegetation loss, secondary vegetation loss, and secondary vegetation gain), protected areas designated as of 2000, elevation, and slope. Since forest restoration is just one of many land-use options for private landowners, a key identifying assumption of this approach is that control cells represent land with a similar opportunity cost of restoration, where opportunity cost is defined as the economic benefit from the next-best alternative use of the restored land. Such uses might include agriculture, pasture, or urban development. Matching covariates are thus chosen to capture factors that are likely to determine key aspects of opportunity cost, according to both prior studies and our interviews with stakeholders. In addition, to ensure matched pairs are politically and geographically similar, we restrict matching to be within municipality. We do not match on distance to city as there is little variation in this variable within municipalities. As a robustness check, we also implement coarsened exact matching38. We discuss this in greater detail in Supplementary Methods.

Our main sample excludes untreated cells within a 500 m buffer of treated cells in the set of matched controls. The purpose of this is to limit the possibility of contamination of estimates from spatial spillovers. To check robustness, we also conduct estimation on a sample that excludes cells within a 1000 m buffer (Supplementary Table 14) and one that includes matches from all possible cells within the 5 km buffer (Supplementary Table 13).

Matching produces a plausible control group according to balance of covariates and pretreatment trends. Supplementary Table 5 shows covariate means, normalized differences and variance ratios for the groups after matching. Net recovery in the pre-treatment period is nearly identical across groups. In the treatment period, the two groups diverge. In the treatment group, about 0.155 hectares are recovered, net of loss. This is 3.8 times larger than that of the untreated group on average (see also Fig. 3). Starting forest and non-forest cover are very similar for both periods. Land characteristics and protected area overlap are also well-balanced. On average, the treated and untreated groups have similar slope, elevation, and distance to city. Overall, normalized differences between pre-treatment means show good balance after matching, as defined by a difference below |0.25|39. Supplementary Fig. 2 shows the change in normalized differences between treated and control groups before and after matching. For all matching covariates, standardized differences in means move closer to zero after matching. Supplementary Table 5 also shows that variance ratios of pre-treatment values between treated and untreated groups are all very close to one (indicating distributional similarity) and quite far from ratios that might indicate distributional “imbalance” (ratios like 2 or 0.540).

Difference-in-differences estimation with matched sample

After matching, we apply difference-in-difference regressions with cell fixed effects to estimate the average treatment effect on the treated. This combined spatial matching and difference-in-differences approach is in accordance with similar quasi-experimental causal analysis in spatial forest contexts23,24,25,26 and draws from established causal inference methods for conservation interventions41,42,43. It has been shown that this strategy can replicate experimental estimates44.

Equation (1) represents the main estimation strategy:

$$R_{imt}=\beta_0+\beta_1 D_iT_t+\beta_2 T_t+{{{\mathbf{V}}}_{imt}} \alpha +\lambda_i+\epsilon_{imt}$$
(1)

where i denotes cell, m municipality, and t time period (the pre-treatment or treatment decade). The \(R_{imt}\) outcome measures proportional forest recovery, calculated as net forest recovery in period t divided by the initial non-forested area at the beginning of period t, for unit i in municipality m. \(D_{i}\) is a measurement of treatment (either binary for any part of a unit enrolled in restoration, or the proportion of the cell overlapping with a restoration parcel), and Tt is a dummy, which equals 1 if an observation is in the treatment period. \({{\bf{V}}}_{imt}\) includes temperature and precipitation as the time-varying controls. Cell fixed effects are given by λi. Cell fixed effects eliminate the influence of characteristics that do not change over time. These include important determinants of land productivity like slope, elevation, and bioclimate. \(\epsilon_{imt}\) denotes the residual. β1 estimates the treatment effect on the treated group, β2 estimates the effect of the time trend.

Standard errors are clustered at the municipality level, allowing cells to be spatially correlated within municipality and across time. We cluster standard errors because it is possible that a single landowner may manage multiple restoration parcels and PACTO implementers may work on several nearby parcels. To check this assumption, we downloaded available property boundaries for the state of Bahia. Not all PACTO parcels could be associated with a identifiable property, but for those that could, on average there were 3-4 parcels per property. Clustering accommodates the possibility of multiple parcels by allowing for correlation of the errors within the clustering unit. Therefore, clustering at the municipality level is a more conservative approach than clustering at the property level. Further, we cannot cluster at the property level since counterfactual units are not grouped by property.

In some specifications we set the fixed effect at the level of the municipality. In this case, Eq. (2) is used for estimation:

$$R_{imt}=\beta_0+\beta_1 D_iT_t+\beta_2 T_t+\beta_3 D_i+{{{\mathbf{V}}}_{imt}}\alpha+\gamma_m+\epsilon_{imt}$$
(2)

The interpretation of β1 and β2 are the same as above, but now β3 estimates the difference between treated and untreated groups in the pre-treatment period. In the case of municipality fixed effects (γm) we also include a number of time-invariant spatial controls in the term (\({{\bf{V}}}_{imt}\)) which in the first specification would have been absorbed by the cell fixed effect. These include elevation, slope, distance to nearest city, population of that city, land surface, bioclimate and geology categories, as well as the precipitation and temperature as above.

Stakeholder interviews

In addition to the quantitative analysis, we conducted semi-structured interviews with stakeholders engaged at different levels of the PACTO initiative (see Supplementary Discussion and Supplementary Note 1). The questions included in the interviews were driven by a theoretical framework. The theoretical framework considers the incentives that landowners face as they are choosing how to allocate their land, which are determined by the private benefits and costs of restoration combined with institutional enforcement (Supplementary Discussion). The framework and the interviews contextualize our statistical results, partly explain landowners’ restoration motivations, and help identify mechanisms for heterogeneity in treatment effects.

The purpose of the field interviews was to discuss benefits and costs of restoration with stakeholders in order to understand key barriers or facilitating conditions for restoration and to understand the potential role of PACTO in increasing restoration activity. The leadership of PACTO was both interviewed and helped researchers make contact with other restoration stakeholders. Eleven individuals were interviewed, sometimes jointly, over nine separate meetings in Bahia, Brazil between late-August and mid-September of 2023. All interviews, save one, were conducted in person. Interviewees consisted of two government stakeholders, two NGO stakeholders, one company representative, two landowner consultants, and four landowners. All interviewees are directly associated with forest restoration and PACTO (Fig. 2). Participant identities were anonymized as per IRB regulations for research involving human subjects.

Conversations were semi-structured and took place through an interpreter. They were recorded to be able to verify responses after the interview ended. Subjects included institutional information, experience with restoration, benefits and costs of restoration, legal structures related to land use, and opinions regarding barriers to restoration in Brazil. Government officials, individual producers, and large companies received different variations of the questions. The interview protocol is in Supplementary Note 1. Interview results also informed our choice of matching covariates by identifying key context-specific factors likely to facilitate or be barriers to landowner adoption of restoration projects. Finally, they led to the generation of testable hypotheses about heterogeneity in program effectiveness.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.