Editor Jonah Busch
The effects of protected area systems on ecosystem restoration: a quasi-experimental design to estimate the impact of Costa Rica's protected area system on forest regrowth
Article first published online: 17 JAN 2013
©2013 Wiley Periodicals, Inc.
Volume 6, Issue 5, pages 317–323, September/October 2013
How to Cite
Andam, K. S., Ferraro, P. J. and Hanauer, M. M. (2013), The effects of protected area systems on ecosystem restoration: a quasi-experimental design to estimate the impact of Costa Rica's protected area system on forest regrowth. Conservation Letters, 6: 317–323. doi: 10.1111/conl.12004
- Issue published online: 8 OCT 2013
- Article first published online: 17 JAN 2013
- Accepted manuscript online: 17 DEC 2012 05:26AM EST
- Manuscript Accepted: 29 NOV 2012
- Manuscript Received: 8 AUG 2012
- impact evaluation;
- causal effects;
- IUCN categories
- Top of page
- Supporting Information
Global efforts to protect forest biodiversity and ecosystem services rely heavily on protected areas. Although these areas primarily aim to prevent losses from deforestation and degradation, they can also contribute to restoration. Previous evaluations of protected area impacts focus on avoided deforestation and fires. In contrast, we focus on the additional regrowth induced by Costa Rica's renowned system of parks and reserves. We use a quasi-experimental empirical design to control for confounding baseline characteristics that affect both regrowth and the assignment of protection. Between 1960 and 1997, an estimated 13.5% of previously unforested lands inside protected areas reforested because they were afforded protection. The level of additional regrowth does not vary by the strictness of protection. As in previous studies of protected area impacts on avoided deforestation, estimators that do not account for nonrandom assignment of protection can overstate protected areas’ impacts on regrowth by nearly double.
- Top of page
- Supporting Information
National parks and reserves are critical components of efforts to protect forest biodiversity and ecosystem services (MEA 2005). Studies that estimate the effects of these protected areas on environmental outcomes focus on avoided deforestation, avoided fires, or net changes in forest cover (e.g., Andam et al. 2008; Cropper et al. 2001; Deininger & Minten 2002; Curran et al. 2004; Naughton-Treves et al. 2005; Oliveira et al. 2007; Joppa et al. 2008; Sims 2010; Joppa & Pfaff 2010, 2011; Nelson & Chomitz 2011). However, given the substantial losses of biodiversity and ecosystem services globally in recent decades, regrowth1 is also increasingly seen as an important policy objective (Young 2000; Chazdon 2008). Regrowth can, for example, increase habitat for endangered species and contribute to carbon sequestration (Silver et al. 2000). Notably, additional reforestation, but not avoided deforestation, is eligible for emission reduction credits within the Kyoto Protocol's Clean Development Mechanism.
Two studies focus explicitly on the effects of protected areas on regrowth.2 Triantakonstantis et al. (2006) compare forest cover growth before and after a park in Greece was established, observe higher growth after protection was assigned, and attribute this higher growth to protection. The study does not, however, address potential bias from other cotemporaneous, confounding factors that also affect forest cover. Helmer (2000) compares regrowth inside and outside protected areas in southern Costa Rica and observes more secondary forests inside than outside, but fails to consider the non-random placement of protected areas. Failure to account for nonrandom protected area location can bias estimates of impacts (Andam et al. 2008; Andam et al. 2010; Joppa & Pfaff 2010). Globally, protected areas tend to be established on less accessible or productive land; the so-called “rock and ice” or “high and far” phenomenon (Joppa & Pfaff 2009). If protected areas are located on land that is less likely to be used for productive purposes in the absence of protection, then a simple comparison of regrowth on protected and unprotected lands will overstate the impact of protected areas on regrowth. Attributing regrowth to protected area establishment requires the selection of suitable unprotected areas to estimate the counterfactual regrowth that would have occurred on protected lands in the absence of protected area designation.
Although Costa Rica's renowned protected area system was primarily established to prevent deforestation and degradation (Allen 2001), not all protected lands were forested at baseline. To select suitable comparison lands for these protected lands, we follow the quasi-experimental design used by Andam et al. (2008). We select a set of matched unprotected sites to represent the counterfactual regrowth rates of protected sites, and then calculate the effect that formal protection had on regrowth between 1960 and 1997.
Recognizing that not all protection is equal, we unpack “protection” by using International Union for Conservation of Nature (IUCN) categories to measure how regrowth impacts vary by the strictness of protection. Evidence about the environmental effects of different kinds of protection is important for decision-making, particularly if future studies find that the social effects of different protected area land classes differ (e.g., strong land-use restrictions may lead to more adverse impacts on local communities). Theory does not provide clear predictions on how environmental impacts vary with the strictness of protection. Holding all things equal, stricter protection implies fewer anthropogenic disturbances and thus a higher probability of regrowth. However, establishing strictly protected areas is politically more difficult on accessible, productive lands; i.e., stricter protection may be assigned to lands that are less desirable for productive uses. This spatial pattern of establishment has uncertain effects on regrowth: on one hand, more strictly protected unforested lands are more likely to passively regrow in the absence of protection but, on the other hand, humans would be less likely to make costly investments in actively restoring these forests in the absence of protection. Thus although there may be greater disturbances on less strictly protected lands, there could be greater additional regrowth on these lands compared to more strictly protected lands.
Our results indicate that between 1960 and 1997, an estimated 13.5% of previously unforested lands inside protected areas would have failed to reforest in the absence of protection. Naïve estimators that do not control for the nonrandom location of protected areas substantially overstate the impacts (22.8%). The estimated additional regrowth induced by protection is greater in less strictly protected areas: 24% of the protected land experienced regrowth because of protection compared to 13% in protected areas with stricter land-use rules. However, we fail to reject the null hypothesis that these estimates are equal (α = 0.10).
- Top of page
- Supporting Information
We use Geographic Information Systems to build a geospatial data set of relevant geographic and biophysical conditions. We first establish the baseline forest cover conditions in 1960, before the establishment of Costa Rica's protected areas. We partition all of Costa Rica's land that was not forested in 1960 into 3 hectare parcels (the minimum mappable area). We randomly select 20,000 parcels that were not forested. After 4,187 land parcels are excluded for various reasons (see Supporting Information), 15,813 parcels remain to comprise the sample for analysis, of which 1,219 were protected before 1997.
The outcome measure is defined as the presence or absence of forest in 1997. A parcel receives a 1 if it is forested or zero if it remains unforested. A benefit of using the minimum mappable area as the unit of observation (as opposed to, for example, a political or park boundary) is the ability to precisely attribute changes in forest cover within a given area to either avoided deforestation or additional regrowth.
We are interested in estimating protection's effect on additional regrowth between 1960 and 1997. To mitigate potential bias stemming from unobserved changes in the protected area selection process over time, we follow Andam et al. (2008) and break our analysis into two cohorts: (1) protected areas established before 1979, and; (2) protected areas established between 1981 and 1996 (see Andam et al. 2008 for the reasons behind the choice of these date cutoffs). For cohort 1, a parcel is considered protected if it was located within a park established before 1979. For cohort 2, a parcel is considered protected if it was located within a park established between 1981 and 1996, and remained unforested between 1960 and 1986 (see Supplementary Information for details on cohort 2 sample).
Our aim is to estimate the impact of protection on regrowth for protected parcels; i.e., the average treatment effect on the treated (ATT). To estimate the ATT, one must answer the question, “How much regrowth would have occurred on protected parcels had the protected areas not been established?” The difference between the answer to this counterfactual question and the observed (actual) regrowth is the ATT. To estimate the counterfactual regrowth, we use the matching design of Andam et al. 2008 (also used in Andam et al. 2010; Ferraro & Hanauer 2011; Ferraro et al. 2011). A matching algorithm attempts to select unprotected parcels that are observably similar to protected parcels across key baseline covariates that jointly determine selection into protection and regrowth. In other words, after effective matching, the protected and unprotected parcels are observably similar in their distributions of observable characteristics known to affect regrowth in the absence of protection (i.e., factors known to affect land use). Under the assumption that there are no systematic unobservable differences among the matched protected and unprotected parcels in characteristics that affect regrowth in the absence of protection, the expected regrowth of the matched unprotected parcels represents the expected regrowth of the protected parcels had they not been protected. Thus the difference between regrowth on the protected and matched unprotected parcels is an unbiased estimator of the ATT (for more conceptual detail, see Ferraro 2009; for an explanation for our preference of a matching design over a regression design, see Supplementary Information).
Like Andam et al. (2008), we wish to select comparison parcels that are equivalent, on average, in terms of four land-use capacity classes, distance to major cities, distance to the nearest roads, and distance to the nearest forest parcel at baseline. These characteristics capture the most important factors that historically determine land use in Costa Rica (see Supplementary Information of Andam et al. (2008) for details on how these measures and the forest cover measures were calculated). Before any matching, protected parcels are much more likely to be located on unproductive lands and, on average, they are farther from roads and cities and closer to forest edges than unprotected parcels (Tables S1 and S3 in rows labeled “unmatched”).
We select the matching algorithm that achieves the best covariate balance after matching for our sample. As in Andam et al. (2008), we use multiple measures of the differences in the covariate distributions between protected and unprotected parcels: the difference in means, the mean difference normalized by the variance of the protected observations, and the average distance between the two empirical quantile functions (values > 0 indicate deviations between the groups in some part of the empirical distribution). If matching is effective, these measures should move dramatically toward zero. Given the central role of agriculture in Costa Rica, we particularly want good balance on land-productivity classes. In our sample, Mahalanobis covariate matching yields the best covariate balance, with measures of balance at or near zero (Tables S1 and S3 in rows labeled “matched”).
As an additional control on confounding factors in our estimate of the ATT, we use calipers to improve covariate balance. Calipers define a tolerance level for judging the quality of the matches; if a protected parcel does not have a match within the caliper (i.e., available controls are not good matches), it is eliminated from the sample. Calipers reduce bias, but at the cost of estimating avoided deforestation on a subsample that may not be representative of the population of protected parcels. Like Andam et al. (2008), we use calipers that restrict matches to units within 0.5 standard deviations of each covariate. Fewer than 6% of the protected parcels are dropped by the calipers; on average, they are more remote parcels on poor-quality land (for covariate balance results see Tables S2 and S4).
With all estimators, we further control for bias that can remain after matching in finite samples by using a postmatching bias-correction procedure that asymptotically removes the conditional bias in finite samples (Abadie & Imbens 2006). To characterize the precision of each of our estimates, we calculate heteroskedasticity robust standard errors (Abadie et al. 2004; Abadie & Imbens 2006; Imbens & Wooldridge 2009). These standard errors allow for heteroskedasticity both within and across treatment arms by calculating conditional variances via a secondary matching algorithm that matches units within treatment arms (treated units to treated units, etc.). All matching and estimates were performed in R (Sekhon 2007).
Despite our efforts to control for observable sources of bias, protection and regrowth may exhibit correlation in the absence of an effect of protection because of failure to match on a relevant but unobserved covariate. In our analysis, the main concern is that protected parcels may be unobservably more likely to regrow than their matched controls. To examine the degree to which uncertainty about hidden biases in the assignment of protection could alter the conclusions of our study, we use Rosenbaum's (2002) recommended sensitivity test (see Supplementary Information for details and additional robustness checks).
In addition to testing the average effect on regrowth from all protected areas, we wish to estimate the average effect of protection conditional on the strictness of the protection. To measure the strictness of the regulatory protection, we use IUCN management categories. Costa Rica has protected areas in the following categories: Ia, I, II, IV, and VI. We define “strictly protected” as categories Ia, I, II, and IV (biological reserves, wildlife refuges, and national parks), and “less strictly protected” as category VI (protected zones and forest reserves). We then run two separate analyses to test for differential impacts on regrowth according to strictness of protections. We only present the results for the first cohort (protected areas established before 1979) because of the lack of variation in the categories used after 1981 (most were strictly protected).3
First, we match the two strictness subgroups separately, estimate the ATT for each subgroup, and then perform a statistical test of equality on the two subgroup estimates (Welch's t-test). When estimating the ATT for a subgroup, we exclude protected parcels from the other subgroup to avoid matching protected parcels of disparate strictness. The less strictly protected parcels are, on average, located on more productive lands and are closer to roads, forest edges, and cities.
Second, we estimate the ATT for strictly protected parcels compared to a counterfactual state of less-strict protection (the ATT described in the previous paragraph uses a counterfactual state of no protection). In other words, for the parcels assigned strict protection, we estimate their regrowth had they instead been assigned less-strict protection. We then estimate the ATT for less-strictly protected parcels using a counterfactual state of strict protection (see Supplementary Information for details and covariate balance Tables S5–S8 for all subgroup analyses).
- Top of page
- Supporting Information
Table 1 presents the estimated effects. The numbers can be multiplied by 100 and interpreted as percentages of protected parcels that regrew as a result of protection. Row 1 of Table 1 presents the simple differences in means that is calculated by comparing regrowth on protected and unprotected parcels without adjusting for baseline conditions (i.e., without matching). The next two rows present the results from the matching estimators that attempt to remove observable sources of bias. The estimate in the first column and second row implies that 18.3% of the nonforested lands that were protected before 1979 experienced regrowth that would not have taken place in the absence of protection. For protected areas established after 1981, the estimate is 3.9%. The estimates with caliper adjustments in Table 1 are similar, and both matching estimates are substantially lower than the estimates from the simple difference in means in row 1.
|Method||(1) Protected Before 1979||(2) Protected 1981-1996|
|Difference in means||0.308***||0.0649***|
|Matching with calipersa||(0.0373)||(0.0213)|
|[N outside calipersb]|||||
|N Potential controls||14,622||11,952|
Results from Rosenbaum's test of hidden bias are presented under the estimated treatment effects in Table 1. If, for example, an unobserved covariate caused the odds ratio of protection to differ between protected and unprotected parcels by a factor of as much as 2.1, the 90% confidence interval of the first matching estimate for the pre-1979 cohort would still exclude zero. The gamma values close to 1 for the cohort 2 estimates imply high sensitivity of these estimates to the potential presence of hidden bias.
Figure 1 depicts how the results from each estimation strategy translate to hectares of regrowth because of protection by cohort.4 The overwhelming proportion of estimated regrowth because of protection can be attributed to protected areas established before 1979.
Table 2 presents estimates from the heterogeneous treatment analysis. The first column implies that 13.2% of the nonforested lands that were protected before 1979 in IUCN classes I, II, or IV experienced regrowth that would not have taken place in the absence of protection. For less strictly protected areas (class VI), the estimate is higher: 24.4%. Although both estimates are statistically different from zero at the P = 0.01 level, we are unable to reject the null hypothesis that they are equal (P = 0.11). The second and third rows present results from matching protected parcels of differing strictness. For example, had strictly protected parcels been assigned to less-strict protection, 8.1% of the parcels that regrew would have failed to regrow. Neither estimate is significantly different from zero, implying that we cannot reject the null hypothesis that the level of strictness does not affect regrowth on protected parcels.
|Strictness of protection|
|Analysis||Strict IUCN I, II, & IV||Less strict IUCN VI||Difference|
|Matching strict to less-strict||–||–||0.081|
|Matching less-strict to strict||–||–||−0.044|
- Top of page
- Supporting Information
We find that the establishment of protected areas in Costa Rica increased the amount of regrowth observed between 1960 and 1997 beyond that which would have otherwise occurred. Furthermore, there is little evidence that more strictly protected areas induced greater levels of regrowth.
Our results highlight the pitfalls of empirical designs that fail to control for the non-random location of protected areas. A naïve estimator that fails to control for systematic, observable baseline differences between protected and unprotected lands greatly overstates the amount of regrowth attributable to the establishment of protected areas by nearly 70% (Table 1).
By studying both avoided deforestation and additional regrowth, one can generate a more complete picture of the effects of protected areas. Figure 2 combines our results with results from Andam et al. (2008) to present the total effect induced by protected areas on forest cover in Costa Rica through 1997. The large differences between the estimates from naïve estimators and matching estimators emphasize again how inaccurate simple empirical designs can be. Decomposed total effects like the ones portrayed in Figure 2, when combined with insights from heterogeneous impact analyses like Ferraro et al. (2011), establish a foundation with which scholars and practitioners can begin to estimate the effects of protected area systems on ecosystem services. Carbon sequestration impacts, for example, require estimates of both avoided deforestation and additional regrowth.
As noted by Andam et al. (2010), Costa Rica is not representative of all developing nations and thus the effects we find may not be generalizable to other nations. Moreover, our analysis does not elucidate the specific mechanisms through which protected areas encouraged regrowth nor does it characterize the ways in which impacts vary as a function of sociopolitical and biophysical contexts (for an example of the latter related to protected area effects on avoided deforestation and poverty, see Ferraro et al. 2011). As emphasized by others (Andam et al. 2010; Ferraro et al. 2011; Miteva et al. 2012), multination research to understand protected area mechanisms and heterogeneous responses is a priority. Without this understanding of why and under what conditions protected areas affect regrowth, we cannot hope to better design protected areas to achieve their objectives or predict how they may affect regrowth in other locations and time periods.
If, however, the patterns seen in Costa Rica are representative of a larger set of nations, they have important management implications. For example, protection is most often assigned to areas with little recent disturbance. Yet our results suggest that protection in Costa Rica induced a greater amount of additional regrowth than avoided deforestation per 100 ha protected (13.5 ha versus 7 ha), implying that more effort to quantify the net returns to protecting disturbed versus undisturbed habitat may be warranted. Furthermore, the lack of a statistically significant difference in the amount of regrowth on strictly protected areas versus less strictly protected areas would imply that if decision makers wish to increase the amount of regrowth through formal protection, the specific rules governing the protected area may be less important than the act of protection itself.
Our estimates of past regrowth impacts are not perfect predictors of the future regrowth that will result from new and old protected areas. Nevertheless, they give some indication of the potential contribution of protected areas to biodiversity conservation and ecosystem service flows through additional regrowth.
- Top of page
- Supporting Information
The findings, interpretations, and conclusions expressed in this article are entirely those of the authors. They do not necessarily represent the view of the World Bank, its Executive Directors or the countries they represent. The authors thank three reviewers for constructive comments and the people who created the data for the original studies from which we draw (see Andam et al. 2008), including Arturo Sanchez-Azofeifa and the Earth Observation Systems Laboratory, University of Alberta and Juan Robalino.
- Top of page
- Supporting Information
- 2004). Implementingmatching estimators for average treatment effects in stata. Stata J., 4, 290-311. , , & (
- 2006). Large sample properties of matchingestimators for average treatment effects. Econometrica, 74, 235-267. & (
- 2001). Green phoenix: restoring the tropical forests of Guanacaste, Costa Rica. Oxford University Press, New York, USA. (
- 2008). Measuring the effectiveness of protected area networks in reducing deforestation. Proc. Natl. Acad. Sci. U.S.A., 105, 16089-16094. , , , & (
- 2010). Protected areas reduced poverty in Costa Rica and Thailand. Proc. Natl. Acad. Sci. U.S.A., forthcoming. , , , & (
- 2008). Beyond deforestation: restoring forests and ecosystem services on degraded lands. Science, 320, 1458-1460. (
- 2001). Predicting the location of deforestation: the role of roads and protected areas in North Thailand. Land Econ., 77, 172-186. , & (
- 2004). Lowland forest loss in protected areas of Indonesian Borneo. Science, 303, 1000-1003. , , et al. (
- 2002). Determinants of deforestation and the economics of protection: an application to Mexico. Am. J. Agric. Econ., 84, 943-960.Direct Link: & (
- 2009). Counterfactual thinking and impact evaluation in environmental policy. New Direct. Evaluat., 122, 75-84. (
- 2011). Protecting ecosystems and alleviating poverty with parks and reserves: ‘win–win’ or tradeoffs? Environ. Resour. Econ., 48, 269-286. & (
- 2011). Conditions associated with protected area success in conservation and poverty reduction. Proc. Natl. Acad. Sci. U.S.A., 108, 13913-13918. , & (
- 2000). The landscape ecology of tropical secondary forest in montane Costa Rica. Ecosystems, 3, 98-114. (
- 2009). Recent developments in the econometrics of program evaluation. J. Econ. Lit., 47, 5-86. & (
- 2009). High and far: biases in the location of protected areas. PLoS One, 4, e8273. doi:10.1371/journal.pone.0008273. & (
- 2010). Reassessing the forest impacts of protection: the challenge of nonrandom location and a corrective method. Ann. N.Y. Acad. Sci., 1185, 135-139. & (
- 2011). Global protected area impacts. Proc. Royal Soc. B, 278, 1633-1638. & (
- 2008). On the protection of “protected areas”. Proc. Natl. Acad. Sci. U.S.A., 105, 6673-6678. , & (
- Millennium Ecosystem Assessment (MEA) (2005). Ecosystems and human well-being: policy responses. Island Press, Washington, DC.
- 2012). Evaluation of biodiversity policy instruments: what works and what doesn't? Oxford Rev. Econ. Policy, 28, 69-92. , & (
- 2005). The role of protected areas in conserving biodiversity and sustaining local livelihoods. Annu. Rev. Environ. Resour., 30, 219-252. , & (
- 2011). Effectiveness of strict vs. multiple use protected areas in reducing tropical forest fires: a global analysis using matching methods. PLoS One, 6, e22722. & (
- 2007). Land-use allocation protects the Peruvian Amazon. Science, 317, 1233-1236. , & et al. (
- 2002). Observational Studies. Springer Series on Statistics, New York, USA. (
- 2007). Multivariate and propensity score matching software with automated balance optimization: the matching package for R, http://sekhon.berkeley.edu/papers/MatchingJSS.pdf (visited Jan. 10, 2008). (
- 2010). Conservation and development: evidence from Thai protected areas. J. Environ. Econ. Manage., 60, 94-114. (
- 2000). The potential for carbon sequestration through reforestation of abandoned tropical agricultural and pasture lands. Restor. Ecol., 8, 394-407. , & (
- 2006). Forest re-growth since 1945 in the Dadia forest nature reserve in northern Greece. New Forests, 32, 51-69. , & (
- 2000). Restoration ecology and conservation biology. Biodiv. Conserv., 92, 73-83. (
We use the term “regrowth” to capture forest cover increases on previously unforested land. We do not distinguish among the channels of regrowth. Protection can lead to regrowth by controlling fires and stopping agriculture and forest product extraction. Protection can also lead to more active restoration efforts such as planting trees and seeds to further facilitate the natural recolonization process (i.e., managed restoration). Science-based, low-intensity efforts to facilitate natural recolonization are widespread in Costa Rica (Carlos Manuel Rodriguez, per. comm., 2012; Allen 2001).
Studies that use net change in forest cover as the outcome combine avoided deforestation and regrowth in one measure. When only net forest cover is observed, the results are consistent with an infinite number of combinations of deforestation and reforestation rates (over a given time, within a given area).
Moreover, after 1981, some new category IV protected areas included substantial amounts of private land for which government control was often limited. The choice of strictness definition for these protected areas could be disputed.
Calculated by multiplying the treatment effects from Table 1 by the amount of unforested land protected. According to GIS calculations, 1,132,668 ha of land was protected during our study period (663,802 ha in cohort 1 and 495,886 ha in cohort 2), of which 149,648 was unforested (100,326 ha in cohort 1 and 49,412 ha in cohort 2).
- Top of page
- Supporting Information
Disclaimer: Supplementary materials have been peer-reviewed but not copyedited.
Supporting Information—restoration effects of protected area systems
Unforested area calculations
1981–1997 protected sample
Heterogeneous treatment analysis
Tests for sensitivity to hidden bias
Matching versus regression designs
Additional robustness checks
Table S1: Covariate balance for pre-1980 protection analysis, no calipers.
Table S2: Covariate balance for pre-1980 protection analysis, with calipers.
Table S3: Covariate balance for analysis of protection established between 1980 and 1996, no calipers.
Table S4: Covariate balance for analysis of protection established between 1980 and 1996, with calipers.
Table S5: Covariate balance for Strictly protected areas
Table S6: Covariate balance for Less-Strictly protected areas
Table S7: Covariate balance for Matching Strict to Less Strictly protected areas
Table S8: Covariate balance for Matching Less Strict to Strictly protected areas
Please note: Wiley Blackwell is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.