Introduction

Approximately one-quarter of India’s land area comprises common-pool resources (Chopra and Gulati, 2001), sustaining the livelihoods of over 300 million rural people to varying degrees (Basu, 2014). These resources play a crucial role in rural landscapes, providing essentials such as fuelwood, fodder, timber, medicinal herbs, oils, and resins (Agarwal, 1997). A systematic review identified 34 ecosystem services generated by intact common-pool resources in India, estimated to be worth $90.5 billion/year ($2108/ha) (Sandhu et al., 2023). Yet, since the middle of the 20th century, the commons in India have been in a state of decline, both in terms of area and quality. The reasons are multiple, including population pressure, mechanization, land reform programs accelerating private land ownership, and the undermining of customary common property management regimes in favor of more centralized state management approaches (Narain and Vij, 2016; Thapliyal et al., 2019). This raises questions about what can be done to arrest the degradation and promote the restoration of common-pool resources.

There are several noteworthy initiatives that are rising to this challenge. Government programs, such as the Mahatma Gandhi National Rural Employment Guarantee Act (MGNREGA), focus on developing community assets through afforestation and water conservation (MoHA, 2024). Others, such as the National Rural Livelihood Mission (NRLM), work to strengthen community institutions for sustainable land management (MoRD, 2021). In addition, community-driven efforts under the Forest Rights Act (FRA) empower forest-dwelling communities to manage their resources, and local initiatives in states like Rajasthan emphasize pasture land restoration (GoI, 2006). There are further various initiatives undertaken by Non-governmental Organizations (NGOs). An example is the Watershed Organization Trust, which focuses on promoting integrated watershed management, combining water conservation with soil erosion control (WOT, 2021).

The Foundation for Ecological Security (FES) is another example. It has been working since the early 2000s to support local communities to both assert their rights over and protect and restore the common-pool resources on which they depend. Drawing heavily on the work of Elinor Ostrom (1990) and the CGIAR Systemwide Program on Collective Action and Property Rights (CAPRi, 2010), a central tenet of FES’s work is that the ‘tragedy of the commons’ (Hardin, 1968) is not inevitable. Indeed, there are many examples where local people have put in place governance mechanisms to manage common-pool resources effectively (Ostrom, 2008).

FES’s core intervention model comprises three complementary components, leading to two high-level impacts—improved ecological health and more resilient livelihoods (Supplementary Fig. 1). The first component involves supporting communities to secure their rights to the commons. This work ranges from mapping common-pool resources and facilitating conflict resolution to strengthening local capacity for rights claiming and supporting village institutions to file and negotiate ordinances. The second component focuses on facilitating collective action and strengthening village institutions for both the restoration and long-term management of the commons. Recognizing that much of the commons is degraded and requires significant (financial) investment to promote its restoration brings us to the third component of FES’s intervention model. The focus here is to link village institutions to government social programs, including MGNREGA. Since 2001, FES has directly implemented its core intervention model in over 7000 villages.

This article presents the results of an impact evaluation of FES’s core intervention model implemented in two districts, Bhilwara and Pratapgarh, in Rajasthan state. To mitigate bias, we employed Propensity Score Matching (PSM) to identify and compare 24 villages where FES had intervened from 2002 to 2015 and 24 villages where it did not (i.e., control villages). Using multiple data collection methods, including surveys and Earth Observation, we comprehensively investigate the rollout of FES’s intervention model and its ecological impacts. We further hypothesize that the strength of village-level commons institutions is a significant mediating mechanism giving rise to the ecological effects we evidence and use mediation analysis to test its congruency with the variation in our data.

Methods

A key challenge to overcome in our evaluation is that FES—understandably as a development organization—did not randomly select the villages in which it intervened. Hence, finding significant differences in our outcomes of interest among units (e.g., households and common land areas) in villages where it worked and where it did not may be indicative of pre-existing baseline differences or differences in how these outcomes evolved over time, independent of FES’s intervention.

In order to address potential selection bias, we therefore first sought to understand how FES selected the villages in which it intervened and how local people engaged with its program. FES informed us that it used several criteria to prioritize its village selection process (see Supplementary Table 4). Once villages were selected for intervention, FES then worked with the concerned communities to map their common-pool resources, which were then prioritized for restoration and rights claiming. Given that the entire village and its inhabitants were targeted by FES’s intervention model, participant self-selection did not take place as would be the case, for example, in a job training program. Consequently, program placement bias (Ravallion and Wodon, 1999) is the key type of bias we were challenged to mitigate. Our causal identification approach is summarized in Supplementary Fig. 4.

Village-level propensity score matching

A key component of our causal identification strategy, therefore, involves comparing units within villages targeted by FES and those residing in other non-intervention villages located in the same districts that are statistically similar vis-à-vis FES’s targeting criteria. The validity of our results, therefore, rests significantly on the Conditional Independence Assumption (CIA) (Morgan and Winship, 2015); that is, we need to assume that FES did not systematically target villages based on some other consideration(s), i.e., ‘unobservables’, particularly those correlated with our outcomes of interest. Given FES’s interest in collecting baseline data on a large number of villages to support another prospective impact evaluation design, we were in the fortunate position of being able to do two rounds of matching, one before data collection using secondary data and one after using more refined primary data. Given feasibility considerations, we limited our study in Rajasthan to two districts, Bhilwara and Pratapgarh.

In the secondary data matching round, the universe of potential non-intervention villages was restricted to those that FES was interested in expanding into (n = 960), thereby meeting our secondary objective of compiling baseline data for these villages. We conducted three rounds of one-to-one Propensity Score Matching (PSM)—a procedure that matches units on the basis of their conditional probability of being treated given a set of observable characteristics (Rosenbaum and Rubin, 1983)—to obtain a ratio of three matched new villages for everyone treated village (112 villages in total, 28 intervention villages, and 84 potential comparison villages). The objective of PSM (and related matching approaches) is to reduce selection bias by improving covariate balance. Overall, we were successful in this regard (see Supplementary Table 5).

Following data collection (see below), we took advantage of the fact that we had three times more potential comparison villages than intervention villages. However, despite efforts undertaken during both village matching and data collection, we dropped 12 potential comparison villages due to the absence of common land, as well as four intervention villages, given their proximity (<3 kilometers) to several matched non-intervention villages. In the second matching stage, we used more refined and village-specific data for FES’s targeting criteria (see Supplementary Table 6). Our objective—improved covariate balance—was also achieved during this second round (see Supplementary Table 7). This included significant improvements in the results of our joint statistical significance tests, coupled with improved statistical balance at the individual covariate level. Moreover, this village-level covariate balance translates into household-level covariate balance also, further increasing our confidence that our two-stage village matching exercise was successful in mimicking village-level random assignment (Supplementary Table 8).

Difference-in-differences estimation

Nevertheless, given the absence of random assignment, we cannot guarantee that our village matching approach completely eliminated program placement bias. Consequently, for outcome indicators for which baseline data can be reliably reconstructed—particularly through the analysis of historical satellite imagery—we complement the above with difference-in-differences estimation (Blundell and Costa Dias, 2009). The validity of the impact estimates generated through this approach rests significantly on the parallel trend assumption (Khandker et al., 2010). That is, we must assume that our outcome indicators would have evolved in the same way (both in terms of direction and magnitude) had FES’s program never existed.

Data collection

From January to March 2020, we visited the matched treated and untreated villages to digitally map common land and list households. In the initial phases of FES’s work in the intervention villages, local leaders and others were consulted to prioritize one common land area for restoration. To replicate this in the matched control villages, local leaders and other residents were requested to prioritize a specific common land area to serve as the focus of future protection and restoration work. While all common land in both sets of villages was mapped, such land initially prioritized in the intervention villages and newly prioritized in the non-intervention villages were earmarked for the ecological data collection exercise described below.

LDSF primary data collection in prioritized common land

The Land Degradation Surveillance Framework (LDSF) methodology (Vågen et al., 2015) typically follows a standardized sampling design, where clusters are selected within 10-by-10-kilometer sentinel sites, and then 10 0.1 hectare plots are randomly selected within each cluster. Each plot is then divided into four 0.01-hectare sub-plots from which field measurements are subsequently taken. We adapted this sampling approach to meet the unique needs of our study. Specifically, we treated each prioritized common land area as a sampling cluster. We then randomly sampled 10 plots (including their associated four sub-plots within each). We then took inventories of all the trees, shrubs, and saplings in each of the 40 sampled sub-plots. In addition, we collected data on other LDSF indicators, including soil erosion prevalence, infiltration capacity, and soil properties.

It is important to note that our study is situated in a tropical thorn and dry deciduous forest context where what are now grazing lands were historically covered by both grasses and trees. Consistent with Tewari and Arya (2004), we argue that the presence of native tree species is a good indicator of ecological health in this context, given that their decline or absence is linked to worsening ecological conditions.

Village leader (informant) and household surveys

As the LDSF data collection was taking place, we geared up to administer two surveys, one directed to village leaders and another to male and female representatives from households residing in the matched villages. We developed these two survey instruments using Open Data Kit (ODK) (Hartung et al., 2010). Drafts were then reviewed and piloted in two non-study villages in Udaipur District, Rajasthan, revised accordingly, and translated into Hindi.

Soon after, in late March 2020, the Government of India instituted a series of lockdown measures to contain the COVID-19 pandemic. Face-to-face data collection did not resume until 14 January 2022, which lasted until 15 April of this same year. Morsel India, a professional data collection firm, was contracted to carry out the data collection exercise. Its senior members of staff received an online orientation on the two survey instruments. Then, in each study site, these staff trained teams of experienced enumerators over a four-day period, which included practice runs (pretests) of the survey instruments. During data collection, the Morsel India field teams first met with village leaders to obtain consent for carrying out the survey in the village in question.

While Morsel implemented its own data quality checking procedures, we complimented these with our own. In addition to checking whether the household sampling protocol was followed, this entailed three strategies: (a) the use and review of timestamps, both for the overall survey and for specific modules; (b) verifying enumerator visitation to the selected villages through daily reviews of captured household geocodes; and (c) daily reviews of audio audits. The latter were captured as audio files for randomly selected modules of the surveys. Links for these audio files and results for the other two checks were uploaded online. The audio recordings were reviewed and rated by members of FES’s Study Team. Identified data quality concerns were immediately communicated to the Morsel field teams for correction.

Measurement

Measuring the rollout of FES’s intervention model

To assess the rollout of FES’s intervention model at the village level, we developed the Commons Restoration Action Index (CRAI). Aligned with FES’s core model, the CRAI ranges from 0 to 1, encompassing three dimensions (Rights, Institutions, and Action), each with four sub-dimensions represented by binary indicators. All dimensions and their respective indicators are weighted equally, with weights of 1/3 and 1/12, respectively (Table 1).

Table 1 The Commons Restoration Action Index (CRAI).
Full size table

The binary indicators together, when positive, signify progressive levels of achievement under their respective dimensions. Under the Rights dimension, for example, rights claiming first takes place, followed by formal recognition of these rights, with additional points for having more than one type of right explicitly recognized (e.g., rights related to access, exploitation, management, and exclusion) or the existence of a formal document recognizing these rights. We obtained data to construct the CRAI through the administration of the village informant survey.

Measuring encroachment on common land

In Rajasthan, pasture or grazing land constitutes the predominant type of common land, overseen by Gram Panchayats (village councils) tasked with its management. As per the Rajasthan Land Revenue Act (1956), this land is designated for communal use and grazing purposes. However, in practice, these lands are subjected to persistent encroachment, e.g., for farming and mining. This poses a critical challenge for the state (Gaur et al., 2018). To assess the extent of this challenge, village leaders, guided by community resource maps, were queried about each accessible common land area in or near their respective villages. Villages were coded as having experienced encroachment if such instances were reported.

We corroborated and augmented these self-reported findings with Earth Observation. As explained above, our data collection included geo-tracing prioritized common land in both intervention and control villages. Using Google Earth, we examined the extent each common land polygon had been encroached through agriculture and/or mining activities using a simple ordinal scale—none, minor, significant, and extreme.Footnote 1

Remote sensing analysis and predictive modeling

The primary data collected using the LDSF, including data on soil variables, were then used to train machine learning models (Breiman, 2001) to predict and map indicators based on Landsat satellite remote-sensing data at 30 m resolution. Satellite reflectance data were extracted for each LDSF sampling plot and used as covariates in Random Forest ensemble prediction models to predict and map tree cover (%), soil erosion prevalence (%), and SOC (g/kg) at 0-20 cm depth (Vågen et al., 2013; Winowiecki et al., 2021). Predictions were made based on annual composite images using Landsat 5 and 7 for the year 2000 and Landsat 8 for 2020, after cloud masking of all available input images for each year.

Perceptions of survey respondents

During the household survey, the sex of the respondents for each randomly sampled household was also randomly determined, thereby enabling us to generate sex-disaggregated results for individual-level indicators. In one module, respondents were asked the extent they agree with the following statements: (1) The conditions of common land in our village has improved over the last 10 years; and (2) The availability of products accessible from common land in our village has improved in the last 10 years. While the data were collected on a Likert scale, we transformed the results into binary indicators and compared the percentages of men and women who agreed to these statements among our treatment groups. We present the results in this paper to assess the extent to which the perceptions of local people corroborate with our main results.

Data analysis

Ordinary least squares regression

Given that our matching strategy worked well, we initially compared mean outcome indicator values between the intervention and non-intervention villages directly using ordinary least squares (OLS) regression, without the inclusion of additional covariates. Given that the village is our pseudo-unit of random assignment, we clustered our standard errors at this level. However, as further explained below, comparing the intervention and control villages directly yielded null results in most cases. It was only when we decomposed the intervention villages into relatively older and newer groups that we found significant differences in favor of the former. Given that we did not implement our PSM matching procedures on the basis of intervention timing, we, therefore, included measures of FES’s targeting criteria in our OLS models. As an additional robustness check, we further ran models that included both the matched and unmatched control villages.

Mediation analysis

Moreover, to test the viability of a hypothesized mechanism through which FES’s intervention model gave rise to our identified treatment effects, we conducted statistical mediation analysis (MacKinnon, 2008) implemented with Stata’s sem (structural equation modeling [SEM]) command (StataCorp, 2017). This analysis evaluates the extent to which the variation in the data shared between an independent variable and one or more hypothesized mediator variables predicts variation in an outcome variable of interest. If a significant proportion of this shared variation proves predictive, then the variation in the data supports—but does not conclusively prove—the hypothesized mediating mechanism.

Recall that the CRAI was specifically developed to measure the rollout of FES’s intervention model. Consequently, its potential role as a mediating mechanism was our first candidate. However, while it is positively correlated with both our older intervention village dummy variable and estimated gains in tree cover, it failed our mediation tests (see below). However, when we preformed mediation analysis using only the CRAI’s institutional dimension, the results were positive, motivating us to explore it as a potential mediating mechanism further.

Consequently, we took the four dummy variables that make up this dimension (Table 1) and included three others as complementary indicators to strengthen our measurement approach.Footnote 2 These additional variables include: (1) VI meets at least once per month; (2) VI has its own by-laws; and (3) VI has an executive committee or equivalent. We then tested the inter-item correlation among all seven variables and found it to be high (Cronbach α = 0.79). We then put these binary indicators on a tetrachoric matrix and ran principal component analysis (PCA) on it. We constructed a commons institution index based on the resulting first principal component, which accounts for 66% of the variables’ shared variation.Footnote 3

We further included the proportion of respondents from each village that reported that their village has effective rules for governing and managing the commons into the mediation model. Our reasoning is that if the general population is aware of such rules and views them to be effective, there is likely a strong institution in place governing the commons, even if this is informal or based on traditional systems of governance.

Results

We first investigate the rollout of FES’s intervention model in both the intervention and control villages using the CRAI. Given that we found significant differences between intervention villages targeted before 2011 and after 2010 for key ecological outcome measures (see below), we decompose our results accordingly (Fig. 1). We make several observations. First, the overall average index score is significantly higher for both the older and newer intervention villages (0.25 and 0.21, respectively) compared to the control villages (0.12). We are not surprised that the CRAI is above zero for the control villages; we expected that some of them would have initiated action on the commons, either independently or through the assistance of other organizations. Second, the average score for both intervention village cohorts and, by extension, their relative difference vis-à-vis the control villages are lower than we expected. Our third observation pertains to the Rights dimension; very few village informants in both the intervention and control villages reported pursuing rights-claiming efforts. We revisit this below.

Encroachment on common land

Village leaders from the control villages were far more likely to report common land encroachment. Specifically, 69% of leaders in the control villages reported that at least one common land area in their village had been encroached, compared to 22% in the intervention villages (p < 0.01). We corroborate this through Earth Observation. We find that observable signs of encroachment are significantly lower in the intervention villages, with an ordered logistic regression coefficient of −1.33 (p < 0.05). We further observe visual differences between intervention villages treated before 2011 and after 2010, with significantly lower levels of observable encroachment in the former (see Supplementary Fig. 2).

Ecological indicators

To evaluate whether FES’s intervention model resulted in improvements in the ecological conditions of common land, we compare prioritized common land areas of the matched intervention and control villages. We do this via a) field-based indicators of land health derived through the implementation of LDSF and b) changes in the status of similar measures derived through predictive LDSF calibrated models run on Landsat 2000 and 2020 30-m resolution imagery. For both the field- and remote-sensing-based measures, we again find differential treatment effects in favor of the older (pre-2011) intervention villages. Hence, we highlight these estimated effects in our presentation below.

Fig. 1: Rollout of FES’s intervention model in intervention and non-intervention villages based on the Common Restoration Action Index (CRAI).
Fig. 1: Rollout of FES’s intervention model in intervention and non-intervention villages based on the Common Restoration Action Index (CRAI).
Full size image

Each colored bar represents each indicator’s weighted contribution to the index, which ranges from zero to one. The thicker the bar, the greater the number of villages that scored positive on the indicator in question.

For the field-based measures, we observe significant differences in favor of the older intervention villages for all four biophysical measures, i.e., tree and shrub density and diversity (Fig. 2). This is not the case for the newer intervention villages, where either the distributions are similar (Shannon Shrub Diversity Index) or at a lower level than those of the non-intervention villages. This pattern holds, particularly for the tree density indicator, following our implementation of three statistical models (Supplementary Table 1).

Fig. 2
Fig. 2
Full size image

Violin plots comparing intervention and control prioritized common land areas vis-à-vis land health indicators derived through LDSF field-based data collection, including results of T-tests for mean differences between the distributions.

The number of trees and shrubs per hectare is significantly higher in the prioritized common land areas of the older intervention villages, with average intervention effect estimates ranging from 89 to 131. However, our results are only statistically significant for the tree density indicator. We observe further that tree and, to a lesser extent, shrub diversity is higher in the prioritized common land of the older intervention villages, but the intervention effect coefficients are not statistically different from zero.

A key issue in the above analysis is that—despite our matching efforts based on FES’s targeting criteria—the above-noted differences may have already existed prior to FES’s engagement in the intervention villages. Our second set of indicators—estimated changes in tree cover, vegetation cover, soil erosion prevalence, and soil organic carbon (SOC) from 2000 to 2020—overcomes this challenge. Indeed, we find that all four indicators were statistically balanced between our originally matched intervention and control villages at baseline (p > 0.3), providing further confidence in the success of our two-stage village matching effort in mitigating programme placement bias. We therefore focus our analysis on how these indicators changed over time, again disaggregating the intervention villages by intervention timing (Fig. 3). We visually observe positive results in the older intervention villages.

Fig. 3
Fig. 3
Full size image

Kernel density plots comparing intervention and non-intervention prioritized common land areas vis-à-vis changes (2000 to 2020) in the status of land health indicators derived through remote sensing.

The results of our statistical models (Supplementary Table 2) are consistent with these visual observations, where we find relatively greater gains in tree cover (5.9% to 9.5%) and, to a lesser extent, overall vegetation cover (3.8% to 6.7%) in the older intervention villages. There were greater decreases in estimated soil erosion prevalence in these villages also (–1.4% to −3.22%). The relatively greater gains in estimated SOC—which is recognized as a slow-moving variable (Smith, 2004)—are, however, washed out in the two models that include covariates. Nevertheless, we find strong evidence that FES’s intervention model improved the ecological conditions of common land in the older intervention villages.

Perceptions of survey respondents

We observe that significantly more men in both the older ( ≥ 25%) and newer intervention villages ( ≥ 11%) reported that conditions of common land had improved over the last 10 years, respectively, compared to their counterparts in the control villages. This is largely consistent with the results above. Interestingly, we observe no statistically significant differences among our three intervention groups with respect to the perceptions of female respondents.

Exploring mechanisms for gains in tree cover

We performed mediation analysis using three SEM models: (1) older intervention and matched control villages without covariates; (2) older intervention and matched control villages with covariates; and (3) older intervention and both matched and unmatched control villages with covariates (Table 2). Our total effect estimates comprise both the direct effect—the variation in tree cover gains that the older village intervention dummy variable explains independent from the variation it shares with the two mediator variables—and the indirect effect—this same variation that is explained by the shared variation between the intervention dummy and these mediators. The indirect effect estimates are statistically significant across all three models. We, therefore, find that the variation in our data is consistent with the institutional dimension of FES’s intervention model as being at least partly responsible for the positive ecological impacts we evidence. We acknowledge, however, that our two mediator indicators are unlikely to fully measure the construct of commons village institutional strength.

Table 2 Mediation analysis with commons village institutional strength (as measured both by a PCA-derived commons institution index & perceived effectiveness of rules governing commons) as a hypothesized mechanism between FES intervention rollout & tree cover gains.
Full size table

Discussion and conclusion

Land degradation adversely affects agricultural productivity and life-sustaining ecosystem services, such as climate and water regulation, biodiversity conservation, and carbon storage (Petrosillo et al., 2023). With approximately one-quarter of the global land area exhibiting signs of degradation (Hossain et al., 2020), initiatives seeking to restore degraded land (or, even better, preventing such degradation in the first place) are high on the policy agenda, with 2021 commencing the United Nations Decade on Ecosystem Restoration (Meli et al., 2023).

However, credible evidence is lacking on what interventions work where and at what cost in both reversing and preventing land degradation, spurring demand for well-conducted impact studies (Malan et al., 2022). Indeed, policy makers and practitioners are challenged with what to invest in and upscale. To complicate matters further, generating such evidence can be challenging. First, land restoration programs tend to purposively target the most degraded areas, resulting in non-random program placement and, by extension, challenging the ability of evaluators to reliably estimate the counterfactual. Second, most initiatives fail to capture quality baseline data vis-à-vis both intervention and potential comparison locations, limiting the potential for using quasi-experimental approaches, such as difference-in-differences estimation. Third, key restoration effects, e.g., increases in tree cover and soil organic carbon, are likely to only fully manifest several years following project closure (Constenla-Villoslada et al., 2022).

In addition to providing evidence on the effectiveness of FES’s intervention model, our study makes a significant contribution to the literature in overcoming these challenges. Specifically, we demonstrate how different types and sources of data can be integrated together in the pursuit of a strong and informative quasi-experimental design. While data derived from remote sensing is being increasingly used to evaluate land restoration initiatives (see, for example, Meroni et al., 2017; Constenla-Villoslada et al., 2022; Sacande et al., 2021), we combine these data with both primary survey data (collected through LDSF field measurements and village and household surveys) and secondary data (e.g., census data), to both strengthen our estimation of the counterfactual and enable us to provide (via mediation analysis) a plausible explanation of how the differential effects we estimate for the older treated villages manifested.

Indeed, we found strong and consistent evidence—triangulated through several independent data sources—that FES’s intervention model improved the ecological conditions of prioritized common land, particularly in the eight older intervention villages. Being a key challenge in Rajasthan, we found that this land was less likely to have been encroached on (as reported by village informants and assessed through Earth Observation). The prioritized common land of the older intervention villages is further characterized by significantly higher tree density and diversity. This land further experienced greater gains in tree and, to a lesser extent, vegetation cover between 2000 and 2020. We further found that soil erosion decreased to a greater extent in these common lands as well. Finally, male survey respondents from these villages were at least 25 percentage points more likely to report that the conditions of common land in their villages had improved and that the availability of products from this same land had increased.

However, a key question is why FES’s intervention model only generated ecological impacts in the older, as opposed to newer, intervention villages. One hypothesis is that trees in the older intervention villages had more time to grow and produce larger canopies, as compared with their relatively newer counterparts. However, we rule this out, given that we do not observe a monotonic (much less linear) relationship between intervention timing and changes in tree cover (see Supplementary Fig. 2). For example, we estimate that the prioritized common land of one intervention village entered in 2015 experienced an over 21% increase in tree cover over the 20-year period. This is compared to a 9.4% increase in three villages that were entered in 2011.

The explanation that fits our data best relates to the results of our mediation analysis. Specifically, we find evidence that commons-related institutions were stronger in the older intervention villages, as compared with both the control villages and newer intervention villages. Specifically, the average difference for our common institutional strength PCA index between the control villages on the one hand and older and newer intervention villages on the other is 0.59 (p < 0.05) and 0.25 (p = 0.28), respectively. For the percentages of survey respondents reporting effective rules governing commons in their villages, these same numbers are 49% (p < 0.1) and 22% (p = 0.37). In other words, the effect sizes for the older intervention village vis-a-vis our proxy measures for common institutional strength are more than double those of the newer intervention villages. The reason why this is the case is an area for further investigation.

Moreover, while there are shortfalls in the ability of these proxy measures to precisely measure the construct of common institutional strength, the variation in the data they share with our older intervention village dummy variable explains a large proportion (42% to 59%) of our effect estimates on gains in tree cover over the 20-year period. The variation in our data, therefore, supports the hypothesis that commons-oriented institutions were stronger and more effective in the older intervention villages, resulting in better protection and management of prioritized common land and, in turn, improvements in its ecological health. This is an important finding for ongoing efforts to protect and restore common land in India and beyond. With increasing pressure to go to scale and achieve more with less, investing in the softer aspects of the development process, such as community institutional strengthening, may get side-lined, despite their critical importance in both achieving and sustaining impact.