Implications of heterogeneous impacts of protected areas on deforestation and poverty

Abstract

Protected areas are a popular policy instrument in the global fight against loss of biodiversity and ecosystem services. However, the effectiveness of protected areas in preventing deforestation, and their impacts on poverty, are not well understood. Recent studies have found that Bolivia's protected-area system, on average, reduced deforestation and poverty. We implement several non-parametric and semi-parametric econometric estimators to characterize the heterogeneity in Bolivia's protected-area impacts on joint deforestation and poverty outcomes across a number of socioeconomic and biophysical moderators. Like previous studies from Costa Rica and Thailand, we find that Bolivia's protected areas are not associated with poverty traps. Our results also indicate that protection did not have a differential impact on indigenous populations. However, results from new multidimensional non-parametric estimators provide evidence that the biophysical characteristics associated with the greatest avoided deforestation are the characteristics associated with the potential for poverty exacerbation from protection. We demonstrate that these results would not be identified using the methods implemented in previous studies. Thus, this study provides valuable practical information on the impacts of Bolivia's protected areas for conservation practitioners and demonstrates methods that are likely to be valuable to researchers interested in better understanding the heterogeneity in conservation impacts.

1. Introduction

Protected areas are a popular policy instrument in the global fight against loss of biodiversity and ecosystem services. The abundance and importance of protected areas have spawned interest in understanding their performance from conservationists, economists and policy makers alike. Despite the ubiquity of protected-area systems, scientific evidence related to their environmental and social impacts is weak [1,2]. The dearth of empirical evidence hinders resolution of an important international debate: do ecosystem conservation goals conflict with poverty alleviation goals in developing nations [2–4]?

Although the evidence base on the socioeconomic and environmental impacts of protected areas is still at a fledgling stage, significant strides have been made in recent years. Most studies tend to either focus on the impacts of protected areas on socioeconomic outcomes (e.g. [5–8]) or environmental outcomes (e.g. [9–13]). However, relatively few studies examine the joint environmental and socioeconomic impacts of protected areas (exceptions include [14]), and only a handful address the heterogeneity in the joint impacts of protected areas (e.g. [15–17]).

In this study, we use quasi-experimental methods to examine the heterogeneous impacts of Bolivia's protected-area system on deforestation and poverty. In doing so, we build from two previous studies which found that Bolivia's protected areas avoided a modest amount of deforestation [18] and reduced poverty in surrounding communities [6].¹ The motivation to move beyond the average impacts estimated in these previous studies is twofold. First, if treatment effects are not constant in the population, then average treatment effects will mask potential variations in response to treatment from subgroups of the population. Therefore, in spite of the fact that the environmental and socioeconomic impacts of Bolivia's protected areas have been modest, on average, there may be large impacts within certain areas or subgroups [19]. Second, we echo the arguments of Ferraro & Hanauer [20] that evidence-based policy design can greatly benefit from understanding the conditions under which conservation efforts have been most and least successful in the past. Such information will allow conservation planners to formulate what Manski [21] terms conditional empirical success (CES) rules. CES rules select interventions that maximize expected impacts based on observable covariates [21, p. 75]. In the context of ecosystem protection, CES rules are developed by first understanding the heterogeneous impacts of interventions conditional on observable biophysical and demographic characteristics. For example, a characteristic that is correlated with deforestation pressure, such as slope, might be used to spatially target the establishment of future protected areas. In this example, a planner interested in achieving both avoided deforestation and poverty reduction would want to understand the covariation of these two outcomes as a function of slope.

We estimate the impacts of Bolivia's protected areas that were established between 1991 (the year in which protected areas became recognized by law [6]) and 2000 on deforestation between 1991 and 2006, and poverty between 1992 and 2012. Our study provides several important contributions to both the understanding of the impacts of Bolivia's protected areas and to the empirical approaches through which researchers can estimate heterogeneity in the impacts of conservation policies. In terms of understanding the impacts of Bolivia's protected areas, we provide two important contributions. First, compared to [6], we make use of an updated (2012 as opposed to 2001) and more disaggregated geospatial socioeconomic dataset. This allows us to obtain more updated and precise estimates of the impacts of Bolivia's protected areas on poverty. Second, we use multiple estimation strategies to show that the impact of Bolivia's protected areas on deforestation and poverty has exhibited significant heterogeneity across multiple socioeconomic and biophysical moderators. From a methodological perspective, we extend the analyses suggested by Ferraro et al. [16] by using a multiple variable locally weighted scatter plot smoothing (i.e. three-dimensional LOESS) estimator that allows us to study the joint variation in conditional average treatment effects across multiple dimensions of biophysical characteristics. We show that the three-dimensional LOESS approach highlights areas of variation in the impact of protected areas that are not captured by simple two-dimensional LOESS or partial linear model (PLM) approaches. Therefore, our study provides valuable practical information on Bolivia for conservation planners and policy makers, and highlights new methods for researchers interested in obtaining a more accurate picture of the impacts of conservation policies.

2. Data

Previous studies indicated that the establishment of Bolivia's protected areas resulted in reductions in deforestation [18] and reduced poverty [6]. To estimate the heterogeneity in these impacts, we draw from those previous studies and employ two distinct datasets: the deforestation data and the socioeconomic data. We conduct all geographical information systems (GIS) calculations in ArcMap v. 10.x.²

(a). Deforestation data

We use data from [18] to analyse the impacts of Bolivia's protected areas on deforestation between 1991 and 2006. All data were obtained from Conservation International, Bolivia. The unit of analysis is a 30 m land parcel. To mitigate concerns of spatial dependence, we sample 20 000 parcels that were forested (more than 30% canopy cover) at baseline (i.e. prior to the establishment of protected areas, 1991). We define each parcel as forested or deforested in 2006 based on whether or not the parcel lies within the boundaries of the 2006 forest cover. Our outcome is calculated as the difference in forest cover status between 2006 and 1991. Thus, we have a binary outcome where a parcel that has been deforested receives a value of 1 and a parcel that remained forested receives a value of 0. Parcels are defined as protected if they fall within the boundaries of a protected area that was established prior to 2000. Control parcels were never protected. According to these definitions, we are left with 3780 protected and 16 220 unprotected (control) parcels.

To capture the determinants of deforestation pressure and the location of protected areas, we follow [16,18] and calculate the following covariates (all measured at baseline) for each parcel: distance to a major city, distance to nearest road, elevation, slope and distance to the edge of the forest (see table 1 and the electronic supplementary material for more detail).

Table 1.

Summary statistics for deforestation and socioeconomic data.

covariate	description	status	mean	median	s.d.	min	max
deforestation data
distance to major city	distance to nearest state capital city (km)	unprotected	216.94	203.75	106.21	0.26	593.28
		protected	278.57	227.49	163.28	24.60	596.39
distance to road	distance to nearest road in 1991 (km)	unprotected	42.73	31.19	40.20	0.003	248.35
		protected	45.19	33.97	33.64	0.03	147.33
elevation	elevation of parcel (m)	unprotected	436.38	226.00	516.64	89.00	4138.00
		protected	636.61	323.50	636.33	86.00	4004.00
slope	slope of parcel (degrees)	unprotected	3.72	1.00	7.19	0.00	50.00
		protected	7.91	2.00	9.91	0.00	52.00
distance to forest edge	distance from parcel to edge of forest in 1991 (km)	unprotected	0.93	0.45	1.28	0.02	14.81
		protected	1.24	0.74	1.44	0.02	10.68
socioeconomic data
poverty index 1992	poverty index measured in 1992	unprotected	−1.25	−1.68	1.74	−4.01	5.86
		protected	−1.22	−1.67	1.86	−3.64	5.34
average distance to city	avg. distance from each 1 ha parcel within each canton (km)	unprotected	99.55	88.42	64.74	2.57	565.71
		protected	139.02	117.89	118.17	5.38	589.06
elevation	average elevation of canton (m)	unprotected	2959.05	3542.58	1291.03	128.40	4840.51
		protected	2295.52	2444.22	1454.96	136.62	4838.71
slope	average slope within canton (degrees)	unprotected	20.85	19.44	15.07	0.79	66.93
		protected	27.09	29.64	15.79	1.30	60.46
forest cover 1991	per cent of canton covered by forest in 1991	unprotected	0.15	0.00	0.29	0.00	0.99
		protected	0.39	0.34	0.38	0.00	0.98
roadless volume	sum of the product of the area and dist. to road for each parcel within canton	unprotected	2.24 × 108	1.37 × 107	1.24 × 109	1.26 × 104	1.92 × 1010
		protected	1.17 × 109	4.39 × 107	5.65 × 109	2.69 × 104	5.41 × 1010

(b). Socioeconomic data

We extend the data of Canavire-Bacarezza & Hanauer [6] to measure the impact of Bolivia's protected areas on poverty between 1991 and 2012. The main source of our census data is the Bolivian National Statistical Office (INE), from which we obtained information at the canton level (penultimate level of disaggregation). During the period of our analyses, Bolivia contained 1385 cantons with an average area of 772.281 km² (range: 1.783–65 423.568 km²).

We define our unit of analysis as the canton, which we delineate spatially through the creation of a map (shapefile) of cantons using GIS (see the electronic supplementary material for details). As in previous studies [5,6,15,16,22], we create a temporally comparable asset-based poverty index using data from the 1991, 2001 and 2012 censuses (see the electronic supplementary material for greater detail on the poverty index).³ The poverty index allows us to compare poverty across units and across time. We define our outcome as poverty in 2012, thus a positive difference in the poverty index between protected and unprotected units indicates that protected areas had a poverty-alleviating effect on surrounding communities.

As in [5,6], we define a canton as protected if at least 10% of the canton is covered by one or more protected areas.⁴ A canton is designated unprotected if less than 1% of its area is covered by a protected area.⁵ According to these definitions, we are left with 106 protected and 1110 unprotected cantons.

To capture the determinants of poverty and the location of protected areas, we follow [5,6,15,16] and calculate the following covariates (all measured at baseline) for each canton: poverty index, average distance to a major city, average elevation, average slope, per cent of the canton covered by forest and roadless volume (a measure of infrastructure development; table 1).

3. Study design

We use several distinct estimators to assess the heterogeneous impacts of Bolivia's protected areas. Each estimation strategy comprises multiple steps, and each begins with one-to-one nearest neighbour matching. The first stage matching serves two purposes: (i) it allows us to estimate the average treatment effect on the treated (ATT) of protected areas on deforestation and poverty (similar to [6,18]) and (ii) it provides a quasi-experimental set of matched pairs on which we can implement various estimators of heterogeneity (i.e. we use matching to preprocess the data [23,24]; see the electronic supplementary material for more details on matching).

(a). Matching and ATT estimators

We follow [6,18] in our first stage matching strategies. As in those studies, the final chosen matching algorithm is based on the one that provides best post-match balance across the key covariates believed to jointly impact the establishment of protected areas and outcomes.⁶ For the deforestation sample, we use inverse variance weighting with replacement, and post-match bias adjustment to estimate the ATT. For the poverty sample, we use one-to-one genetic matching [25] with replacement, and post-match bias adjustment [26,27] to estimate the ATT.

The fact that protected areas have the potential to affect multiple units (parcels or cantons) raises concerns regarding the statistical precision of our ATT estimates. The concern is that if the outcome is systematically correlated across units according to some grouping, then the precision of our estimates will be overstated. This is a particular concern in the deforestation analysis because we are not able to observe the relevant management units, which is one of the reasons that we randomly sample parcels from across the forested landscape.⁷ Therefore, we have numerous forest parcels in our analyses that lie within the same protected areas, which may result in correlated outcomes among these units. Because cantons represent a relevant management unit we have fewer concerns in the poverty analyses. However, there are instances in which multiple cantons are affected by the same protected area, thus the concern over statistical precision remains.

To account for the effects of intragroup correlation in outcomes, we run a post-match regression for each analysis in which we calculate cluster-robust standard errors. In other words, we take the protected and matched unprotected units from the matching process described above and then run a regression of outcomes on an indicator of protection and all other covariates using this matched set.⁸ We define group clusters as follows: (i) for the deforestation analysis, we cluster protected parcels by protected area and we cluster unprotected parcels by locality (the ultimate disaggregation of the socioeconomic unit) and (ii) for the socioeconomic analysis, we cluster protected cantons by protected area and unprotected cantons by canton (i.e. clustering does not affect the unprotected cantons). See the electronic supplementary material for more details on the clustering and estimation of standard errors.

(b). Moderating covariates

A moderating covariate is a baseline characteristic that is unaffected by treatment but affects a unit's response to a treatment [28]. In the context of Bolivia's protected areas, we are interested in estimating heterogeneity across four potential moderators: (i) slope, (ii) distance to major city, (iii) baseline poverty and (iv) baseline proportion of indigenous peoples. Slope, distance to major city and baseline poverty are chosen for the theoretical reasons discussed below and to offer comparability to previous studies [16,17,29]. Baseline proportion of indigenous peoples within a canton is chosen because the Bolivian government is particularly interested in understanding if and how indigenous populations are differentially affected by the establishment of protected areas. Summary statistics for all covariates can be found in table 1.

(i). Slope

Slope has been shown to be a good proxy for both the opportunity cost of land (in terms of agricultural productivity [15]) and pressure from deforestation [1,16,30]. Thus, if protection effectively prevents extraction of resources and other production-based uses, then when protected areas are sited on flat (low slope) land, one might expect to see poverty exacerbation and high rates of avoided deforestation (similar to the results in [15,16]). The potential impacts of protected areas as slope changes are not well founded in theory, and thus remain an empirical question. Previous studies (e.g. [16,17]) have shown that the impacts of protected areas can vary nonlinearly with slope, and this variation can differ across countries. For the deforestation analyses, we calculate the slope of each 30 m land parcel. For the socioeconomic analyses, we calculate the average slope (in degrees) based on the slope of each 100 m land parcel within a canton.

(ii). Distance to major city

Distance to city captures the variation in the opportunity cost of land and deforestation pressure as a function of distance to major markets [16,30]. In theory, proximity to a major market should increase the opportunity cost of land (e.g. lower transportation costs, etc.) and deforestation pressure. However, as with slope, the functional relationship between distance to cities and the impacts of protected areas is not well understood and has been shown to be nonlinear [16,17]. The eight state capital cities are considered to be major cities in our analyses. For the deforestation analyses, we calculate the distance from each 30 m land parcel to the nearest major city. For the socioeconomic analyses, we calculate the average distance from each 100 m parcel within a canton to the nearest major city.

(iii). Baseline poverty

A serious concern regarding the establishment of protected areas is their potential to exacerbate poverty in surrounding communities [2–4,31]. Further concern has been raised regarding the potential for protected areas to create or enhance poverty traps [32]. In the following [16], we estimate the impact of protected areas conditional on baseline poverty in order to look for evidence of differential impacts on poor communities.

(iv). Baseline indigenous population

In addition to environmental conservation, Bolivia's protected areas are also thought to serve as safeguards of cultural landmarks, limiting the effects of development on indigenous groups. Indigenous peoples are the largest segment of the population in rural areas and policies related to protected areas affect them directly [33]. Because this ethnic group has been marginalized in the past, the Bolivian government has a particular interest in understanding the impacts that protected areas have had on indigenous populations. We estimate the impact of protected areas conditional on the proportion of the baseline population within each canton that identifies as indigenous in order to look for evidence of differential impacts on indigenous communities.

(c). Heterogeneity estimators

We employ three distinct estimators to explore the heterogeneity in the impacts of Bolivia's protected areas: (i) locally weighted scatter plot smoothing (LOESS), (ii) PLM and (iii) multidimensional LOESS (three-dimensional LOESS). Each estimator allows us to explore the non-parametric continuous relationship between our moderators and outcomes in a different way. Prior to running any of the heterogeneity estimators, we use matching, as described above, to ensure that our protected and unprotected units are comparable.

(i). Locally weighted scatter plot smoothing

LOESS estimates the unconditional non-parametric relationship between two variables. The data are smoothed using the tri-cube weighting function [34,35], using only observations within a user-defined (local) span. The result, in our context, is a non-parametric function that describes the relationship between our moderators and outcomes. For each analysis, we estimate a LOESS function for the protected and unprotected units. The conditional ATT (i.e. the ATT conditional on the value of the moderator) is then determined by taking the difference in the predicted outcomes between protected and unprotected units at each value of the respective moderators (see the electronic supplementary material for greater detail).

(ii). Partial linear model

PLM estimates the non-parametric relationship between two variables after using regression to control for other influential variables [36,37]. The PLM estimator is performed in two steps. In the first, a linear regression is used to control for the impact of all covariates (other than the moderator of interest) on the outcome. In the second, LOESS is used to estimate the non-parametric relationship between the moderator and outcome, net the effects of all other covariates. The benefit of this approach is that it allows us to conduct inference along a continuum of moderator values (e.g. slope) while holding constant the potentially complementary or countervailing covariates (e.g. distance to a major city). For each analysis, we run the PLM on protected and unprotected units and the conditional ATT is estimated based on the difference in predicted outcomes between protected and unprotected units (see the electronic supplementary material for greater detail).

(iii). Precision estimates for locally weighted scatter plot smoothing and partial linear model

Pointwise confidence bands for LOESS (and therefore a PLM that uses LOESS for the non-parametric portion of the estimator) are typically calculated using the standard errors of the fit. Just as intragroup correlation in outcomes introduces concern regarding the precision of our ATT estimates, failure to account for such correlation in our heterogeneity estimators can potentially lead to pointwise confidence bands that are too narrow. To account for clustering in our LOESS and PLM estimators, we use a bootstrapping method suggested by Cameron & Miller [38] (see the electronic supplementary material for greater detail). All precision estimates in figure 1 are based on this cluster-robust method. Electronic supplementary material, figure S4, provides a comparison of the cluster-robust estimates to those of the default pointwise confidence bands based on the standard error of the fit.

Figure 1.

Results from LOESS and PLM estimators. In each panel, the dashed and dotted lines represent the conditional relationship between outcomes and moderators for protected and matched unprotected units, respectively. The solid lines represent the estimated difference between protected and matched unprotected units (i.e. the conditional average treatment effect on the treated (ATT) for avoided deforestation and poverty impact, respectively). The green (red) shaded areas around the solid lines represent the 95% pointwise confidence band (CB) for the deforestation (poverty) ATT estimate. The hash marks spanning the x-axis at the bottom of each panel denote the location of a protected (green) or matched unprotected (red) unit. Owing to the high degree of overlap along many portions of the range, the green hashes appear more prominent as the top layer.

(iv). Three-dimensional locally weighted scatter plot smoothing

Three-dimensional LOESS estimates the non-parametric joint relationship between several variables. We use three-dimensional LOESS to estimate the variation in our outcomes across the joint distribution of key moderators. For instance, in contrast to the PLM, which allows us to estimate the heterogeneity of protected-area impacts across the continuous range of distance to city while controlling for the impact of slope (and vice versa), three-dimensional LOESS allows us to estimate the heterogeneity of protected-area impacts across the continuous joint values of both distance to major city and slope. Intuitively, three-dimensional LOESS should allow us to more accurately capture the variation in outcomes across our moderators. For instance, if the impact of slope on deforestation is different for a parcel that is located near a major city as opposed to far, three-dimensional LOESS will capture this variation, but PLM will tend to average this variation out.

For each analysis, we run a three-dimensional LOESS to estimate the joint impact of distance to major city and slope on our outcomes. A separate three-dimensional LOESS is run for protected and matched unprotected units and the conditional ATT is based on the differences between those two estimated surfaces (see the electronic supplementary material for greater detail).

4. Results

(a). Average impacts

Table 2 presents the results from the first stage ATT analyses. The deforestation matching results are the same as those presented in [18], which indicate that Bolivia's protected areas that were established prior to 2000 prevented approximately 1.9% of the protected forest from being deforested. Although this seems like a modest number, given that over 9 million hectares of Bolivia's forest falls within a protected area (at baseline), our estimates suggest that 193 000 ha of forest would have been deforested had the protected areas not been established. Similar to [6], we find evidence that, on average, Bolivia's protected areas that were established prior to 2000 reduced poverty in surrounding communities between 1991 and 2012. Specifically, the poverty index was approximately 0.341 points higher, on average, in protected cantons compared to matched unprotected cantons, which equates to a moderate 0.2 effect size [39]. Our results differ slightly from the previous study because we have more recent census data that are measured at a more disaggregated level. Balance results for the deforestation and socioeconomic matched sets can be found in the electronic supplementary material, tables S1 and S2, and figures S2 and S3. The post-match regression estimates, in which we calculate cluster-robust standard errors, are qualitatively similar to the matching results for both analyses and remain statistically significant at similar levels to the matching analyses.

Table 2.

Results from first stage matching for deforestation and poverty samples. Γ represents maximum Gamma at which estimates are still significant at 10% level according to sensitivity to unobserved heterogeneity test. See the electronic supplementary material for more detail. Asterisks (*, **, ***) represent significance at the 10%, 5% and 1% level, respectively.

estimator	deforestation			socioeconomic
	protected Y1|D = 1	counterfactual	ATT	protected Y1|D = 1	counterfactual	ATT
matching	0.009	0.028	−0.019**	1.765	1.424	0.341*
	3780a	3782a	0.008b	106a	106a	0.194b
			Γ = 4.41			Γ = 0
post-match	0.009	0.032	−0.023***	1.765	1.324	0.441*
regression	3780a	1872a	0.009c	106a	86a	0.245c
potential controls	16 220			1110

^aNumber of observations.

^bAbadie and Imbens heteroskedasticity robust standard errors.

^cClustered standard errors.

Taken together, our first stage results indicate that Bolivia's protected areas are associated with so-called ‘win–win’ outcomes, on average. In other words, on average, Bolivia's protected areas reduced both deforestation and poverty. However, as noted in §1, estimates of the average impacts may mask areas or groups that responded negatively or particularly well (in terms of both deforestation and poverty) to the establishment of protected areas. Thus, we are interested in characterizing the potential heterogeneity in impacts.

(b). Heterogeneous impacts

Figure 1 presents the two-dimensional (LOESS for poverty and PLM for other moderators)⁹ results and figure 2 presents the three-dimensional LOESS results for the slope and distance to city moderators.

Figure 2.

Results from three-dimensional LOESS. (a,d) and (b,e) show the relationship between deforestation/poverty and the joint distribution of slope and distance to city for protected and matched unprotected units, respectively. (c,f) show the heterogeneity of protection (i.e. the conditional ATT) in terms of deforestation/poverty across the joint distribution of slope and distance to city.

(i). Poverty traps

If protected areas were causing or exacerbating poverty traps, we would expect to observe differentially negative socioeconomic responses to protected areas among communities that were poor at baseline. In figure 1a, we find no evidence that Bolivia's protected areas are associated with poverty traps. In general, we see the expected positive relationship between poverty in 1991 and 2012 for both protected and unprotected communities (upward-sloping dashed and dotted lines). However, there is no significant heterogeneity in the impact of protected areas, and thus no evidence that protected areas had a differential impact on poor communities.

(ii). Indigenous population

Figure 1b presents estimates from a PLM of poverty on percentage of the population within the canton that identifies as indigenous. Overall there is a very slight negative relationship between per cent indigenous and poverty for both protected and unprotected communities (the dashed and dotted lines). However, there is no significant difference in the impact of protected areas on poverty between cantons with very low and very high percentages of indigenous people. Overall, there does not appear to be evidence that protected areas had a differential impact on indigenous populations.

(iii). Slope and distance to major city

Partial linear model results. Figure 1c,d (e,f) present the results of the moderating effects of slope (distance to major city) on deforestation and poverty, respectively. We observe that protection on low-slope land (slope of 5° or less) is associated with the highest levels of avoided deforestation and poverty reduction. Although the poverty results are not statistically significant, along this range we observe poverty alleviation upwards of four times greater than the average impact (table 1) and avoided deforestation nearly three times greater than on average. As expected, the relatively high avoided deforestation on low-slope land is driven by the relatively high rate of deforestation on unprotected parcels (the dotted line in figure 1c), whereas deforestation within protected areas appears relatively constant across the range of slope (the dashed line in figure 1c).

We also observe significant heterogeneity in the impacts of protection as a function of distance to a major city (figure 1e,f). However, the distances associated with the greatest poverty alleviation (more than 200 km) are not the ones associated with the most avoided deforestation (less than 200 km). Again, we observe our priors on deforestation pressure playing out in figure 1f. On land near major cities, we see large conditional ATT estimates (over three times the unconditional ATT) driven primarily by the high rates of deforestation on unprotected parcels. Interestingly, we observe the least desirable effects on poverty at near and intermediate distances. As discussed above, the opportunity cost of protection is expected to be highest near markets, thus insignificant or even negative impacts on poverty would not be counterintuitive. However, the fact that we find the least desirable impacts of protection on poverty at intermediate distances is counter to results from Costa Rica and Thailand [16], which suggests that the mechanisms through which protected areas affect poverty are perhaps different in Bolivia.¹⁰

Three-dimensional locally weighted scatter plot smoothing results. Figure 2 presents the results from the various three-dimensional LOESS analyses. Our expectations regarding deforestation pressure are more fully realized through the three-dimensional LOESS estimates, which clearly capture more of the heterogeneity in protected-area impacts. Recall that we expect high pressure of deforestation on low-slope land and land that is near major cities. The three-dimensional LOESS allows us to capture these effects jointly. For instance, we clearly observe the highest levels of deforestation, inside and outside of protected areas, on land that is both flat and near a city. This is one of the benefits of estimating impacts across the joint distributions of our moderators: it allows us to capture heterogeneity that is lost in two-dimensional smoothing. Take, for example, the predicted deforestation on unprotected parcels with zero slope. In figure 2b, we observe significant variation in deforestation (on zero slope land) depending on how far the flat parcel is from a major city. However, this variation is not captured in figure 1c. Instead, the two-dimensional PLM essentially averages across the variation in deforestation at different distances to cities on the zero slope parcels (imagine taking the average deforestation values across the band of cells where the three-dimensional LOESS surface runs along the distance to major city axis in figure 2b). The results of the three-dimensional LOESS indicate that avoided deforestation (figure 2c) was greatest on flat land near cities—upwards of 15% avoided deforestation, which is nearly an order of magnitude greater than the ATT. While we observe avoided deforestation on low-slope land at all distances to major cities, there is essentially no impact of protection once the land becomes sufficiently steep (approx. more than 10°), regardless of distance to major city.

Figure 2d–f illustrates the high degree of heterogeneity in poverty and the impact of protection on poverty across the joint distribution of slope and distance to major city. In figure 2f, we see that when protected areas are placed on flat land near cities, they have the potential to exacerbate poverty. Similar to the PLM results, we observe relatively high poverty alleviation far from cities. However, the highest degree of poverty alleviation is seen on steep land near cities. This might suggest that communities that do not face the typical agriculture/protection trade-off (because they are on land that is not suitable for agricultural cultivation) may benefit from protection when they are near cities that provide potential tourists.

In juxtaposition, figure 2c,f offers interesting insight to the joint impact of Bolivia's protected areas. According to these three-dimensional LOESS results, the areas in which protected areas were most effective in preventing deforestation (flat land near cities) were the areas in which protection had the most negative socioeconomic impacts. This type of stark trade-off is not observed in the PLM results. We do, however, observe areas of the joint slope–distance to major city distribution in which protected areas have benefited the environment and the local communities. Protected areas appear to have been effective in preventing deforestation on low-slope land (less than 10°) at all distances to major cities. We also observe some of the greatest poverty alleviation distant from cities on low-slope land.

When we compare where protection has been most successful with where protected areas have tended to be placed, the reason for the modest success of Bolivia's protected areas, on average, becomes more apparent. The most avoided deforestation and poverty alleviation occurred near cities (on land with differing characteristics). However, protected areas tend to be placed relatively distant from major cities, where they have been less successful from a joint environmental and socioeconomic perspective.

5. Discussion

Previous studies found that Bolivia's protected areas were modestly successful in preventing deforestation and reducing poverty in surrounding communities [6,18]. We use sophisticated econometric methods to estimate how the joint environmental and socioeconomic impacts of protected areas varied across key socioeconomic and biophysical moderators. We find that there was significant heterogeneity in the impacts of protected areas that was not captured in previous studies that estimated the average impacts of Bolivia's protected areas. Our results indicate that, in Bolivia, the impacts of protected areas are nuanced and thus do not conform to the boilerplate arguments for or against protected areas in developing nations (e.g. [2]).

From a socioeconomic perspective, we find no evidence that protected areas created or exacerbated poverty traps. In fact, some of the most positive responses to protected areas are observed in communities that lie below the average baseline poverty level. This is consistent with findings from Costa Rica, which should help assuage concerns that, because poorer populations might tend to depend on the environment for sustenance or as a backstop, poorer populations are at greatest risk of socioeconomic harm from the establishment of protected areas. Our results also provide no evidence that protected areas differentially affected indigenous populations. This is of particular interest to the Bolivian government, which sees protected areas as a mechanism to preserve indigenous culture.

Despite the fact that we find no evidence of poverty traps, we do find evidence of trade-offs across the landscape. For instance, according to the results from the three-dimensional LOESS, the biophysical characteristics that are associated with the most avoided deforestation are the ones associated with the worst socioeconomic impacts from protected areas. Indeed, conditional ATT estimates for protected areas on flat land near cities are nearly an order of magnitude larger than the average avoided deforestation in protected areas throughout the country. However, this is also the type of land on which we observe the potential for protected areas to exacerbate poverty. Throughout all of our results we see evidence in support of the simple theory that the opportunity cost of land has significant moderating effects on the impacts of protected areas.

Previous studies from other countries have estimated the heterogeneous impacts of protected areas using two-dimensional non-parametric and semi-parametric estimators. Those studies and estimators provide valuable information on the heterogeneity of protected areas. However, by comparing the results across several non-parametric and semi-parametric estimators (including the ones used in previous studies), we show that estimating the heterogeneity in protected area impacts across the joint distribution of moderators captures more of the variation in impacts. Thus, future studies that wish to estimate the heterogeneous impacts of protected areas in other countries would benefit from following or expanding upon the methods we employ.

In any estimation strategy, a researcher must be concerned that all relevant controls are included in the analysis. In our matching-based strategy, unbiased estimates of the ATT hinge on our ability to balance all the relevant covariates that jointly determine where protected areas are established and outcomes. While we do argue that we achieve good balance across the covariates we measure, it is difficult to argue that we can measure all of the variables associated with the siting of protected areas. However, we feel comfortable arguing that because our covariates are chosen to capture, among other things, the opportunity cost of land, the confounders that we do not observe are likely correlated with the covariates for which we do account; thus mitigating concerns of bias. Further, we are most interested in estimating how the impact of protected areas varies according to our moderators. The general trends we estimate are unlikely to be affected, even if there is bias in our primary ATT estimates.

It also should be noted that our results are from but one country. And while certain themes (e.g. moderating effects of the opportunity cost of land) are consistent with those from Costa Rica and Thailand, there are certainly differences in the shapes and magnitudes of the effects. The fact that we only have results on the detailed impacts of protected areas for a handful of countries highlights the need to further build the evidence base on protected areas. We believe that the best way in which to do this is on a country-by-country basis, using methods similar to those presented here.

Authors' contributions

M.M.H. conceptualized study design, performed statistical analyses, analysed results and wrote article. G.C.-B. conceptualized study design, gathered data, analysed results and wrote paper. Both authors approved the final version of the article.

Competing interests

We declare we have no competing interests.

Funding

We received no funding for this study.