· web viewwe also looked at human population density from the socioeconomic data and applications...

Click here to load reader

Upload: phamdat

Post on 19-May-2018

214 views

Category:

Documents


1 download

TRANSCRIPT

Supporting Information: J Cinner et al. Comanagement of coral reef social-ecological systems

Data collection methods

We studied 42 independent comanagement arrangements spanning five Indo-Pacific countries: Kenya, Tanzania, Papua New Guinea, Indonesia, and Madagascar. Some comanagement arrangements included multiple villages, leading to a total of 65 surveyed villages. Sites were purposively selected to represent a variety of governance systems and user characteristics, as well as social, economic, and political settings. We used purposive sampling to ensure variation in independent variables and to allow for causal inferences with relatively low levels of bias (1). Whereas purposive sampling is an appropriate strategy for exploratory studies such as this (1), caution should be used when interpreting our results because the villages were not selected randomly. A priori assumed success was not part of our site selection criteria but we did select one to two sites in each country where the authors had long-enduring (i.e. 10-25 years) relationships as research scientists and were intimately familiar with the local ecology, culture, and institutions (these sites were not necessarily considered successful). The exception to our purposive sampling strategy was Kenya, where sites were randomly selected from a list of 33 pilot comanagement institutions. Random selection was feasible in Kenya because a comprehensive list of comanagement sites was available and the relatively small coastline meant it was feasible to study any site on the list. These conditions were not present in the other countries we studied and we consequently used purposive sampling. In these types of linked social-ecological systems studies, it is not necessarily feasible to have strict control sites (i.e. there may not be a similar unmanaged site with social and ecological characteristics that are both comparable). Likewise, given the long history of some of the management systems we have studied (e.g. formal recognition since the 17th century 2), before and after management experiments were clearly not possible. Nevertheless, we have done everything possible to reduce bias from our sampling strategy. This type of field-based comparative study has important advantages over meta-analyses of published literature (3) because it is more likely to fit the statistical assumptions of haphazard or random selection of treatments and not subject to publication bias (i.e. only publishing studies of comanagement where there is a significant positive result).

To gather information and triangulate results in each study site, we employed a combination of household surveys, semi-structured interviews with key informants (community leaders, resource users, and other stakeholders), and analyses of secondary sources such as population censuses (Table S1). All interviews were conducted in the local language by highly trained scientists from the respective countries. In total, we conducted 1374 household surveys (of which 960 were resource users), 53 key informant interviews, 54 community leader interviews, and 51 organizational leader interviews. These interviews provided information on 22 covariate attributes relating to the local governance system; the social, economic, and political setting; and the socioeconomic characteristics of resource users in each community (Fig. S1, Table S1).

Fig. S1. Social-ecological system attributes, grouped according to classifications described by Ostrom (4, 5). The social-ecological system is organized into several key sub-components, including: the governance system (GS), consisting of the institutional design and the operational rules in use; key socioeconomic characteristics of the resource users (U), which incorporates issues such as poverty and dependence on resources; the social, economic and political setting (S) in which they both operate; and the social and ecological outcomes (O) resulting from the interactions of the other sub-components. Bold font indicates components according to Ostroms classification, followed by a summary of how they have been measured in this study (further details in Table S1). Dotted lines indicate potential inter-relationships between system components (4, 5), although co-linearity measures were minimal (SI).

* Denotes the variable was not included in the final analysis

Key informant interviews were conducted using a semi-structured interview form in their native languages. Key informants were selected using non-probability sampling techniques, such as convenience sampling (for example, a respondent may be approached during resource use activities) (6). Between one and five key informants were interviewed per comanagement site. Community walk-throughs were also used to identify new issues and verify responses from surveys (7). Secondary sources, such as previous studies, local by-laws, population censuses, etc. were examined to gather information on operational rules in use, and population size (Table S1).

Two sampling strategies were used to target resource users. In Kenya and some Indonesian villages, where it was not feasible to survey enough resource users using household surveys within the community (due, for example, to low numbers of fishing households), respondents were randomly selected from lists of resource users provided by comanagement leaders. Lists were cross-referenced with other fishermen for accuracy. In all other sites, households were systematically surveyed, whereby a sampling fraction of every ith house (e.g. 2nd, 3rd, 4th) was determined by dividing the total village population by the desired sample size (6). Variance from the systematic sample was assumed to be equal to the estimated variance based on a simple random sample (8). For the purposes of this study, we only examined resource users, so households that did not engage in marine resource use were dropped from further analysis. After dropping non-resource user households, the average number of resource users surveyed per community was 23, but ranged between 7 and 75.

Operationalization of indicators

We collected data on 22 institutional and socioeconomic covariates (16 of which remained in the final results) and three comanagement outcomes. These indicators were developed to operationalize Ostroms diagnostic framework for social-ecological systems (4, 5) (Fig. S1, Table S1). This framework organizes a social-ecological system into several key sub-components, including: the governance system (GS); key socioeconomic characteristics of the resource users (U); the social, economic and political setting (S) in which the system operates; aspects of the resource and resource system, and the social and ecological outcomes (O) resulting from the interactions of the other sub-components (Fig. S1, Table S1). Our study focuses on a single resource system, shallow-water coral reef fisheries, to standardise for a number of resource characteristics (e.g. the productivity, mobility, and predictability of the resource being managed) thought to influence the success of common property institutions (5) (Table S1).

Governance System (GS)

Governance system covariates were primarily those associated with Ostroms institutional design principles (9).

Operational rules. All comanagement arrangements provided members with de jure or de facto rights to develop and enforce at least one of the following operational rules: fishery closures, gear restrictions, temporal closures, and restricted access of non-members. Each operational rule was a covariate in the analysis, except for gear management, which was related to at least one other variable.

Existence of clearly defined resource boundaries was recorded as yes or no, depending on whether all key informants agreed that there has never been confusion about the boundaries. In cases where there has been confusion, the interviewee was probed for further details and where possible, court cases or other documents were examined. Likewise, clearly defined membership was also recorded as yes or no depending on whether key informants agreed that there has not been confusion over who are members of the organization involved in comanagement.

Collective choice rules. Respondents participation in resource management decision-making was examined by asking them how they participated in decisions about resource management and classifying responses as: 0= no attendance; 1= passive attendance (e.g. attend meetings but do not talk or participate) and; 2= active participations (e.g. elected role, actively voice opinions in meetings etc.).

Monitoring and sanctioning processes. A comanagement site was considered to have graduated sanctions when key informants and community leaders affirmed that sanctions increased with the severity or number of offences committed. The effectiveness of conflict resolution mechanisms was examined by asking key informants and organizational leaders whether mechanisms were in place to resolve conflicts between organizational members. If so, they were asked whether they had successfully resolved none, few, half, most, or all of the conflicts. These responses were treated as a five point ordinal scale of the effectiveness of resolving conflicts. Having no conflict resolution mechanisms in place was considered a lower ordinal level than one that was in place, but was unable to effectively resolve any conflicts.

User characteristics (U)

Number of users. Population size was taken from local census information or gathered from our surveys by multiplying the average household size by the total number of households in the community. We also looked at human population density from the Socioeconomic Data and Applications Center (SEDAC) gridded population of the world database (available Online http://sedac.ciesin.org/gpw/global.jsp). Geographic coordinates of field sites were overlaid on the gridded population database. Grid cells were 4.66 km2. When a field site was near the border of two grids, those grids were averaged to give a mean population density. However, population size co-varied with population density. Since common property theory suggests that group size (9, 10), rather than group density is important to the success of commons institutions, we retained population size in the analysis.

Socioeconomic attributes of users. We examined a multivariate index of material style of life. Material style of life is a composite measure of wealth, based on the presence or absence of household possessions, used as an indicator of relative wealth or social status in a community (7). To ensure that our indicators of wealth were comparable and culturally appropriate across nations, in each country key informants were asked to describe the house of a rich person and the house of a poor person. This resulted in 13 items that were common across the study countries, including electricity, television, VCR, fan, radio, the type of walls, roof, and floor, etc. (Table S2). These items were reduced into a single index of material style of life using principal component analysis (Table S2). The first component explained 40% of total variance. We also recorded the highest level of formal education attained by respondents (recorded in years).

History of use. We examined the number of years that the comanagement arrangement had been operational as an indicator of the history of involvement in collective action. In the few instances when the system had been in place as long as key informants could remember, it was given the value of 60 years (the approximate time that many village elders could first recall the system).

Leadership. Leadership was examined by asking respondents to rank the degree to which they trusted their local leader on a five point Likert scale (11).

Norms/social capital. As indicators of social capital, we examined levels of trust, the frequency of participation in community events, and group homogeneity. We asked resource users to describe their level of trust of other community members on a 5 point Likert scale. We also examined how frequently resource users participated in community events, such as feasts, ceremonies, celebrations, etc. The frequency of participation in community events was standardised to the mean for each country. Thus we examined whether respondents participated in more of less than the average number of community events in the context of their specific country. After adjusting by country, there were six outliers with high values. These outliers were capped at 5, which was the upper limit of the remaining 894 cases. As an indicator of group homogeneity, we also examined the proportion of resource users that were migrants. Migration was examined by asking respondents where they were born. Respondents born in other communities were considered migrants.

Importance of resource. We examined two aspects of dependence on fishery resources: occupational diversity and primary dependence on marine resources. Respondents were asked to list all of the jobs they do to bring food or money into the house, and then rank these in order of importance (12). We grouped occupations into the following categories: fishing, selling marine products, tourism, farming, cash crops, gleaning, salaried employment, the informal sector, other, and none (see 12 for details). Livelihood diversity was considered to be the total number of different occupation categories extant in the household. Households were considered primarily dependent on marine resources if they ranked fishing, selling marine products, or gleaning as the primary occupation.

Knowledge of social-ecological system (SES)/mental models. We examined the degree to which respondents recognized that humans were causal agents of change in marine systems by asking an open-ended question about what could increase the number of fish in the sea. Respondents that mentioned human-related agency were given a one, whereas respondents that mentioned non-human (e.g. fatalistic or meta-physical) mechanisms were given a zero (13). This covariate is abbreviated as human agency in Fig. 4.

Social, economic, and political setting (S)

Markets. To provide an indication of the level of market access for marine resources, we examined the distance between the community and the provincial market (kilometres). Several studies have found this indicator to be significantly related to in situ ecological conditions on coral reefs as well as both the length and the trophic level of catch (14-16).

Outcomes (O): measures of comanagement success

We examined two social (perceived impacts on livelihoods and reported compliance) and one ecological (in situ levels of reef fish biomass) comanagement outcomes.

Social performance measures. Responses about the perceived impact of the comanagement arrangement on respondents livelihoods were grouped into positive, negative, or neutral/dont know. Resource users were also asked to gauge their perceptions of the level of compliance within their community on a four point Likert scale (full compliance, some people break rules, most break rules, all break rules) for each operational rule limiting resource use (11). In cases where compliance for different operational rules were different (e.g. better compliance for a protected area than a gear restriction), the scores were averaged across rules (scaled to between zero (none) and 100% (full) compliance). Reported compliance is a good indicator of the degree to which resource users within a community follow the rules in situations where the quality and consistency of infringement reporting is unreliable (17). Although this indicates individual perceptions of community-level behaviour, there is some evidence from, for example medical studies on doping behavior, that individual behaviour is somewhat correlated with (and thus can be indicated at a population level from) perceptions of the behaviour of others (18).

Ecological performance measures. In situ ecological conditions of fished resources were examined to see how comanagement coincided with ecosystem status. Fish biomass (kg/ha) was measured in 27 of the comanagement sites using underwater visual census (UVC). The abundance of all diurnally active, non-cryptic, reef-associated fishes was quantified at each site, and their size (cm total length) estimated. Biomass was calculated for fish >10 cm total length, as the abundance of smaller bodied fish is typically underestimated using visual census (19). Pomacentrids were excluded from biomass estimates because the size cut off meant that certain species had been excluded in some locations. Large subsets of the dataset have been standardized in this way for previous studies (20, 21). To compare comanagement ecological performance measures with other management systems, we also examined 26 permanently settled areas that largely lacked local scale management (open fisheries) and 14 no-take fisheries closures across the countries we examined. This incorporated all available reef fish biomass data across different management systems in our study countries (22, 23) that were directly comparable to our study sites. Although the approach of comparing outcomes from different management strategies at a single point in time does not explicitly address causality, long-term studies of ecological recovery (e.g. 24) and the consistency of these findings across our study sites (22, 23) and throughout the literature more broadly (25, 26) lend credible support to the notion that these differences in ecological conditions may be due to management.

To determine the likelihood that levels of fish biomass in the comanaged sites were overharvested, we used the concept of multi-species maximum sustainable yield, which suggests that in multi-species fisheries, sustainable harvest levels will be achieved at some proportion of unfished biomass (B0) (23, 27). Stock-recruitment models for temperate species regularly estimate such values and, based on these and known population growth models, we used 0.25-0.5 B0 as a conservative but theoretically defensible window of multispecies maximum sustainable yield. Providing additional empirical support, a study of 300 reef sites from the western Indian Ocean found that ecosystem impacts of fishing became very apparent when stocks were fished below 0.25 B0 (23). In this present study, fish biomass estimates were separated into three gross categories; those below 0.25 B0, between 0.25-0.5 B0, and >0.5 B0. This ordinal classification also allowed for some objectivity in evaluating the status of the communitys fishery resources relative to one of their expected goals of maximizing yields, which could not have been achieved by using interval-level biomass data. Estimates of B0 were 1,200 kg/ha based on broad regional surveys (23, 24).

We acknowledge the considerable heterogeneity in coral reefs and human communities and the importance of long-term and before and after evaluations of these social-ecological systems. Our multicultural and interdisciplinary research team has conducted long-term and geographically broad studies of ecological and fisheries dynamics recovery trajectories and associated human use impacts, social and cultural institutions, and socio-economic conditions of tropical coastal communities (e.g. 24, 28-30). In each of these countries, over the course of 10 to 25 years, this team has used long-term studies in combination with the complementary snap-shot approach presented here.

Statistical Analysis

We used Ostrom's diagnostic framework (4, 5) to build a series of Bayesian hierarchical models that quantify the relationship between our 22 measured covariates and three dimensions of comanagement success (Table S3). This initially included 11 household covariates and 11 at the comanagement site level. We examined covariates and responses for co-linearity and this resulted in population density and gear restrictions being removed from future analyses.

The remaining comanagement level covariates were used to model the intercepts for each comanagement site in the household-level part of the model. In other words, the average response at each comanagement site was determined by comanagement-level factors and the household-level covariates were modelled against these averages (Table S4). This structure effectively handled comanagement level differences in each response, allowing us to infer why these differences may occur and permitting cross-comanagement site comparisons of household-level variables.

For each response (i), non-binary covariates (xj) were standardized by subtracting their mean and dividing by two times their standard deviation:

. (1)

This effectively places all covariates on a -0.5 y=to +0.5 scale that is approximately binary (spans 1) and allows direct evaluation of their relative effects on a given response; i.e. binary and continuous variables can be readily compared on the same scale (31). Means and standard deviations were calculated across all observations, as covariates were developed so as to be comparable among different societies.

Our modelling strategy was to include all covariates where possible, in a series of candidate models that included a priori combinations of comanagement level covariates representing distinct SES components from Ostrom's framework (4, 5) (Fig. S1). At minimum we fit one parameter for every 10 replicates to avoid overfitting. This was particularly important among the ecological state of fish biomass models, because this response was available at only 27 locations; household-level covariates were included as comanagement unit averages. Importantly, covariate priors were defined as being Multivariate Normal to structure potential covariance among various covariate factors or between Multinomial components.

Once the candidate model sets had been defined we determined the appropriate statistical distribution (family) for each response based on the form of the data and our study objectives, with livelihoods being split into winners and losers relative to neutral (Multinomial); perceived compliance being evenly-spaced integer scores from 0 to 3 (Binomial), indicating none, 1/3, 2/3, or full compliance; and fish biomass status being scored as below (0), within (1) or above (2) the BMMSY window (Ordinal multinomial). We examined posterior model fits and, where appropriate, ran supplemental models that included a revised set of informative comanagement level covariates so as to include as many informative variables as the data could tolerate. Finally we ran a series of a posteriori models, using subset combinations of variables to check that none of our candidate model parameters were affected by potential overfitting. We relied however on the a priori models in the majority of cases as these models included theoretically important covariates that had no discernable effect.

All models were run using Markov-chain Monte-Carlo, with 80,000 iterations with a 60,000 iteration burn-in, in the PyMC package for Python (32). Models were assessed for convergence using multiple chains that were examined by eye for stationarity. Model fit was assessed using Bayesian p-values, according to the Freeman-Tukey test statistic (33):

(2)

where is a vector of model parameters, j indexes each species, nj is the observed response, and ej is the expected value, according to the model. The Bayesian p-value is obtained directly from the MCMC scheme; in contrast to conventional p-values, either large (>0.975) or small (