Collaborative management partnerships strongly decreased deforestation in the most at-risk protected areas in Africa since 2000

Significance

Protected areas (PAs) are vital for nature conservation but often fall short in Sub-Saharan Africa due to funding and institutional challenges. Since the 2000s, collaborative management partnerships (CMPs) have emerged to tackle these issues. As of 2023, we identified 127 CMPs across 16 countries. Our counterfactual analysis reveals that tree cover loss, a proxy for habitat protection, was, on average, 55% lower in CMP-managed PAs compared to similar PAs without CMPs over 22 years. In areas with low historical human impacts, CMPs had no effect. However, in areas with high historical anthropogenic disturbance, CMPs decrease tree cover loss by 66%. This shows that PAs can be effective at preserving natural habitats, and that conservation interventions can have highly variable effects.

Abstract

Collaborative management partnerships (CMPs) between state wildlife authorities and nonprofit conservation organizations to manage protected areas (PAs) have been used increasingly across Sub-Saharan Africa since the 2000s. They aim to attract funding, build capacity, and increase the environmental effectiveness of PAs. Our study documents the rise of CMPs, examines their current extent, and measures their effectiveness in protecting habitats. We combine statistical matching and Before-After-Control-Intervention regressions to quantify the impact of CMPs, using tree cover loss as a proxy. We identify 127 CMPs located in 16 countries. CMPs are more often located in remote PAs, with habitats that are least threatened by human activity. Our results indicate that, on average, each year in a CMP results in an annual decrease in tree cover loss of about 55% compared to PAs without CMPs. Where initial anthropogenic pressure was low, we measure no effect. Where it was high, we see a 66% decrease in tree cover loss. This highly heterogeneous effect illustrates the importance of moving beyond average effect size when assessing conservation interventions, as well as the need for policy makers to invest public funds to protect the areas the most at risk.

Protected areas (PAs) are central policy instruments to conserve natural habitats, biodiversity, and carbon sinks (1–3). However, their effectiveness at protecting ecosystems is highly variable (4–6), and challenged by the structural lack of funding (7, 8), limited management capacities (9), weak institutions and governance (10, 11).^* In Sub-Saharan Africa, the home of 13% of global species richness and about 20% of global forest cover, the environmental performance of PAs has been limited (18–20).

Collaborative management partnerships (CMPs) between government bodies and nonprofit organizations have emerged as a solution to increase funding and management capacities of existing PAs (21–24). Three broad categories of CMPs can be distinguished along a spectrum of increased sharing and delegation of authority from governments to nongovernmental organizations (NGOs): financial–technical support, comanagement, and delegated management (25). In financial-technical arrangements (FT), NGOs provide financial support and technical advice without having a formal long-term role in governance or management decision making. FT arrangements have predominated across Sub-Saharan Africa from the 1960s to the 2000s (26) Comanagement and delegated management arrangements, which started in the late 1990s, grew significantly since the mid-2000s. They include a sharing of management and, often, governance of the PA between the NGO and the Government (24). The median funding for PAs with comanagement is 2.6 times higher than baseline State budgets. This number grows to 14.6 for delegated management arrangements (23). However, though CMPs appear to have positively contributed to biodiversity conservation outcomes (27), there is a dearth of hard data to demonstrate their additionality (28–30).

Here, we provide a quantitative evaluation of whether CMPs increase the effectiveness of PAs at conserving habitats, as proxied by satellite-measured tree cover loss (31), across all of Sub-Saharan Africa between 2001 and 2023. We specifically focus on the types of arrangement that delve the furthest in terms of cooperation and delegation and that are growing the fastest: comanagement and delegated CMPs.

Building on previous efforts (21, 24), we first map all existing CMPs across Sub-Saharan Africa. Our work notably includes Madagascar in this literature—a biodiversity hotspot that is home to 50% of the comanagement and delegated CMPs on the continent. Second, we use counterfactual statistical models to quantify the causal effect of establishing CMPs on the protection of natural habitats in terrestrial PAs. Because CMPs were conceived to primarily reinforce the management of existing PAs, and not to create new PAs, our empirical strategy compares the evolution of tree cover loss between PAs with and without CMPs. These PAs without CMPs, which we include in the control group, may receive some FT support from donors and NGOs. However, their management remains fully under the responsibility of Governments.

Understanding whether CMPs are effective at better protecting natural habitats—a necessary condition to halt the collapse of biodiversity (32)—is a central question to orient conservation planning in low-income countries, including across the African continent (33). As governments, NGOs, and donors are considering renewing existing agreements and signing new ones, there is a growing necessity to understand the conservation benefits of these arrangements (27). Beyond the nature-conservation sphere, our results contribute to understanding the effect of Public–Private Partnerships more broadly, which is especially relevant given their proliferation since the 1980s in low- and middle-income countries (34).

1. Results

1.1. CMPs across Africa.

We first document the evolution of the spatial distribution of comanagement and delegation agreements, from their beginning to the end of 2023 (Fig. 1A).

Fig. 1.

CMPs for PA across time and space. (A) Temporal evolution of PAs with CMPs. Left: Share of PAs with CMP among all the PAs from the same 16 countries that have a similar designation. Center: Percentage of area of PAs with CMPs compared to the area of all PAs in the same 16 countries that have a similar designation. Right: Share of countries with PAs with CMPs among the 16 countries that have CMPs by the end of 2023. (B) Spatial distribution of current PAs with CMPs and possible control PAs. PAs with CMPs in Sub-Saharan Africa are color coded in yellow to purple based on the year of establishment of the CMPs. PAs without CMPs, that have the same designation of the ones with CMPs and located in the same 16 countries, are represented in gray.

The first comanagement arrangement documented in our database, Kasanka National Park (NP) in Zambia, was established in 1990 between the Department of National Parks and Wildlife Service and a British-registered NGO, Kasanka Trust. The early 2000s saw further growth of CMPs with the establishment of agreements in Mozambique (Cabo de São Sebastião in 2000), Tanzania (Grumeti and Ikogorongo Game Reserves in 2002), Malawi (Mejete Wildlife Reserve in 2003), and the Democratic Republic of the Congo (the Garamba landscape in 2005, comprising a NP and three hunting reserves). The number of CMPs continued to increase steadily, reaching 25 PAs with CMPs in 2014.

In 2015, 64 new CMPs were established. Sixty-three of them were located in Madagascar where CMPs facilitated the establishment of new PAs. This contrasts with the rest of the continent where CMPs contributed to the reinforcement of existing PAs.

By the end of 2023, 127 PAs located in 16 Sub-Saharan African countries were under a comanagement or delegated CMP model (Fig. 1B and SI Appendix, Table S1). Out of these, 5 CMPs were in Marine PAs and 122 in terrestrial PAs. The 122 terrestrial PAs with CMPs represent 5% of all similarly designated PAs within these 16 countries and 35% of the area under protection. This surface represents 979,387 km².

All but one PA with a CMP were initially managed by national authorities. Most were designated as NPs of IUCN category I or II. Only one of the PAs with CMP is registered as a community reserve (Lac Télé in the Republic of Congo).

Forty-eight partner organizations participate in co- and delegated management agreements. Forty-six of them are not-for-profit organizations. Two are for-profit organizations and both operate in Madagascar.

Of the 48 partner organizations, 27 international NGOs manage 96 CMPs, and 21 national organizations manage 31 CMPs. The African Parks Network, for which CMPs are their raison d’être, has the largest portfolio with 26 CMPs. Other international conservation NGOs, such as the Wildlife Conservation Society (19 CMPs), not only manage PAs through CMPs in these 16 countries but also provide FT support and manage integrated conservation and development programs in other countries. CMPs run by national organizations are mostly found in Madagascar.

1.2. Differences in Location between PAs with and without Collaborative Management Agreements.

First, we compare countries that have CMPs with those that do not. Comanagement and delegated CMPs are found in 16 out of 48 countries in Sub-Saharan Africa: Angola (1 CMP), Benin (3), the Central African Republic (6), the Democratic Republic of Congo (11), the Republic of Congo (5), Madagascar (63), Mozambique (9), Malawi (4), Niger (1), Nigeria (1), Rwanda (2), South Sudan (5), Chad (6), Tanzania (2), Zambia (6), and Zimbabwe (2). These CMPs are predominantly located in central Africa and parts of southern and east Africa. However, few CMPs are established in West Africa and several countries that are prominent for biodiversity conservation do not have any comanagement or delegated CMPs, including some where FT support has long been implemented. This applies to Cameroon, Kenya, South Africa, and to a lesser extent Tanzania and Namibia which only host three CMPs in total.

In SI Appendix, Fig. S1, we show that CMPs are located in countries with a lower GDP per capita than countries without CMPs ($1,481 vs. $3,789, $P$-value $=$ 0.007). Furthermore, CMPs are located in countries with a lower basic state function score (SI Appendix, Fig. S1) than countries without CMPs—a variable that measures the ability of states to keep a monopoly on the use of force and to provide basic public services (35). This difference is, however, not statistically significant [5.5 vs. 6.25 on a 1 to 10 scale (most functioning), $P$-value $=$ 0.25].

Second, we focus on the 16 countries that have CMPs and explore the differences between PAs with and without CMPs. Within these countries, we identify 2,572 PAs without CMPs that share a common designation status with the PAs with CMPs. PAs with CMPs are, on average, significantly larger than those without CMPs (5,539 km² vs. 515 km², $P$-value $<$ 0.001, SI Appendix, Table S2). This difference is partly explained by the presence of some particularly big PAs among those with CMPs (e.g., Termit et Tin-Toumma in Niger and Ouadi Rimé-Ouadi Achim in Chad have an area of over 90,000 km² and 80,000km², respectively).

Still within these 16 countries, we divide PAs into a regular grid of 0.05 $\times$ 0.05 degrees (approximately 5 $\times$ 5 km, Section 1) and compare the ecological characteristics and the initial anthropogenic pressures between gridcells from PAs with CMPs and gridcells from PAs without CMPs (gray triangles in Fig. 2 and SI Appendix, Table S3). Gridcells from PAs with CMPs are more likely to be tropical and subtropical moist broadleaf forests, while grasslands and savannas are underrepresented in CMPs. Regarding anthropogenic pressures, gridcells belonging to PAs with CMPs are on average significantly further away from cities and villages and have less initial population than those without CMPs. This suggests that PAs with CMPs are located in areas that face, on average, less anthropogenic pressure than other PAs. This is coherent with the fact that in 2000, gridcells from PAs with CMPs had a larger forest cover and less croplands than those without CMPs. These results suggest that the selection of CMPs is subject to location bias.

Fig. 2.

Differences in observable characteristics between gridcells from PAs with CMPs and gridcells from PAs without CMPs. In this figure, data come from the final database used for the impact evaluation. Starting from the original database, we dropped cells that had no pixel of initial tree cover (using a 25% tree cover threshold), PAs with an area smaller than 25 km² (corresponding to the area of a gridcell in the impact evaluation), and control PAs that had a different designation status from PAs with CMPs. The final database contains 94 PAs with CMP and 991 PAs without CMPs—corresponding to 19,707 gridcells before matching. After matching, 7,224 gridcells were kept.

1.3. Impact of Collaborative Management Agreements on Tree Cover Loss.

Finally, we use our gridcell database to analyze whether the establishment of CMPs decreased annual tree cover loss measured through satellite images between 2001 and 2023 (31).

We proceed in two steps to quantify possibly causal impacts of CMPs. First, we correct the location bias using statistical matching. We find an appropriate match for 62% of gridcells with CMPs (SI Appendix, Table S4). Nevertheless, no match could be found for gridcells from 15 PAs with CMPs (SI Appendix, Fig. S2). After matching, no difference in observable characteristics remains between gridcells in PAs with and without CMPs based on commonly used thresholds of similarities (36) (Fig. 2, green circles and SI Appendix, Table S4). Second, we estimate a Before-After-Control-Intervention event-study model, allowing for staggered entry in the treatment, heterogeneous treatment effects, and country and biome specific time dynamics (37). Our main specification uses the percentage of tree cover loss as a dependent variable. We construct it by dividing, for each year $t$ between 2001 and 2023, the number of hectares of tree cover loss during year $t$ by the number of hectares of tree cover at the beginning of year $t$. Baseline results are displayed in Panel A of Fig. 3, where we compare the difference in tree cover loss between treated and control gridcells, both before and after the creation of CMPs.

Fig. 3.

The formal establishment of a CMP is followed by a decrease in deforestation compared to similar PAs without CMPs, notably in cells with an high initial level of anthropogenic pressure. (A) Baseline estimate of the effect of CMP. (B) Heterogeneity of the effect between cells facing initially low anthropogenic pressures (green) and high anthropogenic pressures (purple). Dots represent the point estimate of the impact (difference in the tree cover loss between treated and control units) for each year ranging from 10 y before and after the establishment of a comanagement or delegated agreement to 16 y after its establishment. Lines represent the 95% CI of the estimates. The effects are detailed in SI Appendix, Tables S5 (baseline results), S6 (low anthropogenic pressure areas), and S7 (high anthropogenic pressure areas). The model is estimated using the procedure developed by de Chaisemartin and d’Haultfœuille (37).

In the 10 y before the establishment of CMPs, the difference in tree cover loss is small. However, in four of the five years before the establishment of a CMP, tree cover loss is already between 3 and 11 percentage points lower in gridcells that will become CMPs than in similar gridcells that will not become CMPs.

After the establishment of CMPs, tree cover loss in PAs with CMPs becomes lower than in matched gridcells from PAs without CMPs. The effect is gradual and increases with the number of years spent under a CMP. In the first four years of implementation, the magnitude of the difference in tree cover loss between gridcells with and without CMPs is similar to the one that existed in the last few years prior the establishment of the CMP. After five years, this difference in tree cover loss increases. Eventually, the point estimates become less accurate as sample size decreases.

The size of the impact is notable. The aggregated average treatment effect (ATE) is $-$0.14 percentage points each year (95% CI $=$ [$-$0.276; $-$0.009]). In comparison, the average tree cover loss of the matched control observations is $-$0.25 percentage points. Hence, each year spent under CMPs decreased tree cover loss by 56% (Fig. 3).

1.4. Heterogeneity.

We explore how the effect of CMPs varies across different contexts. The previous section highlighted the existence of a strong location bias of CMPs toward areas with an initial lower anthropogenic pressure. We construct an index of anthropogenic pressures for each gridcell using the covariates from the matching process. We then compare the effect of CMPs on tree cover loss for gridcells below and above the median value of our of anthropogenic pressure index.

Where anthropogenic pressure is low, so is tree cover loss. The average yearly tree cover loss rate in control gridcells is 0.10 percentage points in low anthropogenic pressure areas. This is over 50% lower than in all control PAs (0.25 percentage points). Consequently, the additional effect of CMPs is null (Panel B in Fig. 3 and SI Appendix, Table S6).

Where anthropogenic pressure is high, so is tree cover loss. The average yearly tree cover loss rate in control gridcells is 0.39 percentage points in high anthropogenic pressure areas. This is three times higher than in low anthropogenic pressure areas, and 50% higher than in all control PAs. In these highly pressured areas, CMPs decreased annual tree cover loss by 0.26 percentage points on average (SI Appendix, Table S7). This corresponds to a 66% decrease in annual tree cover loss.

The heterogeneity of the impact of CMPs for the individual variables that we use to construct the anthropogenic index is explored in SI Appendix, Figs. S3–S5. The effect of CMPs is particularly strong in gridcells located at lower elevation, in those with lower tree cover in 2000, more croplands, higher population, closer to villages and cities, and in flatter areas.

1.5. Robustness.

We present robustness checks for our baseline estimate. Event studies are displayed in SI Appendix, Figs. S6–S21.

First, we show that being two times stricter (caliper $=$ 0.25SD) or two times more permissive (caliper $=$ 1SD) in the selection of matched controlled gridcells does not change the baseline results. When matching is stricter (SI Appendix, Fig. S6), the aggregated ATE ($-$0.394; 95% CI $=$ [$-$0.803; 0.016]) becomes larger relatively to the tree cover loss in controlled gridcells (0.18 percentage point). When our matching is more permissive (SI Appendix, Fig. S7), the aggregated ATE ($-$0.1; 95% CI $=$ [$-$0.202;0.002]) becomes relatively smaller. We also show that results are robust when conducting matching with replacement, meaning that a same gridcell can be the control gridcell for several treated ones (SI Appendix, Fig. S8), when forcing pairs to belong to the same country (SI Appendix, Fig. S9), when dropping trends from the BACI regression (SI Appendix, Fig. S10), when calculating annual tree cover loss rate using the year 2000 as a reference instead of a rolling rate (SI Appendix, Fig. S11), and when weighting gridcells by the areas of the PAs (SI Appendix, Fig. S12). In this last case, deforestation before the establishment of CMPs is quite different in controlled and future-treated observations, which complicates the causal interpretation of the results.

1.5.1. Madagascar.

CMPs in Madagascar were used to expand the network of PAs and not to strengthen existing PAs. Because of the singularity of this context, the effect of CMPs in Madagascar can differ from the rest of the continent, and taking PAs as controls might be less appropriate. We estimate our baseline model by dropping observations from Madagascar (SI Appendix, Fig. S13). The results become relatively larger ($-$0.146; 95% CI $=$ [$-$0.285;$-$0.006], Mean untreated $=$ 0.161).

1.5.2. Alternative estimator.

We re-estimate our baseline model using the econometric approach developed by Sun and Abraham (38) which is conceptually similar to the method of de Chaisemartin and d’Haultfœuille (37) mobilized in our paper. The results obtained with this alternative estimator are extremely close to those from our baseline estimate (ATT $=$ $-$0.17 (95% CI $=$ [$-$0.315;$-$0.024], Mean untreated $=$ 0.253, SI Appendix, Fig. S14).

1.5.3. Alternative grid size.

In spatial analysis, the choice of the geographic unit, which is often arbitrary, can impact the results (39). We test the robustness of our findings using a coarser spatial grid (10 km $\times$ 10 km). We are able to construct a control group that is extremely similar to treated observations (SI Appendix, Fig. S15). The effect size from the postmatching BACI regression is significant, and larger than with our main grid ($-$100% of tree cover loss on average). Estimates are, however, less precisely estimated (SI Appendix, Fig. S16).

In addition, we assess the robustness of the results by using a finer grid of 2.5 $\times$ 2.5 km. Here too, we are able to construct a convincing control group (SI Appendix, Fig. S17). We also find that CMPs are associated with lower tree cover loss (SI Appendix, Fig. S18). However, the pretrends are now entirely convincing in this analysis, likely because, at this spatial scale, 40% of cells experienced no deforestation over two decades—making an ordinary least square model less suitable for these data.

1.5.4. Spatial autocorrelation.

Our main estimate accounts for spatial autocorrelation by sampling 1/3 of grid cells. As a robustness test, we use all available grid cells instead of sampling them, and adjust SE by clustering them at different spatial resolution: a 50 $\times$ 50 km grid (SI Appendix, Fig. S19), a 100 $\times$ 100 km grid (SI Appendix, Fig. S20) and at the park level (SI Appendix, Fig. S21). A 50 $\times$ 50 km grid provides smaller SE than our main model, hence increasing the statistical significance of the results. A 100$\times$100 km grid provides SE that are comparable to the ones from the main estimate, which leave the results unchanged. Finally, clustering the SE at the park level leads to a decrease in the precision of the SE during the latest years of the treatment for which we have a small number of clusters. The aggregated impact becomes significant only at a 10% level.

2. Discussion

PAs, a cornerstone instrument of conservation policies, can deliver important environmental outcomes (6, 40). However, their ability to do so in underresourced, highly threatened areas of Sub-Saharan Africa is often limited (16, 17). CMPs are ambitious approaches to inverse the dynamic of habitat destruction in African PAs. Their number has been increasing rapidly and they are attracting important funding. Our results show that these CMPs have delivered a large decrease in tree cover loss (56% per year on average).

Our results show that tree cover loss starts to decrease in the four years preceding the creation of CMPs. Although the magnitude of this decrease is small, it is statistically significant. The presence of some anticipation effect is a likely interpretation of this result. Indeed, the formal establishment of a CMP is the end-result of a process that takes several years. The CMP toolkit, coordinated by the World Bank (24), presents nine detailed case studies in which the CMPs took between 1 and 4 y to be finalized. As a consequence, the decrease in tree cover loss that we observe in the years prior to the official CMP creation date could reflect that some activities already started in the PAs prior to the official establishment of the CMP.

Furthermore, the literature has for long established that PAs are often located in areas that are more remote than unprotected areas (13). In these remote areas, pressures on habitats are generally low and the presence of PAs does not result in better protection (5). Our results highlight that gridcells with CMPs are subject to an even stronger location bias, as CMPs tend to be primarily established among the most remote PAs. Some of the CMPs are even too remote for us to be able to pair them with similar PAs without CMPs despite considering a large pool of over 1,000 control PAs. Indeed, while the average gridcell of a control PA is located 12 h away from the closest city of more than 500,000 inhabitants, several gridcells within Salonga, Lomami, and Chinko NP (three CMPs in the D.R. Congo and the Central African Republic) are located as far as 46 h away. Similarly, the average control gridcell is located 5 h from the nearest village of 5,000 or more inhabitants while several parts of Lomami and Maiko NP are located over 12 h from such a village. As a consequence, we show that the establishment of CMPs that faced below-median anthropogenic pressure but for which we were still able to find a correct match, led to a precisely estimated null effect on tree cover loss.

Not all CMPs are located in remote areas. Virunga NP in the Democratic Republic of Congo is a good illustration, as over five million people live directly around the NP. The travel time between some gridcells of Virunga NP and the city of Goma (one of the fastest growing cities on the continent with already over 1.5 million inhabitants) is 10 min. Likewise, some gridcells from Kahuzi-Biega NP, also in the Democratic Republic of Congo, are located 42 min away from the large city of Bukavu.

While protecting PAs that face high anthropogenic pressure is challenging, our results highlight that CMPs have a positive and large effect on habitat conservation specifically in the gridcells that faced above-median anthropogenic pressure. This suggests that the positive impact of CMPs may even become greater if NGOs are willing to take on the management of these tough contexts—and get donors to understand the higher associated risk of not succeeding. More generally, our results highlight that the average effect of conservation policies can hide large heterogeneity, making the average a poorly informative indicator for policy makers.

In addition, our results illustrate that it can take several years before conservation policies turn into large measurable effects. In our study, the effect of CMPs was limited over the first five years of their establishment. This effect slowly increased from years 6 to 10 and increased even after 10 y. This result raises two important points. First, it is key that wildlife authorities engage in long-term partnerships. Second, long-term rather than short-term evaluations are crucial to assess accurately the effect of conservation policies.

We do not mean to present CMPs as being the only policy option or as being uniformly effective. Examples of CMPs that do not appear to be delivering effective management exist (23). These appear to be due to a combination of poorly designed partnership agreements and lack of proficiency on the part of the NGO and wildlife authorities (25). Thus, it is important for governments to select partners wisely and to monitor outputs such that steps can be taken in the event of nonperformance. Furthermore, the importance and effectiveness of management of natural habitats by local communities and indigenous peoples is well evidenced (41). We do not see CMPs as an alternative, but rather as a complement, to community involvement. This can take the form of community conservancies where legislation recognizes community land rights.

Our research opens lines for further work. While our paper analyzes the effect of CMPs on habitat conservation, PAs actually face a confluence of threats going beyond deforestation, such as wild meat hunting and high value poaching (42–45). Due to the large demand for wild game meat across the continent and the high value of the commodity, it is likely that the threat from poaching is insidious and more difficult to control than that of deforestation. Thus, while well-financed and well-managed CMPs could effectively limit both deforestation and wild game meat poaching, less well-managed and financed CMPs may deliver on limiting deforestation without effectively controlling wild game meat poaching. More research into the impact of CMPs on poaching and other threats is therefore needed. It would also be important to study the effect of CMPs on species abundance and diversity (46, 47), and on the ability to safeguard critical ecosystem services.

Finally, it would be interesting to analyze whether the encouraging outcomes of CMPs in terms of habitat protection are achieved in synergy with or at the expense of local socioeconomic development. Underfinanced and understaffed PAs can be either an opportunity or a burden for local communities (48). Some community members may benefit from conducting unsustainable activities within PAs that are vital for their subsistence (49) and enforcing laws may negatively impact these households (50). Furthermore, underfinanced and understaffed PAs are also prone to larger-scale natural resource trafficking, sometimes coordinated by armed groups (51), which has direct negative effects on local communities. Many CMPs have invested in supporting alternative livelihoods, developing social infrastructure, and promoting tourism and security, which can have important knock-on effects for local economies (23). A critical issue is whether CMPs or state-run models generate greater resources for engaging and benefiting communities. So far, this question has mostly been explored through measuring attitudes and perceptions, with mixed results (52–54). A recent paper provides a quantitative assessment using satellite-derived measures of poverty for the CMP of one organization, African Parks Network (30). Further research is required to better understand the cost–benefit of CMPs on the well-being of local communities and on macroeconomic outcomes (55).

3. Materials and Methods

3.1. CMP Database.

We built on previous analyses to create an updated database of PAs with CMP management (21, 24, 56). We categorized PAs as falling under a “CMP” when they are jointly managed by a State authority and a private partner, or when their management has been exclusively delegated to a private partner by legal mandate. Following Fitzgerald (24), we did not consider PAs that receive only a Financial or Technical (FT) support from a private partner as a CMP. This is because FT support represents a more heterogeneous set of approaches, and it is hard to distinguish different levels of FT support between PAs. Most PAs in our database previously received FT support prior the establishment of a comanaged or delegated CMP.

We updated existing lists using expert knowledge to have an exhaustive list of CMPs as of December 2023. This update fills an important gap in terms of spatial coverage, notably for Madagascar that was absent or underreported in previous assessments. We georeferenced our database of CMPs using primarily the World Database of Protected Areas (57) (version: April 2024). When PAs with CMPs were missing in the WDPA or when their shapefiles were imprecise, we used shapefiles provided by local or national authorities. This was the case of Chinko in the Central African Republic, Gorongosa NP in Mozambique, and all PAs in Madagascar for which the precision of shapefiles in the WDPA is debated (58, 59). Our final database results in 127 PAs with CMPs. For each of them, we determined the year of establishment of the first agreement. The database is available in SI Appendix, Table S1 and in an open repository in a tabular format (linked to be provided when published).

3.2. Impact Evaluation.

We quantified the environmental impact of CMPs by examining changes in tree cover loss over 23 y compared to a control group comprising similar PAs but without CMPs. Tree cover loss is measured annually from 2000 to 2023 using version 1.11 of the Hansen (31) dataset. A 30 m $\times$ 30 m pixel is defined as “lost” in Hansen when at least 50% of its initial tree cover is lost. Given Hansen’s limitations in accurately capturing data in arid regions, we concentrated on pixels that had a minimum canopy cover of 25% in the year 2000. We tested the robustness of the estimates when using a more conservative threshold of 50% of canopy cover.

3.2.1. Inclusion criteria for CMPs in the impact evaluation.

Out of the 127 PAs with CMPs, we excluded from the impact evaluation a) five PAs that are primarily classified as marine PAs in the WDPA because tree cover loss is not a relevant criteria, b) two PAs for which a CMP was signed before or in 2000 in order to keep only PAs for which we have tree cover loss observations before and after the intervention, c) 17 PAs that have an area below 25 km² (all in Madagascar, the size of our unit of observation), and d) four PAs that had less than 1 ha of forest in 2000 using the 25% threshold.

3.2.2. Selection of possible control PAs.

Next, we constructed a control group of PAs without CMPs. Using the WDPA and the Madagascar shapefile of PAs, we selected terrestrial PAs without CMPs that are located in the same countries as PAs with CMPs, that are larger than 25 km², that had more than one hectare of forest in 2000, and that have a similar designation as PAs with CMPs (e.g., National Park, Game Reserve, Hunting Reserve). We obtained a total of 991 PAs that we consider as possible controls.

3.2.3. Grid.

PAs included in the study have heterogeneous areas (min $=$ 26 km², max $=$ 45,824 km², sd $=$ 6,310 km², SI Appendix, Table S2. To obtain observations that are comparable, we followed the literature and subdivided PAs into a regular grid (60). Working at the cell and not a 30 $\times$ 30 m pixel helps limit nonclassical measurement error which can be found in remote sensing data for deforestation (61).

We created a grid of approximately 25 km² over each PA (5 $\times$ 5 km). We clipped the grid to the PA borders and kept pixels that are inside the PA. As a consequence, grid cells that are bordering the PA delimitation can have an area smaller than 25 km². To limit disparities in terms of areas between grid cells, we combined adjacent gridcells that have an area smaller than 12.5 km²; and excluded cells that have an area smaller than 12.5 km² and which could not be combined with another adjacent cell. Finally, we randomly sampled 1/3 of these grids to limit spatial autocorrelation (62). We summarized this procedure in SI Appendix, Fig. S14. Our final database contains 17,953 cells in which we follow deforestation annually between 2001 and 2023.

3.2.4. Outcome variable and covariates.

We measure tree cover loss using the Global Forest Watch (GFW) data v1.10 (31). The data are available on a yearly basis from 2001 to 2023 and are produced with a consistent methodology at a global level. We extracted tree cover loss data for each grid cell using the gfcanalysis package in R (63). We calculated the rolling deforestation rate, i.e., the annual share of tree cover that is lost, as our main outcome of interest. The GFW data have been shown to underreport deforestation in certain contexts (61) and to be less precise in areas with low tree cover densities (64). It has the advantage of not being specific to certain biomes (e.g., ref. 65).

We constructed a set of time-invariant covariates using other spatial data. For each cell, we calculated the average elevation and slope using AWS Open Data Terrain Tiles (66) and the elevatr R package (67), the dominant biome (68), population in 2000 (69), travel time to the nearest village or town of 5,000 or more inhabitants and travel time to the nearest city of 500,000 or more inhabitants (70), and the earliest estimate of the surface of croplands available for our study period, a composite image between 2000 and 2003 (71). We also computed the Euclidean distance between the centroid of each cell and the border of the PA, and the area of each PA. We extracted precipitation data using the CHIRPS database (72) and calculated a yearly Standard Precipitation Index.

3.2.5. Statistical model.

In the absence of a true experimental framework, we combine prematching techniques to ensure comparability between treated and comparison PAs at baseline, along with panel regression analysis incorporating fixed effects. This methodology, which combines panel regression with preprocessing matching techniques, has been demonstrated to approximate a true experiment more effectively (73, 74). It has been employed in various prior studies examining conservation policies (e.g., ref. 60).

We used statistical matching to correct for the possible difference in time-invariant observable characteristics between PAs with and without CMPs. We implemented a nearest-neighbor matching without replacement and using a Mahalanobis distance in matchit (75). We set a caliper of 0.5 SD to keep acceptable pairs. We matched on the following variables: the initial forest cover in 2000, the Moran I statistic of forest cover in 2000, the distance between the cell and the boarder of the PA, population in 2000 in the gridcell, the distance between the gridcell and the nearest village, the distance between the gridcell and the nearest large city, the surface of croplands in the gridcell in 2000–2003, the average elevation, and average slope. We imposed that pairs of treated and control observations belong to the same biome,^† but not necessarily the same countries as forcing pairs to belong to the same country would decrease by two the number of observations for which we can find an appropriate match (SI Appendix, Fig. S9).

Second, we estimated the impact of CMPs on annual tree cover loss using a difference-in-difference (DID) model, which are also known as before-after-control-intervention (BACI) (76), on the matched dataset. When more than two time periods are available, such as in our case, DID models are usually estimated using a two-way fixed effects (TWFE) model. We estimated nonnormalized event study a model of the form:

$$\begin{array}{cc}De{f}_{i,t}& ={\alpha }_i+{\gamma }_t+{\displaystyle \sum _{k=-10}^{-2}{\beta }^k·CM{P}_{i,t}^k}+{\displaystyle \sum _{k=0}^{16}{\beta }^k·CM{P}_{i,t}^k}\\ & +{\displaystyle {\sum }_c{\tau }_1·Countr{y}_c\times Yea{r}_t}+{\displaystyle {\sum }_b{\tau }_2·Biom{e}_b\times Yea{r}_t+{ϵ}_{i,t},}\end{array}$$

where $De{f}_{i,t}$ is the deforestation rate in gridcell $i$ during year $t$, ${\alpha }_i$ is a gridcell fixed-effect, ${\gamma }_t$ is a year fixed-effect, $CM{P}_{i,t}^k$ is an indicator for unit $i$ being $k$ periods away from initial treatment at year $t$. The first summation captures the effect time periods leading up to the treatment (placebo) and the second summation captures the time period following treatment (effect). We included country-year and biome-year trends to capture more finely time-varying unobserved factors.

Recent advances have demonstrated that TWFE can assign negative weights to observations when the treatment was staggered and that the treatment effect was heterogeneous [e.g., varying through time, (77)]. These negative weights lead to biased estimates of the average treatment effect of an intervention. In particular, the classic TWFE approach provides contaminated estimates of time-dynamic effects (37).

New methods were developed to correct negative weights and provide unbiased estimates of average treatment effects. In particular, the approach of De Chaisemartin and d’Haultfoeuille (37) consists of computing cohort-specific weights, and aggregates treated cohort $\times$ relative time to treatment dummies into an unbiased average treatment effect (ATE) for each period and the entire posttreatment period. This approach relies on a weakened parallel trend assumption compared to the canonical BACI model. We rely on their procedure and package to estimate treatment effect (78). We report nonnormalized results. All coefficients are long-difference compared to the outcome from one year prior to the establishment of CMP.

3.2.6. Heterogeneity.

We test the heterogeneity of the effect of CMPs depending on whether the gridcell is exposed to low or high initial anthropogenic pressures. To that end, we construct an index for anthropogenic pressure that add or subtract eight standardized continuous variables:

$$\begin{array}{cc}\hfill \text{Index}& =\text{Population}-\text{travel time to the nearest city}-\text{travel time to the}\hfill \\ \hfill & \text{nearest village}-\text{initial forest cover}+\text{initial cropland cover}\hfill \\ \hfill & -\text{slope}-\text{elevation}-\text{distance to the boarder of the PA}\hfill \end{array}$$

Next, we determine whether a gridcell is above the median of the index in the matched dataset (high anthropogenic pressure) or below this median (low anthropogenic pressure).

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

All analyses were conducted using R software (version 4.2.3). Data and codes (Database of Collaborative Management Partnerships in Africa) are available on https://zenodo.org/records/14195855 (79). All other data are included in the manuscript and/or SI Appendix.