Research ArticleSOCIAL SCIENCES

Conditional cash transfers to alleviate poverty also reduced deforestation in Indonesia

See allHide authors and affiliations

Science Advances  12 Jun 2020:
Vol. 6, no. 24, eaaz1298
DOI: 10.1126/sciadv.aaz1298

Abstract

Solutions to poverty and ecosystem degradation are often framed as conflicting. We ask whether Indonesia’s national anti-poverty program, which transfers cash to hundreds of thousands of poor households, reduced deforestation as a side benefit. Although the program has no direct link to conservation, we estimate that it reduced tree cover loss in villages by 30% (95% confidence interval, 10 to 50%). About half of the avoided losses were in primary forests, and reductions were larger when participation density was higher. The economic value of the avoided carbon emissions alone compares favorably to program implementation costs. The program’s environmental impact appears to be mediated through channels widely available in developing nations: consumption smoothing, whereby cash substitutes for deforestation as a form of insurance, and consumption substitution, whereby market-purchased goods substitute for deforestation-sourced goods. The results imply that anti-poverty programs targeted at the very poor can help achieve global environmental goals under certain conditions.

INTRODUCTION

Two of the great global challenges of the 21st century are to reduce poverty and slow tropical deforestation (1). Yet, there is no consensus on the causal relationships between efforts to alleviate poverty and changes in ecosystems (28). Recently, scientists have argued that it is possible to protect habitats and biodiversity without interfering with, and potentially helping, poverty alleviation goals (912). However, less attention has been paid to whether targeted poverty alleviation programs, with no direct connection to environmental goals, can produce positive conservation outcomes. Science-based evidence for positive outcomes is important because so much international and philanthropic aid is aimed at poverty alleviation. If that investment can also advance conservation, the implication would be globally important because biodiversity and deforestation are disproportionately located in regions with high levels of poverty (13, 14).

Historically, the links between poverty and the environment were studied at two scales: the macroeconomic scale of gross domestic product per capita (15, 16), and the local scale, where communities meet basic needs by exploiting ecosystems (17). These studies typically suffer from one or more shortcomings: (i) weak designs for causal inference; (ii) use of broad indicators of economic growth, rather than of poverty itself; and (iii) failure to isolate the interventions through which poverty changes.

Here, we provide mechanistically grounded, causal evidence that a national-scale, anti-poverty program in the ecologically and economically important country of Indonesia also reduced deforestation as a side benefit. The program uses conditional cash transfers (CCTs), an increasingly popular way to transfer income to poor households. By making the transfers conditional on taking specific actions, usually related to education and health, CCTs aim to enhance human capital and thereby reduce the intergenerational transfer of poverty. Over half of the 30 tropical countries with the most forest cover have CCTs, and others are planning to create them. One well-designed study (18) reported that CCTs in Mexico increased deforestation and argued that this impact arose after the transfers increased the consumption of land-intensive products. Our study finds that different mechanisms appear to dominate in an Indonesian context and these mechanisms could potentially dominate in many parts of Asia due to the importance of rice as a staple crop and growing market access in rural areas.

Study context

Indonesia is one of the biologically richest countries and among the top 10 biodiversity hotspots, with the greatest area affected by poverty (13). It has the third largest area of tropical forest and one of the highest deforestation rates (19, 20), accounting for about 1 in 5 ha of total gross tropical forest loss between 2000 and 2012 (21). Mainly due to this loss, Indonesia is one of the world’s largest emitters of greenhouse gases (22). Rural villages contribute to this loss through agriculture and timber removal (23).

Starting in 2008, rural forested villages began participating in Program Keluarga Harapan (PKH; “Family Hopes Program”), which targets poor households. PKH transfers cash conditional on households meeting health and education obligations. Transfers are approximately 15 to 20% of recipients’ consumption and are received quarterly for between 6 and 9 years (24). Program evaluations report that the transfers are received by poor households and bring them above poverty line expenditure levels (25). Our study comprises 7468 rural forested villages exposed to PKH between 2008 and 2012 across 15 provinces (see Materials and Methods and fig. S1). By 2012, 266,533 households in these villages were receiving transfers.

Study design

CCT programs can affect deforestation through multiple mechanisms (Fig. 1). Among poor agricultural households whose consumption and production decisions are intertwined, and whose access to credit and insurance is limited, cash transfers could change consumption and investment (26). Consumption of forest-intensive goods may increase, which could exacerbate deforestation if those goods are supplied by clearing forests (18), or reduce deforestation if those goods are supplied by conserving forests (16). Cash can also allow households to substitute forest-intensive goods with market-produced goods. The promise of regularly receiving cash could also help households smooth their consumption over time, which could affect how they use forests (e.g., make them less likely to resort to deforestation when they experience negative shocks). CCTs can also affect deforestation through the conditionality obligations themselves or by indirectly encouraging compliance with forest-use laws through reciprocity or greater government scrutiny [for a complementary view of how CCTs can affect environmental outcomes, see (27)].

Fig. 1 The impact of CCTs.

Directed acyclic causal graph in which a single-headed arrow (➔) represents a potential causal pathway between two variables. To identify the causal effect of CCTs on deforestation, an empirical design must eliminate the confounding effects of time-varying and time-invariant community characteristics that are correlated with both deforestation and exposure to CCTs. Such confounders could mask or mimic a causal effect of CCTs on deforestation.

The net effect of CCTs is thus an empirical question, to which the answer requires a design that facilitates unbiased estimation of counterfactual rates of deforestation in the absence of the CCTs. Such estimation is challenging when the cash transfers are not randomized across villages. Care must be taken to control for confounding factors that affect both participation and deforestation (Fig. 1). In Indonesia, for example, PKH and non-PKH villages differ in important economic and environmental attributes (table S1), including their pre-PKH deforestation trends.

To estimate PKH’s causal effect on deforestation, we combined geo-referenced administrative data, remotely sensed panel data on annual forest cover loss (21), a field-verified understanding of how PKH was phased in over time, and statistical methods to estimate counterfactual deforestation rates in the absence of PKH (Materials and Methods). The way in which PKH was phased in over time can be exploited to credibly estimate its causal effects on deforestation in participating villages. PKH targeting was a function of time-invariant subdistrict and household characteristics and of pressure exerted by the central government on PKH administrators to roll out the program evenly across provinces within the annual budget constraint [see the Supplementary Materials (SM)]. This pressure created some variation in exposure to PKH that was unrelated to deforestation; i.e., earlier and later cohort villages have similar PKH exposure propensities, but their enrollment timing differs because of factors not systematically related to deforestation. Another way to express this idea is that, on the basis of our understanding of how the program was rolled out, we assume that, across the country, there are villages that are similar in terms of their poverty, infrastructure, and other potential confounding variables, but some were enrolled early and some enrolled late. Consistent with this claim, earlier and later exposed villages are relatively homogeneous in terms of observable baseline characteristics (table S1), including pre-PKH deforestation trends (fig. S2). Using only the sample of villages that are eventually exposed to the PKH helps control for unobservable factors that determine exposure to PKH and are held in common by all PKH villages. Using panel data on annual forest cover loss further helps control for time-invariant confounders, both observable and unobservable, as well as factors that change over time but not across villages (e.g., national economic and political events). For further control, and to increase the precision of the estimates, we also control for changes in precipitation and temperature over time. Because inferring causality from correlations is challenging in nonexperimental contexts, we also subject our design to a battery of robustness checks.

RESULTS

On average, exposure to PKH is associated with a 30% reduction in forest cover loss [95% confidence interval (CI), 10 to 50%] (Fig. 2), a conclusion that is robust to how "forest" is defined (table S4) and the administrative level at which we conduct the analysis (i.e., villages or the subdistricts into which groups of villages are nested; table S10). If we were to use the count of pixels of forest lost rather than the area in hectares lost (table S4), exposure to PKH is associated with a 16% reduction in forest cover loss (95% CI, 9 to 23%).

Fig. 2 Estimated impact of PKH on forest cover loss in participating villages.

Black bars depict the estimated average effects. Whisker lines depict 95% CIs. SE estimates are clustered at the subdistrict level. In the middle bar, the sample is restricted to participating villages that have primary forest in 2000, and the estimated reduction in forest cover loss is for all forest areas in these villages. In bar at right, the analysis is restricted to the primary forest area in the same villages. For the two bars at right, estimation using the phase-in design is preceded by matching (Materials and Methods).

After augmenting our data with data on primary forest in 2000 (20), we estimate that almost half of PKH’s impact comes from primary forest (Fig. 2). The estimated effects in villages without primary forests are smaller (table S5). The PKH’s estimated effect increases as the density of participating households per hectare of forest increases (table S3). Disaggregating villages by cohorts suggests that the PKH’s deforestation-reducing effect increases with longer exposure, or perhaps that the villages exposed earlier were more responsive to the PKH in terms of reducing deforestation (table S3).

To provide evidence that forest loss trends were not driving exposure to PKH (i.e., reductions in forest cover loss were “leading” PKH exposure rather than “lagging” it), we decompose the PKH’s estimated effect into year-specific effects, with both leads and lags (Fig. 3). PKH impacts are not discernible in villages before PKH exposure (before year 0 in figure), but as soon as a village is exposed to PKH, impacts are discernible, albeit with more noise than in the original specification or the lag specification without leads.

Fig. 3 Estimated impact of PKH on forest cover loss for years before, during, and after implementation, 2001–2012.

Black squares depict the estimated year-specific, average effects during the 4 years before PKH implementation (leads), the year of implementation, and the 4 years after implementation (lags). Whisker lines depict 95% CIs. SE estimates are clustered at the subdistrict level.

We find no statistical evidence in favor of spillovers to non-PKH, neighboring villages (table S13). We do, however, detect evidence that PKH’s impact is moderated by variables known to be correlated with deforestation (tables S7 and S8). In other words, the PKH’s effects on deforestation appear to be heterogeneous, a feature that we exploit later to shed light on the mechanisms through which PKH could reduce deforestation. The estimated effect in Fig. 2 ignores the potential effect of PKH on forest regrowth, but there is little evidence of regrowth in the study villages during pre-PKH and post-PKH periods (see SM).

Like all nonexperimental designs, our design is potentially subject to hidden biases. We assess these potential biases through a decomposition of our estimator, alternative model specifications, alternative estimators, and sensitivity tests to hidden bias (Materials and Methods). First, as described elsewhere (28), our phase-in design estimator can be decomposed into a weighted average of all two-group/two-period difference-in-differences (DDs) estimators in the data. Thus, if the PKH’s effects were time-varying, our phase-in estimator could be biased. By time-varying effects, we mean that the PKH’s impacts on forest cover grow or shrink the longer a village is exposed to the program. The bias associated with this pattern in the impacts increases with increases in the weights that our estimator assigns to the DDs that contrast later cohorts to earlier cohorts. In our study, the weights assigned to the later-to-earlier comparisons are small (fig. S3), implying that if a bias exists, it is small. Furthermore, the patterns in Fig. 3 and table S3 suggest that, if anything, the PKH’s impact grows over time. That trend would imply that the bias in our estimator is toward zero, not away from it (for details, see SM).

Next, we try alternative estimation strategies (tables S9 and S10). First, to further control for potential unobserved, time-varying confounders, we allow for forest cover trends to differ by village and by province (Fig. 4). We find that the estimated effect of the PKH is smaller, in absolute value, than the original estimate by about 20%, but still well within the CI of the original estimate. Second, we preprocess the data by covariate matching, which makes early and late cohort villages even more similar in terms of observable, time-invariant, and time-varying attributes that are believed to be correlated with deforestation trends in the absence of PKH (table S1). The estimated effect is about 8% larger in absolute value than the original estimate (Fig. 4). Third, we use an alternative estimator, the lagged dependent variable model, which assumes that exposure to PKH is determined by time-varying confounders that are correlated with prior forest cover loss; e.g., villages could be more likely to be exposed to PKH after a negative shocks that affected how much land they could clear in the prior year. The estimated effect is about 2% smaller in absolute value than the original estimate (Fig. 4). Fourth, we use an alternative estimator, the generalized synthetic control (GSC) model, which allows one to estimate treatment effects as they evolve over time, without imposing linearity. The GSC method creates a reweighting scheme among never-treated villages to predict the post-PKH counterfactual trends in forest loss among the PKH villages. This estimator yields an estimated treatment effect that is larger in absolute value than our original estimate (table S9) and, like the year-specific estimates (table S3), suggests that the PKH’s effect may grow over time. Overall, the alternative estimation strategies yield results that are consistent with the claim that exposure to PKH is associated with reductions in deforestation.

Fig. 4 Assessing rival explanations for estimated PKH impacts.

If the estimated impact in Fig. 2 reflects the true PKH impact, rather than hidden biases in our design, we would expect that (i) the estimates from the flexible time trend (village linear time trend and province by year dummy variables), the matched design, and lagged cover loss design imply an impact similar to the estimated impact in Fig. 2, and (ii) the estimate from the sensitivity test, which is a stress test that assesses how much our estimate would change if there were a powerful unobserved confounder in our design, implies a deforestation-reducing effect of the PKH, albeit smaller than the estimate in Fig. 2. Estimated percent reduction is calculated by dividing the estimated reduction in forest cover loss by the (weighted) average annual forest cover loss in the villages that provide the estimate of counterfactual loss. Black bars depict the estimated average effects. Whisker lines depict 95% CIs. SE estimates are clustered at the subdistrict level.

Our final robustness check, a sensitivity test to hidden bias, is most informative. Our estimated reduction in deforestation by the PKH could have been generated by one or more unobserved, time-varying confounders that were negatively associated with early exposure to PKH and positively associated with forest loss. Even after adjusting for a strong form of this potential bias (Materials and Methods), the estimated impact remains negative, statistically significant, and about half the size of our original estimate (Fig. 4 and table S12).

To show that empirical design matters, we contrast our estimated impact with estimates from popular designs in the sustainability science literature for estimating causal relationships (Materials and Methods). These designs are more limited in their ability to eliminate the effects of confounding variables (29). Estimates from these designs range from a 71% decrease in forest loss to a 60% increase (table S6), highlighting the importance of careful control for confounding variables when estimating the impacts of anti-poverty programs on the environment.

DISCUSSION

Estimating causal effects from observational data is challenging. Nevertheless, the analysis suggests that the estimated negative correlation between PKH and forest loss can plausibly be interpreted as reflecting a causal relationship. To provide evidence on which mechanisms were operative (Fig. 1), we use experimental and nonexperimental data and predictions from theory (Materials and Methods). We use data from a 2007 randomized pilot of the PKH, which was largely in agricultural communities with little or no forest, to assess the likelihood that the conditionality mechanisms were operative. In the evaluation of that pilot, no effects on child labor (agricultural or otherwise) were detected. Moreover, the detectable health effects were small in the third year and waned by the sixth year, a pattern that does not match the pattern of the PKH treatment effect, which appears immediately in the first year and persists over time (see SM for more details). Thus, although the labor and health responses of households in the pilot may have differed from the responses of households in our sample, the evidence does not support the hypothesis that PKH affected deforestation through conditionality obligations.

In contrast, the evidence from our sample of 7468 villages more strongly supports two other channels: consumption smoothing, whereby cash substitutes for deforestation as a form of insurance for poor households, and consumption substitution, whereby market-purchased goods substitute for deforestation-sourced consumption goods. As in other nations, deforestation in Indonesia serves as an investment in insurance (or a coping strategy) against weather shocks: when the rice season rains are delayed, poor farmers anticipate lower yields and deforest more land to compensate for the higher probability of a negative consumption shock. PKH is most effective at reducing deforestation in years when a village experiences negative rainfall shocks (table S14), a pattern consistent with transfers acting as a consumption smoothing mechanism. Moreover, PKH is most effective at reducing deforestation in communities with better access to markets (table S14), a moderating effect that is consistent with recipients using PKH cash to substitute deforestation-sourced goods with market-purchased goods, but not consistent with changes in demand for forest-intensive products or with changes in productive investments.

Using the same data, we find weak evidence for the reciprocity channel: The likelihood of citizens voting in favor of an incumbent candidate in the 2008–2012 period increases with the proportion of a district’s villages that is exposed to PKH (table S15). That relationship, however, does not necessarily imply that villagers were more likely to obey government regulations. With regard to villagers obeying government regulations via the scrutiny channel, we do not have empirical evidence for or against the scrutiny channel. Nevertheless, multilateral evaluations of PKH and respondents in our field interviews report that the facilitators who monitor conditionalities and payments had trouble doing their core mission because of insufficient funds, large travel distances, and other constraints. Their ability to also serve as enforcers of forest use rules was therefore limited (see SM for more details).

There are, of course, other potential mechanisms through which PKH may have affected forest outcomes. For example, a recent study reported that migration led to greater reforestation in Nepal (30). In the PKH context, cash transfers could have been used to facilitate the migration of some family members, which could have reduced deforestation through, for example, cash remittances and reductions in the local labor supply. With new data and creative testing strategies, future studies may identify other mechanisms through which PKH affected forest cover outcomes.

CONCLUSION

Does reducing poverty have unavoidable environmental costs? In Indonesia, we find evidence that, under some conditions, the answer to this question is “No.” Although the PKH program was not designed to protect the environment, it appears to have reduced deforestation by about 1/10 of an SD. For comparison, in a review of payments for environmental service programs designed explicitly to reduce deforestation, the median effect size was 0.12 SD, and those programs have been criticized for not reaching the very poor (31).

Moreover, the mechanisms through which the program appears to have achieved its environmental impact are not unique to Indonesia. Throughout the tropical world, poor households depend on deforestation as a means to acquire consumption goods and to provide insurance against negative shocks. Whether our results generalize to other settings will depend on role of these mechanisms and other competing ones. Nevertheless, our study thus offers some hope that global efforts to eradicate extreme poverty need not have unavoidable environmental costs.

Although the PKH was not an environmental program, one can evaluate how the benefits of PKH’s avoided CO2 emissions compare to its implementation costs (Materials and Methods). Depending on assumptions about the persistence of PKH’s effects on deforestation, the economic value from reduced carbon emissions alone could cover the PKH implementation costs (Table 1).

Table 1 Benefit-cost analysis of PKH for carbon storage.

The social benefits of delayed CO2 emissions are compared to PKH costs, both measured per MT of averted CO2 (Materials and Methods). No other benefits from the PKH program are included (e.g., from poverty alleviation or non-CO2, ecosystem services). In the first four scenarios, all CO2 is assumed to be emitted immediately after the trees are cut; delayed emissions would increase the benefit-cost ratios. The scenarios vary assumptions about how long the avoided CO2 emissions are delayed. Given current forest cover loss rates, the average village will have deforested all forests unaffected by the PKH in 50 years. Scenario 2 assumes that villages begin deforesting the PKH-protected forests at the historical rate after 50 years. Scenario 3 assumes a delay for only the maximum PKH eligibility period, while scenario 4 assumes a delay for the minimum eligibility period.

View this table:

Our study does not, however, imply that conservation funds should be redirected to poverty alleviation. Drawing such a conclusion requires analyses, similar to Table 1, for a larger set of environmental and social values across a range of environmental and anti-poverty interventions. Our study, with its emphasis on better design for causal inference and on elucidating multiple mechanistic pathways, offers one exemplar for future research in that direction.

MATERIALS AND METHODS

To ensure sufficient detail to evaluate and replicate our results, we provide a detailed SM section. All data and code used in this study can be found at https://osf.io/uzaxg/. Here, we provide an outline of the key aspects of Materials and Methods.

Determination of PKH assignment across villages (section S1)

The key task in any observational design is to uncover the treatment assignment mechanism. To understand how the PKH was implemented, we used program documentation from government and nongovernment sources, supplemented with field surveys with PKH program administrators in Jakarta, and field surveys of heads of villages in forested areas in Bali, Sumatera, and Kalimantan.

Sample (sections S2a and S2b)

We use a panel data set that runs from 2001 to 2012. We end the period in 2012 for two reasons: (i) the eligibility criteria used to construct the list of PKH-eligible households changed in 2013, and (ii) the way in which tree cover was measured from satellite data changed in 2013. Causal inferences are most credible if the criteria for determining eligibility do not change during the study period. Our final sample consists of 15 provinces, distributed across Sumatera, Kalimantan, Bali, and West Nusa Tenggara Islands, representing 53% of Indonesia’s forests with 75% or more canopy cover in 2000 and accounting for over 80% of the forest cover loss between 2000 and 2012 (fig. S1). After applying inclusion and exclusion criteria for villages, we had a sample of 26,574 villages, of which 7468 villages received the PKH between 2008 and 2012, and 19,106 villages had never received the PKH by the end of 2012. The number of villages that are newly exposed to PKH each year is 1251 in 2008, 307 in 2009, 633 in 2010, 1624 in 2011, and 3653 in 2012.

Data (section S2c)

Our outcome variable is forest cover loss, for which our measures come from Hansen et al. (21). To avoid including losses from non-native forests (e.g., rubber plantations), we defined “forests” as pixels with tree cover canopy density greater than 75%, with a robustness check using a 30% threshold (table S4). To designate forest in our sample as “primary,” we use polygons of Indonesia’s primary forest in 2000 from Global Forest Watch (GFW), which are based on data from Margono et al. (20). To calculate regrowth, we use Vegetation Continuous Fields data from Song et al. (32). Our treatment variable is a binary variable indicating whether a village was exposed to the PKH program in a given year. Its value was obtained by merging Indonesian village boundaries data from Statistics Indonesia with the village indicator of PKH recipients from The National Development Planning Ministry (BAPPENAS). We also use alternative definitions of treatment in various measures of “exposure intensity,” which are based on BAPPENAS annual data on the number of participants in each village. To capture temporal variation in weather, we use data on average annual surface temperature and average precipitation during the wet season (end of September until end of December) from NASA. We obtained polygon data of protected areas from World Database of Protected Areas in 2016. Palm oil, wood fiber, and logging concessions polygons were obtained from GFW. For the subdistrict attributes that were used to assign PKH eligibility, we obtained data from the National Team for the Acceleration of Poverty Reduction. For elevation and slope data, we used SRTM 90m Digital Elevation Database version 4.1. For road data, we use data from the Socio Economic Data and Applications Center at Columbia University. For distance measures, we used ArcPro 2.4.

Estimand and estimators (sections S2d to S2g)

The estimand (the average treatment effect on the treated), the main estimator, and the required assumptions for inferring causality from the estimated correlations are described in detail in section S2d. In order to estimate the effect of the PKH on deforestation in participating villages, we use a linear, additive, two-way fixed-effects panel data estimator, using only the sample of 7468 villages that were exposed to the PKH sometime between 2008 and 2012 (equation in section S2d). We used the same estimator when the unit of analysis was the subdistrict level (N = 521). Variants of these estimators were used to estimate how the PKH effect varies by type of forest (sections S2e and S2f) and by intensity of exposure (section S2g). To implement the same estimator with counts of pixels rather than hectares, we use an inverse hyperbolic sine transformation (because of the zero forest loss values in the panel). We implemented these estimators using the “xtreg fe” command in Stata 16 MP. All statistical tests are two-sided. Detailed results are found in tables S3 and S4.

Traditional designs and estimators (section S3)

We implement (table S6) (i) a “before-after” design, which uses longitudinal data to infer the effect of PKH by detecting a shift in deforestation between pre-PKH and post-PKH periods; (ii) a “with-without” design, which seeks to infer the effect of PKH by looking at post-PKH cross-sectional differences in deforestation between PKH and non-PKH villages; and (iii) a sophisticated version of the “before-after-control-impact” design, also called a DD design, which seeks to combine the designs of (i) and (ii) to infer the effect of PKH by contrasting the trends in deforestation for PKH and all non-PKH villages; we call this design the fixed-effects panel data design using the full sample of villages. We implemented these estimators using Stata 16 MP.

Heterogeneity in PKH impact (section S4)

To test for the heterogeneity of treatment effects, we interact the PKH variable with characteristics that are believed to be moderators of forest use (tables S7 and S8). Other tests of heterogeneity are used in the mechanism analysis.

Rival explanation robustness checks (section S5)

We conduct five supplementary analyses to probe the credibility of the assumptions required to infer causality from correlation in the main analysis. To decompose our two-way, fixed-effects estimator into its constituent 2 × 2 DDs and weights, we use the DDtiming command in Stata 16 MP (fig. S3). To allow for unit-level time trends and fully flexible province-by-year effects (table S10), we use a variant of our two-way, fixed-effects estimator in which we add linear and/or quadratic unit-level time trends and interaction terms of the year and province dummy variables. To preprocess the sample using matching, we use the GenMatch command in R (tables S1, S9, and S11). We then use the matched data with our two-way, fixed-effects estimator from section S2. To estimate the lagged dependent variable (lagged forest cover loss) model, we use the xtreg command in Stata 16 MP (table S9). Assuming that lagged forest cover loss is positively associated with exposure to PKH and positively associated with current forest cover loss, this estimator and the fixed-effects estimator provide bounds on the true PKH impact. To implement the GSC estimator (33), we used the gsynth command in R (3.6) (table S9). To test for the sensitivity of our results to hidden bias (an unobservable confounder), we followed a method introduced by Altonji et al. (34) and further developed by Oster (35). We assume that the relationship between PKH and an omitted, unobserved confounding variable can be recovered from the relationship between the PKH and the observable variables in our model (e.g., village fixed effects). Using the areg command followed by a post-estimation psacalc command in Stata 16 MP, we then explore the conditions under which the omitted variable could substantially change our conclusions, as well as calculating a lower and upper bound on the treatment effect if this unobserved variable were to explain all unexplained variation in exposure to PKH and forest cover loss (table S12).

Spillovers (section S6)

To detect spillovers from PKH villages to neighboring non-PKH villages, we use the sample of 19,106 never-treated villages. Using ArcPro 2.4, we identify whether a never-treated village shared administrative boundaries with a PKH village. Using the main fixed-effects panel data estimator and two ways of assessing the intensity of exposure to neighboring PKH villages, we estimate the localized spillover effects on forest cover loss (table S13).

Mechanisms (section S7)

To shed light on mechanisms, we combine theory, the results from evaluations of randomized 2007 pilot PKH program (25), and additional analyses with our nonexperimental data. We explore the implications for each mechanism listed in Fig. 1, and we test those implications against our observational and experimental data (tables S14 and S15).

Benefit-cost analyses (section S8)

To calculate the metric tons of CO2 equivalent emissions per hectare in the PKH communities’ forests, we use GFW data. We estimate the average aboveground biomass (AGBM) from provinces in our sample with canopy cover greater than 30% in 2000 and then multiply that value by 0.5 to obtain carbon per ha, and then by 3.67 to convert it to CO2 per ha. We conservatively estimate that 75% of the AGBM is lost soon after clearing in the form of greenhouse gas emissions, and the remaining 25% is permanently sequestered in wood products. We use government and multilateral donor data to determine the average cash transfers and the average cost of delivering the cash and monitoring conditionalities per village per year. We calculate the value of delaying a metric ton of CO2 equivalent emissions using a formula that takes into account the social cost of CO2 (assumed to be USD $31, the U.S. Environmental Protection Agency’s estimate for the year 2010 using a 3% discount rate social cost of CO2), the duration from deforestation to carbon emissions, the effective discount rate, and the delay in deforestation induced by the PKH program.

SUPPLEMENTARY MATERIALS

Supplementary material for this article is available at http://advances.sciencemag.org/cgi/content/full/6/24/eaaz1298/DC1

This is an open-access article distributed under the terms of the Creative Commons Attribution-NonCommercial license, which permits use, distribution, and reproduction in any medium, so long as the resultant use is not for commercial advantage and provided the original work is properly cited.

REFERENCES AND NOTES

Acknowledgments: P.J.F. and R.S. thank J. Alix-Garcia, A. Balmford, P. Kareiva, C. McIntosh, P. Stephan, B. Meiselman, K. Sims, C. Weigel, and S. Wunder for helpful comments on earlier versions and V. Yulaswati, BAPPENAS, for PKH data. Funding: R.S. gratefully acknowledges financial support from the UNCE project (UNCE/HUM/035) and AYSPS Dissertation Fellowship. Author contributions: P.J.F. and R.S. designed the study, interpreted the results, and wrote the final manuscript. R.S. supervised data collection and data analysis. Competing interests: The authors declare that they have no competing interests. Data and materials availability: All data needed to evaluate the conclusions in the paper are present in the paper and/or the Supplementary Materials. Additional data related to this paper may be requested from the authors. All data and code used in the analysis are available at https://osf.io/uzaxg/.
View Abstract

Stay Connected to Science Advances

Navigate This Article