Measuring the importance of CES
To develop the measure for CES, we applied the Common International Classification of Ecosystem Services (CICES) developed from the work on environmental accounting undertaken by the European Environment Agency (Version 4.3; Haines-Young and Potschin 2012). CICES aims to provide a comprehensive classification by introducing the idea of a five-stage hierarchical structure to describe and measure ecosystem services: sections, divisions, groups, classes and class types. To account for context, place and culture, we needed to operationalize an ES classification for the Baltic Sea context. We used the 11 ES classes for CES from CICES as a starting point (see column two in Table 1). From these classes we developed a more detailed specification of the ES classes for the Baltic Sea ecosystem. These Baltic Sea examples are presented in column three of Table 1.
Table 1 CICES-based description of cultural ecosystem services provided by the Baltic Sea To deduce the relative importance of CES we designed a question in which respondents had to distribute 100 points according to the importance they attach to the various CES provided by the Baltic Sea. Using such a relative measure allows us to deduce the importance of a particular CES relative to other CES. However, to apply such a question, the number of CES categories had to be kept reasonably low, the categories had to be sufficiently distinct from one another and they had to be phrased such that they were easily understandable for lay people. For this reason, the ES used in the survey were reduced to seven categories following Kandziora et al. (2013). Moreover, the wording was shortened and kept as simple and understandable as possible (column three in Table 1).
This operationalization of CES was tested with Latvian and Finnish focus groups of citizens. The groups discussed the reasons for which they value the Baltic Sea based on the presented CES. In addition, the method for distributing points was tested and found to work reasonably well, even though adding up the points was found slightly difficult with pen and paper. In the final internet version of the survey, points were added up automatically so that respondents did not have to carry out any calculations themselves. The seven categories as described in the last column of Table 1, plus an option for “other reasons not mentioned in the list”, were included in the pilot survey and in the final survey. For full phrasing of the survey measure, including the question, see Appendix S1.
To test the assumption that perceived CES from the Baltic Sea region are shaped by the more general human–nature relationship of a respondent (DeGroot and Steg 2008), we measured respondents’ ecological attitudes using the NEP (Dunlap et al. 2000). The NEP measure with a five-point scale (from totally agreed to totally disagree), encompasses statements with the following facets: (1) balance of nature, (2) limits to growth, (3) risk of an eco-crisis, (4) anthropocentrism and (5) humans’ ability to control nature. The final NEP measure is produced from the sum of these statements.
Information regarding recreational use of the Baltic Sea during the previous 3 years made it possible to separate nonusers from users. Mapping the home location of the respondents allowed us to calculate the distance from home to the Baltic Sea, and also to define the urbanization level of the home location. In follow-up questions respondents provided information on gender, age, education, income, occupation and household size.
Survey implementation
To develop the CES measure and to test it in the Baltic Sea context, new survey data were collected in Finland, Germany and Latvia to reveal the diverse benefits to human well-being from the Baltic Sea. The survey was designed with international cooperation in 2015–2016. Pre-testing included expert reviews by researchers in environmental valuation and marine ecology, focus groups (one in each country) and a pilot survey in each country in June–July 2016.
The final survey was implemented between November 2016 and February 2017. It was targeted at residents of each country using stratified random sampling that was representative of the population in each country. Stratifying, for instance, according to age, gender, location and educational level ensured representative samples of the national populations. For Germany, coastal regions were oversampled to increase the share of Baltic Sea visitors in the final sample. The primary data collection method used in Finland and Germany was computer-assisted web interviews (CAWI) with internet panels (Table 2). The implementation method in Latvia combined computer-assisted personal interviews (CAPI) and CAWI to ensure a representative sample of respondents in all age groups (including older age groups for whom internet use is insufficient in Latvia to achieve the representativeness by CAWI) was obtained. The CAPI were conducted at the respondent’s place of residence. Altogether, 4,800 respondents answered the survey, with a little over 2000 respondents in Finland and in Germany, and around 760 in Latvia. The average response time was around 20 min.
Table 2 Survey implementation Table 3 presents descriptive statistics for the final data and the corresponding national statistics. There were slightly more men in the German and Latvian data sets compared to the national population. In all three data sets, average age was somewhat higher than the average of the populations, while the share of respondents with higher education was slightly lower.
Table 3 Descriptive statistics for survey respondents, and corresponding national statistics The survey included 41 questions and 6 sections. After an introduction to the survey and the Baltic Sea, respondents’ recreation visits were mapped and visit information collected. Reponses towards changing environmental conditions were collected. The importance of CES from the Baltic Sea was measured before the set of background questions.
Statistical methods
The dependent variable in the statistical analysis was the whole CES measure consisting of eight elements based on survey responses to the importance of CES, named as: recreation, landscape, inspiration, education, spiritual, historic, habitat and other services. These were denoted R1, R2, R3, R4, R5, R6, R7 and R8, respectively. Any vector x with non-negative elements x1, …, xD representing percentages of the total is subject to the obvious constraint \( x_{1} + \cdots + x_{D} = 100 \). As these individual services were subject to this constraint, and only have importance in relation to each other, we have a compositional dependent variable. Compositional Data Analysis is the standard statistical method used when data contain information about the relative importance of parts of a whole, typically with a fixed sum (Aitchison 1986). In the statistical modelling, the aim was to explain the differences in the compositions, that is, the relative importance of the CES between individuals. Compositional data pose certain restrictions with respect to traditional statistical analyses, and thus special treatment is required (Aitchison 1986). We performed an isometric log-ratio (ilr) transformation on the dependent variable and performed compositional data analysis using the ilr-transformed variable as the dependent variable [see Egozcue et al. (2003) and Ahtiainen et al. (2015) for more details]. We fitted several models with predictors describing the respondents’ sociodemographic characteristics (e.g. gender, age, household size, occupation, income, education), geography (e.g. country, distance to coast, category of place of residence), recreation (user or nonuser) and human–nature relationship (NEP scale) (see Table 4).
Table 4 Explanatory variables used in the models Zero elements within compositions were assumed to be so called ‘rounded zeros’. We applied the multiplicative replacement strategy presented by Martin-Fernandez et al. (2003), generating imputations for zeros from a continuous uniform distribution unif (0.001, 0.0049).
Due to the compositional nature of the dependent variable, a univariate modelling approach was not meaningful. Thus, we applied a multivariate ANCOVA to examine the joint significance of the influence of each predictor on the composition response. As model selection tools, we used fivefold cross-validation R2 values and multivariate ANOVA type III p values based on an F-distribution approximation of Pillai’s trace. Pillai’s trace takes values in the range of [0,1], where a larger value indicates higher significance of a predictor. We first fitted a model with the main effects for all potential predictor variables, including distcoast, caturb, nonuser, male, country, age, sumnep, hhsize_under, income_class, low_edu and fulltime (see Table 4 for definitions). In addition, we added interaction terms between country and nonuser, and between country and sumnep. To select the final model we dropped each predictor for which the p value was greater than 0.05. With the remaining predictors, we found a model that maximizes the cross-validated R2 value.
Beyond multivariate ANCOVA, we used clustering techniques for the CES composition. Hierarchical cluster analysis, applicable with compositional data, was used to illustrate the clustering of CES with each other. The distance matrix was constructed using Euclidean distance. For clustering with respect to variables, the distance matrix was calculated from the variation matrix. The clustering was performed using the Ward method. We also used the same method to cluster the respondents based on their CES compositions to define and describe the respondent groups that had similarities in the perceived CES.
The statistical analyses were performed with R software (R Core Team 2017) and the package compositions version 1.40-1 (Van den Boogaart et al. 2014).