in

Community-based rangeland management in Namibia improves resource governance but not environmental and economic outcomes

Theory of change

At the heart of the of CBRLM’s theory of change is the assumption that improvements in the ecological sub-system provide a sustainable resource base for increased livestock production and marketing34. The ecological sub-system, however, depends on a functioning economic sub-system because herd owners must be able to destock quickly in response to adverse ecological circumstances. The theory holds that the most important constraint on the economic sub-system is unproductive herds and low-quality cattle because farmers are unwilling to sell their cattle when they command low market prices. Therefore, improvements in rangeland grazing management need to be complemented by improvements in information and access to livestock markets, herd structures, and animal husbandry practices.

Crucially, changes to the ecological, economic, and livestock sub-systems rely on effective community governance and collective-action capacity in CBRLM communities. This is because rangeland grazing management practices can be easily undermined by non-participating herd owners inside or outside the GA. The theory therefore calls for investments at multiple levels of the social-ecological system to ensure that improvements in certain program areas are not undermined by failures in others34. The CBRLM implementers believed that previous rangeland development programs were undermined by a failure to account for the linkages among sub-systems, which motivated them to design a more holistic intervention34.

Intervention components

CBRLM was a multi-faceted package of administrative, educational, financial, and technical support. Implementation of the package was designed as an experimental treatment to assist in project assessment. To select study areas for evaluation, GOPA identified 38 RIAs with sufficiently low density of people, livestock, and bush cover to enable the implementation of new group-grazing plans, one of the core treatment components. The evaluation team randomly assigned 19 RIAs to treatment and 19 RIAs to control (see Randomization for details). GOPA implemented CBRLM in up to seven GAs within each treatment RIA.

Mobilization

GOPA conducted pre-mobilization meetings with TAs and other stakeholders in the second half of 2010 to identify GA communities most likely to participate in CBRLM34. Early mobilization efforts focused on soliciting community buy-in for the cornerstone principles of CBRLM, including community-planned grazing, combined herding of cattle, and efficient livestock management. There is also substantial evidence from qualitative surveys that some community members were motivated to participate in the CBRLM by prospects for water infrastructure development by GOPA34.

While almost 100 GAs were initially mobilized for the project, by 2014 GOPA was targeting resources and support towards 58 GAs based on community receptivity and the discretion of CBRLM management. In each GA, GOPA worked principally with households owning 10 or more cattle, although other community members benefitted from participation in a “Small Stock Pass-on Scheme” and a variety of training activities, which are described below.

Rangeland grazing management

The core aim of CBRLM was to shift how communities approached livestock grazing, forage conservation, and risk management by encouraging two key practices: planned grazing and combined herding. Planned grazing entails rotating a community’s cattle to a new pasture on a regular basis in accordance with a written plan. The goal was to preserve grass for the dry season and allow grazed pastures more time to recover. Combined herding entails grouping many owners’ cattle into one large herd and herding them in a tight bunch. This practice is meant to concentrate animal impact on rangeland, minimize cattle losses, and increase the likelihood that cows are exposed to bulls, thus increasing the pregnancy and calving rates of the entire herd. The scientific and practical rationale behind these practices is reviewed in Supplementary Note 2.

GOPA staff developed grazing plans with each participating community and taught them planned grazing and combined herding via field-based training sessions. These followed a “training of trainers” approach in which GOPA recruited field facilitators from each community, taught them the principles of CBRLM, and tasked them with training their fellow participating pastoralists.

Livestock management

GOPA taught participants some best practices in animal husbandry, including structuring herds to maximize productivity (by increasing the proportion of bulls and reducing the proportion of oxen and cattle over the age of 10 years), providing vaccinations and supplements, and deworming34. Additionally, to support the introduction of more bulls into herds, the project implemented a “bull scheme” in which participating communities were given the opportunity to collectively buy certified breeding bulls at a subsidized price. Communities were meant to repay the cost of the bulls either with cash or in-kind trades of goats. Goats collected in this repayment process fed into the small stock pass-on scheme under which participating community members nominated households to receive goats from GOPA. GOPA requested that communities nominate households that owned few or no livestock and were led by youth and/or women. When GOPA received goats as payment for loaned bulls, they would pass them on to nominated households. The recipients were then expected to pass on the offspring of the goats they received to other disadvantaged households.

Cattle marketing

CBRLM also sought to increase participants’ marketing of cattle to generate revenue from livestock raising and encourage offtake of unproductive animals34. Community facilitators and project experts provided participating herd owners with information about market opportunities and ideal herd composition, and encouraged flexible offtake in response to forage shortages. In 2013, GOPA invested in the development of regional livestock cooperatives that held local auctions and helped farmers transport their animals to markets. Finally, GOPA invested in identifying international export opportunities for CBRLM farmers to Zimbabwe and Angola, although these were generally not successful31.

Community development

The project sought to institutionalize community-level governance to organize and enforce collective activities like planned grazing, water point maintenance, and financing of livestock inputs. The central management unit of each GA was a new Grazing Area Committee consisting of five to 10 elected community members. The project encouraged participating communities to collectively cover operational expenses in their GA through a GA fund managed by the committee. Among these expenses were the payments to herders, costs of diesel for water pumps and maintenance of water infrastructure, financing collective livestock vaccination campaigns, and any other collective expenses that would support operation of the GA. CBRLM supported every GA fund with a 1:1 matched subsidy. The matched subsidy was limited by a ceiling amount determined by the estimated number of cattle in a GA. GOPA also instructed committees to maintain “GA record books” to track grazing plans, record meeting minutes, and keep logs of community members’ participation and financial contributions.

Water infrastructure

GOPA upgraded water infrastructure at a total of 84 sites throughout the NCAs to facilitate planned grazing and combined herding. Water infrastructure improvement included minor upgrades like water tanks and drinking troughs, and larger investments such as the installation of diesel and solar pump systems, the drilling and installation of boreholes, and the construction of pipelines, deep wells, and a large earthen dam31.

Intervention timeline

The timeline for major components of the research process and CBRLM roll-out is illustrated in Supplementary Fig. 1. The research team conducted the random assignments and the implementation team began community mobilization in early 2010. Formal enrollment in CBRLM began in early 2011. The program implementer conducted mobilization in two waves: they mobilized 11 of 19 RIAs in 2010 and the remaining 8 RIAs in 2011. The evaluation team conducted qualitative data collection to inform the design of social and cattle surveys prior to project end 2014; social surveys in 2014 and 2016; rangeland surveys in the wet and dry seasons of 2016; a cattle survey in 2016; and a household economic survey in 2017.

Cumulative GA-level implementation is illustrated in Supplementary Fig. 2. The project implementer first formally reported enrollment and field visits in April 2011. The implementer achieved nearly full targeted enrollment (50 GAs) by November 11, although some grazing areas were added or subtracted thereafter. Mobilization exceeded enrollment because some grazing area communities chose not to participate in the program and some enrolled in the program and then dropped out. The program averaged between 25 and 50 field visits per month over the project period. A field visit consisted of a week-long community meeting about grazing-plan development and implementation, animal husbandry and budget training, and marketing opportunities.

Randomization

The unit of randomization is the RIA, an intervention zone with a locally recognized boundary. Each RIA falls under the jurisdiction of a single local governing body, known as a Traditional Authority (TA). As noted above, RIAs contain five to 15 GAs where a community of producers share water and forage resources. Grazing areas do not have legally defined boundaries. A herd owner’s ability to move among GAs is variable.

GOPA mapped 41 RIAs prior to randomization. Three contiguous RIAs in the north-central region, composed of two treatment RIAs and one control RIA, were omitted from the study post-randomization because reexamination of baseline density of bushland vegetation deemed them unviable for CBRLM implementation. These are the three RIAs without sampled GAs in Fig. 1. The other 38 RIAs were randomly assigned to either receive the CBRLM treatment (19 RIAs) or serve as controls (19 RIAs).

The randomization was stratified by TA to ensure that at least one RIA was assigned to the treatment in each TA. The research team then re-randomized the sample units until seven variables were balanced (a p value of 0.33 or higher for an omnibus f test of all seven variables) between treatment and control: (1) Presence of forest; (2) number of households; (3) number of cattle; (4) cattle density per unit area; (5) quality of water sources; (6) presence of community-based organizations (CBOs); and (7) overlap with complementary interventions (see Supplementary Table 1). For future researchers, we recommend re-randomizing a set number of times and choosing the re-randomization with the highest balance35. These variables and indicator variables for TA are included as covariates in all analyses.

Sample selection

In the original sampling strategy, the project implementer was asked to predict the GAs where they would implement the project if the RIA were assigned to treatment. However, there was limited overlap between the GAs that the implementer predicted and the GAs where CBRLM was ultimately implemented. Therefore, the evaluation team devised a revised sampling strategy in 2013, which proceeded in four steps:

  1. 1.

    Map GAs in sampled RIAs: The evaluation team traveled to all 38 RIAs and worked with TAs and Namibian Agricultural Extension (AE) officers to map all the GAs in each RIA. The team mapped 171 GAs in control RIAs and 213 GAs in treatment RIAs.

  2. 2.

    Collect pre-program data on GAs: The evaluation team collected information on pre-program characteristics of each GA from interviews with TAs and AE staff, the Namibian national census36, and the Namibian Atlas37. The latter has a geo-referenced database on climate, ecology, and livestock for the nation.

  3. 3.

    Predict CBRLM enrollment for treatment GAs: The researchers used these data in a logistic regression to predict the probability that each GA would enroll in CBRLM and would adopt the CBRLM interventions based on pre-program characteristics. For example, the model found that GAs with more existing water infrastructure, strong social cohesion, and adequate cell phone service were more likely to be enrolled in the program. The variables used to predict CBRLM adoption were: (1) Presence of water installations (yes/no); (2) carrying capacity of the land (above/below the regional median); (3) community’s readiness to change (high/very high); (4) community’s social cohesion (high/very high); (5) spillover effects from neighbors; (6) quality of herders and herder turnover; (7) presence of members of the Himba ethnic group; (8) the TA’s readiness to change; (9) cell phone coverage; and (10) primary housing material (mud, clay, or brick).

  4. 4.

    Generate sample of GAs in treatment and control RIAs: The evaluation team applied the statistical model (above) to all GAs in the sample and set a cut-off point to separate GAs that were likely to adopt the CBRLM program vs. those that were unlikely to do so. In treatment RIAs, the model predicted 52 GAs, of which 37 were formally enrolled in CBRLM and 15 were not. In control RIAs, 71 GAs met or exceeded the cutoff; they offer the best counter-factual estimate of which GAs would have enrolled in the program had their RIA received treatment.

Data collection

The names, survey questions, and variable constructions for all outcomes included in the analysis are available at the AEA RCT Registry (ID number: AEARCTR-0002723). See Supplementary Methods for a list of definitions of variables depicted in Fig. 2 and 3.

Social surveys

Social surveys were intended to assess the effect of CBRLM on community behaviors, community dynamics, knowledge, and attitudes. All data were collected using electronic tablets with the SurveyCTO software38.

The primary unit of analysis for household respondents is the manager of the cattle kraal (holding pen). Researchers conducted surveys with kraal managers, rather than heads of households, for three reasons. First, many kraals contain cattle owned by multiple households, and decisions about grazing practices, cattle treatment, and participation in grazing groups are generally made at the kraal level. Second, many cattle-owning households do not directly oversee the day-to-day activities of their cattle (many live outside the GA), and so would be unable to answer questions about key outcomes, such as livestock management behaviors and community dynamics39. Finally, enrollment in CBRLM occurred at the kraal, rather than household, level.

In 2014, the research team worked with local headmen and other community members to generate a complete census of kraals in every sampled Grazing Area (GA) that contained 10 or more cattle at the start of the program (an eligibility requirement for enrollment in CBRLM). The research team randomly sampled up to 11 community members for participation in the 2014 kraal manager survey. Surveys were conducted in the manager’s local language and lasted ~45 min. Alongside the 2014 survey, teams of two surveyors visited all grazing areas where at least one respondent reported participating in a community grazing group or community combined herd to corroborate reported behaviors through direct observation.

To assess the persistence of CBRLM’s effects on behaviors, community dynamics, knowledge, and attitudes, the research team conducted a follow-up survey of kraal managers in 2016, two years after program end. The survey team randomly sampled two additional kraals in each grazing area to account for the possibility of attrition. The 2016 survey lasted approximately one hour on average, and included an expanded list of questions about governance, social conflict, and collective action as well as new survey modules on cattle marketing, cattle movement, and livestock management. In 2017, the research team randomly sampled three kraals in each grazing area to conduct direct observation audits of key rangeland grazing-management behaviors.

To assess the effects of CBRLM on economic outcomes, the research team conducted a household-level survey in 2017, three years after program end. The survey instrument asked detailed questions on topics that could not be answered by kraal managers, such as household consumption, income, food security, and savings. To select households for this survey, during the 2016 survey the research team asked kraal managers to list all households that owned cattle in the manager’s kraal, then randomly selected one household from each kraal. Alongside the 2017 survey, the research team conducted an in-depth survey with the local headman of all 123 GAs in the sample. The headman survey focused on historical background about the grazing area, as well as the headman’s perceptions of rangeland and livestock issues.

Cattle data

The cattle component was intended to assess effects of CBRLM on cattle numbers, body condition, and productivity. The variables of key interest involved the average liveweight and body condition, calving rates, and average market value of cattle, as well as overall herd structures.

The data collection protocols closely followed standards from livestock assessments elsewhere in Sub-Saharan Africa40. The research team randomly selected up to six kraals in each GA to participate in the cattle survey. The survey team mobilized selected herds during multiple community visits to ensure all herds were accounted for. Herd owners were compensated for the costs of rounding up animals and weighed cattle received anti-parasite treatment (“dipping”)41. A total of 19,875 cattle from 669 herds were weighed.

The data-collection process for each herd proceeded in six steps. First, surveyors worked with herd managers to round up all cattle that regularly stayed in the selected cattle kraal. Once cattle had been brought to the designated location for data collection, they were passed through a mobile crush pen and scale. As each animal passed through the crush pen, a survey team member recorded the animal type (i.e., bull, ox, cow, calf) and used a SurveyCTO randomizer to calculate whether the animal was randomly selected for assessment. The random number generator was set to randomly select approximately 30 cattle from each herd for weighing. If the animal was selected, the survey team kept the animal on the scale and recorded its weight and body condition. A semi-subjective 1–5 scale, commonly used by livestock buyers in the NCAs (see Supplementary Fig. 3), was adjusted to a 0–4 scale used to determine formal market pricing. The team then placed the animal in a neck clamp and estimated the animal’s age by dentition (but extremely young calves were aged visually). Each animal was marked as it moved through the crush pen to ensure that it was assessed only once. In addition to assessing randomly selected animals, the survey team weighed and aged all bulls in the herd. The cattle survey yielded average cattle weight, age, and body condition for 19,875 animals across all treatment and control GAs, as well as estimates of calving rates, ratios of bulls to cows, and ratios of productive to unproductive animals.

Rangeland data

The rangeland ecology research was intended to assess treatment effects on vegetation and soil surface conditions. Full research details, including field technician training protocols, are available elsewhere42. The data collection approach followed methods commonly used in Africa43,44. Extended definitions of variables depicted in Fig. 3 and Table 2 are available in the “Supplementary Methods” section.

The rationale for how the ecological variables presented in Fig. 3 translate into assessments of rangeland condition or health is based on forage and soil characteristics from a livestock production perspective25. The highest quality forages for cattle on rangelands are perennial grasses, since annual grasses are more ephemeral in terms of nutritive value and productivity. Herbaceous forbs often have the poorest forage quality for large grazers because of their low fiber content and risks of containing toxic chemicals. When rangelands are degraded by over-grazing, perennial grasses are reduced and replaced by annual grasses and forbs. This trend reflects animal diet selectivity that favors consumption of the perennial plants. Reversing such trends via management interventions can be difficult. The main option is to reduce grazing pressure and hope that perennial grasses can outcompete annuals and become reestablished over time. Another option is to implement a grazing rotation that allows perennial grasses to recover after a grazing period.

Increases in annual grasses are documented to occur as one outcome of chronic overgrazing in Namibia45,46. In 2016, annual grasses were 5-times more abundant than perennial grasses in our study area. When over-grazing occurs, most plant material is harvested and less is available for the pool of organic matter (OM) for the topsoil. Less OM (e.g., plant litter) on the soil surface means that more soil is also exposed to wind and rain, accelerating erosion. The GAs in our research occur on various soil types and landscapes, some of which are more susceptible to erosion than others. Silty soils on slopes are vulnerable to erosion, for example, while sandy soils on level sites are less vulnerable25.

On-the-ground sampling was conducted in all 123 selected GAs along an 800-km zone running West to East. Elevations ranged from 750 to 1700 masl (West) and 1050 to 1120 masl (East). Within each sampled GA, up to 12 1-ha (square) sampling sites were initially chosen using coordinates generated randomly from latitude and longitude coordinates in a satellite image of the GA47. About 17% of sites were later removed from the sample based on their close proximity to landscape disturbances or inaccessibility by field technicians. Overall, 972 sites were analyzed in the wet season and 885 in the dry season of 2016, two years after the implementation phase of CBRLM had ended.

The geographic center-point for a sampling site was generated using a spatially constrained random distribution algorithm applied to the satellite image, and the field team navigated to the center-point coordinates using GPS technology. The team took photographs and recorded descriptive information including elevation, slope, aspect, other landscape features, vegetation type, dominant plant species, soil type, soil erosion, and degree of grazing or browsing pressure, and proximity to high impact areas such as trails, water points, and villages.

At the center point, the survey team then established two perpendicular transects, each 100 m in length and crossing at the middle. The resulting four, 50-m transect lines ran according to each cardinal direction (N, S, E, W) as determined with a compass. Technicians then placed 1-m notched sampling sticks at randomized locations along each transect line and recorded what plants or other materials (i.e., stone, wood, leaf litter, animal dung, etc.) were located under or above the notches of the sampling sticks. These data points were tabulated to calculate percent cover for various categories of vegetation; there were n = 200 data points per site based on 40 stick placements and 5 notches per stick. This method enabled precise calculation of cover values for herbaceous (i.e., grass, forb) and diminutive woody plants (i.e., small shrubs, seedlings, saplings, etc.). Tree cover was estimated from point data collected via a small adjustment in the approach42. Herbaceous species were identified in wet seasons but not in dry seasons due to senescence during the latter.

Quadrat sampling supplemented the notched stick approach. Random placements of a 1-m2 quadrat frame within the sampling site allowed for 20 estimates of a soil surface condition score ranging from 1 (poor) to 2 (moderate) or 3 (good)42. Poor was indicated by smooth soil surfaces, absence of litter, having poor infiltration and signs of erosion such as rills, pedestals, or terracettes; good was indicated by rough soil surfaces, abundant litter, seedlings evident, and lack of evidence of erosion. Herbaceous biomass was estimated in the quadrats and weighed to estimate herbaceous biomass.

Statistics

Index creation

Index construction for socioeconomic variables was composed of several steps48. For each response variable we first signed all component variables such that a higher sign is a positive outcome, i.e., in line with CBRLM’s intended impacts. Then we standardized each component by subtracting its control group mean and dividing by its control group standard deviation. We computed the mean of the standardized components of the index and standardized the sum once again by the control group sum’s mean and standard deviation. When the value of one component in an index was missing, we computed the index average from the remaining components. See Tables 3–6 for index components.

Calculation of average treatment effects

The estimate of interest is the Average Treatment Effect (ATE), or the average change in an outcome generated by assignment to CBRLM. We estimate the ATE using standard Ordinary Least Squares regression and control for variables used in stratification. Regressions for rangeland outcome variables include a unique set of controls, including rainfall over the project period, rainfall in the year of data collection, grazing area cattle density, grazing area ecological zones, and a remote-sensing estimate of pre-project biomass. The core model takes the form:

$$hat{Y}=alpha +{beta }_{1}T+{{{{{boldsymbol{beta }}}}}}{{{{{bf{X}}}}}}$$

(1)

where T represents treatment assignment and X represents pre-treatment covariates used to test for balance during re-randomizations. The results capture the intention-to-treat (ITT) effect rather than the effect of treatment-on-treated (TOT). ITT is more appropriate than TOT in this context for two principal reasons. First, it is more relevant for policymakers – the effect of policies should account for imperfect compliance. Second, “uptake” is not well-defined, and certainly not a binary concept, for CBRLM since many communities and community members complied partially, complied with some but not all components, and complied for some but not all of the time.

Standard errors and p values

We report two-tailed p values for all analyses. For each outcome, we show the two-tailed p value from a standard Ordinary Least Squares (OLS) regression with standard errors clustered at the level of the RIA, the unit of randomization49. We also calculate two-tailed p values using Randomization Inference (RI). To calculate RI p values, we re-run the randomization procedure (described above) 10,000 times and generate an Average Treatment Effect (ATE) under each hypothetical randomization. The p value is the percent of re-randomizations that generate a treatment effect that is either equal to, or larger in absolute value than, the true ATE.

Multiple hypotheses correction

We calculate q values to account for families of outcome indices with multiple hypotheses50. The q value represents the minimum false discovery rate at which the null hypothesis would be rejected for a given test. We pre-specified five families of indices:

  1. 1.

    Behavioral outcomes (all in 2014): Grazing planning, Grazing-plan adherence, Herding practices, and Herder management.

  2. 2.

    Behavioral outcomes (all in 2016): Grazing planning, Grazing-plan adherence, Herding practices, and Herder management.

  3. 3.

    Primary material outcomes: Cattle herd value (2016), Herd productivity (2016), Household income (2017), Household expenditures (2017), Household livestock wealth (2017).

  4. 4.

    Secondary material outcomes: Time use (2017), Resilience (2017), Female empowerment (2017), Diet (2017), and Herd structure (2016).

  5. 5.

    Mechanisms: Collective Action (2014, 2016), Community Governance (2014, 2016), Community disputes (2014, 2016), Trust (2014), Self and community efficacy (2014, 2017), and Knowledge (2016).

Heterogeneous treatment effects analysis

We are interested in whether the effect of CBRLM was impacted by lower rainfall in some grazing areas during the project period. We evaluated heterogeneous treatment effects by rainfall in grazing areas using a variety of measures of rainfall, including aggregate rainfall during the project period and deviation in aggregate rainfall from the ten-year mean during the project period.

For simplicity, Supplementary Tables 5 and 6 present the results of analysis of the interaction between treatment and a binary indicator of low rainfall. To construct this indicator, for each GA we first compute the absolute difference between mean rainfall during the project and mean rainfall during the 10 years prior (2000–2010). We divide the absolute difference by mean rainfall during the 10 years prior to produce a relative (%) difference. We then determine the median relative difference over all GAs. For each GA, we assign the value 1 to the low rainfall indicator if the relative difference for the GA is less than the median relative difference over all GAs; we assign 0 otherwise. The results are consistent when we use alternative rainfall measures.

Spillovers analysis

Because CBRLM grazing areas were more likely to experience external incursions by cattle herds from outside the community, we test for spillovers. Specifically, we are interested in whether control grazing areas near treatment areas were affected by having a treatment grazing area nearby. We conducted the spillovers analysis only on control group grazing areas. For each control group grazing area, we measured the distance to the border of the nearest treatment grazing area. We created a binary measure taking the value 1 if the distance between the control group grazing area and nearest treatment group grazing area is below the median distance, and 0 otherwise. We find no evidence of spillover effects. The results are presented in Supplementary Table 7.

Ethical considerations

Approval for this study was obtained from the Institutional Review Boards at Yale University (1103008148), Innovations for Poverty Action (253.11March-001), and Northwestern University (STU00205556-CR0001). The program was conceived, designed, and implemented by the Millennium Challenge Account compact between the Millennium Challenge Corporation and the Government of Namibia. The research team did not participate in program design or implementation. Communities and individual farmers were informed that they were free to withdraw from participation in evaluation activities at any time. The random assignment of the program was appropriate given the uncertainty around the program’s effect, and the Government of Namibia committed to implementing the program in control areas if the evaluation showed positive results.

The research team took a number of steps to ensure the autonomy and well-being of study participants. First, we designed the survey and data collection protocols after considerable qualitative field work to ensure that questions about sensitive issues (e.g., cattle wealth, cattle losses, attitudes towards the Traditional Authority) were phrased appropriately and did not engender adverse emotional or social consequences. Second, all survey activities were reviewed and approved by the MCA compact, Regional Governors, and Traditional Authorities. Third, surveys were conducted with informed consent and in private to ensure that information remained private and respondents were as comfortable as possible during the survey. Finally, the research team disseminated findings on market prices and rangeland condition to communities and regional Agriculture Extension Officers.

We received no negative reports about the community reception of the survey from surveyors during the evaluation. Two cows were injured during the cattle weighing exercise, and the owner was financially compensated in line with a compensation agreement made with all farmers prior to the cattle weighing exercise.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.


Source: Ecology - nature.com

3 Questions: The future of international education

Advancing public understanding of sea-level rise