The combination of human exposure, extreme weather events and lack of adaptation strategies to cope with flood-related impacts can potentially increase losses not only on infrastructure but also on human lives. These impacts are usually difficult to quantify due to the lack of data, and for this reason most of the studies developed at the national scale only include the main characteristics that define the societal or individual predisposition to be affected, resist, adapt or recover, when exposed to a flood.
The main objective of this work was to develop a flood social susceptibility index for the continental Portuguese territory based on the most representative variables able to characterize different influencing factors. This index is a component of the national vulnerability index developed in the scope of Flood Maps in Climate Change Scenarios (CIRAC) project, supported by the Portuguese Association of Insurers (APS).
The main results showed that the proposed index correctly identified populations less prepared to avoid flood effects or able to cope with them, mostly concentrated in rural inland areas with lower income and education levels when compared with the coastal region between Viana do Castelo and Setúbal.
The number of natural disasters as well as the number of people affected by them has been increasing in the last decades, showing that societies are currently more vulnerable and exposed to these phenomena (Ge et al., 2013). Extreme climate events are responsible for 80 % of the damage caused by those natural disasters worldwide, with floods affecting more than a billion people in the last decade and causing thousands of deaths every year (Vörösmarty et al., 2013). In Europe, floods, together with windstorms, are the most frequent natural disaster and their damages correspond to a third of total economic losses related to these types of phenomena (EEA et al., 2008; IPCC, 2012).
In the last decades the frequency and intensity of natural extreme events has been increasing (Ge et al., 2013) as a result of climate-change-induced changes in climatic patterns, which, most likely, will be aggravated in the next years (e.g. Hov et al., 2013; IPCC, 2012).
For this reason, vulnerability assessment techniques are becoming a fundamental tool in flood risk management, helping to define more effective risk reduction strategies and promoting societal disaster resilience (Birkmann, 2006). The concept of vulnerability was introduced in the 1970s in the context of social sciences and was originally oriented to the risk perception related to catastrophes (Birkmann, 2006). Currently, there are several definitions derived from the different application scopes of application of the scientific communities behind them (Veen et al., 2009, Thywissen, 2006).
In general, vulnerability can be defined as the loss potential of assets or individuals when exposed to a natural disaster of a certain magnitude (Ionescu et al., 2009; Cutter et al., 2000; Schanze et al., 2006). This definition covers several vulnerability dimensions, namely, physical, social, economic, politic, cultural and environmental that, when aggregated with a physical component (Thywissen, 2006), form a composed vulnerability index (see e.g. Balica et al., 2012; Sebald, 2010). This scope has been expanding to include nowadays concepts such as coping capacity and resilience (Armaş and Gavriş, 2013). The work presented here refers solely to the social component of this composed index.
Nowadays, there are still many difficulties to determine the flood loss potential due to the lack of data to estimate the affected area and their associated costs, mainly at the national level. For that reason, most of the studies developed at this scale only include the main characteristics that define the societal or individual predisposition to be affected, resist, adapt or recover, when exposed to a flood (Ge et al., 2013; Armaş and Gavriş, 2013). In the opinion of the authors of this paper, this characterization, also adopted here, is better suited to define flood social susceptibility (FSS) and therefore the developed index was designated as a social susceptibility index (SSI). Nevertheless the adopted methodology derives from the existing bibliography on flood vulnerability indexes.
There are usually two different methodologies to evaluate flood social vulnerability: (a) the SoVI (social vulnerability index) model and (b) the SeVI (social vulnerability assessment using spatial multi-criteria analysis) model. The first was developed by Cutter et al. (2003) and uses principal component analysis (PCA) to select the most representative indicators to compose the final index, without providing different variable weights. Since its formulation, this method has been widely used in the United States and more recently in Europe, becoming the standard vulnerability assessment method (Armaş and Gavriş, 2013; Ge et al., 2013). The second is based on a multicriteria analysis developed by Saaty (1980) called analytical hierarchical process (AHP). This method combines expert evaluation and statistical methods to determine the relative weight for each variable.
The main objective of this work is to develop a SSI for the Portuguese territory based on the approach initially proposed by Cutter et al. (2003) and further developed by Fekete (2010). Although there are some studies in other European countries to develop national flood vulnerability indexes, in Portugal there is only one published social vulnerability index for some municipalities, implemented by de Oliveira Mendes (2009), that includes both natural and technological risks and does not differentiate floods.
The results presented here are part of a composed flood vulnerability index
for continental Portugal developed in the scope of the CIRAC project (Flood
Risk Mapping in Climate Change Scenarios
Continental Portugal, situated in the southwest of Europe, is part of the
Iberian Peninsula and occupies an area of 89 015 km
Characterization of the study area –
Variables used in this study (with the exception of the percentage of urban area all data were obtained from Statistics Portugal).
According to the 2011 census data (INE, 2011), its number of inhabitants increased approximately 2 %, between 2001 and 2011, from 9 869 343 to 10 047 083, which represented a decrease in the growth rate, when compared to the 5 % registered in the previous decade. From the 278 municipalities, 171 in 2001 and 198 in 2011 have registered a decrease in population, contributing to an imbalance in population spatial distribution (INE, 2001), with an overall movement from rural to urban municipalities. In the last decades, the migratory movements from inland to coastal areas within the Portuguese territory, together with emigration, mostly from rural areas during the 1970s, and, more recently, the emigration phenomena to urban areas, first from the Portuguese former colonies (starting from 1976 onwards), and, in the last decade, from EU eastern countries, Brazil and Asia contribute to this tendency.
In parallel, other demographic phenomena have intensified in Portugal. On the one hand, according to the 2011 census, the double aging of the population process, characterized by a decrease in youth population and an increase in older aging groups, has continued to strengthen in the last 40 years. The total dependency index, defined by ratio between the sum of the population in the 0–14 and over 65 age groups and the active population, defined by the 15–64 age group, has increased 4 % in the last decade, supported solely by the 21 % growth in the older population.
On the other hand, in the last 10 years, two factors had a positive evolution: education and income. Regarding the first, the percentage of people with higher education almost doubled, going from approximately 6 to 12 % (INE 2011), while the percentage of people with no education or only basic education cycles completed (first to sixth grade) decreased from 67 to 57 %. Nevertheless, there is still a significant regional imbalance in the evolution of the Portuguese population educational level, with more highly educated people usually more concentrated in the coastal urban municipalities. As for average annual income, statistics show an increase from EUR 7294 in 2000 to EUR 10 838 in 2011. The spatial distribution of average income also highlights the same coastal/inland differences shown for other indicators.
Unemployment rate is another important socioeconomic factor to characterize flood social vulnerability in continental Portugal. In the last 10 years, this rate has risen significantly from 6.8 to 13.2 %, mostly after the 2008 crisis, after 20 years of low and stable values
In summary, this characterization shows a slow growing and aging country with increasingly lower birth rates, higher education and higher income. Also highlighted by these indicators is the existence of significant regional inequalities between the densely populated, more educated and richer coastal urban areas and the depopulating, less educated, poorer inland rural regions. This snapshot of the continental Portuguese territory will surely be reflected in the social vulnerability index described in the next sections.
Table 1 presents the 39 variables used initially in this study, providing information on its origin, production year, the abbreviation used in this study to label them, as well as information on the indicator group they represent and a first evaluation of its role in flood social susceptibility characterization. This evaluation is represented by one or two minus signs in the case of variables that increase social susceptibility; one or two plus signs if a variable decreases it; and one minus and one plus signs, where variables can play both a positive and negative role in flood social susceptibility. The evaluation of each indicator was made by the authors, following a similar analysis made in the work of Feteke (2010). Nevertheless, as in any variable selection process, there is some degree of subjectivity that should be taken in consideration when evaluating the results of this flood social susceptibility index. Regarding the label, it should be noted that the abbreviations of the final normalized variables used in the composition of the index are equal to the ones presented in the table but with the prefix “NORM”.
The selection of indicators took into account their ability to characterize the relevant socioeconomic (e.g. age, income, dependence) and built environment characteristics (building age and typology) for flood social susceptibility assessment in the different parishes of the continental Portuguese territory.
Whenever possible, data sets of similar origin were used to assure input data
homogeneity in the development of the final index. For that reason most of
the selected data refer to the 2001 census. The 2011 census was not
included in this study because only provisional data were available at the
time. In the authors' opinion, although this is a limitation of the study,
it does not compromise the results presented here. In the last 10 years only
the magnitude, not the spatial distribution, of each parameter within the
Portuguese territory has changed significantly, rendering the comparison
between the different parishes still valid. Whenever the required indicators
were not available through 2001 census data, alternative data sets were used,
available in the statistical yearbooks published by Statistics Portugal
(INE, 2010a, b, c, d, e) or by other
governmental sources (IGP, 2010). All the values were originally
provided at parish level, except in the cases indicated in the table
footnotes, where calculations had to be performed to adjust to this scale.
In the specific cases of the dependency ratios the values were calculated
based on the 2001 census and refer to the following:
youth dependency ratio (IND_DJ) – defined by ratio
between the sum of the population in the 0–14 age groups and the active
population, defined by the 15–64 age group; aged dependency ratio (IND_DI) – defined by ratio
between the sum of the population in the over 65 age groups and the active population; total dependency ratio (IND_DT) – the ratio between
the sum of the population in the 0–14 and over 65 age groups and the active population.
The methodology adopted to develop the Portuguese flood social vulnerability
index was based on the work of Fekete (2010), and it is comprised of
four main stages:
pre-selecting census data variables that could better
describe social vulnerability to floods in continental Portugal (Table 1)
and characterizing their role and influence; using principal component
analysis to define the variables or group of variables that better represent
the different components of flood social susceptibility; aggregating
those variables into indicators, according to the components defined in the
previous step (This aggregation takes into account the role and influence in
flood social susceptibility of the variables (subtracting the sum of the
negative ones from the sum of the positive variables).); composing the
final index by summing the different components. This methodology follows
the SoVI model, an approach perceived as more appropriate for this study,
since it provides a less subjective selection procedure of the most
representative variables in large data sets.
The variable pre-selection step consisted of an analysis made by the authors, comparing the statistical data sets available for the Portuguese territory with the most relevant factors, identified in previous studies (e.g. Vörösmarty et al., 2013; Fekete, 2010; Azar and Rain, 2007; Cutter et al., 2003), influencing flood social susceptibility: age, income, education, urban/rural background and building function/typology.
After arriving at the final set of variables, shown in Table 1, a PCA was
performed, using SPSS 20, to reduce data set dimensionality to the variables
that summarize the main characteristics of flood social susceptibility
(Field, 2007). In parallel, analysing the variables with higher loadings
within the main final components variables can help derive a set of
indicators that define a social susceptibility profile (Fekete, 2010).
Before performing the PCA, a standardization procedure was implemented to
render the variable values between different parishes comparable. The
standardization reference values differed, according to the different
variables:
building construction and typology variables were normalized
by the total number of buildings; family-income-related data sets by the
total number of families; employed and unemployed population variables by
the total number of economically active people; the not economically
active population by the 2001 total population; the foreign population
variables and the number of people receiving guaranteed minimum income were
divided by the 2010 total population; the percentage of social housing
buildings by the 2010 total number of buildings; monthly net average wage
and average annual pensions were not normalized because they already
averaged values; all gender, age and education variables were normalized
by the total number of residents; and the total, aged and youth
dependency ratios, percentage of urban area and population density are
already normalized values. All the reference values are given at the parish
scale for the same year of the data set being normalized.
After standardization, a variable correlation matrix was computed to
identify cases of extreme multicollinearity, defined as the variables pairs
with an absolute value of the Pearson's correlation coefficient
The PCA was applied with the remaining variables using a full model approach
(all variables included) in a Varimax rotation with Kaiser normalization to
maximize the sum of the variances of the squared loadings of each variable
across the different components, providing a higher loading in a specific
component and lower on the remaining. This method provides a clearer
interpretation of the correspondence between variables and components. The
selection of the final set of variables was established on three criteria
based on PCA outputs:
The overall Kaiser–Meyer–Olkin measure of sampling adequacy (KMO statistic)
(Kaiser, 1974) should be higher than 0.5 (Hutcheson and Sofroniou, 1999). This
statistic provides a general measure of the adequacy of the collected data to
perform a factor analysis, based on their correlation matrices. A value higher
than 0.5 is considered to be the minimum value to consider that the included variables
share a significant common variance and therefore can be further reduced through factor
analysis. If the KMO value is lower, individual variables should be dropped,
preferentially the ones with lower communality values, a measure of how well
each variable is represented in the different components. The diagonal values of the anti-image correlation matrix should also be
greater than 0.5. The anti-image correlation matrix contains the negative of the
partial correlation coefficients between each pair of variables. The diagonal of
this matrix provides the individual KMO statistics; when one of its values is below
the 0.5 threshold, one of the two variables involved should be excluded since this
means that they are not well factored into the principal components (Feteke,
2010). The off-diagonal values of the anti-image correlation matrix, representing
the negative of the partial correlations between variables, should be as small as
possible in a good factor model (Field, 2007). A threshold value of 0.6 was established
for this study (Feteke, 2010). If lower values are found one of the involved variables should be excluded.
These three criteria were applied in the order they are presented in this
paper and whenever one variable was excluded, the PCA was reprocessed, since
removing one variable changes the final model and it is necessary to
recalculate all statistics.
After arriving at a final model, the final set of principal components was chosen based on an evaluation of the eigenvalues, a measure of the standardized variance associated with a particular factor. Only the components with an eigenvalue higher than 1 were included as flood social susceptibility indicators. Each variable was attributed to one of those specific components, based on their highest loading value. A lower threshold loading value of 0.5 was defined to consider that a certain variable is strongly factored into a component. The final flood social susceptibility indicators were identified by interpreting the final variables groups of each component and their respective signs.
From the variables contained in each component/indicator, only two variables
with a positive influence on flood social susceptibility and two with a
negative influence were chosen to be included in the index, based on their
highest loadings. To arrive at the final values per parish of each of the
identified indicators, the values of the corresponding variables were
aggregated by calculating the difference between the averaged sums of the
variables with positive and negative influence, as can be seen in Eq. (1)
(adapted from Feteke, 2010):
The final step was to aggregate the different indicators into the final
flood susceptibility per parish index by summing the values of all
indicators. Since all indicator values could theoretically vary from
Variable pairs within the correlation matrix with extreme
multicollinearity (
This results section is divided into two parts. The first focuses on the description of the main PCA results that established the set of indicators and variables introduced in the final index. The second discusses the index's capability to characterize flood social susceptibility index across the Portuguese territory and the main reasons behind its spatial distribution.
As described in the Methods section, the first variable selection step was
to compute a correlation matrix based on the normalized variable values to
identify cases of extreme multicollinearity ( Some variables often refer to very similar age groups:
the aged dependency index (IND_DI), the retired persons
and pensioners (NORM_IR_PR) and the
traditional families with people 65 and older (NORM_FCPMA65); One variable is included in a broader one and can be the main
responsible factor for its variance:
the youth dependency index (IND_DJ) and the resident
population between 5 and 9 years old (NORM_R5_9); the traditional families with people with younger than 15
(NORM_FCPME15) and the resident population between 0 and 4
years old (NORM_R0_4) and 5 and 9 years old
(NORM_R5_9); the total dependency ratio (IND_DT) and the resident
population over 65 years old (NORM_R65) The two variables are inversely correlated, as is the case of the following:
the resident population over 65 and residents between 20 and 65
years old, since areas with a higher percentage of active population,
usually have a smaller percentage of residents in the older age groups
(typically the parishes located around cities) and vice versa (like the
rural areas)
Since for all these cases maintaining the two variables would not add any
extra information to the final model, one of the variables was excluded
(variables marked in grey in Table 2). Preference was given, on the one hand, to
variables with a broader scope and, on the other hand, a focus on flood-susceptible age groups (such as children and the elderly). An example is
the selection of the dependency ratios and the traditional families'
indicators over the different age groups of the resident population. The
only exception was the exclusion of the age dependency ratio
(IND_DI), because it was already highly correlated with other
broad variables such as the total dependency ratio (IND_DT)
and the traditional families with people 65 and older
(NORM_FCPMA65). By adopting this strategy it was possible to
exclude a wider number of variables and maintain only the more transversal
ones with useful information in flood social susceptibility. Nevertheless,
it should be noted that this type of analysis is subjective and therefore
open to different interpretations.
Apart from the age-related variables, only three other collinear pairs were
found, all inversely correlated, meaning that they are complementary
variables:
exclusively residential buildings (NORM_ER) and mainly
residential buildings (NORM_PR); traditional families without unemployed members (NORM_FCP0) and
traditional families with one unemployed members (NORM_FCP1); not economically active population (NORM_IR_SAC) and employed population (NORM_IR_EP).
For each of these pairs the maintained variable was either the one with a
higher representativity in the Portuguese territory (a and c) or a higher
information content regarding flood social susceptibility (b).
This step excluded 11 variables, which meant only 28 were introduced into the PCA.
The first full model approach PCA provided an overall KMO statistic of approximately 0.7, well above the 0.5 minimum threshold referred to in the Methods section. This means that the variables have some common variance, and therefore the data set can be reduced using a factor analysis method like the PCA. This value progressively increased to a final value of 0.86 as the variables with individual KMO statistics lower than 0.5 were removed in a recursive way, following the order given in Table 3. Three of removed variables refer to building typology (NORM_EORE, NORM_EPAT and NORM_EARG): this is not surprising since most of the variables in the data set refer to socioeconomic characteristics of either individuals or families which might not correlate well with building-related variables. The remaining variables refer to income/unemployment (NORM_IRD1E, GMMTCO and NORM_IRDNE), one to education (NORM_IRQA_110) and another to building function (NORM_IRQA_110). Although any of these variables could help characterize flood social susceptibility, the decision to remove them took into consideration that other variables could provide similar information, like, for instance, in the case of building typology, the “buildings with concrete structure” (NORM_EBAR) variable.
Finally, as shown in Table 3, the off-diagonal values exclusion criteria also reduced the number of variables included in the final model. As in previous steps, the selection of the excluded variables within each pair took in consideration their relative territorial representativeness and their importance to characterize flood social susceptibility. For instance, the decision to keep the variable “residents with secondary education” (NORM_IRQA_200) and exclude the variables “residents with third cycle of basic education” (NORM_IRQA_130) and “residents with higher education” (NORM_IRQA_400) was based on two reasons: (a) it is a broader variable than NORM_IRQA_130 since it represents all stages of secondary education, and (b) in the opinion of the authors it represents a more significant cut-off education group, regarding social susceptibility to floods than NORM_IRQA_400.
Excluded variables due to low individual KMO values (< 0.5) taken from the diagonal of the anti-image correlation matrix.
Variable pairs with off-diagonal anti-image correlation matrix values > 0.6. In grey are the excluded variables based on this criterion.
After arriving at a set of the most representative variables to include in
the final model, the PCA was recalculated. From all the calculated
components, three were selected to define the main flood social
susceptibility indicators that will compose the SSI (Table 5). These three
components were the only ones with eigenvalues higher than 1, explaining
approximately 63 % of the total data set variability. Table 5 shows the
correspondence between original variables and components based on their
higher loadings. The definition of the three flood social susceptibility
indicators represented by these components resulted from an interpretation
of their main variables:
Regional conditions included most of the education variables
(NORM_IRQA_001, NORM_IRQA_120, NORM_IRQA_200,
NORM_IRQA_300) as well as an income variable
related to average annual value of pensions (VMAP), a population density
variable (DENS_POP) able to differentiate urban and rural
areas and a building typology variable that identifies areas with higher or
lower presence of concrete-based buildings. As referred to above in the
description of the study area, all these variables can help characterize the
significant regional inequalities between less susceptible coastal urban
areas and the more vulnerable inland regions. Furthermore, those variables
can also help distinguish, within the inland areas, some important urban
areas from the remaining rural territory. The assumption of a higher
vulnerability in inland regions is mainly associated with lower education and
income levels and higher distance to institutions that provide assistance
during and after flood events. Age includes all variables related to more susceptible age groups
(the children – NORM_FCPME15 – and the elderly –
NORM_FCPMA65) as well as the more resilient (active
population – NORM_IR_EP). Social exclusion is defined by variables characterizing the lower income
(NORM_RSI_Total, NORM_Edif_habit_Social) or possibly less integrated
immigrant communities (NORM_Imigrantes_Varios).
Finally, for each indicator, up to two variables with a positive influence
on flood social susceptibility and two with a negative influence were
selected to determine its final value. The selection was based on the
highest loadings present in each indicator and in the interpretation of the
role each variable played regarding flood social susceptibility (negative or
positive influence). Table 6 shows the following: (a) the first indicator uses two
different positive variables (higher value, lower susceptibility) to
characterize education and income (residents with secondary education
(NORM_IRQA_200) and average annual value of
pensions (VMAP)) and only one negative variable (higher value, higher
susceptibility) to characterize the presence of populations with lower
education (residents with no qualification, NORM_IRQA_001); (b) in the age indicator the selected positive
variable is related to the presence of people in active age, usually less
susceptible to floods and the two negative variables are related to the
existence of higher susceptible age groups (children under 15 and elderly
over 65 years old); (c) the social exclusion indicator is composed of two
negative indicators related to the presence of immigrant lower-income
communities, which is understandable since it is an indicator aimed at
characterizing highly vulnerable populations.
Final components and their corresponding variable loadings. The name given to each component was based on the interpretation of the flood social susceptibility characterization given by the variable group that composes it.
Final set of variables included in each indicator that composed the final flood SSI.
The maps with the results, per parish, of each indicator and the aggregated
index are shown in Figs. 2 and 3. All indicator maps use a common
scale of equal 0.1 intervals between
Maps of the three flood social susceptibility indicators
for the continental Portuguese territory:
Flood social susceptibility index (SSI) for the continental Portuguese territory.
The regional conditions indicator, related to education and income variables, expresses the significant regional inequalities described in the Study area section. The lower susceptibility values are concentrated in the Setúbal–Viana do Castelo coastal axis and along Algarve coastline (see Fig. 2). Those correspond to the more developed Portuguese regions, where the population has higher education and income levels. The major inland urban centres, where most of the youth population of the surrounding rural areas migrated to in search of better work conditions, also present low susceptibility values. The higher susceptibility values are associated with rural inland areas with a more fragile economy and an aging population.
This territorial dichotomy is also present in the age indicator, although the higher values are mostly focused in the centre and North inland regions, due to a lower presence of individuals in active age and a higher incidence of elderly rural populations. In the northern part of Alentejo the aging population factor is partially absorbed by the higher presence of people in active age.
Finally the social exclusion indicator shows a more limited territorial influence, concentrated in the southern regions with a high incidence of low income and immigrant communities.
The SSI index compiles the partial information given by its indicators, highlighting, as expected, the coastal/inland differences and showing a higher ability to cope with floods in the more populated and developed coastal urban centres along the Atlantic coast. Within those areas, the metropolitan regions of Lisbon and Oporto have the lowest SSI values, mainly due to their higher per capita incomes and education and lower unemployment. Higher social susceptibility values are located in the poorer inland regions, with a focus on the north and centre eastern quadrant and the northern and southern part of Alentejo.
The main objective of this work was to develop a flood social susceptibility index for the continental Portuguese territory based on the most representative variables able to characterize different influencing factors such as age, income, education and building typology. This goal was achieved effectively using a PCA-based methodology to reduce the original set of 42 variables to eight, representing three indicators used in the final index: regional conditions, which aggregated income and education variables; age with parameters related to susceptible age groups; and social exclusion characterizing particularly susceptible very low-income and immigrant communities. The PCA-based technique avoided successfully most of the subjective selection processes based on expert analysis methodologies that can add bias to the final index, based on personal assumptions. Nevertheless some degree of subjectivity is unavoidable in different steps of this methodology, namely in the definition of the role given to each variable to characterize flood social susceptibility. An optimization of this process could only be achieved by the existence of flood effect validation data for the Portuguese territory, since it would corroborate the selection of the final set of variables included in the index and their respective role.
The use of a restrict set of variables contributed to index simplicity and consequently to its transparency, as shown in the straightforward interpretation of the results given in the previous section. In general, the index correctly identified populations more socially susceptible to floods, mostly concentrated in rural inland areas with lower income and education levels, when compared with the coastal region between Viana do Castelo and Setúbal.
Nevertheless, as referred to above, this index would benefit in the future from
a validation procedure similar to the one developed by Feteke (2010). This
study correlated questionnaire answers given by people affected by floods in
Germany with the variables in the main PCA components to choose the
variables to include in the index. The main reason not to pursue this
methodology in the work presented here was the lack of systematized
information on flood events in Portugal. Future integration with the results
of projects like DISASTER (GIS database on hydro-geomorphologic disasters in
Portugal: a tool for environmental management and emergency planning –
We acknowledge (a) Ângela Antunes for her contribution and (b) APS – Portuguese Association of Insurers, who funded the project. Edited by: V. Artale Reviewed by: two anonymous referees