Poverty alleviation resettlement in China reduces deforestation

Significance

Poverty-driven deforestation has raised widespread concerns globally. This study deepens scientific understanding of how structured resettlement programs have contributed to forest conservation. We assess the ecological impact of China’s Poverty Alleviation Resettlement program by comparing changes in forest quality before and after the policy for participants and nonparticipants, and find a significant decline in county-level deforestation and households’ logging behaviors among program participants having broad implications on forest ecosystems and climate change. The study reveals the multifaceted mechanisms by which poverty alleviation through relocation mitigates deforestation, in addition to remoteness from forest, involving market access, off-farm employment, and income. It highlights that comprehensive support measures that help change households’ livelihoods and lifestyles are pivotal to the success of resettlement programs.

Abstract

Deforestation frequently accompanies poverty, yet various antipoverty programs in many countries have exhibited mixed results in addressing deforestation. Poverty Alleviation Resettlement (PAR) stands out as one of the few government-led resettlement programs designed to alleviate poverty, offering comprehensive follow-up support for quality of life and employment after relocation. Our study uncovered empirical evidence of the PAR program’s impact on forest quality. Through a multiperiod difference-in-differences analysis of remote sensing and household survey data, we found that the PAR program significantly curbed deforestation in the participating counties and reduced forest-clearing activities among the resettled households, even those relocating to rural villages. Mechanism analysis revealed that the program discouraged deforestation by enhancing market accessibility, fostering nonfarm employment opportunities, and elevating income levels. The study underscores that altering livelihood strategies and lifestyles is essential for resettlement programs to effectively mitigate ecological degradation.

Poverty eradication and sustainable forest management are both critical in the 17 United Nations Sustainable Development Goals, in which the two are closely linked, as the poorest often reside in and around forests for their livelihoods. An estimated 1.6 billion people live near forests, and 71.3% of them are from low- or middle-income countries, as classified by the World Bank (1). Of the world’s 689 million extremely poor people, approximately 40% are located in forest and savannah areas (2, 3). Forests offer the poor not only food, shelter, and fuel but also cash income (through sales of wood products) and employment, providing social safety nets to protect against climatic and economic risks (4).

Despite the importance of forests for the poor, deforestation often coincides with poverty (5, 6). Although the pace of forest loss has slowed over the past decade, the estimated rate of deforestation between 2015 and 2020 is still 10 million hectares per year, which is equivalent to approximately 36 football fields per minute (7). Unsurprisingly, the poor contribute to tree removal, and global evidence has shown a strong impact of poverty on deforestation (8–10). In addition to the unsustainable use of forest provisioning services, such as the illegal extraction of fuel, construction, or money, subsistence farmers also tend to clear forests for agriculture (11). Once the vulnerable residents destroy the forests, it becomes difficult to recover resources in the short term, thereby pushing them deeper into poverty. Consequently, there is a vicious cycle of poverty in which deforestation plays a fundamental role and which generates a giant negative externality in biodiversity conservation and climate change (12, 13).

To break the cycle of poverty and deforestation, various antipoverty programs have been implemented in many ecologically and economically important countries (6, 14). However, the literature has found ambiguous effects of increased income on deforestation (15–17). Intuitively, poverty reduction helps break the trade-off between forest conservation and local livelihoods. In Indonesia, conditional cash transfers targeted at poor households cause an estimated 30% drop in tree cover loss in rural forested villages, on average (6). Nevertheless, for the poorest, income growth may relax capital constraints, and thus encourage forest clearing (8). Evidence from Peru suggests that small increases in the income of the poorest accelerate deforestation (18). In Laos and Cambodia, pure income-increasing financial support does not significantly affect poor households’ deforestation decisions, unless their education and social status improve (19). It appears that the poorest people dependent on forests need additional help in addition to income transfers.

What about moving? Unfortunately, many of the trapped rural poor spontaneously outmigrate to other pristine forest frontiers, only to repeat the tragic loop again (20–22). However, it remains unclear whether sponsored resettlements with policy interventions can effectively alleviate deforestation-related poverty. China, a densely populated developing country with a vast territory, has long adopted resettlement as a tool for poverty alleviation in ecologically fragile areas (23). In 2015, the Chinese government started a new stage of Poverty Alleviation Resettlement (PAR) as a major means of achieving its ambitious goal of eliminating absolute poverty across the country by 2020. This state-led voluntary resettlement program was designed to move the extremely poor (officially identified with per capita income <2,300 yuan at constant 2010 prices) out of harsh and hazardous areas that were deemed unable to support sustainable livelihoods. Most participants were relocated to centralized resettlement sites in cities and towns, while some new homes were still in rural villages with access to farmland (24). In addition to housing with better roads, education, and healthcare, employment assistance services were provided as follow-up support measures (25, 26). According to China’s National Development and Reform Commission, more than 9.61 million poor people had participated in the PAR program by the end of 2020.

Although a few studies have evaluated the PAR program in China, none has examined its ecological impact. Existing PAR evaluations are mainly from the perspective of individual welfare, focusing on income, quality of life, employment, and so on (27–32). For instance, an empirical study in an impoverished county of China using the full-sample administrative poverty population panel data during 2014-2018 at the individual level found that the PAR program significantly increased the participants’ income by 9.61%, primarily due to wage income growth. This study verified that the PAR program achieved its basic goal, i.e., poverty alleviation. However, despite the significant spatial coupling of PAR participation and ecologically vulnerable areas, to the best of our knowledge, no empirical assessment of the program’s effect on the natural environment, especially forests, has been conducted.

In this study, we conducted a quantitative analysis of the effect of China’s PAR program on deforestation using econometric techniques for causality identification. Combined with high-resolution satellite imagery data on forest quality, we adopted national longitudinal datasets of PAR participation at both the county and individual levels, with multiple cross-sections before and after the launch of the program. Our sample included 1,273 participating counties covering 95% of the entire population relocated by PAR, together with 1,491 uninvolved neighboring counties for comparison. We found that the PAR program significantly reduced deforestation in the participating counties, where most relocated individuals moved within their original counties. Further heterogeneity analysis showed that PAR had larger ecological effects in counties with less forest endowment and higher budget revenue. In addition, focusing on another sample of 1,352 PAR-participating households resettled in batches over different years, the migrant individuals mitigated their forest-clearing behavior after resettlement, even for those who moved to rural villages. Apart from moving away from the forest, channels through which PAR can discourage deforestation include improved market access, nonfarm employment opportunities, and income levels. This study provides fresh evidence that China’s resettlement program, aimed at alleviating poverty, contributes to forest conservation. This highlights the effectiveness of resettlement as an alternative strategy to combat deforestation and poverty, by not only physically moving the poor away from the forest but also changing their livelihood strategies and lifestyles.

Results

Impact on County-Level Deforestation.

Deforestation was generally mitigated after PAR implementation in participating counties, where forest change was implied by the Normalized Difference Vegetation Index (NDVI), which is commonly used for assessing overall forest quality (33, 34). Our study focused on intracounty resettlement (relocation to places within the same county), which is the dominant form of PAR implementation (99.86% of the total resettled population). Fig. 1 illustrates the change in forest quality for participating and nonparticipating counties across China; that is, the average NDVI measures in the pre-PAR period of 2011–2015 (Fig. 1 A and B), in the PAR period of 2018–2020 (Fig. 1 C and D), and the difference in NDVI between the two periods (Fig. 1 D and E). For the participating counties, the NDVI increased by 0.029 on average, indicating a significant overall reduction in deforestation. As shown in Fig. 1E, the central regions, in particular, exhibited significant improvements in forest quality. In contrast, for nonparticipating counties, Fig. 1F illustrates noticeable deforestation, especially in the northeastern provinces and coastal areas along the Bohai and Yellow Seas.

Fig. 1.

Temporal and spatial trends of NDVI. (A) Average NDVI of participating counties in the pre-PAR period 2011-2015. (B) Average NDVI of nonparticipating counties in the pre-PAR period 2011-2015. (C) the average NDVI of participating counties in the PAR period 2018-2020. (D) the average NDVI of nonparticipating counties in the PAR period 2018-2020. (E) the difference in average NDVI between the PAR period and the pre-PAR period of participating counties. (F) the difference in average NDVI between the PAR period and the pre-PAR period of nonparticipating counties.

After controlling for confounding factors using a multiperiod difference-in-differences (DID) model, we found that the impact of PAR implementation on forest quality at the county level was positive and statistically significant. In our DID design for 2011–2020, the treatment group included all 1,273 counties that participated in PAR, whereas the control group included 1,491 counties. As the PAR implementation year varied among counties from 2016 to 2020 (Fig. 2), the pre- and postprogram periods were tailored for each county, and the treatment occurred between 2018 and 2020 (2016–2017 was absent owing to data unavailability). We found that PAR implementation led to an increase of 0.004 in the NDVI (or 2.15% in SD units) for an average participating county, which was statistically significant at the 1% level (Table 1, Column 1). A more positive change in the forest quality of a participating county before and after PAR implies less deforestation or more afforestation than in a parallel county.

Fig. 2.

County-level PAR implementation by year. (A) Shows the number of participating counties in each year. (B) Shows the resettled population of urban and rural resettlement in each year.

Table 1.

Estimated impacts of the PAR program on county-level NDVI

	(1)	(2)	(3)	(4)	(5)	(6)
Dependent variable	NDVI	NDVI	NDVI	NDVI	NDVI	NDVI
Sample	Whole	PSM	Participating counties		PSM
PAR implementation (0/1)	0.004***	0.025***
	(3.503)	(4.844)
Total resettled population (10,000 people)			0.003***		0.008***
			(4.958)		(3.114)
Resettled population to urban areas (10,000 people)				0.005***		0.013***
				(6.320)		(3.463)
Resettled population to rural areas (10,000 people)				0.000		0.001
				(0.342)		(0.294)
Precipitation in March (1,000 mm)	0.059***	−0.094***	−0.011	−0.010	−0.091***	−0.091***
	(5.808)	(−2.759)	(−0.965)	(−0.870)	(−2.689)	(−2.689)
Precipitation in June (1,000 mm)	−0.003	0.062***	0.032***	0.031***	0.065***	0.065***
	(−0.454)	(3.136)	(3.759)	(3.635)	(3.291)	(3.275)
Precipitation in September (1,000 mm)	0.022***	−0.084***	−0.014	−0.016*	−0.087***	−0.088***
	(2.813)	(−3.083)	(−1.462)	(−1.713)	(−3.133)	(−3.187)
Precipitation in December (1,000 mm)	0.156***	0.239***	0.175***	0.178***	0.235***	0.238***
	(10.168)	(6.713)	(11.989)	(12.155)	(6.707)	(6.761)
Temperature in March (°C)	−0.002***	−0.001	0.001**	0.001**	−0.001	−0.001
	(−9.998)	(−0.804)	(2.217)	(2.157)	(−1.183)	(−1.200)
Temperature in June (°C)	0.000	−0.001	−0.001	−0.001*	−0.001	−0.001
	(0.247)	(−0.860)	(−1.630)	(−1.667)	(−0.829)	(−0.860)
Temperature in September (°C)	−0.000	−0.003**	0.002***	0.002***	−0.004**	−0.004**
	(−0.495)	(−1.963)	(3.639)	(3.650)	(−2.233)	(−2.225)
Temperature in December (°C)	−0.002***	−0.006***	−0.003***	−0.003***	−0.006***	−0.006***
	(−6.518)	(−4.339)	(−7.150)	(−7.129)	(−4.622)	(−4.632)
Nighttime light intensity [nW/(cm2×sr)]	0.002***	0.000	0.003*	0.003*	−0.001	−0.001
	(4.983)	(0.360)	(1.724)	(1.711)	(−0.652)	(−0.652)
Year fixed effects	Yes	Yes	Yes	Yes	Yes	Yes
County fixed effects	Yes	Yes	Yes	Yes	Yes	Yes
Number of treated counties	1,273	639	1,273	1,273	639	639
Number of counties	2,764	1,278	1,273	1,273	1,278	1,278
Number of observations	22,112	10,224	10,184	10,184	10,224	10,224
R2	0.193	0.132	0.276	0.276	0.130	0.130

Note: Robust t statistics are in parentheses. All regressions include a constant. ***P < 0.01, **P < 0.05, *P < 0.1.

Using the propensity score matching (PSM) technique to select similar counties for constructing treated and control groups for our DID analysis, we found that the positive effect of PAR on NDVI was still robust, and the magnitude increased. Based on each county’s historical forest quality and government budget, we identified 639 participating and 639 nonparticipating counties by 1:1 nearest neighbor matching with replacement. The results show that the change in NDVI for a participating county after PAR was 0.025 (or 13.44% in SD units) greater than that of the other counties (Table 1, Column 2). A larger result indicates that the effect of PAR on NDVI is more significant when comparing matched counties with similar initial statuses.

Various tests demonstrate the validity of our PSM-DID model. We conducted an event study analysis, indicating that the parallel-trend assumption held for our PSM sample, as there was no significant difference in NDVI before PAR between the matched participating and nonparticipating counties (Fig. 3 and SI Appendix, Table S1). In addition, the NDVI in participating counties started to significantly exceed that in other counties 2 y after implementation, illustrating a lagged effect of PAR on the forest. Besides, considering that the PAR implementation occurred at different times and might hinder causal inference, we conducted the Goodman-Bacon decomposition of the DID estimator and parallel-trend sensitivity test as well as alternative robust multiperiod DID methods addressing variation in treatment timing (SI Appendix, Figs. S1 and S2 and Tables S2 and S3). These tests and robust estimators consistently showed that our results were not biased by the policy rollouts.

Fig. 3.

Impacts of the PAR program on county-level NDVI after PSM: event study analysis. There was no significant difference in NDVI before PAR between the matched participating and nonparticipating counties, confirming the parallel-trend assumption. Period -1 is the baseline of the analysis and not shown. Corresponding regression results are presented in SI Appendix, Table S1.

We further conducted placebo tests and verified that the impact we found did not occur by chance. Replicating our DID analysis using 500 fake treatments, where treated counties and implementation years were randomly assigned, we found that the resulting estimate of the treatment effect, as well as its P-value, followed a normal distribution, and the mean was not significantly different from zero (SI Appendix, Fig. S3). The randomized inference confirmed that the change in NDVI in the participating counties resulted from PAR implementation.

Focusing on the size of migration in each county (Fig. 4) to reflect the differences in PAR implementation, we found that the resettled population in a county was positively associated with forest quality. For all 1,273 participating counties, the results showed that an increase of 10,000 in the migrated population via the PAR program increased the county-level NDVI by 0.003 (Table 1, Column 3). The outmigrated population, with a mean of 7,384 for participating counties, was resettled within the same county in our sample; of this, 62.5% were relocated to urban areas (towns) and the others to rural areas (villages). Further dividing the resettled population by migration destination shows that the ecological impact of PAR is mainly sourced from urban resettlement (Table 1, Column 4), which potentially changes rural households’ livelihoods and lifestyles through reduced deforesting behaviors.

Fig. 4.

Resettled population of the PAR program. The resettled population depicted in the map refers to the intracounty resettlement population (resettled to places within the same county).

To compare with other nonparticipating counties and further address endogeneity, we adopted the formerly constructed PSM sample to investigate the effect of the migrant population on forests. These results were consistent with our previous conclusions. Compared with nonparticipating counties, an additional 10,000 resettled residents led to a significant increase of 0.008 in the NDVI in participating counties (Table 1, Column 5). Similarly, the estimated effect of urban settlement was significant, and the magnitude was larger than before because the control group was different.

A series of sensitivity analyses demonstrated the robustness of our results, based on a model of the resettled population affecting forests. First, we utilized different matching methods for PSM to select nonparticipating counties as the control group, including kernel matching and local linear regression matching (SI Appendix, Table S4). Second, we considered other criteria to determine sample counties rather than matching, such as selecting nonparticipating counties adjacent to participating counties as the parallel, or only selecting participating and nonparticipating counties adjacent to each other as the treatment and control groups (SI Appendix, Table S5). Third, we restricted our sample to the participating counties and constructed treatment dummies based on the resettled population (SI Appendix, Table S6). For instance, in Column 1, a county was defined as treated when its resettled population was above the 15% quantile of all participating counties. Across all robustness checks, the results resembled previous ones, and our conclusions held that the PAR population positively affected the NDVI at the county level. In addition, we adopted another commonly used measure of forest quality, Vegetation Continuous Fields (VCF), to estimate our baseline county-level model. The results confirmed our conclusions using NDVI (SI Appendix, Table S7). Using PSM, the change in percentage of tree cover for a participating county after PAR was 2.05% greater than that of the other counties (SI Appendix, Table S7, Column 2).

Heterogeneity Analysis.

We found heterogeneous effects of PAR implementation on forests across counties dependent on their natural and economic characteristics, highlighting that resettlement programs, as a potential policy tool for ecological conservation, need to adapt to local conditions. Based on our preferred model that uses PSM-DID to examine the impact of the resettled population on NDVI (Table 1, Columns 5 to 6), we further investigated whether the impact varies across counties with different initial forest statuses by categorizing the sample into three groups according to the tertiles of counties’ NDVI in 2010 (the lowest, middle, and highest). Table 2 shows that the ecological effect of PAR was more significant and of greater magnitude in counties with less forest endowment. For counties with a low level of forest quality before our study period, the resettled population significantly increased the county-level NDVI, even for those who relocated to rural areas. In contrast, the effect of PAR on NDVI was not significant for counties that initially had a high level of forest quality, and rural resettlement there may have potentially exacerbated deforestation.

Table 2.

Estimated heterogeneous impacts of the PAR program on county-level NDVI for counties with different initial forest quality

	(1)	(2)
Dependent variable	NDVI	NDVI
Low-NDVI group × Total resettled population (10,000 people)	0.072***
Low-NDVI group × Resettled population to urban areas (10,000 people)	(3.979)	0.097***
		(3.243)
Low-NDVI group × Resettled population to rural areas (10,000 people)		0.047*
		(1.930)
Medium-NDVI group × Total resettled population (10,000 people)	0.010***
	(3.384)
Medium-NDVI group × Resettled population to urban areas (10,000 people)		0.011**
		(2.402)
Medium-NDVI group × Resettled population to rural areas (10,000 people)		0.009
		(1.528)
High-NDVI group × Total resettled population (10,000 people)	−0.001
	(−0.361)
High-NDVI group × Resettled population to urban areas (10,000 people)		0.006
		(1.546)
High-NDVI group × Resettled population to rural areas (10,000 people)		−0.008*
		(−1.954)
Control of county-level characteristics	Yes	Yes
Year fixed effects	Yes	Yes
County fixed effects	Yes	Yes
Number of treated counties	639	639
Number of counties	1,278	1,278
Number of observations	10,224	10,224
R2	0.134	0.135

Note: Robust t statistics are in parentheses. All regressions include a constant. ***P < 0.01, **P < 0.05, *P < 0.1.

Similarly, we analyzed how counties’ initial levels of economic development affected the environmental impact of PAR, grouped by public budget revenue per capita in 2010. Table 3 shows that PAR in counties with higher budget revenues had a greater and more significant impact on forests. The ecological effect of PAR in the poorest counties was generally insignificant, although urban settlement was still effective across all groups.

Table 3.

Estimated heterogeneous impacts of the PAR program on county-level NDVI for counties with different initial budget revenue per capita

	(1)	(2)
Dependent variable	NDVI	NDVI
Low-income group × Total resettled population (10,000 people)	0.006
	(1.243)
Low-income group × Resettled population to urban areas (10,000 people)		0.013**
		(2.147)
Low-income group × Resettled population to rural areas (10,000 people)		−0.003
		(−0.462)
Medium-income group × Total resettled population (10,000 people)	0.004**
	(1.992)
Medium-income group × Resettled population to urban areas (10,000 people)		0.009***
		(3.016)
Medium-income group × Resettled population to rural areas (10,000 people)		−0.003
		(−0.930)
High-income group × Total resettled population (10,000 people)	0.045***
	(2.693)
High-income group × Resettled population to urban areas (10,000 people)		0.061***
		(2.788)
High-income group × Resettled population to rural areas (10,000 people)		0.032
		(1.206)
Control of county-level characteristics	Yes	Yes
Year fixed effects	Yes	Yes
County fixed effects	Yes	Yes
Number of treated counties	639	639
Number of counties	1,278	1,278
Number of observations	10,224	10,224
R2	0.131	0.132

Note: Robust t statistics are in parentheses. All regressions include a constant. ***P < 0.01, **P < 0.05, *P < 0.1.

Impact on Household-Level Fuelwood Logging.

By analyzing the impact of PAR on household fuelwood logged per capita, we demonstrated that resettlement reduced households’ deforestation behavior from a microperspective. We utilized a panel dataset at the household level for 2015, 2016, and 2018, where 1,352 sample households were located in Concentrated Contiguous Destitute Areas (CCDA) in Central and Western China (SI Appendix, Fig. S4). All our sample households voluntarily participated in PAR and were scheduled to be resettled by the end of 2020, although the actual resettlement time for each household varied between 2016 and 2020. Therefore, we considered households that were resettled during the sample period as the treatment group (823 households) and those that had not yet been resettled by the end of 2018 as the control group (529 households). Through a DID analysis, we found that the participating households significantly reduced deforestation, and their fuelwood logged per capita decreased by 105 kilograms of coal equivalent (kgce) (Table 4, Column 1).

Table 4.

Estimated impacts of the PAR program on fuelwood logging and clean energy consumption at the household level

	(1)	(2)	(3)	(4)	(5)	(6)
Dependent variable	Fuelwood logged per capita	Fuelwood logged per capita	Fuelwood logged per capita	Clean energy consumption per capita	Clean energy consumption per capita	Clean energy consumption per capita
Sample	Whole	Households resettled to rural areas	Households resettled to urban areas	Whole	Households resettled to rural areas	Households resettled to urban areas
PAR implementation (0/1)	−105.071***	−63.037**	−194.972***	13.067***	9.259***	24.690***
	(−4.256)	(−2.278)	(−5.493)	(5.140)	(3.526)	(5.275)
Household size (person)	−38.683***	−40.015**	−40.401**	−4.979***	−4.785***	−3.557**
	(−2.632)	(−2.467)	(−2.051)	(−3.842)	(−3.629)	(−2.035)
Labor ratio (%)	1.260	0.819	2.134*	−0.060	−0.006	−0.144
	(1.290)	(0.764)	(1.682)	(−0.712)	(−0.071)	(−1.145)
Female labor ratio (%)	−1.632	−1.674	−1.227	0.030	−0.031	0.160
	(−1.221)	(−1.242)	(−0.702)	(0.258)	(−0.279)	(0.963)
Year fixed effects	Yes	Yes	Yes	Yes	Yes	Yes
Household fixed effects	Yes	Yes	Yes	Yes	Yes	Yes
Number of treated households	823	546	277	823	546	277
Number of households	1,352	1,075	806	1,352	1,075	806
Number of observations	4,056	3,225	2,418	4,056	3,225	2,418
R2	0.083	0.065	0.089	0.189	0.174	0.187

Note: Robust t statistics are in parentheses. All regressions include a constant. ***P < 0.01, **P < 0.05, *P < 0.1.

Moreover, even for households resettled in other rural areas within the county, their fuelwood logged per capita decreased by 63 kgce (Table 4, Column 2). The impact of rural resettlement on household behavior implies that the PAR program not only simply moved the poor away from the forest but also changed their livelihoods and lifestyles with additional support, leading to a reduction in deforestation. By contrast, urban resettlement resulted in a larger ecological effect due to move-in dwellings farther away from the forest, and the participants were more likely to switch to nonfarm employment (Table 4, Column 3).

Meanwhile, we also found that the participating households’ clean energy consumption per capita, including electricity, liquefied petroleum gas (LPG), and natural gas, increased compared with that of the control group, which is consistent with their reduced logging behavior (Table 4, Column 4). Likewise, even those resettled in rural areas significantly increased the clean energy consumption, and such effect for urban resettlement was even greater (Table 4, Columns 5 to 6).

The DID analysis also passed parallel-trend and placebo tests. Before migration, there was no significant difference in fuelwood logged between the participating and nonparticipating households, but only the participating households showed reduced deforestation after migration (SI Appendix, Table S8). In addition, 500 fabricated treatments that randomly selected treated households and migration years exhibited a normal distribution of the estimated treatment effect and its P-value, demonstrating that the ecological impact of PAR on household fuelwood logging we found was not an occasional result (SI Appendix, Fig. S5).

Further examination revealed the mechanisms through which households’ participation in PAR abated logging, involving distance from forest, market access, off-farm employment, and income. Basically, the PAR program significantly increased the distance to the closest forest for participating households by 11 km on average, or 2.5 times the sample mean (Table 5, Column 1). Meanwhile, the program shortened the distance to the closest market by 4 km or 42% of the sample mean (Table 5, Column 2). Moreover, the PAR-participating households’ per capita off-farm income and total income increased by 265 yuan (17.9%) and 603 yuan (15.5%), respectively (Table 5, Columns 3 and 5). Being farther away from the forest made deforestation less feasible; moreover, the resettled households had better access to the market and were provided with more off-farm employment opportunities and alternative energy sources. With additional support from the PAR program, such as skill training and employment agencies, participating households began to adopt nonagricultural jobs with higher wages. Consequently, the deforestation of cultivated land was mitigated, and the increased opportunity cost discouraged logging for livelihoods. Finally, the increased overall income enabled households to afford clean energy as a substitute for fuelwood, further reducing the need for logging. However, the PAR program did not bring additional transfer income to the participants compared with other registered poor households (Table 5, Column 4). While the PAR program did not directly issue transfer payments, other transfers from general poverty alleviation policies did not differ between the PAR participants and nonparticipants.

Table 5.

Estimated impacts of the PAR program on distance to forest, market access, and income at the household level

	(1)	(2)	(3)	(4)	(5)
Dependent variable	Distance to the closest forest	Distance to the closest market	Off-farm income per capita	Transfer income per capita	Income per capita
PAR implementation (0/1)	10.904***	−4.080***	265.303*	7.735	602.696***
	(10.249)	(−7.870)	(1.905)	(0.114)	(3.388)
Household size (person)	0.786	0.194	−40.037	−191.579***	−537.430***
	(1.635)	(0.811)	(−0.563)	(−7.438)	(−6.269)
Labor ratio (%)	0.060**	0.003	8.551**	−5.570**	0.401
	(2.246)	(0.174)	(2.180)	(−2.377)	(0.072)
Female labor ratio (%)	−0.055*	−0.006	0.122	4.706	5.957
	(−1.815)	(−0.290)	(0.020)	(1.619)	(0.853)
Year fixed effects	Yes	Yes	Yes	Yes	Yes
Household fixed effects	Yes	Yes	Yes	Yes	Yes
Number of treated households	823	823	823	823	823
Number of households	1,352	1,352	1,352	1,352	1,352
Number of observations	4,056	4,056	4,056	4,056	4,056
R2	0.139	0.067	0.095	0.201	0.167

Note: Robust t statistics are in parentheses. All regressions include a constant. ***P < 0.01, **P < 0.05, *P < 0.1.

Discussion

This study quantitatively investigates the impact of the PAR program, a resettlement initiative aimed at enhancing the living standards of residents in impoverished regions, on deforestation across China. We utilized satellite images matched with county-level statistics and household surveys to demonstrate that intracounty resettlement contributes to improving forest quality. We examined the effects of PAR implementation and the extent of resettlement on NDVI at the county level and found heterogeneous effects for counties with different initial characteristics. We further evaluated the effects of PAR on deforesting behavior at the household level, distinguishing between resettlement in urban and rural areas, and explored possible mechanisms for relocated households to reduce logging. Overall, the findings suggested that a systemic resettlement program with follow-up support can help mitigate deforestation.

Our county-level results showed that the ecological effects of the PAR program have tangible benefits in mitigating deforestation. Comparing counties that participated in the PAR program with nonparticipating counties, we found an additional increase in NDVI of 0.025 on average after resettlement, which reflects reduced deforestation activities and augmented vegetation coverage. Moreover, our event study analysis revealed that the NDVI in the participating counties only showed a significant upward trend 2 y after resettlement, as ecological restoration is a gradual and long-term process. The time lag suggested that policy evaluations of resettlement initiatives should fully consider changes in the time dimension to better understand their long-term ecological impacts. In addition, focusing on intracounty resettlement, we found that the ecological effects mainly stem from resettlement in urban areas. As urban destinations generally have higher levels of economic development, infrastructure, and environmental awareness, these factors facilitated the transformation of livelihood patterns among the resettled individuals, leading to a reduction in their reliance on and a detrimental impact on the natural resources of their original habitats.

The county-level analysis also illustrated significant heterogeneity in the impact of PAR implementation. Counties that initially possessed more vulnerable natural environments exhibited greater ecological enhancement after resettlement. This finding can be attributed to the inherent dependence of these areas on natural resources, in which the livelihoods of local residents were more likely to be tied to deforesting activities. Consequently, the program relocating individuals away from their previous means of subsistence resulted in a more pronounced decrease in direct pressure on the surrounding forests. We also found that participating counties with higher levels of economic development and fiscal revenue yielded greater ecological benefits. These relatively developed areas, with stronger financial support, could provide resettled individuals with improved living conditions and diversified employment opportunities, thereby accelerating the transformation of their livelihoods and reducing their dependence on forests.

Our household-level results were consistent with the overall findings at the county level, further confirming the positive effects of the PAR program on forest conservation. However, notably, even households relocating to other rural areas within the county significantly reduced fuelwood logging after resettlement. In addition, we found a shift in the energy structure of the resettled households, especially for the rural resettlement, in that they used cleaner energy to replace traditional fuelwood as an energy source. Apart from facilitating relocation, the PAR program also provides better housing conditions and infrastructure as well as financial, employment, educational, and medical support. This comprehensive support system effectively helps the resettled households adapt to a new lifestyle with diversified options of livelihoods and, therefore, reduces logging activities, even if the new homes are still located in rural areas with potential access to forest resources.

The numerical results at the household level hold profound implications for understanding the ecological impacts of the PAR program. As we found that the program resulted in a decrease in fuelwood logged per capita by 105 kgce, it is equivalent to 184 kg of wood given the conversion factor of 0.571 kgce/kg (SI Appendix, Table S10). Since the fuelwood logged and used by rural residents is mainly hardwood, of which the specific gravity is approximately 0.8 tons/m³, 184 kg of fuelwood is further equivalent to 0.23 m³ of stranding trees (35). According to the statistical data of the China Statistical Yearbook 2019, the timber stock of forest land in China was approximately 58 m³/ha. Assuming the average rotation period is 8 y for hardwood species (35), the annual growth of timber is 7.25 m³/ha. Therefore, the annual saving of standing trees by an average resettled individual is equivalent to the annual growth of wood on 0.0317 ha of forest land. Considering the total resettled population of PAR was 9.61 million nationwide, it implies that approximately 2.21 million m³ of standing trees were saved each year after the program, which is equivalent to the annual growth of wood on 304,637 ha of forest land.

The trees saved and the consequent reduction in carbon emissions brought further benefits from a climate change perspective. Based on the reduction in fuelwood logged per capita of 184 kg per year, or a total of 1.77 million tons resulting from the national program, we further consider that a) the moisture content of wood is 15%, b) half of the dry weight is carbon, and c) a ton of carbon is equivalent to 3.67 tons of carbon dioxide according to the United States Environmental Protection Agency. As a result, it implies that 2.76 million tons of carbon dioxide emissions have been saved each year from burning fuelwood due to PAR. Using the estimate of the social cost of global carbon emissions from Rennert et al. (36), i.e., approximately 185 US dollars per ton of carbon dioxide emitted into the atmosphere, the annual economic value of reduced fuelwood carbon emissions due to PAR is around 511 million US dollars, not to mention the extra carbon absorbed by the standing trees that were saved. Considering the various and long-term benefits of the preserved forest ecosystems in addition to carbon stock, the monetary value of PAR’s ecological effects will be even greater.

Through mechanism analyses at the household level, we revealed multiple channels through which the PAR program reduces deforestation in addition to moving away from the forest. The first was by improving market access. The resettled households were not only physically distant from the forest but also closer to the market. Proximity to the market brought more employment opportunities unrelated to logging and more options for alternative energy with related services, reducing the resettled households’ reliance on forest resources.

The second channel related to off-farm employment prospects. Increased nonfarm income after resettlement, reflecting more available jobs and higher wage levels in nonagricultural employment, amplified the households’ opportunity cost of farming. Faced with more stable and substantial income sources from off-farm work, the resettled farmers no longer relied on felling trees as their main means of livelihood. Meanwhile, the higher opportunity cost of farming reduced cultivation activities and demand for arable land, thereby mitigating forest clearing for agriculture.

The third channel arose from the total income effect. After resettlement, the increased overall income level of the households brought about more comfortable economic conditions. As a direct impetus to reduce deforestation, richer households could more easily afford clean energy and reduced fuelwood consumption as an inferior good. In addition, better financial status brought households an overall improvement in education, quality, and environmental awareness, providing them with an intrinsic motivation to cherish the ecological environment and reduce deforestation.

This study has important policy implications for developing countries in terms of poverty alleviation and ecological protection. First, the organized implementation of relocation policies has contributed to forest conservation. Government-led resettlement plans help guide the poor away from ecologically fragile areas and reduce direct damage to forest resources by human activities. Second, policymakers should pay attention to the provision of support measures as key to a successful resettlement program. Resettlement is not only a change in geographical location but also a profound change in livelihood patterns and lifestyles. The government needs to provide all-round support, including employment assistance, industrial support, and social security, to ensure that resettled households smoothly adapt to new environments, helping them achieve sustainable development of their livelihoods and reducing their dependence on forest resources.

This study makes two main contributions to the literature. First, it provides evidence of the ecological impact of relocation policies on poverty alleviation. For a nationwide study across China focusing on both county and household levels, we adopted empirical methods that effectively address potential endogeneity issues to causally demonstrate the positive effect of PAR on deforestation. This study evaluates the ecological benefits of the PAR program in China and provides insights into nature conservation for other relocation programs in developing contexts. Second, this study reveals the various mechanisms by which poverty alleviation through relocation slows deforestation. In addition to the simple perspective of physical relocation, we explored multiple dimensions, such as market availability, nonagricultural employment, and total income effects, to empirically analyze how resettlement policies change the livelihoods of the poor and thus reduce damage to natural resources. These findings will help inform conservation policies in other developing countries facing similar challenges of poverty-related ecological degradation.

Despite the important contributions of this paper, we acknowledge that there are certain limitations. Owing to the lack of unavailable data, our sample is not continuous year by year, which may affect the accurate capture of the short-term effects of the PAR program to some degree. However, the given dataset structure made us pay more attention to the long-term impacts of the resettlement policy. Through event study analysis, we identified the time-lag effect of the program, which indicates that the ecological impact of resettlement requires time accumulation and transmission processes to fully manifest. This enriches our understanding of dynamic changes in policy effects and highlights that policymakers should fully consider the long-term impact when evaluating resettlement programs, especially their ecological benefits.

Materials and Methods

In this study, we assessed the impact of the PAR program and its mechanisms using two sets of data: county level and household level. At the county level, we combined PSM and a multiperiod DID model to estimate the impact of PAR on forest quality. We designed a quasi-natural experiment at the household level and applied a DID model for evaluation. Below, we introduce the two datasets and then discuss the empirical strategies for analysis at the two levels. The processed data and code utilized in this paper are publicly available for download at https://doi.org/10.6084/m9.figshare.28235258.

Data.

Policy context.

The PAR program has been one of the key components of China’s targeted poverty alleviation policies since 2013, targeting officially identified poor households based on a per capita net income below 2,300 yuan (at constant 2010 prices) who were registered in the National Poverty Alleviation Information System. Among the identified poor, eligible for relocation are those in areas with a) harsh natural conditions, b) national planning restrictions, c) high development costs, and d) frequent disasters. At the county level, the PAR policy was implemented in 1,273 counties out of 2,764 in China and lasted for 5 y. The county-level Implementation implied that the local government started to organize the poor who were interested in the program to relocate with investment on centralized construction of resettlement housing and infrastructure. Following the county-level policy implementation, the participating households did not relocate immediately at once but instead moved in batches year by year, depending on the construction plans of new housing.

Beyond the physical act of moving, according to the National “13th Five-Year Plan” for Poverty Alleviation Resettlement formulated by China’s National Development and Reform Commission in 2016, the PAR program encompassed comprehensive support measures, including: a) infrastructure provision such as housing, water, electricity, roads, and internet in resettlement areas; b) access to public services ensuring equal education, healthcare, pension, and cultural opportunities after relocation; and c) employment assistance fostering modern agriculture and labor service economy by enhancing skills training, job creation, exploring asset-based poverty alleviation, etc., thereby augmenting livelihoods and incomes of relocated households (see more details in SI Appendix). Note that these measures, especially those regarding individual development, are principally aimed at poverty alleviation rather than ecological restoration such as deforestation mitigation.

County-level data.

To assess the effects of the PAR program on forest quality, we compiled data from multiple sources to create a county-level dataset. This dataset contained 8-y panel data before (2011–2015) and after (2018–2020) the PAR program for 1,273 participating counties and 1,491 nonparticipating counties. China’s PAR program has relocated 9.61 million impoverished individuals from 1,386 counties in total, among which 9.59 million were relocated within their respective counties, accounting for 99.86% of the total resettled population. The rest 13.4 thousand mainly resided in peripheral areas of the original counties and migrated short distances to adjacent counties. Our sample solely considered the cases of intracounty resettlements. The final sample covered 92% of the PAR-participating counties and more than 97% of all counties across 31 provinces in mainland China (excluding Hong Kong and Macau). Note that except for PAR participation, there was no difference between the two groups of counties regarding the general policies of targeted poverty alleviation across China. The main components of the county-level data included forest quality, PAR program implementation, climate, nighttime lights, and budget revenues.

Forest quality was measured by the NDVI, which indicates the health and density of vegetation and is commonly used for assessing overall forest quality. We acquired monthly NDVI raster data at a spatial resolution of 1 × 1 km² from 2010 to 2020 using the data product of MOD13A3 from NASA’s Moderate Resolution Imaging Spectroradiometer (MODIS), which can be accessed at https://modis.gsfc.nasa.gov/data/. We combined the NDVI raster data with the county boundary data for 2020 to extract the annual NDVI for each county, using the maximum value synthesis method. The alternative measure, VCF, which indicates the proportion of tree cover within a given pixel, was obtained from the raster data of MOD44B Version 6.1 Vegetation Continuous Fields yearly product from MODIS (37) with a spatial resolution of 250 × 250 m².

Data on the implementation of the PAR program were obtained from the National Poverty Alleviation Information System, established by the State Council Leading Group Office of Poverty Alleviation and Development. The data included the annual total resettled population and the resettled population in different areas of each county from 2018 to 2020. Our sample encompassed a population of 9.40 million, constituting 97.86% of the total resettled population of the program. Note that the individual participants relocated in batches year by year after the PAR policy was implemented at the county level, according to the construction plans of new housing. In addition, the PAR participants were resettled in either rural or urban areas within their original county (Fig. 2). Rural resettlement refers to the resettlement of impoverished households in their own administrative villages, new migrant villages, or nearby villages, whereas urban resettlement refers to their resettlement in nearby towns, county towns, and industrial parks. The destination was chosen by the households or through negotiations with the government based on local construction plans.

Monthly climate data, including the average precipitation and temperature, were important control variables in our county-level analysis. First, daily precipitation and temperature data were collected from Daily Timed Data From Automated Weather Stations of China Meteorological Data Service Centre (https://data.cma.cn/en) from 2011 to 2020. Then, the data were interpolated into raster data at a spatial resolution of 0.1° × 0.1°, using the inverse distance weighted interpolation method. Finally, the daily raster data were combined with the county boundary data to aggregate and derive the monthly climate data at the county level.

We also controlled for the annual nighttime light intensity, which is commonly used as a measure of regional economic development. The annual nighttime light data in 2011–2020 were obtained from Chen et al. (38), who built extended National Polar-orbiting Partnership Visible Infrared Imaging Radiometer Suite (NPP-VIIRS) like nighttime light data through a new cross-sensor calibration from the Defense Meteorological Satellite Program Operational Linescan System nighttime light data and a composition of monthly NPP-VIIRS nighttime light data. These nighttime light data, combined with county boundary data, were used to estimate the average annual nighttime light intensity for each county.

Budget revenue per capita, measuring residents’ income, was used as a covariate for PSM and to construct interaction terms for the heterogeneous analysis. The data were sourced from the China Statistical Yearbook (county level) compiled by the Department of Rural Socio-Economic Surveys of the National Bureau of Statistics of China.

Household-level data.

The household-level data used in this study were sourced from a three-wave survey of 16 counties participating in the PAR program in eight provinces. The survey involved face-to-face interviews. The first wave was conducted in 2016, followed by tracking surveys in 2017 and 2019. Note that the reported household information from the three waves was for 1 y before the survey year, that is, in 2015, 2016, and 2018. Specifically, the data comprised: 1) household resettlement time and destination; 2) household energy consumption, measured by the quantity of each type of energy, particularly for fuelwood logging; and 3) household demographic and socioeconomic characteristics, such as household size, labor ratio, household income from different sources, and distance to the closest market.

Stratified random sampling was used to construct the sample. Considering the geographical distribution of the resettled population of the PAR program as well as the representativeness of different types of poverty-stricken areas regarding terrain, eight provinces were selected as sample provinces: Hunan, Hubei, Guangxi, Yunnan, Guizhou, Gansu, Sichuan, and Shaanxi. During the 13th Five-Year Plan period in China, the resettled population of these provinces accounted for 78.43% of the total population participating in the PAR program. Subsequently, the two major participating counties in each province were selected based on their scale of resettlement and geographical representativeness. The final sample covered 15 counties within five of China’s CCDA, including the Wuling Mountains, Yunnan-Guizhou-Guangxi rocky desertification area, Qinba Mountains, Wumeng Mountains, and Liupan Mountains, as well as a nationally designated key county for poverty alleviation outside these areas (SI Appendix, Fig. S4). In each selected county, 2 to 3 townships were randomly chosen. Administrative villages and households participating in the PAR program were then randomly selected based on official rosters provided by local governments. The baseline survey conducted in 2016 included 2,185 households, with follow-up surveys tracking 1,864 households in 2017 and 1,725 in 2019. After missing data and untracked households were removed, the final household-level analysis used a balanced panel of 1,352 households across the three waves, reflecting the household characteristics in 2015, 2016, and 2018. All of these officially identified poor households equally benefited from other general poverty alleviation polices during the sample period regardless of the status of relocation.

We focused on the household-level fuelwood logged and clean energy consumption per capita. Clean energy comprises electricity, LPG, and natural gas. These two variables of household $i$ were measured in kgce, which makes different types of fuelwood and energy consumption comparable and additive. We calculated these variables using the conversion formula $Y_{\mathit{it}}={\sum }_{i=1}^j{\eta }_j{e}_{\mathit{ijt}}$, where ${\eta }_j$ is the conversion coefficient of standard coal equivalent for energy subcategory $j$ (SI Appendix, Table S10), and $e_{\mathit{ijt}}$ is the consumption of energy $j$ measured in physical units, such as fuelwood in kg and electricity in kWh.

Distance to the closest market and household income were adopted to conduct the mechanism analysis. Off-farm income per capita and total income per capita were adjusted to the 2015 constant price to address inflation, using the consumer price index for rural residents. The definitions and descriptive statistics for all household-level variables are presented in SI Appendix, Table S11.

Empirical Methods.

In this subsection, we first introduce the PSM-DID model for county-level estimation and then the quasi-natural experiment design for household-level estimation. Stata 17.0 was used for the data analysis.

County-level PSM-DID model.

We combined PSM and the multiperiod DID method to estimate the impact of the PAR program on the county-level NDVI. First, PSM was conducted using covariates from the first year of the sample (2011). As the PAR program primarily targets impoverished areas where the local environment is unable to support the population, two major factors influencing whether a county participated in PAR are its original ecological endowment and socioeconomic development level. All sample counties are represented as the universal set $A=\left\{T,C\right\}$, where the treatment group (T) contains PAR-participating counties and the control group (C) contains nonparticipating counties. The probability of a county participating in PAR is modeled as follows:

$$P=Pr\left\{k\in T\right\}=\varnothing \left\{{\mathit{NDVI}}_{k,2009},{\mathit{NDVI}}_{k,2010},R_{k,2011}\right\},$$ [1]

where $P$ represents the probability of county $k$ participating in PAR, and $\varnothing \left\{·\right\}$ denotes the cumulative normal distribution function. The covariates ${\mathit{NDVI}}_{k,2009}$, ${\mathit{NDVI}}_{k,2010}$, $R_{k,2011}$ refer to the NDVI values 1 and 2 y prior to 2011, the starting year of our sample, and the budget revenue per capita in 2011 for county $k$, respectively. This probability model allowed us to estimate the predicted probability $P\left(·\right)$ of a county participating in PAR for all the counties in A. Then, PSM was used to pair each county in T with a county in C of a similar predicted probability. Thus, the PSM process generated a matched control group $C_p$ that consisted of counties with ecological endowments and socioeconomic development levels comparable to those in the treatment group T. Descriptive statistics of the control variables for the treatment and control groups after matching are presented in SI Appendix, Table S9.

Based on the propensity score-matched sample $A_p=\left\{T,C_p\right\}$, a multiperiod DID model with year and county fixed effects was constructed, as specified below:

$$Y_{\mathit{kt}}=a_0+a_1{\mathit{PAR}}_{\mathit{kt}}+X_{\mathit{kt}}\beta +{\delta }_t+{\xi }_k+{\epsilon }_{\mathit{kt}},$$ [2]

$$Y_{\mathit{kt}}=b_0+b_1{\mathit{TotalPop}}_{\mathit{kt}}+X_{\mathit{kt}}\beta +{\delta }_t+{\xi }_k+{\epsilon }_{\mathit{kt}},$$ [3]

$$Y_{\mathit{kt}}=c_0+c_1{\mathit{RuralPop}}_{\mathit{kt}}+c_2{\mathit{UrbanPop}}_{\mathit{kt}}+X_{\mathit{kt}}\beta +{\delta }_t+{\xi }_k+{\epsilon }_{\mathit{kt}},$$ [4]

where $Y_{\mathit{kt}}$ is the forest quality of county $k$, measured by its NDVI in year $t$. We included three types of policy intervention variables in Eqs. 2–4: a) a dummy variable for policy intervention ${\mathit{PAR}}_{\mathit{kt}}$, where ${\mathit{PAR}}_{\mathit{kt}}=1$ if county $k$ implemented PAR in year $t$ (i.e., the first batch of participants was resettled in the county), and ${\mathit{PAR}}_{\mathit{kt}}=0$ otherwise; b) total relocated population ${\mathit{TotalPop}}_{\mathit{kt}}$, which represents the total resettled population of county $k$ in year $t$; and c) population relocated in different resettlement modes, where ${\mathit{RuralPop}}_{\mathit{kt}}$ and ${\mathit{UrbanPop}}_{\mathit{kt}}$ are the populations resettled to rural and urban areas, respectively. $X_{\mathit{kt}}$ denotes a set of county-level climate and socioeconomic control variables, including monthly precipitation, monthly average temperature, and annual nighttime light intensity. ${\delta }_t$ and ${\xi }_k$ represent year and county fixed effects, respectively. $a_1$, $b_1$, $c_1$, and $c_2$ are the coefficients of the key policy intervention explanatory variables to be estimated.

To test the matching quality of PSM and the parallel-trend assumption, we conducted an event study to estimate the annual changes in county-level NDVI before and after resettlement, as follows:

$$Y_{\mathit{kt}}=d_0+{\displaystyle \sum _{m=-9\&m\ne -1}^2}{\lambda }_m{D}_{\mathit{kt}}^m+X_{\mathit{kt}}\beta +{\delta }_t+{\xi }_k+{\epsilon }_{\mathit{kt}},$$ [5]

where $D_{\mathit{kt}}^m$ is a set of dummy variables around the year of resettlement, and $m$ denotes the difference between the current year $t$ and the starting year when PAR was implemented in county $k$. In particular, $D_{\mathit{kt}}^m$ equals a) 1 for states in the $\left|m\right|$th year before the PAR implementation when $m<\; <0$, b) 1 for states in the $m$th year after the PAR implementation when $m\ge 0$, and c) 0 otherwise. The year before the starting year (m = −1) is the baseline scenario. Therefore, ${\lambda }_m$ is the dynamic effects of the PAR implementation. $X_{\mathit{kt}}$ and ${\xi }_k$ are defined the same as in Eq. 2. Moreover, since the policy rolled out across counties rather than occurring simultaneously and thus our DID estimator varied in treatment timing, we a) conducted the Goodman-Bacon decomposition (39), b) tested the sensitivity of the parallel trends assumption by imposing restrictions on the posttreatment violations (40), and c) adopted other robust multiperiod DID estimators (41, 42) to confirm the validity of our empirical design. In addition, we conducted a placebo test to further examine the reliability of the empirical results. By randomly selecting fictional treatment counties and implementation times, we repeated the PSM-DID model estimation 500 times to obtain the distribution of their coefficients and P-values.

To further test the robustness of the empirical results, we considered alternative strategies to determine the treatment and control groups in several ways. First, we adopted different PSM methods, including kernel matching and local linear matching. Second, we constructed treatment and control groups according to geographic location, as geographic proximity implies similar economic and resource endowments. Specifically, we considered the following strategies: a) nonparticipating counties adjacent to participating counties were selected as the control group (SI Appendix, Fig. S6B); and b) participating and nonparticipating counties directly adjacent to each other were selected as the treatment and control groups, respectively (SI Appendix, Fig. S6C). Third, we focused only on the participating counties as the whole sample and constructed a treatment based on the resettlement population. Specifically, the treatment group was formed by counties with resettlement populations above the 15%, 20%, or 25% quartiles (SI Appendix, Fig. S6 D–F).

Considering that the effect of PAR on forest quality may vary among counties with different natural endowments and economic foundations, we conducted a heterogeneity analysis. We split the sample based on the tertiles of the NDVI or budget revenue per capita of the counties in 2010. The PSM-DID model was then estimated separately for each subsample.

Household-level DID model.

At the household level, this study designed treatment and control groups according to resettlement time across households. To construct comparable treatment and control groups, all the sample households were scheduled to be relocated by the end of 2020, but actually moved in different years. In each year, we regarded the households that were relocated as the treatment group and those that had not yet been relocated as the control group. Consequently, both the treatment and control groups were officially registered as impoverished households living in similarly remote and inhospitable areas.

Based on a quasi-natural experiment at the household level, we employed a DID model controlling for household and time-fixed effects to estimate the impacts of PAR implementation on household deforestation behaviors and their mechanisms. The following model is specified:

$$Y_{\mathit{it}}=e_0+e_1{\mathit{Resettled}}_{\mathit{it}}+H_{\mathit{it}}\gamma +{\delta }_t+h_i+{\epsilon }_{\mathit{it}},$$ [6]

where $Y_{\mathit{it}}$ denotes the fuelwood logged per capita by household $i$ in year $t$. In an alternative specification, $Y_{\mathit{it}}$ denotes the clean energy consumption per capita. For the mechanism analysis, $Y_{\mathit{it}}$ represents the distance to the closest forest, distance to the closest market, off-farm income per capita, transfer income per capita, and income per capita of household $i$ in year $t$, respectively. ${\mathit{Resettled}}_{\mathit{it}}$ is a dummy variable with a value of 1 if household $i$ was resettled in year $t$. $H_{\mathit{it}}$ is a vector of household characteristics, such as household size and labor ratio. ${\delta }_t$ represents year fixed effects that control for the national-wide shocks and trends to all households. $h_i$ represents household fixed effects that capture time-invariant household characteristics. ${\epsilon }_{\mathit{it}}$ is the error term. $e_1$ is our parameter of interest to be estimated.

Supplementary Material

Appendix 01 (PDF)

Appendix 02 (PDF)

Appendix 03 (PDF)

Appendix 04 (PDF)

Data, Materials, and Software Availability

Anonymized (.dta) data have been deposited in figshare (10.6084/m9.figshare.28235258) (43).