The role of location and social impacts of protected areas on subjective wellbeing

Greece is one of the most important countries in Europe in terms of biodiversity including one of the highest numbers of endemic vascular plants in the region (Dimopoulos et al 2016). According to the latest NATURA 2000 Barometer (2020) 27.3% of terrestrial area and approximately 20% of the national marine area is protected under the Natura 2000 network. There is very limited evidence regarding the effectiveness of Greek PAs and their ecological and socio-economic outcomes. The most recent valuation of the effectiveness of Greek PAs was conducted in 2011 (Vokou et al 2014). The study showed that there is no effective management in several of the PAs. Maragou and Christopoulou (2012) have also suggested that NGOs have played a significant role in managing Greek PAs covering existing gaps in the legislative and monitoring framework. It should be noted that the Greek government has recently passed a legislation allowing multiple uses in certain categories of PAs such as the Habitat/Species Management Areas (Law 4685/2020, FEK 92/A/7-5-2020).

Our study explores for the first time how PAs in Greece have influenced wellbeing levels of people living inside or near PAs. In particular, we researched subjective wellbeing levels of local communities living inside or near four Greek PAs covering in total 409 021 hectares (table 1 and figure S1 supplementary material available online at https://stacks.iop.org/ERL/15/114030/mmedia): Prespes National Park (NP), Messolonghi NP, Samaria NP, Tzoumerka NP. These four areas were selected based on the following criteria: a) they are all designated as National Parks thus having a similar designation status; b) all research areas have communities which live inside or near their borders directly impacted by the PA designation and c) human activities within the four parks are similar allowing comparisons between the PAs (these activities are defined in the designation legislation for each area, see also section 2.2). Having similar human activities allowed the research team to use the exact same questionnaire in all areas, especially regarding perceptions of social impacts of the PAs.

A questionnaire was developed and distributed in communities living inside the four PAs and the surrounding areas (maximum 20 km distance from the PA). Participants were selected randomly based on their location. Initially the total population living inside or near the PA was defined (within the 20 km distance) by identifying all villages in the area. The desired sample size from each village was subsequently determined in order the sample to be geographically representative of the actual population. Researchers visited each village and randomly selected participants to fill in the questionnaire until the defined sample quota was achieved (table 1). Demographics of the sample were checked throughout the selection process in order to ensure that the characteristics of the sample are similar to the ones of the sampling frame. The questionnaire was distributed during summers 2016 (Prespes, Samaria), 2017 (Tzoumerka,) and 2018 (Messolonghi) after it was tested in a small sample of respondents. A total of 818 questionnaires were completed through face to face interviews. Due to missing values in certain indicators the final sample retained for the statistical modeling was 772.

The questionnaire aimed to assess subjective wellbeing along with a number of indicators that may explain wellbeing levels according to existing literature (Pretty 2003, Ostrom 2009, Rollero and De Piccoli 2010, Hommerich and Tiefenbach 2018, Jones et al 2018, Bennett et al 2019) (table 2). The main categories of indicators are explained below.

• a)  
  Subjective wellbeing was measured by asking respondents how satisfied they are with their personal quality of life in a 10-point Likert scale in accordance to the Organisation for Economic Co-operation and Development (OECD) guidelines (OECD 2013) and other large-scale surveys (ESS, 2018, Eurostat 2019) (indicator 1, table 2)
• b)  
  Explanatory parameters included:

  • Social impacts: In this paper we adopt the definition of social impacts used by Jones et al (2020) as the intended and unintended social consequences, both positive and negative, which occur because of the designation of a PA and any social change processes invoked by a PA. This definition focuses on the subjective assessment of social impacts and allows each impact to be potentially interpreted both as a cost and benefit depending on the perception of each individual. Three categories of social impacts were assessed in the study each one measured through three questions (indicators 2–4, table 2): impact on social aspects (recreation, social networking, cultural activities), impact on economic aspects (trading, employment and construction industry) and impact on activities directly linked with the use of natural resources (livestock, agriculture/farming and fishing). The final three indicators used in the analysis were calculated via a Principal Component Analysis (PCA) (using SPSS 26.0, see Table S1) reducing the initial nine questions to three factors. This was in order to facilitate the modelling process.
  • Frequency of doing specific activities in the national park (e.g. recreation, hunting, fishing, indicator 5, table 2).
  • Social networks (frequency of meeting other people, indicator 6)
  • Perceptions regarding the environmental quality in the area (indicator 7)
  • Socio-economic attributes (gender, age, years of residence in the area, occupation and educational level, indicators 8–11 - information on income was also collected but 50% of respondents refused to provide their details)
  • Social trust referring to the level of trust to other people (indicator 12)
  • Place attachment referring to how important is the area for the respondents (indicator 13)
  • Information regarding the location of the respondents recorded through geographical coordinates of the village of residence (indicator 14)

We run Bayesian regression models using the WinBUGS software (Lunn et al 2000) to identify the parameters determining subjective wellbeing. All 13 explanatory parameters (indicators 2–14, table 2) were introduced in the models. Markov chain Monte Carlo (McMC) techniques were used for the Bayesian inference and parameter estimation. In order to explore whether the location of respondents also has an impact on wellbeing, a spatial component was introduced to the regression models. Suitable kernel functions (Duncan et al 2017) were used approximating the distance of the respondent's village from the centroid of each PA.

We adopted a g-prior type of approach for the specification of prior densities in the parameters of the covariates of interest. This was in order to avoid the assumption that fixed-effects covariates are independent. This allowed us to account for potential correlations among the covariates of the regression model when selecting statistically significant covariates. The Zellner's informative g-prior (Zellner 1986) was utilized, being appropriately adjusted for the generalized linear models (i.e. Poisson and negative binomial) (Bové and Held 2011).

A total of nine models were fitted to the data. These models compared three specifications under three variations. Specifications included the Gaussian distribution (suitable for continuous response variables), the Poisson distribution and the negative binomial (NB) distribution (the latter two being more suitable for Likert type data) (Lindén and Mäntyniemi 2011, Cameron and Trivedi 2013, Malesios et al 2018). The three variations included a model ignoring the spatial information (non-spatial), except for the categorical variable of the PA, a model with spatial component specification via an inverse distance power function (spatial 1) and a model with a spatial component specification via a negative exponential decay function (spatial 2). Since the inverse distance power function has a fatter tail when compared to the negative exponential function, these two alternative specifications allowed us to test longer vs. shorter distance effects on wellbeing, depending on which of the two candidate functions had better fit to the data. Model selection was carried through the use of the deviance and the deviance information criterion (DIC) (Spiegelhalter et al 2002). A detailed description of the modeling approach is included in the supplementary materials.

Finally, for the visual representation of potential spatial patterns in subjective wellbeing, an inverse distance weighting (IDW) methodology was used (i.e. a deterministic interpolation method that derives the value of a variable at a new location using values obtained from known locations) with the QGIS software (QGIS, 2015).

The best model was the one which assumed a normal distribution for subjective wellbeing (Gaussian) including a spatial component that accounted for longer distance effects (inverse distance decay function—spatial 1 model) (table 3). The evidence from applying the model comparison methodology in this paper suggests that a modeling strategy based on the use of the Gaussian distribution works well and provides robust results, when compared to the Poisson and NB specifications. This outcome comes in contrast with the widely accepted argument that the Poisson or NB models would probably work better for this type of data (Greene 2008). A possible explanation for this outcome is the presence and/or magnitude of skewness of the dependent variable (Malesios et al 2018).

Table 3. Goodness-of-fit statistics for alternative models (dependent variable: subjective wellbeing).

Model	Mean deviance ($\bar D$)	DIC
Gaussian (non-spatial)	3156.24	3218.13
Gaussian (spatial 1)	3155.23	3147.49
Gaussian (spatial 2)	3157.54	3219.34
Poisson (non-spatial)	3384.98	3426.85
Poisson (spatial 1)	3384.58	3391.15
Poisson (spatial 2)	3384.55	3425.03
NB (non-spatial)	3407.65	3448.32
NB (spatial 1)	3405.56	3437.08
NB (spatial 2)	3385.94	3426.41

Upon selection of the best fitted model, to test the quality of its fit we look at the normal probability plot (QQ-plot) of the residuals (see figure S2 in the supplement). Inspection of Figure S2 indicates that there are only few moderate deviations from normality. We also checked for autocorrelation through visual inspection of autocorrelation plots and found acceptable autocorrelation levels for all parameters.

The parameters estimates of the best selected model (Gaussian spatial 1), after the implementation of the g-prior selection process, are described in table 4 (including also the corresponding 95% credible intervals, see also figure 1). Similar results in terms of covariate selection have been observed in all remaining fitted candidate models. The most important indicators determining wellbeing are perceived impacts on social aspects (indicator 2) and perceived impacts on activities directly linked with the use of natural resources (indicator 4) (table 4). Other indicators that explain wellbeing, but to a lower extent, are the impact on economic aspects (indicator 3), place attachment (indicator 13) and social trust (indicator 12).

Zoom In Zoom Out Reset image size

Regarding the role of the geographical location on wellbeing levels (indicator 14), the statistical significance of the Spatial 1 variation in all three specifications is an indication of positive effect on wellbeing levels of those residing closer to the center of the PA than for those further apart. The comparative better fit of the spatial component models including the inverse distance specification, in comparison to the alternative inverse exponential component, suggests that the spatial distance effect is stronger in long distances, in comparison to shorter distances.

According to the IDW methodology spatial dependence exists between subjective wellbeing and the location of the respondent in relation to the PA. The obtained estimates are visualized via heatmaps for each PA in figure 2. They include the locations of the villages of respondents along with the exact location of the PAs' centroid. Darker colour in the maps represent higher wellbeing levels. The results reveal that in all research areas the location of the individual in relation to the PA plays a significant role in subjective wellbeing. This connection is particularly evident in Prespes and Tzoumerka where wellbeing levels are reduced as participants live further out of the center of the PA.

Zoom In Zoom Out Reset image size

Perceived impacts are the most significant predictor for subjective wellbeing (table 4). People who consider that they have benefited more from the PA (perceived positive impacts) also feel that they have a better quality of life compared to those who regard that the PA has impacted them in a negative way. The strongest connection is noted with the impact on social aspects, such as cultural and recreational activities, and the impact on activities directly linked with the use of natural resources. When exploring differences between the research areas these impacts differ significantly (supplementary material tables S2–S3). In Messolonghi people perceive impacts in a more positive way compared to other PAs. Highest scores are observed for the impact on recreational activities in Tzoumerka and Samaria revealing that the PA has been successful in protecting local cultures. The impact on activities directly linked with the use of natural resources is the one that has the lowest scores with two of the PAs (Tzoumerka and Prespes) being particularly negatively affected.

Regarding factors which are not directly connected with the PA social trust is positively associated with wellbeing levels (beta = 0.24, p-value < 0.05, table 4). People who tend to trust others have higher levels of wellbeing. Place attachment has also a positive effect (beta = 0.22, p < 0.05) meaning that those who are more 'attached' to the area they have higher levels of wellbeing. The socio-economic profile of the respondent does not play a significant role in explaining variations in wellbeing levels.