2.1 Regional disaster resilience indicators
As one of the most well-known and widely used frameworks, the disaster resilience of place model had six dimensions; ecological, social, economic, institutional, infrastructure, and community competence (Cutter et al. 2008). However, the ecological was excluded due to “data inconsistency” (Cutter et al. 2010). Further, the institutional was excluded because disaster-related institutional change has been observed to be a very slow process with the changes being consistent for each county in one region. So, in this study, four dimensions were chosen: economic, social, infrastructure, and community competence. In reference to Kusumastuti et al. (2014)’s indicator selection and considering data accessibility at the county level in Sichuan Province, a regional disaster resilience index was built that had five primary indicators and 28 secondary indicators (Table 2), with all resilience effects of these indicators being positive.
Table 2 Indicators after data cleaning to analyze the impact of the Wenchuan Earthquake
Total investment in fixed asset/ CNY, Value-added of private economy/ CNY, Value-added of primary industry/ CNY, Value-added of secondary industry/ CNY, Value-added of tertiary industry/ CNY
Measures the general macro-economic circumstance of each county and the economic structure
Industrial value added/ CNY,
Post and telecommunications income/ CNY,
Retail sales/ CNY,
Tourism revenue/ CNY,
End-of-year loan balance/ CNY,
End-of-year deposit balance/ CNY
Measures the detailed economic performance in each county in several industrial sectors, including industry, post and telecommunications, retail, tourism and finance
Net income of rural residents/ CNY,
Net income of urban residents/ CNY
Measures the urban and rural resident income, distribution, and net income
Road mileage/ km,
Standard road mileage/ km,
Passenger traffic volume 10k ppl/km,
Freight traffic volume 10k t/km
Measures the road infrastructure construction and the corresponding carrying capacity of each county
Education & medical treatment
Primary school num/ PT,
Primary school student num/ PT,
Primary school teacher num/ PT;
Middle school num/ PT,
Middle school student num/ PT,
Middle school teacher num/ PT;
Medical institution num/ PT,
Medical technician num/ PT,
Medical bed num/ PT
Measures the education and medical institutions, relevant staff and facilities provided as well as school students
2.2 Research site and data processing
The case-study example in this study was the 2008 Wenchuan Earthquake, which affected 417 counties (cities, districts) in 10 provinces (cities) of China, and especially Sichuan Province (Dunford and Li 2011), which is an earthquake-prone province with 21 cities and 183 counties. Due to data accessibility, this research mainly focused on 55 counties and 6 of the most seriously affected cities in Sichuan for which there was accessible data from 2005 to 2016. Based on the extent of the devastation, these 55 counties were divided into three groups: extremely-affected counties, heavily-affected counties and not-heavily affected counties: by the State Council. Therefore, in this study there were 9 extremely affected counties, 22 heavily affected counties, and 24 not-heavily affected counties. The research site was as shown in Fig.1.
This study used authoritative panel data extracted from the Statistical Year Book of each county’s statistics bureau. However, as the statistical data varied from county to county and from time to time, due to the different data acquisition policies and the changing statistical policies in different counties, some indicator values were missing. There were also significant value differences; for instance, the GDP in most developed counties, such as Shuangliu near the capital Chengdu, was over 100 times higher than the least developed GDPs. Therefore, in the disaster resilience assessment, the change trends were considered more valuable than the absolute values, for which two data processing methods were implemented.
First, the missing values were imputed from an intermediate bisector line between the global and local lines, after which a variation was added to each imputed value to force the imputed value to follow the shape of the average trajectory (Genolini et al. 2013). However, if there were more than 30% values missing at random of a county for specific indicators, the county was temporarily deleted, which also meant that the samples varied in different sections. Overall, however, even the section with least available samples, i.e., Tourism Revenue for Long-term Disaster-recovering Capability Evaluation, had 16 sample counties, which implied sufficient data in each section from enough counties to allow for a detailed analysis of the relationships between disaster resilience and economic development/devastation extent.
2.3 Resilience Evaluation Models
(a) Correspondence Analysis
To determine the SDRC performance for the Wenchuan Earthquake, correspondence analysis (CA), which has been widely used in the ecological field (Beh and Lombardo 2014; Greenacre 2017), was applied. As an unconstrained ordination analysis method, CA is capable of mapping indicators and samples simultaneously on a 2-dimensional space (biplot), and preserves maximum data features by applying orthogonal component constructions on the distance matrix. Therefore, the resilience indicators and the affected counties could be intuitively obtained by using biplots.
CA can be applied when there are at least 2 rows and 2 columns, no missing data, no negative values and all data is of the same scale (Gauch 1982). Although ordination methods are sensitive to outliers, the ratio transformation in Section 2.2 was able to solve this problem, and the best CA performance was gained by comparing its data feature preservation and ordination results interpretability with three other unconstrained ordination analysis methods: Principle Component Analysis, Principle Coordinate Analysis and Non-metric Multidimensional Scaling.
The biplot analysis logic is the same as when used in ecological evaluations; each row represents the species distribution in a sample plot, with the higher the quantity value of a species, the more preference the species has for the corresponding sample plot. In this research, the higher the ratio, the more the county has positive feedback (a large growth rate) for the corresponding indicator.
The biplot can be explained in several ways according to Gauch (1982):
- The counties and indicators around the origin share the least unique features and represent the most common counties/indicators that have similar properties.
- A relative distance between counties or indicators can be applied to measure the similarities between those counties or indicators; the closer the distance, the more similar they are.
- The correlation between two counties or indicators can be measured by their included angle towards the origin; the smaller the angle, the stronger the correlation.
- The approximate preference between counties and indicators can be obtained via relevant distance; however, the distance value is meaningless.
(b) Gaussian Mixture Model
To further analyze the regional characteristics, cluster analysis was employed to group the county sets so that counties in the same cluster were more similar in some sense to each of those in the other clusters.
Taking the primary industry changes as an example, although many counties were influenced by the Wenchuan Earthquake and have experienced a steady recovery since then, the change trends have not been homogeneous (Fig. 2). Therefore, the Gaussian Mixture Model (GMM), a model-based clustering method and widely utilized heterogeneous group growth analysis method (Reinecke and Seddig 2011), was used to cluster the counties based on the different changing trends.
In this part, GMM is used to cluster the counties with similar change trends under a certain indicator into one group, so as to form several clusters, and then observe the different characteristics of different clusters under this indicator. After the clustering results were obtained, the general trends in each cluster were analyzed, and the counties’ common recovery properties determined in relation to their differing economic development stages and devastation extent.
In this way, the characteristics of different geographical regions can be observed simultaneously, thus forming a horizontal analysis perspective. Then, the change trend of the indicator for several consecutive years is shown in the same plane Cartesian coordinate system, and a longitudinal perspective is obtained. As a result, the comprehensive resilience assessment is realized.
2.4 Research framework
The framework had two parts: preparation and resilience assessment, which had two components: a SDRC analysis using the CA and a LDRC evaluation using the GMM. After the CA, five substantially affected secondary indicators in 2008 were selected to assess their recovery trends as these could possibly take as long a time to recover as the primary indicators. In particular, the three macro-economic secondary indicators selected to assess economic development were the core focus of this study, and none of the transport indicators were selected because the substantial transport reconstruction projects were completed in 2008. The model structure is shown in Fig. 3.
2.5 County clustering and coding
The GMM was used to evaluate the economic development of each county before the Wenchuan Earthquake (from 2005 to 2007). As the GMM took the trends in these 3 years into account, it was much more valid than grouping them according to average value.
The counties were renamed based on their economic development and devastation extent, with the former based on the derived clusters, and the latter based on the official devastation extent certification. For instance, Wenchuan County, which had medium development and was extremely affected by the earthquake, was renamed ME1, and counties with similar properties, such as Dujiangyan City and Shifang County, were sequentially renamed as ME2 and ME3. The code names for the 55 counties are listed in Appendix A. Based on the different economic development levels, the average and underdeveloped counties were called less developed counties; and the extremely-and heavily-affected counties were called severely affected counties.