Impact of protected areas on poverty, extreme poverty, and inequality in Nepal

Protected areas (PAs) are key for biodiversity conservation, but there are concerns that they can exacerbate poverty or unequal access to potential benefits, such as those arising from tourism. We assess how Nepalese PAs influence poverty, extreme poverty, and inequality using a multidimensional poverty index, and a quasi‐experimental design that controls for potential confounding factors in non‐random treatment allocation. We specifically investigate the role of tourism in contributing to PA impacts. Nepali PAs reduced overall poverty and extreme poverty, and crucially, did not exacerbate inequality. Benefits occurred in lowland and highland regions, and were often greater when a larger proportion of the area was protected. Spread of benefits to nearby areas outside PAs was negligible. Furthermore, older PAs performed better than more recently established ones, suggesting the existence of time lags. Although tourism was a key driver of poverty alleviation, PAs also reduced extreme poverty in areas with fewer tourists.


INTRODUCTION
Protected areas (PAs) are key conservation strategies but also have socioeconomic impacts on people living in and around them (Brockington & Wilkie, 2015). PAs limiting anthropogenic activities can harm local economic development (Brockington & Wilkie, 2015), but can also safeguard ecosystem services that local communities depend on, and generate additional sources of income, for example, through tourism (Ferraro & Hanauer, 2015). Some studies find that PAs are linked to high poverty levels (de Sherbinin, 2008;Fisher & Christopher, 2007), but such associations can be confounded because PAs are often located in areas with limited development potential (Joppa & Pfaff, 2009). There are, therefore, growing efforts to assess PA outcomes using techniques This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. that control for this non-random allocation of PAs. Such studies provide increasing evidence that PAs can reduce poverty, albeit with much heterogeneity in effect sizes (Andam, Ferraro, Sims, Healy, & Holland, 2010;Hanauer & Canavire-Bacarreza, 2015;Miranda, Corral, Blackman, Asner, & Lima, 2014;Sims & Alix-Garcia, 2017;Yergeau, Boccanfuso, & Goyette, 2017).
Despite this progress, several topics remain understudied. First, PA assessments have focused primarily on mean poverty outcomes across entire communities (but see Sims, 2010). However, PA's financial benefits may suffer from elite capture (Agrawal & Gupta, 2005), leading to greater inequalities. Assessing the mechanisms through which PAs influence poverty is essential. PAs may increase tourism opportunities leading to improved local income and employment (Walpole & Leader-Williams, 2001). Assessing tourism impacts across large spatial extents is often limited by data availability, and assessments have predominantly used binary proxies (presence or absence of tourism infrastructure, Ferraro, Hanauer, & Sims, 2011). Yet to gain a better understanding of how tourism contributes to local poverty alleviation, it is important to move beyond binary assessments of tourism and consider variation in the intensity of tourism in PAs (Robalino & Villalobos, 2015). Finally, there is substantial spatial variation in the proportion of land surrounding a community that is protected, the duration that it has been protected for, and livelihood opportunities that are constrained by a series of factors, such as slope and elevation that influence agricultural suitability (Gentle & Maraseni, 2012). These factors can influence the magnitude, and possibly even the direction of PA effects on poverty.
Here, we assess how PAs in Nepal influence multiple measures of poverty. We combine national census-derived poverty estimates for 2001 and 2011, and use statistical matching to construct a counterfactual group. We build upon previous research by (a) quantifying how PA status influences measures of extreme poverty and inequality, in addition to overall measures of poverty, (b) using tourism indicators to assess if tourism is an important mechanism through which PAs influence poverty, and (c) testing whether effects of PAs on poverty are moderated by variations in the amount of protected land, time since establishment, and elevation.
Nepal provides a good case study to assess the effects of PAs on multiple poverty outcomes. It is one of the poorest Asian countries (Alkire & Foster, 2011) and has an extensive PA network, covering 20% of the country's land surface. Nepalese PA policies were first characterized by a strict "fences and fines" approach (Heinen & Shrestha, 2006), which denied local people's user rights. However, during the 1990s several important pieces of legislation were passed to promote social welfare including redistribution initiatives to minimize inequality by spending 30 to 50% of PA revenues on community development (Spiteri & Nepal, 2008).

Data
We compiled a high spatial-resolution, national-level data set using 3,845 of Nepal's 3,973 Village Development Committees (VDCs), the subdistrict level administrative unit, as our unit of analysis.

Poverty metrics
We use household health, education, and living standards data from the Nepali national censuses of 2001 and 2011 to develop three multidimensional poverty (MDP) measures based on the MDP index developed by Alkire and Fos-ter (2011): poverty (MDP > 0.33-following the cutoff of Alkire and Foster [2011] for measuring poverty); extreme poverty (MDP > 0.66-this doubles the standard poverty threshold, following other studies [e.g., Lokshin & Ravallion, 2000] and indicates that at a minimum a household is completely deprived in one of the three poverty dimensions and partially deprived across the remaining two dimensions); and inequality-measured as the standard deviation of the incidence of household poverty (Supporting Information Figure S1; Figure 1a). Using alternative thresholds for defining extreme poverty either generates too few VDCs that contain extreme poverty (70% threshold-314 VDCs using 2001 baseline data compared to 1,153 with a 66% threshold) or generates qualitatively identical results and conclusions (60% threshold, Supporting Information Figure S2).

Defining PA treatments
We define protected treatments as VDCs that overlap Nepal's 32 PAs (IUCN categories II-VI, Nepal lacks category I PAs) using the World Database on Protected Areas (IUCN & UNEP-WCMC, 2016; Figure 1b). The vast majority of these are multiple-use PAs. We conduct two separate analyses: one focusing on PAs established before 2001 (the baseline year of our poverty data), and one focusing on PAs established between 2001 and 2011. We conduct this second analysis as a robustness check because PAs established prior to 2001 could affect our baseline measures, although baseline poverty metrics were similar in VDCs that were protected before and after 2001 (see Figure 2). We also defined protected VDCs using two separate definitions: those with (a) at least 10% of their area overlapping with a PA (e.g., Andam et al., 2010;Hanauer & Canavire-Bacarreza, 2015) and (b) at least 70% of the VDC being protected (which is close to the mean percentage overlap for overlapping VDCs-PAs established before 2001 = 65.2%; PAs established after 2011 = 71.4%). VDCs with <1% of their area protected were defined as nonprotected to ensure a clear distinction in the magnitude of protection between control and treatment VDCs.

Tourism metrics
We assessed how PAs with different tourism intensities impacted our outcome variables, using data on official tourism numbers for each PA in 2011 (low < 10,000 visitors; intermediate 10,000-100,000; high > 100,000; Ministry of Culture, Tourism and Civil Aviation, 2013). We also assessed how proximity to a PA entrance and trekking routes (categorized as major or minor; Supporting Information Table S9) contributed to heterogeneity in PA impacts using a mean travel time estimate (weighted by population density) from each VDC to the nearest PA entrance, and major and minor trekking routes (Supporting Information Figure S3).

F I G U R E 1 Poverty and protected areas. (a) Multidimensional poverty in 2011. Each polygon represents a Village Development Committee
(VDC). Data are presented as deciles. Grey areas with red contours represent excluded VDCs (reasons for exclusion include missing data due to armed conflict and instances of inconsistent data from the Nepali department of forests). (b) Schematic map of protected VDCs in Nepal (using the 10% threshold). Data from the world database of protected areas. In our analysis we included 192 VDCs that were protected before 2001 (of which 110 were protected using the 70% threshold definition), and 106 VDCs that were protected between 2001 and 2011 (of which 67 were protected using the 70% threshold)

Confounding factors
We selected a suite of biophysical and socioeconomic covariates based on their potential to influence the outcome or the relationship between treatment and outcomes. These covariates were baseline levels of our poverty measures, slope, elevation, precipitation, VDC area, forest cover, travel time from the VDC to population centers and district headquarters, proportion of the VDC under community forest management and the age of community forestry arrangement, population density, agricultural effort, international migration, and district (Supporting Information Table S3).

Matching and post-matching analyses
We used a combined matching-and regression-based approach to explore the causal link between PAs and poverty outcomes. We model poverty metrics in 2011 while    (2001) of VDCs, thick lines (T) represent treatment, that is, PAs, thin lines (C) represent counterfactual controls without protection. Significance: **** P < 0.001; *** P < 0.01; ** P < 0.05; * P < 0.1 controlling for baseline poverty in 2001 to avoid constructing models that can generate spurious correlations (Brett, 2004). This approach yields similar parameter estimates for our treatment variables as those generated when modeling absolute change (Supporting Information Figure S2). The preprocessing of data using matching methods optimizes the balance of covariates across treated and control units, and is useful when imbalance between treatment and control is an issue for traditional causal inference techniques (Ho, Imai, King, & Stuart, 2007). We used genetic matching with replacement, which performs well when covariates have skewed distributions (Diamond & Sekhon, 2013). We performed all of our statistical analyses in R version 3.3.2 (R Core Team, 2013) using the "Matchit" package (Ho et al., 2007). We used post-matching standardized mean differences of <0.25 as an acceptable balance between treatment and control groups for each covariate (Stuart, 2010, see Supporting Information Figures S3-S5). We then performed an Ordinary Least Squares (OLS) regression to adjust for remaining imbalances in covariate distributions (Ho et al., 2007). When modeling extreme poverty, we implemented a two-step hurdle model (Cragg, 1971) using matched binomial regressions to first model the incidence of extreme poverty, and then OLS regressions to model the magnitude of extreme poverty in those VDCs in which extreme poverty occurs. We first measured the average impact of our treatments (protection) on our response variables (poverty, extreme poverty, and inequality in 2011). We then subset and separately matched PAs in each tourism intensity category (high, intermediate, or low) to assess the impact of tourism intensity. PAs with high tourism levels were all designated before 2001, so we only performed this subgroup analysis on PAs established before 2001. We conducted robustness checks to test for spillover effects from unprotected VDCs adjacent to a PA (defined as the treatment) into unprotected control VDCs that are not adjacent to a PA (Supporting Information Figure S4), and spatial autocorrelation (Supporting Information Figure S5); results are robust to spillover and spatial autocorrelation unless stated otherwise.

Heterogeneity analysis
We assessed if PA impacts were moderated by travel time to the nearest tourism hub (PA entrance, major and minor trekking route) and elevation, which affects livelihood choices (greater range of options in the lowlands, including commercial agriculture) and tourism options (safaris in the lowlands, trekking in the mountains). We used partial linear modeling (PLM; Hanauer, 2015;Yatchew, 1998) to assess heterogeneous impacts along the gradients of our moderating factors following methods described in Ferraro et al. (2011) and Hanauer and Canavire-Bacarreza (2015). In a first step, we controlled for confounding factors using a linear regression. In the second stage, we employed a nonparametric locally weighted scatter plot smoothing to estimate the nonparametric relationship between moderator and outcome. This method allows us to estimate the impact of PAs on our outcome variables as a function of our moderator variables of interest (elevation and travel times to the nearest PA entrance, major and minor trekking routes) while holding other covariates constant.

Average impact on poverty, extreme poverty, and inequality
We found no evidence that PAs exacerbated poverty in Nepal. In fact, matched-protected VDCs (defined using the 10% threshold and established before 2001) had significantly lower poverty in 2011 than unprotected VDCs (coefficient = −0.03, SE = 0.02, P = 0.027; Figure 2a). Poverty was not exacerbated when raising the protection threshold to 70% (coefficient = −0.06, SE = 0.03, P = 0.060; Figure 2a). For PAs established after 2001 we found no evidence of positive or negative impacts of PAs on overall poverty (Figure 3a). Models without matching showed similar patterns (Supporting Information Table S7).
PAs established before and after 2001 reduced the incidence of extreme poverty. For PAs established before 2001, this result was significant for our 10% protection threshold (coefficient = −0.95, SE = 0.38, P = 0.012; Figure 2b) and was accentuated by raising the threshold to 70% (coefficient = −3.51, SE = 1.20, P = 0.003; Figure 2b). For PAs established after 2001, this result was not significant using a 10% protection threshold, but was significant after raising the protection threshold to 70% (coefficient = −2.82, SE = 1.18, P = 0.018; Figure 2b). We found no significant impact of protection on the magnitude of extreme poverty (Supporting Information Table S8). Results from models without matching showed the same patterns for PAs established before and after 2001 (Supporting Information Table S7).
We found no consistent evidence that inequality was influenced by PAs established before or after 2001, using either 10 or 70% protection thresholds (Figure 2c). Models without matching indicate that PAs established before 2001 reduced inequality, while PAs established after 2001 increased inequality (Supporting Information Table S7), but these difference were not significant after controlling for spatial autocorrelation (Supporting Information Table S6).

Tourism intensity
PAs with high tourism levels significantly reduced overall poverty (coefficient = −0.05, SE = 0.02, P = 0.023; Figure 2a), while PAs with low tourism levels had no

Heterogeneity: Travel time to PA entrance and trekking route
Travel time to a PA entrance had no impact on poverty (Supporting Information Figure S7) and inequality (Supporting Information Figure S6), while reductions in the incidence of extreme poverty were greater closer to a PA entrance (Figure 3d). Travel time to a minor and major trekking route moderated the influence of PAs on poverty, with significant reductions only occurring in VDCs close to the trekking route (Figure 3a,b). Incidence of extreme poverty was lower further away from a minor trekking route (Figure 3d), but was not influenced by travel time to a major trekking route (Supporting Information Figure S7). Inequality was not influenced by proximity to major or minor trekking routes (Supporting Information Figure S6).

Heterogeneity: Elevation
Our PLM results do not show significant heterogeneous impacts of PAs on extreme poverty and inequality as a function of elevation (Figure 3f; Supporting Information Figure S6). PAs established before 2001 reduced poverty to a greater extent at low elevations than high elevations ( Figure 3e).

DISCUSSION
Nepali PAs typically reduced poverty, concurring with previous research elsewhere (Andam et al., 2010;Hanauer & Canavire-Bacarreza, 2015). Crucially, PAs reduced extreme poverty without deepening inequalities. This finding is particularly important as creating pathways out of extreme poverty is more difficult than tackling less extreme poverty (Halder & Mosley, 2004). Our findings suggest that PAs are able to provide pathways out of extreme poverty in remote areas, challenging previous evidence that PA policies only benefit community elites (Agrawal & Gupta, 2005).
PAs with high tourism levels reduced poverty without exacerbating extreme poverty and inequality, while PAs with low tourism levels reduced extreme poverty and inequality but had no impact on overall poverty. These results suggest that the poorest receive the greatest benefits from small-scale tourism, contrasting with previous suggestions that tourism increases inequalities (West, Igoe, & Brockington, 2006). We provide further evidence for beneficial impacts from tourism by showing that poverty reductions in PAs only occurred close to trekking routes. This suggests that redistribution policies (that 30-50% of PA revenue is spent on local community development; Heinen & Shrestha, 2006) may not fully address spatial biases in which communities benefit from tourism in PAs. Notably, however, the impact of PAs on reducing extreme poverty increased with distance from minor trekking routes that are typically located in remote areas with little development potential that can benefit from park redistribution policies. Future studies should specifically assess if, where and how these policies influence PA poverty outcomes.
Distance from PA entrances had no impact on extreme poverty inside PAs, but increased extreme poverty outside PAs. This suggests localized negative spillovers, with PA residents living close to PA entrances receiving benefits that people living equally close to entrances outside of the PA miss out on. Other research on PA spillover effects show similar patterns of heterogeneity Robalino, Pfaff, & Villalobos, 2017), with tourism benefits only occurring close to PA entrances (Robalino & Villalobos, 2015). Indeed, our analyses indicate that benefits of protection do not spread to neighboring unprotected VDCs. Redistribution policies might thus need to target communities inside and outside PA more equally.
Time since establishment moderated the effect of PAs on our measures of poverty. PAs established after 2001 did not show the same significant social benefits as PAs established before 2001, although in newer PAs we observe a trend toward lower extreme poverty and inequality. This pattern is expected if there are time lag effects that arise because communities need to adjust to new regulations imposed by PAs and the new opportunities provided by them, and for the tourism industry to develop. The reduced benefits of more recently established PAs are unlikely to be associated with changes in management regimes as these have been constant across all Nepali PAs since the 1990s (Bhattarai et al., 2017). Notably, an increase in the threshold used to define a protected VDC (10-70%) accentuated our main findings. This suggests that communities in VDCs that have restrictions placed on activities across a larger proportion of their land do not experience adverse impacts on poverty metrics, thus larger PAs may deliver greater economic benefits. Finally, impacts of PAs were similar across a wide range of elevations indicating that PAs can deliver socioeconomic impacts even in areas that typically support livelihoods that are less compatible with nature conservation, such as agriculture.
Our study makes a number of important contributions. First, we demonstrate not only that PAs in Nepal reduce poverty and extreme poverty, but that they do so without increasing inequality. These benefits occur even in lowland regions with high capacity for alternative land uses, and when capacity for alternative livelihoods is reduced by protecting larger proportions of land. Second, we find that tourism is a key driver of PA benefits, but that reductions of extreme poverty are possible even in marginalized areas with limited tourism potential. Finally, we find no evidence that socioeconomic benefits of PAs spread to people living outside, but close to, PAs. Addressing this by adjusting PA's revenue redistribution policies, could increase the benefits for these communities and reduce conflict between local communities and PA's conservation objectives (Oldekop, Holmes, Harris, & Evans, 2016). Nepal's PA management policy to promote social welfare via redistribution of PA revenues, gained through tourism and other activities, is similar to policies in other countries including Thailand (Sims, 2010) and Kenya (Walpole & Leader-Williams, 2001), suggesting that our findings may also apply elsewhere.