3. Results
Before starting the cluster analysis, we standardized all variables to account for differences in scale. We conducted all statistical procedures on these standardized variables. As a first step in the cluster analysis, we analyzed correlations among the clustering variables: strong correlation leads to an overrepresentation of the variables in the final clustering solution [
27]. All bivariate correlations fell well below the 0.9 threshold, indicating no potential collinearity issues. There is no clear consensus on whether to opt for a partitional (k-means) or hierarchical clustering method; thus, we pursued both methods, which also served as a test of cluster stability. Since k-means clustering requires a priori selection of cluster numbers, we used the NbClust package in R to calculate a number of different indexes [
28], as indicated in
Table 2.
As evident in
Table 2, there was no single best solution for cluster size: the final choice depended on subjective preference and usefulness of the cluster interpretations. Two clusters would provide little practical use, merely creating two polar opposites; five clusters would seemingly be too extensive given the limited dataset. Thus we tested both three- and four-cluster solutions; four clusters provided clearer interpretation and balanced cluster groups. For our cluster method, we adopted hierarchical cluster analysis using squared Euclidean distance intervals and within-group linkages; we followed that with a nonhierarchical k-means clustering method. The Ward method provided largely similar cluster results. Alternative methods (such as between-group linkage, nearest neighbor, furthest neighbor, centroid clustering, and median clustering) all led to at least one cluster with limited (often only one) cases.
Table 3 compares cluster assignment according to the two methods.
Table 3 clearly indicates that there was significant overlap between the two cluster methods (chi squared = 64.711,
df = 9,
p < 0.000): 80% of cases assigned to cluster 1 and cluster 3 using the k-means algorithm were also assigned to cluster 1 by hierarchical analysis. For cluster 4, the proportion was slightly higher (83.3%); there was slightly more variation with cluster 2 (66.7% overlap). The cluster solution appeared to be stable and comparable; thus, we used the final cluster centers of the k-means method for thematic analysis of the four generated clusters. The F statistics of the analysis of variance (ANOVA) tests provided an indication as to which areas were particularly relevant in the formation of the clusters: higher scores relating to variables held higher importance in distinguishing between clusters. Only one ANOVA test was nonsignificant: role of university after disaster occurrence. It has to be noted that this nonsignificance is not an indication of lack of general importance of disaster prevention in campus sustainability but merely indicates that the variability in disaster prevention scores did not influence cluster membership assignment.
Table 4 provides an overview of the F statistics and average scores per cluster.
Figure 2 provides a graphic profile of each cluster. Cluster 1 (n = 8) combined institutions with limited strategy and implementation of sustainable practices: it scored low for all dimensions. We described this cluster as “lagging behind.” By contrast, cluster 3 (n = 16), which we termed “top of the class”, was characterized by a larger-than-average level of sustainable practices across fields and areas. The other two clusters showed largely negative values, though there were some exceptions. Cluster 2 (n = 9), which we described as “asset driven”, had high scores in the areas of asset management, facility management, land, landscape, and facilities. Cluster 4 (n = 9), which we termed “networkers”, was distinguished by a primary focus on soft approaches to sustainability: outstanding areas here were organization to consider sustainability, research; collaboration between industry, academia, and government; community service, and dissemination of information.
Finally, using discriminant analysis, we tested the quality of the cluster solution; that analysis also provided an answer to the question about which dimensions contributed most to the different clusters [
34]. With discriminant analysis, we constructed a model that provided the best assignment of cluster membership based on a linear combination of quantitative predictor variables (i.e., assessment areas) that best predicted group differences. There will always be one less discriminant function than there are groups (in our case, three discriminant functions were generated); thus, we adopted stepwise discriminant analysis. With this stepwise procedure, we found that the canonical discriminant functions (CDF) could be constructed as a linear combination of just five assessment areas:
where x
1 = financial resource management, x
2 = waste, x
3 = transportation, x
4 = collaboration among industry, academia, and government, and x
5 = disaster prevention.
The first discriminant function (Equation (1)) had an eigenvalue of 3.160 and explained 64.9% of the variance. The second function (Equation (2)) had an eigenvalue of 1.429 with an explained variance of 29.3%. The third function (Equation (3)) had an eigenvalue of 0.281 and explained variance of 5.8%. Wilks’ lambda suggested that all three linear combinations of predictor variables were significantly different among the clusters. However, it was clear from the canonical correlation that Equations (1) and (2) were highly correlated with the clusters (
Table 5).
Figure 3 indicates how well the discriminant functions were able to separate the different clusters. The discriminant functions were linear combinations of the area variables; thus,
Figure 3 provides an indirect answer to the question about the effectiveness of the cluster solution.
Figure 3 uses only two of the three discriminant functions; however, those two functions accounted for most of the explained variance (94.2%). From the plot in
Figure 3, it is apparent that the different observations were quite well separated, with limited overlap between the cluster boundaries. Cross-validation of correctly classified cases based on the derived discriminant function produced promising results. The total proportion of correctly classified cases was 76.2%; in particular, cluster 3 (87.5%) and cluster 1 (75%) were well identified. With the other two clusters, which had less strong profiles, identification was more difficult: 66.7% of cases were correctly assigned in both cases.
4. Discussion
As discussed in
Section 1.2, Japan’s higher education institutions lag behind those in North America and Europe with respect to campus sustainability. It is necessary, therefore, to identify the key factors that support sustainable campus actions in Japan. ASSC has collected information about specific cases of campus sustainability in Japan. From that information, it is possible to identify the respondent institutions that may be regarded as good examples of campus sustainability, allowing an analysis of the crucial factors in the strategy for sustainability.
In this study, we posed the following research question: What are the strategic options of institutions that conform with ASSC? We aimed to identify which fields would be more appropriate for Japanese higher education institutions to begin their campus sustainability actions. Almost no empirical studies have analyzed an extensive number of Japanese cases in this regard. Thus, the results of this study provide a new understanding of how to introduce more institutions to campus sustainability and promote robust growth of such sustainability in Japan.
We used cluster analysis to identify groups of institutions with different strategic options according to their main areas of focus. From the results of our analysis, we were able to identify four groups: lagging behind (n = 8); asset driven (n = 9); top of the class (n = 16); and networkers (n = 9).
Table 4 and
Figure 2 show that the lagging-behind cluster scored low in all areas of ASSC. The top-of-the-class cluster was the opposite: the scores were higher than average in all areas. The findings with the other two clusters (asset driven and networkers) were interesting. Those clusters had largely negative values in most areas; however, each of them had significant positive scores in a particular set of areas (
Table 6). All the prominent areas with the asset-driven and networkers clusters were well represented.
Our analysis identified 16 institutions in cluster 3: they adopted a holistic strategic option to pursue all dimensions of campus sustainability and therefore achieved top-of-the-class status. By contrast, the 18 institutions in clusters 2 and 4 were clearly oriented to certain areas of campus sustainability, i.e., asset management or collaboration (networking); they attained higher scores in those dimensions.
As discussed in other studies [
25,
35,
36], the general barriers for higher education institutions to integrate sustainability are lack of resources and time. It may be assumed that a large institution could overcome such barriers more easily than a smaller one. We surveyed online the scale of each respondent institution by collecting data about the number of enrolled students.
Figure 4 summarizes the distribution of the number of institutions in each cluster according to scale. The three scales are as follows: a small institution had up to 12,000 students; a medium-sized institution had 12,000–24,000 students; a large institution had over 24,000 students.
From this survey about the size of institutions, it was clear that small institutions were clearly dominant in the lagging-behind cluster. However, contrary to expectations, small institutions were also well represented in the top-of-the-class cluster: nine small, five medium-sized, and two large institutions. Proportionally, small institutions were less likely to be in the top-of-the-class cluster (nine of 30) than medium-sized (five of nine) or large institutions (two of three); however, it is remarkable that institutions with only 3000–9000 students had sufficient opportunity to perform well in all dimensions of campus sustainability. As indicated in
Figure 4, institutions that adopted strategic options for asset management or collaboration tended to be small and medium-sized ones.
With respect to the research question, this study provides two key findings:
Both large and small institutions (by adopting holistic strategies) are capable of attaining the top-of-the-class cluster in all dimensions of campus sustainability. Institution size should therefore not be regarded as a limiting factor in pursuing a holistic strategy for sustainability.
By focusing on physical campus management or on collaborations between the institution and external bodies, small institutions are capable of promoting their campus sustainability and potentially develop toward a more holistic approach at a later stage.
The present investigation is one of the first empirical studies on campus sustainability in Japan to analyze an extensive number of institutions. Almost no other studies have adopted a similar quantitative approach concerning strategic options in this regard. Thus, our study can offer new general guidance for Japanese institutions wishing to undertake the initiative in campus sustainability.
The performance of ASSC respondent institutions are analyzed in this study based on their score results without investigating social or institutional contexts of universities. This limits the guidance provided here to recommend two fields, physical campus management or collaborations, to be capable of promoting the holistic campus sustainability. To provide more specific guidance, future research should center on providing a more detailed analysis of the following: the content of strategic options; the methods of implementation; and achievements or failures with implementation among each ASSC respondent institution. In some cases, this type of information about institutions has been collected by ASSC; however, those details were not examined in the present study. A further topic of contemporary societal interest lies in the role of universities in disaster prevention and management. While elements of institutional disaster prevention are currently measured in the ASSC, this study adopted a higher level clustering approach whereby item-specific in-depth discussions were not part of the current scope. It would be of interest to track the evolution of institutional investment in disaster prevention mechanisms against current events and focus on their role within campus sustainability and resilience. Our future work on campus sustainability will involve, first, investigating such details provided by ASSC. Second, it will be necessary to complement any information not provided by ASSC with on-site investigations, including interviews with staff and students of institutions representing each cluster identified in this study. Only after involving these extensive investigations, the best specific strategies for Japanese higher education institutions will be observed.