We conducted a survey of 352 urban green space visitors in three large cities in Paraguay, to capture individuals’ perceptions of physical and social disorder and perceptions of safety in urban green spaces. The sample was a subset of a larger sample of 1119 individuals who participated in a broader survey 41 and from whom we also collected demographic, health, and nature connectedness scores. The urban green space visiting subset of the sample was lower than expected due to higher-than-average rainfall during the data collection period (October 2022) but was large enough to yield sufficient power in our analyses.
We used structural equation modelling to test and refine theoretical frameworks by analysing and evaluating the structural and causal relationships between measured variables, such as physical disorder, green space biophysical characteristics, and perceptions of crime with latent constructs such as social disorder and perceptions of safety. Structural equation modelling allows testing of the direct and indirect effects of pre-assumed and hypothesised causal relationships between measured and constructed variables 53.
Study area
The metropolitan area of Asunción (AMA from the acronym in Spanish) is Paraguay’s largest urban agglomeration comprising 11 cities and housing 59% of the country’s urban population 54. Paraguay is a sub-tropical country with a GDP of USD 41,7 (ranked 100th in the world in GDP terms 55). For this study, we selected three cities in the AMA, namely Asunción, Paraguay’s capital city, Fernando de la Mora, and Luque (henceforth ‘study cities’) containing an estimated combined population of 1,000,000 people. Green spaces in the study's urban areas feature significant tree presence, with tree cover ranging from 60% to 85% of the total land area, falling within the medium to high range 39. In 2019, over 80% of Asunción residents reported a decrease in their perceptions of safety relative the year prior and over a third of the country’s population (34%) feels unsafe at any time of the day 35. Moreover, only 35% of crimes are reported to authorities 35 potentially due to low institutional trust and lack of knowledge of the responsibilities of various officials (e.g., police, judges).
Data collection
We used a structured questionnaire survey of self-reported information to interview adults of at least 18 years old who first opened the door and consented to take the survey in the three study cities (Asunción, Fernando de la Mora, and Luque) in October 2022. The survey was delivered by a market research company (CCR Group Paraguay) (full survey in Supplementary Materials). We used probability sampling and ensured the sample included a range of demographic groups that reflected the study cities’ actual population, stratifying it by age (six age groups), gender (male, female, other), income (quintiles), and the spatial distribution of participants across the city, including formal and informal living settings. Households were selected based on (i) city and (ii) block location. Five households were selected per block with intervals of at least two households between them. Within selected households, individuals were randomly selected to fulfil the four nested stratification criteria noted above. Interviews were conducted until the quota of each stratum was fulfilled. Ten percent of the originally selected households opted not to participate in the survey. In case an individual refused to participate, surveyors sought another household with at least one individual of the same gender, apparent age group, apparent income level, and residential location following the same sampling strategy. Individuals provided written consent to participate in the survey in accordance with the University of Queensland Human Research Ethics Approval, project number 2022/HE000811.
Dependent variables
Perceptions of safety during the day (feeling unsafe): A latent construct comprising three survey measures. Participants were asked to rate on a five-point Likert scale how big of a problem (where 1=not a problem and 5=very big problem) is safety ‘on the way to a public green space’, ‘when alone in the green space during the day’, and ‘in the green space surroundings’. This is an endogenous variable.
Perceptions of safety at night time (feeling unsafe): A latent construct comprising three survey measures. Participants were asked to rate on a five-point Likert scale how big of a problem (where 1=not a problem and 5=very big problem) is safety ‘on the way to a public green space’, ‘when alone in the green space at night’, and ‘in the green space surroundings’. This is an endogenous variable.
Independent variables
Social disorder: Perceptions of social disorder is a latent construct comprising three survey measures. In keeping with the disorder literature 40,56, participants were asked to rate their level of agreement on a five-point Likert scale how big of a problem (where 1=not a problem and 5=very big problem) it is for them when they perceive that other individuals are ‘being harassed, abused, or attacked in the green space’, ‘being a nuisance, causing disturbance, or not behaving the way they are supposed to’, and ‘using alcohol or drugs in public’. This is an exogenous variable.
Physical disorder: Variables were included in the models to represent perceptions of physical disorder. These variables are the responses from participants’ level of agreement on a five-point Likert scale of how big of a problem (where 1=not a problem and 5=very big problem) is ‘poor maintenance of green spaces’, ‘vandalism, broken furniture and sidewalks in green spaces’ (henceforth, ‘vandalism’), and ‘people living in green spaces (henceforth, ‘vagrancy and makeshift housing’). For the model at night, we also included ‘poor illumination’ as an additional indicator of physical disorder.
Perception of crime: Participants were asked to rate on a five-point Likert scale how big of a problem (where 1=not a problem and 5=very big problem) is crime in their neighbourhood green space. We asked participants about their neighbourhood green space since it is assumed that it is the public nature space they are most familiar with and actually or potentially visit the most due to its proximity to the participant’s home.
Crime victimisation: Participants were asked to report if they have been victim of a crime in an urban green space in the last 12 months. This is a dichotomous variable where 1 indicates having been a victim of a crime in a green space and 0 indicates not having been victim of a crime in a green space.
Frequency of green space use: participants were asked to indicate how often they currently visit or pass through an urban green space in a month (never (1), once every 2 weeks (2), once a week (3), 2–3 days a week (4), 4–5 days a week (5) and 6–7 days a week (6)).
Green space biophysical factors:
- Size: We calculated the size of each urban green space using green space spatial data produced by the Technical Planning Office (https://ciudadessustentables.stp.gov.py/) in Paraguay using the landscapemetrics 57 package.
- Tree cover: We calculated the proportion of tree cover within each green space extracting tree cover values from the land cover imagery at 10x10m resolution produced by the European Space Agency in 2020 based on Sentinel-1 and Sentinel-2 using the raster 58 and exactextractr 59 packages in R software 60.
Sociodemographic:
Demographic characteristics are known to affect perceptions of safety in public spaces 61,62 and participants were asked to report their age, gender, and household income.
Statistical analysis
We estimated a series of structural equation models to examine the effects of physical disorder, social disorder, perceptions of crime, crime victimisation, urban green space biophysical characteristics and sociodemographic variables on perceptions of safety in urban green spaces using the lavaan 63 R package. First, we tested a model to estimate the associations between physical disorder, social disorder, perception of crime, green space biophysical factors, and control variables, and perception of safety. Here, we estimate the main effects of physical and social disorder and perception of crime on perception of safety (Models 1,3) (See model equations in Supplementary Material). Second, we fit a fully saturated model to estimate the direct and indirect effects of physical disorder, social disorder, and perceptions of crime on perceptions of safety including green space biophysical factors and control variables (Models 2,4). Models 2 and 4 show us the causal relationships amongst independent variables as well as between independent and dependent variables.
All statistical and spatial analyses were carried out using R Studio. All continuous explanatory variables were standardised. We ran Spearman’s correlations using the car package 64 to check for correlations (all below 0.41; Supplementary Material Table 1). We ensured model parsimony by considering model AIC (Akaike Information Criterion) 65. We evaluated our models using goodness of fit tests (Table1) calculating Comparative Fit Index (CFI), Root Mean Square Error of Approximation (RMSEA), and Standardized Root Mean Square Residual (SRMR) 53 using the lavaan package. We set our criteria for good model fit at CFI > 0.9, SRMR < 0.05, and RMSEA between 0.05 and 0.08. We acknowledge that the CFI in Models 2 and 4 remains slightly below 0.9, yet this value demonstrates considerable improvement when compared to the models without mediation (Models 1 and 3).