In recent decades, national and international efforts to reduce forest loss have had some impact but have not substantially slowed tropical forest loss. Adding climate change to concern about species, such forest losses now account for about one-sixth of anthropogenic greenhouse gas emissions. Climate-related incentives could enhance various programs and policies that affect deforestation, most likely by emphasizing performance with monitoring, reporting, and verification of outcomes. Such emphasis, with incentives, could increase policies’ deforestation impacts [1
Understanding of past policy impact also supports greater future impacts, through better design. It is crucial for countries to know what has (not) worked to date in reducing forest loss, and why. Here, we follow a recent explosion of literature that has provided improved impact evaluations for conservation policies, such as protected areas (PAs—impacts reviewed in [2
]). Improved methods often reduce estimates of average PA impacts, while also clearly demonstrating variation in impacts across space and PA types [3
First, to highlight the actors and incentives relevant for variation in impacts of PAs upon forests, we present conceptual models of individual land-use choices and the political economy of PA siting. They suggest key influences on PAs’ impacts. Our results shed empirical light on those influences, suggesting that the politics of PA implementation in Mexico clearly have shifted over time.
We start with a classic land-use model, taking as given both the PAs’ locations and their enforcement. The clear prediction that typical landscapes vary in their baseline deforestation rates is the motivation for our efforts to control for characteristics of sites that affect individual decisions that yield deforestation. Only by removing the influence of such characteristics, which might well be expected to differ for protected versus unprotected locations [5
], can we infer the impact of the PAs.
As to why PA locations might differ from unprotected sites in relevant characteristics, we highlight political and economic tradeoffs. Those tradeoffs predict variations in both locations and enforcement across PA types and, thus, also varied impacts [6
]. In sum, these models suggest that, depending on the political economic context, either strict or mixed-use PAs can have more impact.
For Mexico, an important basis for comparison is recent work showing no PA impacts in the 1990s ([7
] on ‘paper parks’). Thus, we consider PA impacts on more recent changes in land cover (2000–2005) after a shift in the politics of conservation in Mexico, with a change in administration in 2000 to President Fox, under whom Ernesto Enkerlin shifted management of PAs. In addition, GEF provided stable funds for some PAs (https://www.thegef.org/country/mexico
). The time period we study, coming just after the 1990s, offers a useful window on a possible sharp shift in politics. We estimate the average PA impact and then differences in impacts between strict and mixed-use PAs.
Our application of matching to improve estimates of PA impact—following both of our models—reduces estimated impacts, as hypothesized, relative to when ignoring key location characteristics. Nonetheless, improved estimates show PAs in Mexico did lower land-cover change. Specifically, PAs across Mexico reduced the 2000–2005 loss of land cover by 3.2%, on average.
In comparing PA types, following the political economy model, we find that strict PAs avoid more land-cover change (5.2%) than mixed-use PAs (2.7%). That is intuitive, if enforcement is tougher. However, we show that greater impact from strict PAs is influenced by the relative locations of the PA types. That may seem counterintuitive but it is consistent with the PAs’ sites having been chosen at a time when PA enforcement was low.
The paper proceeds as follows. Section 2
provides background for Mexico and prior literature evaluating policy impacts. Section 3
presents both of the conceptual models just discussed above. Section 4
then describes our data and methods, Section 5
conveys results, and Section 6
2. Background and Related Literature
Mexican forests cover 67 million hectares, about one-third of the country (198 million hectares). Along with agriculture and fishing, forestry accounted for about 5% of the country’s GDP in 2006. Agriculture and forestry are areas where greenhouse gas (GHG) emission can be reduced, having generated about 135 MtCO2
e or 21% of Mexican emissions in 2002 [8
], with two-thirds from the forest sector. Proximate causes of both deforestation and degradation are conversion to grassland, slash-and-burn agriculture, illegal logging, and fire. Underlying forces include a lack of investment in the forestry sector, low income from forest activities, multiple agricultural and livestock activities, uncertainty related to use rights, poverty, and a general lack of opportunities for forest owners [9
]. The drivers are complex and they vary between regions.
It is widely acknowledged that successful forest interventions could not only reduce emissions but also generate ecosystem services, income, and employment—among other co-benefits [8
]. Interventions such as reforestation and commercial plantations are what account for 85% of the state’s proposed mitigation in agriculture and forestry. Success would depend on institutional changes, better public financing, and sustainable forest product markets [8
This direction for policy appears to have some support in Mexico, despite a large illegal timber market, a lack of financial and human resources for operational capability, and (drug-related) insecurity. Mexico currently is emphasizing a national, multi-functional and multi-scale mechanism for monitoring, reporting and verification (MRV) based on remote sensing and ground-based forest inventory methodologies [9
]. It could include early detection for changes in land cover and land use [10
]. Such a mechanism could provide the relevant authorities with more precise measures to address concern about detection of small-scale land-cover changes.
2.1. Evaluating PA Impacts
Joppa and Pfaff [2
] review the PA literature—as do Naughton-Treves [11
], Nagendra [12
], and Campbell et al. [13
]—emphasizing obstacles to inferring PA impacts on forest. This follows global documentation [5
] that, at least on average, the locations of PAs across national landscapes are significantly biased in ways relevant to deforestation rates. They stress that observing fully forested PAs (e.g., [14
]) falls short of observing impacts, as impacts require a comparison to what would have happened without PAs. Given that it is literally impossible to observe such ‘counterfactuals’—scenarios without PAs—one must estimate what would have happened on PA land without protection, using outcomes from unprotected sites. Siting bias means the average outcome for all unprotected lands (as in [16
]) is a poor counterfactual estimate for PAs, since sites are different on average. Comparing to areas around PAs [15
] can help. Yet, without explicitly comparing land characteristics, it is hard to know how similar they are. At least as troubling, those nearby lands could be affected by the establishment of the PA if that generates local spatial spillovers such as ‘leakage’.
Matching methods explicitly compare land characteristics between sites, aiming to increase the similarity of controls to PA observations. Andam et al. [24
]’s application of matching to Costa Rican PAs reduces the estimated impact from about 44%, over decades, to about 11% of the protected area. Joppa and Pfaff [3
] provide analogous results for each of over 100 countries around the globe. Given those average impacts, Pfaff et al. [4
] revisit Costa Rican PAs in subsets, using matching, to confirm predicted variation in PAs’ impacts over the landscape. Impacts are higher near roads and cities and on flat land. Shah and Baylis [25
] also show heterogeneous PA impacts on forest, for Indonesia, while Joppa and Pfaff [3
] confirm such predictable variation in impacts in their global study of PAs.
2.2. Evaluating Conservation Impacts in Mexico
Early conservation evaluations for Mexico included payments by a federal agency to upstream land, which can affect water quality. Munoz [26
] finds significant location bias in payments contracts, lowering impact. Alix-Garcia et al. [27
] extend this with similar results while Alix-Garcia et al. [28
], using additional data, consider both forest and poverty impacts of these hydro-services payments. Recently, Pfaff et al. [29
] and Kaczan et al. [30
] consider institutional adjustments in such payments.
Concerning PAs, Blackman et al. [7
] is analogous to our study except for the time period, i.e., the 1990s, when it was widely asserted that PAs were under-resourced ‘paper parks’ (while Sims and Alix-Garcia 2016’s analogous ongoing work extends further forward in time and compares outcomes of PAs to PES [31
]). Blackman et al. [7
] find no PA effect. Figueroa and Sánchez-Cordero [32
] also find limited effect for the 1990s. Low intention or ability to enforce can affect not only PA outcomes given PA sites but also PAs’ sites because, without enforcement, there is little reason for locals to push back. Mas [33
] and Durán-Medina et al. [34
] find some impact for particular PA sites, as do Honey-Roses et al. [35
], for a site where PES and a PA are combined, while Miteva et al. [37
] examine tenure.
4. Data and Matching Methods
4.1. Land Cover Data
Our data are from a global data set with land cover for 2000 [44
] and land cover for 2005 (ESA) [45
]. The 2000 data, GLC2000, have 23 classifications of land cover. They were reclassified to ‘natural’ or ‘human modified’, the latter including categories 16 (cultivated and managed areas), 17 (mosaic of cropland with tree cover or other natural vegetation), 18 (mosaics of cropland, with shrubs or grass cover), 19 (bare areas), and 22 (artificial surfaces and associated areas). The same was done for 2005 GLOBCOVER300 data. Again, multiple categories are placed into “natural” and “modified”, the latter including categories 11 (irrigated croplands), 14 (rainfed croplands), 20 (mosaic cropland (50%–70%)), 30 (mosaic cropland (20%–50%)), and 190 (urban areas >50%). These datasets were not constructed for precise intertemporal comparison, yet this transformation allows for a reasonable comparison of years. In each dataset, the spatial scale of land observations is 1-km2
The change between the datasets—which we call ‘vegchange’—was calculated after that process. With the transformation, the data in principle track change from a ‘natural’ to a ‘human modified’ landscape and this 2000–2005 change is the dependent variable upon which we focus. We want to study recent changes that can be affected by the presence of PAs (though they are net changes, i.e., from ‘natural’ to ‘human modified’ and vice versa, which matters for indicators of species habitat). Given any issues with comparing datasets, we also analyze spatial patterns in the 2005 land cover.
4.2. Land Characteristics Data
provides descriptive statistics for land characteristics in our analysis. Elevation (m) is from the Shuttle Radar Topography Mission [46
], with slope in degrees from horizontal. The roads and urban areas used to compute distances (km) are from VMAP0 Roads of the World [47
] and the Global Rural Urban Extent data [48
]. While not the best quality, these VMAP0 data are all that is freely accessible to define the global road network.
PAs (protected areas) are from World Database on Protected Areas [49
]. In these data, Categories I–II allow less human intervention, while Categories III–VI tend to be less protected, allowing for multiple uses. For any overlaps of categories, the polygon was categorized into the stricter category. We created three dummy variables: ‘protected’, for any protection (regardless of IUCN category); ‘mixed-use PA’ for IUCN categories III–VI; and ‘strict PA’, if the IUCN category is I or II. In order to check robustness, we also compared categories I–IV versus V–VI (following Joppa and Pfaff [3
], as well as Nelson and Chomitz [40
]). Also, for our time period, we used only the PAs created by 2000.
The World Wildlife Fund classified ecoregions [50
]. Unclassified ecoregions were dropped (n
= 1747) and we created dummy variables for the two ecoregions with the highest frequencies: ‘pineoakdum’ (10.7% of total) if the ecoregion is Sierra Madre Occidental Pine Oak Forest; and ‘chidesertdum’ (15.6% of total) if the ecoregion is Chihuahuan desert. The agricultural suitability is from International Institute for Applied Systems Analysis’ Global Agro-Ecological Zones data [51
]. It uses climate, soil type, land cover, and slope to assign a value to each polygon, ranging from 0 (meaning no constraints) to 9 (severe constraints). We created two dummy variables: ‘low’ suitability for more agricultural constraints (i.e., agricultural suitability score of 8 or higher); and ‘high’ for situations with fewer agricultural constraints, (i.e., agricultural suitability score of 5 or lower). Additionally, as a robustness check, we use the full variable (which does not affect any of the results).
The fire variable, as a continuous metric, simply captures the number of fires from 2001 to 2006. The ‘firedum’ dummy we created takes a value of ‘1’ if that polygon experienced any fires (≥1). Finally, the distance-to-edge variable measures the distance to the edge of a protected area (km) from the polygon in question. A negative distance value indicates that the observation is inside of the PA.
These variables are not expected to fully explain land-cover change or the location of protection. Yet as all influence profit, they predict deforestation as well as the local resistance to protected areas. It is a combination of relevance to protection and deforestation that makes them useful for matching. The data contain approximately 1,935,301 observations (1-km2
polygons of land) and 13 variables. We have run many analyses with random 10% samples to confirm our results are robust to sampling. As noted, Table 1
presents summary statistics for all the aforementioned variables. For our results, we explore their impacts upon land-cover change as well as their significance for locations of PAs.
4.3. Matching Methods
If PAs were sited randomly, PA impacts would be easy to estimate using the differences between deforestation inside and outside of PAs. Deforestation outside would be an unbiased estimate of what would have been occurred inside PAs had there been no protection, as other factors would cancel out. Yet neither PAs in general, nor any PA type, seems to be distributed as if at random. Further, that non-randomness often appears to be along dimensions that can affect deforestation. To isolate PA impact by removing the influences of these differences, we use ‘matching’ methods to improve controls. One ‘matches’ each protected polygon with the most similar unprotected polygon(s) to get as close as possible to ‘apples-to-apples’ comparisons. Thus, PAs are compared not to all unprotected land but, instead, to the most similar unprotected land. Here, we apply propensity-score matching in our many initial analyses and then, for our final results, confirm robustness to covariate matching.
Propensity-score matching assesses the ‘similarity’ of sites using the predicted probability of a polygon being in a PA. PA polygons are then compared to unprotected polygons with similar enough characteristics to yield a similar probability of protection. Probabilities are generated by a probit model using factors in protection and deforestation to explain where protection occurred [52
]. More weight is given, then, to the variables that are important determinants of protection. In covariate matching, similarity is assessed using the distance between polygons in the covariate space. For either method, we matched each treated polygon with the single most similar unprotected polygon.
However, selecting the most similar polygon does not guarantee that controls are, in fact, similar. Thus, we also check explicitly whether the selected unprotected polygons are similar to the protected. We examine balance, i.e., if the characteristics’ values are distinguishable between the protected and matched unprotected observations. Ideally, they should not be. Assuming large differences to begin with, we would expect at least a significant reduction in differences between groups, due to matching.
Given balanced characteristics, the deforestation in matched unprotected sites is an improved estimate of the deforestation that would have occurred at PA locations without the protection. PA impact is calculated as that counterfactual rate minus the observed deforestation (given protection). However, still there will be differences between these groups, in terms of characteristics relevant for deforestation. Thus, our preferred matching estimates involve first matching and then regressions. We refer to the latter as ‘bias adjustment’, as this addresses remaining differences in characteristics.
If the unobservable or omitted factors are correlated with “treatment”, i.e., with where PAs are, that could bias estimates of impact. Matching can control only for the included observable factors. For instance, we do not know anything about the populations in the PAs versus unprotected sites. We suspect that factors we do observe, however, such as the road and city distances, correlate with the unobservables. Thus, given our observables, we cannot be sure of the sign of any residual biases. As a robustness check, we compute ‘Rosenbaum bounds’ to estimate how sensitive are these results.
With two goals in mind, we estimated a suite of PA impacts upon land-cover changes in Mexico. One motivation was to study a period after the 1990s, as conservation politics were alleged to have shifted and prior work had demonstrated that, during the 1990s, PAs functioned as ‘paper parks’. Another was to demonstrate the need to address how land-use decisions imply the possibility of bias in PA impact estimates and how the political economy of public PA choices implies its likelihood.
PAs did reduce losses of natural land cover within their boundaries during 2000–2005. Thus, it would appear that conservation politics shifted, at least as revealed by our impact assessment. Further, it was important in estimating the impacts to check for, and then control for, differences in characteristics of protected (versus unprotected) sites, as they directly influence deforestation. Controlling for them lowers the estimated average impact of PAs by about half. Further, controls generate different adjustments to the estimated impacts for the strict versus the mixed-use PAs.
Across PA types, a combination of siting and enforcement differences yielded different impacts, with strict PAs generating greater reductions in loss within their boundaries than mixed-use PAs. These results may reflect shifts over time in conservation politics, as PA locations seem consistent with a lack of intention to enforce during the 1990s that, as our results suggest, later was reversed.
Extensions could improve on these analyses. Our data for land cover are not ideal and Mexico’s new MRV mechanism may offer options. For moving forward to future analyses, e.g., INEGI has invested in data on roads and vegetation, for example, and the Nature Conservancy has invested in tracking sites and types of protection. Finally, more study of differences in dynamics across space could further contribute to an understanding of the conditions that drive the impacts of protection.