1. Introduction

1Risk and uncertainty have emerged as central themes in social sciences. Risk incorporates probability and consequence components and can be defined with emphasis placed on one or the other of these aspects. Risk can also be seen as a subjective construct that has been invented to help understanding and dealing with hazards, dangers and uncertainties. Risk perception focuses on the various ways in which hazards and uncertainties are transformed by individuals and groups into subjective judgments that subsequently guide their choices and behaviour patterns. Consequently, perceived risk raises subjectivity to a higher level.

2In the 1960s, geographers carried out risk perception surveys to understand the public’s attitude and behaviour to flood scenarios (Burton and Kates 1964, White 1964). Psychologists and cognitive scientists soon after performed risk perception surveys in order to better understand and predict the public’s attitudes to technology and technological hazards (Swoby 1965, Starr 1969). At the beginning of the 1970s, they revealed the large number of biases in the public’s evaluation of risks and probabilities (Tversky and Kahneman 1974). More recently, sociologists focused on the socio-demographic drivers of risk perception such as age, gender and class (Satterfield et al. 2000).

3Over the last 50 years, risk perception studies have grown from these early works to become multidisciplinary field of investigation. However, the development of research did not achieve agreement on the definition of risk perception factors (Finkel, 2008). The main purpose of this paper is to take stock of risk perception studies, trying to test the approaches, hypothesis and models of geographers, psychologists, cognitive scientists and sociologists on the same case study. Then, in order to offer a broader understanding, it aims to explore the relationship between risk perception, residential segregation and exposure to hazards. This leads to propose a shift from location-specific approaches to a spatially and socially based framework.

4In this paper, after reviewing the three main directions in risk perceptions studies, we present the survey design and the method used to challenge them. This is followed by a presentation of the results, highlighting the significance of the respondent’s neighbourhood. Then, we discuss the impacts of exposure to hazards and residential segregation on risk perception. And finally we discuss the construction, use and opportunity of a space-based risk perception framework.

2. Taking stock of risk perception studies

5Social sciences in this sector developed according to three fundamental directions. Firstly, the early geographical approach was an attempt to use risk perception to explain the different types of adjustments to extreme events, with the aim of developing public policy (Kates and Burton 2008). The main findings of their trans-national comparative studies were that inhabitants did not acknowledge that they were living in hazardous environments and that the more they were exposed to these environments, the more they tended to minimize or ‘deny’ risks. The studies also concluded that previous experience and the characteristics of the disasters had a major influence on risk perception. Individuals reduced low risk probabilities to zero risk scenarios. If the area or individuals had experienced a particular catastrophe, there was a greater likelihood of preventative measures being taken, such as increased insurance. However, the memory of disasters diminishes over time and tends to be overtaken by more recent events (Burton, Kates and White 1978). Further work on risk denial was carried out by psychologists who saw the development of ‘optimistic biases’: when asked about their own chances of escaping a disaster, people stated that they were less likely to be affected than their peers when faced with the same situation (Weinstein 1989). It was also determined that respondents tended to rate their personal risk as being less than the general risk (Sjöberg, 2003). But, following in the footsteps of earlier studies by geographers, only location-specific studies were carried out to examine the location-specific relationship between risk perception and exposure to hazards (Lindell 2000, Daniels et al. 1992, Maderthaner et al. 1978). More recent works have focused on industrial hazards, waste facilities and nuclear plants. These have essentially confirmed the various perceptions of people living close to these types of facilities, either by comparing in-depth exploratory interviews (Baxter 2005, Parkhill et al. 2009) or by carrying out surveys based on random samples (Baxter 2009, Venables et al. 2009, Lindell and Hwang 2008) taken at different locations. These location-specific studies continue the ‘optimistic biases’ approach by focusing on the respondents’ distance to hazardous plants. However their discrete approach ignores the overall spatial context as well as the spatial gradients.

6The second direction is represented by the psychometric approach. For psychologists, risk perception is essentially based on the properties of the hazards and the subjectivity of individuals. Their ‘psychometric paradigm’ first assumed that there were no relevant individual differences in terms of risk perception. Basing themselves on multi-dimensional scaling procedures and factor analyses (Fischhoff et al. 1978), aggregated data was used to quantify and predict public risk perception. Psychologists found that two dimensions were important in explaining perceived risk and risk acceptance: ‘novelty’ and ‘dread’ (Slovic, 1987). In other words, people tend to fear hazards that are unfamiliar, cannot be observed or whose impacts are not immediate. They show a tendency to accept hazards that are less unmanageable, uncontrollable, unevenly distributed or deadly. Other dimensions were then introduced, such as trust, morality, affects and the number of people exposed (Slovic 2000), but despite the large number of comparative trans-national studies, results were inconsistent (Boholm 1998). The model has thus being widened, introducing institutional and cultural factors (Slovic 2010), perceived discriminations (Satterfield et al. 2004) or the remote effects of major natural disasters on everyday decision behavior (Västfjäll et al. 2008). However, the wide array of risk subjectivity drivers seems to differ from one study to another. In addition, critics claimed that this model overly relied on small convenience samples and that the analysis of the means rather that the raw data resulted in biases (Gardner and Gould 1989). It was also advanced that the model underestimated socio-demographic attributes, as well as the cultural and social components of the risk (Rippl 2002). When used with larger samples and on raw data, this model only explained one fifth to one third of the total variance, as did individual differences in almost the same proportion (Marris et al. 1997, Sjöberg 2000).

7The third direction in risk perception studies assumes that risk perception reflects the social and cultural context of individuals. ‘Cultural theory’ states that people ‘choose’ to worry about certain hazards and not others, because such worries reflect their social needs and views (Douglas and Wildavsky, 1982). Risk perceptions were assumed to be collective, reinforcing the group’s identity and allowing responsibility to be transferred to ‘others’. Within this context, risk denial was seen as a ‘defence mechanism’ (Lupton, 1999, Joffe, 2003). However, this theory gathered little empirical support: anthropological observations remained anecdotal and quantitative studies failed to explain more than a tenth of the variance (Sjöberg 1997, Brenot et al. 1998, Marris et al. 1998). Sociologists then concentrated their attention on the relationship between risk perceptions and socio-demographic factors such as age, gender and class (Ho et al. 2008, Lindell and Hwang 2008, Adeola 2007, Cebulla 2007, Marshall et al. 2006, Palmer 2003). Despite the recent proliferation of studies concerning socio-demographic attitude predictors regarding environmental and technological risks, results have conflicted with one another, especially insofar as the relative contributions of background variables are concerned (Finkel 2008, Adeola 2007). Some studies have studied the impact of perceived discrimination, vulnerability and contamination, but they did not actually confront residential segregation to hazard exposure (Satterfield et al. 2004, Satterfield et al. 2000). Finally, a few attempts have been made to combine this wide range of contributions with the risk amplification theory and the stigma theory (Kasperson et al. 1988, Kasperson and Kasperson 2005), but these have been based on a reduced quantity of empirical evidence (Sjöberg 2003, Sjöberg 2008).

8Following fifty years of research, most controversies are now focusing on the sampling methods used, the faith that can be placed in mailed surveys with low response rates, biased data analysis and the explanatory power of the various variables and models. Consequently, to try to achieve a more encompassing variance explanation, emphasis should be placed on representative samples and the cross-method analysis of the data.

9To this end, we develop a spatially and socially representative sample that is instrumental to shift from location-specific to space-based surveys.

3. Design of a spatially and socially representative survey

11The main issue of constructing a spatial and social risk perception framework is to build space and neighbourhood as explanatory variables and then to confront them to all other variables and drivers previously discussed in the literature (fig. 1). To this end, we constructed a spatially and socially representative sample (n = 621) of the population in Bucharest over the age of 15 using quotas, by sampling and spatially stratifying this two million inhabitants city into 20 districts. On the basis of data taken from the last available full census (March 2002), we designated these districts by redistributing the 152 census areas into 20 homogeneous groups. The demographic and social homogeneity was constructed on the basis of population breakdowns by gender, age (five-year intervals), education (7 levels), occupation (10 categories), housing types and construction dates (5 categories). We then processed all 152 census districts through Hierarchical Ascendant Classification (HAC) system on the basis of these five controlled variables. This was followed by a final HAC of the previous classifications. We then smoothed the resulting census district typologies in order to increase their spatial cohesion. This resulted in obtaining 20 almost homogeneous districts of a similar weight, each with around 100,000 inhabitants.

12We constructed the sample (n = 621) using quotas to ensure representative sub-samples (n = 31) in each of the 20 homogeneous districts. The representativeness of the sub-samples was qualified by incorporating four controlled variables extracted from the 2002 census: gender, age (five-year intervals), education (7 levels), and occupation (10 categories). During the survey, we also collected three other background variables characterizing the respondents to ensure the representativeness of the sample and the respondent’s residential district. While this sample construction path ensured that the spatial breakdowns were both coherent and logical, it also made them sensitive to residential segregation (fig. 2). Consequently, the social and spatial risk perception framework allows interpreting the respondent’s district in terms of its social and demographic characteristics, as well as its housing conditions (fig. 1).

rated importance	level of protection	time (present compared to pre-1989)
estimated frequency	level of control	space (neighborhood compared to the entire city)

earthquakes	floods	conflagrations	storms
air pollution	soil pollution	water pollution	sewers (sanitation)
high-pressure gas line	industrial plants	petrol stations	lack of green spaces

14Beside these multiple-choice questions, we also used open text questions for a more qualitative appreciation of risk perception. Having collected the background variables, the questionnaire opened then with two open text questions concerning the amenities to be found in the respondent’s residential neighbourhood and a personal definition of risk (‘these days we often talk about risks, what is a risk from your point of view?’). This was followed by two open text questions concerning disaster memory and appropriate behaviour in the event of an earthquake. At the end of the questionnaire, two further open text questions were asked concerning trust in risk communication and measures that could reduce risk in the respondent’s residential neighbourhood. All open text questions have been processed by frequency, regrouping similar answers by theme.

15We then used the collected data to gather an overall appreciation of the nature of the perceived hazards and dangers, as well as the trust and affects of the respondents. Data was subsequently processed through three different explanatory models in order to challenge the main hypotheses and approaches in the literature: the aggregate ratings of extreme events, the significance of individual differences in risk perception, and finally introducing Multiple Correspondence Analysis (MCA). The MCA enables to place emphasis both on the differences between rated phenomena and between respondents, thus summarising the main focus of risk perception. The use of MCA on raw data summarises all the information and highlights connections between variables (Greenacre and Blasius 2006). In the social and spatial risk perception framework, the MCA is used to achieve together an exploration of the expressed differences between hazards and between respondents, to disentangle the social and spatial drivers, and to allow a synthetic mapping of the results. This final step of the social and spatial risk perception framework makes now possible to confront all perceived characteristics, such as trust and emotions, and all individual differences with hazard exposure and residential segregation (fig. 1).

4. Results concerning aggregated data: perceived characteristics of extreme events

16At the aggregated level, the results are validating some of the main hypothesis in the literature, as the link between the rated importance and the estimated frequency of events and the connection between the estimated levels of protection and control. They also settle debates between geographers and sociologists, corroborating the idea that ‘risk denial’ has both spatial and temporal dimensions.

17Most respondents defined risk by associating it with seismic activity (15.8%, SE ± 0.015), traffic dangers and collisions (15.8%), pollution (14.5%, SE ± 0.014) and crime (8%, SE ± 0.011). However, a quarter of respondents failed to provide an answer (26.9%, SE ± 0.018). Only a few respondents technically defined risk as a product of probability, consequences or potential threat (12.5%, SE ± 0.013). Other answers (19%) were linked to economic, domestic or political matters. This confirms that the general public tends to define risk through its everyday or spectacular nature (Lupton 1999). A clear majority declared that the worst disaster they remembered was the March 1977 earthquake (69%, SE ± 0.018) that destroyed almost 30,000 homes (Kinzle et al. 2006). Although a subsequent but much less destructive earthquake struck the city in 1990 and despite the recent increase of the number of floods, these events were mentioned by less than 5% of respondents. These answers saw the emergence of two risk perception poles: urban nuisances and seismic activity.

18Like other studies, the results show a strong link between rated importance and the estimated frequency of hazards (fig. 3a). However, it is striking that these ratings follow the exact opposite of Farmer’s curve (Ackroyd 2005), in which severe events have the lowest probability of occurrence. There is no clear difference between disasters and negative long-term effects, given that the opinions expressed concerning explosions and pollution events are similar to one another. In addition, although storms and floods are the most frequent and recent disasters having occurred in Bucharest, they are considered to be less likely than explosions. This confirms that opinions are based on consequences rather than probability (Slovic and Petters 2006). However, earthquakes are estimated as being severe and infrequent, while industrial accidents are considered less important and more frequent. This corroborates the importance of seismic activity and the general opinion that deindustrialization has led to the elimination of industrial risks. This suggests that more recent events do not seem to have impacted risk perception as much as major traumas such as, in the case of Bucharest, the 1977 earthquake. This contradicts what has been suggested by geographers (Burton, Kates and White 1978).

19The results reveal a further obvious connection between the estimated levels of event protection and control (fig. 3b). Once again, earthquakes stand out. Given that they are judged to be unpredictable and uncontrollable, they are perceived to be beyond their reach. Conversely, pollution, lack of green areas, collapse of sewers and industrial accidents are judged to be both highly predictable and controllable. Unlike what has been argued in previous studies (Slovic 2000), this is not an opposition between familiar and unknown risks, or – given that explosions are rated as having almost the same risk level as floods – between natural and technological hazards. This opposition is more likely to distinguish between what respondents considered to be a risk and what they perceived as a controllable danger.

20The link between space- and time-perceived risk categories is less obvious (fig. 3c). However, post-1989, the more a phenomenon is considered to produce negative events, the less the respondents feel protected. In accordance with other studies, personal risk is generally perceived as lower than general risk (Weinstein 1989, Sjöberg 2003). The only exceptions to this are air pollution and storms, which are both considered as highly threatening. Respondents tended to distance earthquakes and floods as well as industrial plant, filling station and sewer accidents in both time and space. This corresponds to ‘risk denial’ (Burton, Kates and White 1978, Joffe 2003) given that an earthquake struck Bucharest after 1989 and that sewer collapses resulted from many recent floods. In addition, the respondents have declared that since 1989, their perceived level of threat represented by conflagrations, high-pressure gas line accidents, pollution incidents and lack of green areas have increased. Finally, the anxiety level is greater for air pollution and storms rather than for the most frequent or extreme events. Consequently, space-based aggregated data confirm the results of previous models using random samples and location-specific surveys. However, these results rely on means and the aggregated ratings of respondents and this can lead to bias (Gardner and Gould, 1989) and only account for differences between phenomena. According to sociologists, more variance can be explained by focusing on background variables (Finkell 2008, Adeloa 2007).

5. Results concerning individual differences: the significance of respondent’s neighbourhood

21At individual level, the main result is that the major driver of risk perception is the respondent’s neighbourhood, which allowed to map the results of the survey.

23The systematically strong explanatory power of the respondent’s neighbourhood for all the other answers (p < 0.0001) is very clear. As we felt that it could be related to the construction of the sample, we decided that we should test other background variables. We chose to analyze random variables in terms of answers, with maximal, medium and minimum co-variance with the respondent’s neighbourhood (tab. 2): increased soil pollution post-1989 (χ² (19) = 214.59, Cramer’s V = 0.5878), estimated frequency of earthquakes (χ² (38) = 229.02, Cramer’s V = 0.4294) and the severity rating of floods (χ² (38) = 123.27, Cramer’s V = 0.315).

Table 2. P-values related to background variables for three random answers

	Soil pollution post-1989	Estimated earthquake frequency	Importance rating for floods
Age	0.5534	0.2423	0.2346
Gender	0.4590	0.4957	0.7931
Date of arrival in Bucharest	0.8146	0.6419	0.3787
Education	0.4631	0.0386	0.1025
Occupation	0.0389	0.0343	0.0372
Neighborhood of residence	< 0.0001	< 0.0001	< 0.0001

25Once the significance of the respondent’s neighbourhood was established, it became possible to map the collected raw data having excluded any biases. However, mapping out all the answers to survey would have been a fastidious and probably sterile exercise. In addition, having established that many of the answers are interconnected, this process could well have failed to summarise all the information.

6. Results of the MCA: verifying classical hypotheses and allowing multidimensional mapping

26The main results of the MCA were to connect perceived characteristics and individual differences with their social and spatial context, allowing thus a multidimensional mapping of risk perception. The results also validated the model by corroborating the central hypotheses in the literature, especially on the extreme estimations, the link between frequency and predictability, the connexion between prediction and control and their changes over time.

28The first factor revealed a strong opposition between extreme estimations (negative coordinates) and less marked estimations (tab. 3). Rather than being a contrast in the value itself, it is merely an assessment of the intensity level: for example, ‘water pollution is controllable’ (0.45) is in opposition to the answer stating that ‘water pollution is very controllable’ (-1.04). The only exceptions to this are industrial hazards and soil pollution, both of which are only rated as important in the positive coordinates. Modalities expressing the changing perceived frequency of phenomena after 1989 are often close coordinates, as time and space based opinions. The statement of an increasing frequency of air pollution is connected to that of a reduced exposure to industrial accidents. However, pollution is not expressed as an undifferentiated whole: air pollution is considered to be increasing while soil pollution is regarded as to be decreasing. This distinction probably translates the deindustrialization process and the recent conversion of industrial brownfield sites. For respondents, perceived hazard reduction linked to deindustrialization seems to be balanced by increasing environmental concerns. This factor also reveals contrasted opinions concerning the impacts of deindustrialization on risks and soil pollution. As can be seen, this factor distinguishes between the importance placed on technological hazards and pollution, and their evolution over time. The differences between phenomena are based on what psychologists have called the ‘novelty’ factor (Slovic 1987), while it is the differences between respondents that account for their individual scaling levels and their opinions concerning industrial hazards and pollutions, an approach that has been called an ‘attitude toward technology’ (Slovic 2000).

Table 3. Coordinates on the first three MCA factors

factor 1		factor 2		factor 3
industrial hazards are important	0.73	gas line accidents are much less frequent	1.31	water pollution is very controllable	1.20
sewers accidents are important	0.69	conflagrations are much less frequent	1.30	gas line accidents are very predictable	1.17
soil contamination is more frequent	0.59	conflagrations are very infrequent	1.29	gas line accidents are very controllable	1.03
soil contamination is important	0.56	gas line accidents are very infrequent	1.22	air pollution is very controllable	0.97
soil contamination is frequent	0.55	storms are much less frequents post- 1989	1.17	petrol station accidents are very predictable	0.92
less exposed to industrial hazards	0.55	floods are much less frequents post- 1989	0.98	soil contamination is very controllable	0.86
lack of green spaces is predicable	0.53	storms are very infrequent	0.94	storms are much less frequent post- 1989	0.75
gas line accidents are predictable	0.52	petrol station accidents are much less frequent	0.90	soil contamination is very predictable	0.75
industrial hazards are less frequent	0.50	sewers accidents are much less frequent	0.82	water pollution is very predictable	0.69
industrial hazards are not frequents	0.49	much more exposed to petrol stations	0.78	gas line accidents are much less frequent	0.68
air pollution is more frequent post- 1989	0.49	petrol station accidents are very infrequent	0.77	storms are very infrequent	0.65
water pollution is controllable	0.45	sewers accidents are very unimportant	0.68	much less exposed to storms	0.59
industrial hazards are controllable	0.43	concerning earthquakes, failed to answer	0.64	floods are much less frequent post- 1989	0.53
petrol stations accidents are more frequent	0.43	earthquakes are much less frequent	0.61	risk mitigation with political solutions	0.46
industrial hazards are predictable	0.42	not at all scared of violent aggression	0.59	gas line accidents are controllable	0.44
water pollution is more frequent	0.41	petrol station accidents are unpredictable	0.44	sewer accidents are predictable	0.32
air pollution is predicable	0.41	industrial hazards are very infrequent	0.44	water pollution is controllable	0.31
water pollution is predicable	0.40	gas line accidents are unpredictable	0.42
gas line accidents are important	0.35	industrial hazards are much less frequent	0.25
soil contamination is predictable	0.34
water pollution is frequent	0.32
air pollution is much more frequent	-0.40	lack of green spaces is very important	-0.22	water pollution is incontrollable	-0.55
much less exposed to industrial hazards	-0.44	very scared of serious disease	-0.35	neighbourhood derives no advantage	-0.64
industrial hazards are very infrequent	-0.47	conflagrations are frequent	-0.38	soil contamination is unpredictable	-0.69
industrial hazards are much less frequents	-0.48	conflagrations are more frequent post- 1989	-0.41	sewer accidents are unpredictable	-0.74
lack of green spaces is very predictable	-0.50	very scared of violent aggressions	-0.42	water pollution is unpredictable	-0.76
gas line accidents are very important	-0.54	very scared of traffic collision	-0.44	air pollution is very incontrollable	-0.77
air pollution is very predictable	-0.54	water pollution is very predictable	-0.49	gas line accidents are very incontrollable	-0.82
industrial hazards are very unimportant	-0.59	water pollution is very important	-0.51	water pollution is very incontrollable	-0.88
much more exposed to air pollution	-0.64	gas line accidents are very important	-0.51	lack of greens is very incontrollable	-0.95
water pollution is very predictable	-0.74	petrol station accidents are more frequent	-0.56	gas line accidents are very unpredictable	-0.99
industrial hazards are very predictable	-0.79	very scared of work accident	-0.56	industrial hazards are very incontrollable	-1.09
industrial hazards are very controllable	-0.84	petrol station accidents are frequent	-0.58	petrol station accidents are very incontrollable	-1.11
air pollution is very controllable	-0.91	conflagrations are very important	-0.69	sewers accidents are very incontrollable	-1.16
soil contamination is very infrequent	-0.94	petrol station accidents are very important	-0.69	petrol station accidents are very unpredictable	-1.26
soil contamination is much less frequent	-0.98	earthquakes are frequent	-0.73	air pollution is very unpredictable	-1.35
sewer accidents are very controllable	-1.02	gas line accidents are very predictable	-0.82	industrial hazards are very unpredictable	-1.38
soil contamination is very unimportant	-1.04	floods are very important	-0.84	sewers accidents are very unpredictable	-1.55
water pollution is very controllable	-1.04	sewer accidents are very important	-0.86	lack of green spaces is very unpredictable	-1.66
much less exposed to soil contamination	-1.09	storms are very important	-0.88
soil contamination is very controllable	-1.09

29The second factor opposes opinions related to rated importance, estimated frequency and predictability of phenomena (tab. 3). On the one hand, those who considered them rare, less frequent since 1989, but unpredictable (positive coordinates) and, on the other hand, those that considered them frequent, more frequent, but predictable (negative coordinates). Questions concerning personal fears were also to be found in this factor, with modalities translating dread on negative coordinates. This factor opposed estimations related to the same phenomenon: for instance, the answer that ‘accidents connected to petrol stations are less frequent since 1989’ (0.90) was in opposition to the answer that ‘accidents connected to petrol stations are more frequent since 1989’ (-0.56). Failure to answer the two questions on actions to be taken in case of an earthquake also contributed to this factor, but with positive coordinates. This revealed that respondents who considered that phenomena had a low frequency, were becoming less frequent but were unpredictable, also failed to mention instructions for earthquake emergency situations. These connections suggested an opposition between anxious respondents who considered that the worst scenario (extreme event, danger) was inevitable, and more serene respondents who considered that these phenomena, rather than being impossible to assess, were uncertain and unpredictable due to their low probability. This factor accounts for what psychologists have called the ‘dread’ factor (Slovic 1987) and provides a greater amount of information concerning the level of anxiety to be found among respondents than on the different perceptions that they might have to these phenomena.

30The third factor specifically opposes judgments regarding levels of phenomena prediction and control and their assessment changes over time (tab. 3). On the one hand, they are considered highly predictable, highly controllable, highly infrequent and far less frequent since 1989 and, on the other hand, highly unpredictable and uncontrollable, but with some ‘optimistic biases’ (Weinstein 1989). As previously discussed, pair predictability and controllability distinguishes between controlled dangers and uncertainties. Respondents who gave the lowest predictability and controllability estimations also failed to give a definition of risk and stated that their neighbourhood was insufficiently attractive, even though they had no plans to move. Those who acknowledged high risk levels also believed that dangers could be reduced through the implementation of political solutions. This third factor distinguished thus between inhabitants who ‘accepted’ risk, turning it into a controlled danger and proposing solutions, and those who stated that they were ‘captives’ of their neighbourhood. This confirmed that risk ‘deniers’ were clearly aware of the risks (Burton, Kates and White 1978) and that risk denial could well have been part of a defence mechanism (Joffe 2003, Sjöberg 2006).

7. Confronting risk perception, hazard exposure and residential segregation

31The final step of the social and spatial risk perception framework is to offer a synthetic mapping of the results and to confront the perceived characteristics and the individual differences with hazard exposure and residential segregation. This has been facilitated by the very strong explanatory power of the respondent’s neighbourhood and the spatial opposition revealed when mapping the main findings of the previous sections (fig. 4). These strong oppositions were then confronted to hazard exposure and the respondent’s environment. These interrelationships can be appreciated by overlaying the various maps.

32Mapping the coordinates on MCA’s Factor 1 of each neighbourhood shows a clear opposition between the West, where positive coordinates display that residents tended to place more importance on industrial hazards and pollution, and the East, where negative coordinates reveal that deindustrialization and environmental issues are of lesser concern (fig. 5). Factor 2 illustrates an opposition between the South East, where residents expressed low levels of fear but also failed to define risk or review the instructions for earthquake emergency situations, and the North East (fig. 4). Factor 3 reveals an opposition between the city centre, where older residents had a tendency to ‘accept’ risks, and the outlying districts, where respondents seemed to express their ‘captivity’ (fig. 4), being the neighbourhoods with the youngest residents but the poorest living conditions (fig. 1). Other answers completed these oppositions and show the development of a further opposition between the North and the South of Bucharest (fig. 4).

33As previously discussed, while these spatial oppositions may be linked to residential segregation, they are also related to the respondents’ environment. For example, among those who stated that they were more exposed to a lack of urban green areas (47.7%, SE ± 0.020), most lived in the North and the city centre, both of which benefit from a greater number of urban amenities and the best living conditions (fig. 1). This contradiction is confirmed by relating responses to hazard exposure.

34The greatest importance given to industrial hazards and pollution (fig. 5) was given by those respondents living in West Bucharest, despite the fact that the city centre and East Bucharest are the most exposed areas to pollution while the Eastern and Southern neighbourhoods are the most exposed to hazardous materials and industrial hazards. The same also applies to floods: optimistic biases concerning judgments on floods and sewers malfunctions appear to be highest among respondents living in the neighbourhoods that are flooded most often. Consequently, it would appear that the opinions of respondents are directly opposed to their degree of exposure. As developed by sociologists (Joffe, 2003), this confirms that risk ‘acceptance’, being the assertion that consequences may not be so unbearable, and risk ‘denial’, being the belief that the probabilities of adverse events are insignificant, may both be defence mechanisms. As initially suggested by geographers, the more urban residents are exposed to hazards, the more they are likely to deny risk (Burton, Kates and White 1978).

35On the contrary, optimistic biases are very limited regarding earthquake exposure. The centre of Bucharest is built on sedimentary layers that produce maximal seismic wave acceleration (fig. 6). Locally, conditions are worse given previous earthquake exposure. Earthquakes of a power greater than 7 on the Richter scale risk striking the city almost twice a century (Kienzle et al. 2006). In addition, the city centre is characterized by ancient buildings that have remained standing during the last two major earthquakes but which have not been consolidated since. Consequently, respondents from the city centre are right to assert that they are more exposed to the consequences of earthquakes than those living in other neighbourhoods. We have seen that seismicity holds a special position in the way that the inhabitants of Bucharest perceive risk. This is linked to the yearly commemoration of the March 1977 earthquake and the earthquake exposure and building damage information policy that is well relayed by the media (Armaş 2006). As can be seen, risk communication can be efficient when it targets the most exposed populations and is regularly repeated over time. But, as previously discussed, it can also mask or take precedence over other hazards or threats, leading for example in Bucharest to the ‘denial’ of industrial hazards.

8. Conclusion

36The social and spatial risk perception framework confirms the main hypotheses in the literature; which is per se a validation of the model. Besides, it also reveals at each step the pre-eminence of space and the unsurpassed explanatory power of the respondent’s neighbourhood in risk perception. Furthermore, that space-based framework generates a multidimensional mapping of risk perception for a given set hazards, allowing to confront it with hazard exposure and residential segregation. This could subsequently be used to make comparisons with risk and vulnerability assessment maps developed by experts and help to specifically focus risk-awareness campaigns on those neighbourhoods where there are considerable disparities between perceived risk by respondents and assessed risk by experts.

37The comparison between this space-based framework and previous location-specific surveys reveals that space is a factor that cannot be overlooked. To analyze the connection between risk perception and hazard exposure, location-specific methods are used to choose random spots from which to construct their sample. Consequently, they focus more on the distance to the source of hazard than on the characteristics of the exposed area itself. Conversely, the use of MCA on a socially and spatially representative sample explains risk perception through the combination of two sets of explanations, being the differences between individuals and the differences between hazards. These are both mapped together and compared to hazard exposure as described in the risk assessment maps. In addition, the space-based framework reveals that the respondent’s neighbourhood, and not the distance to flooding zones or to hazardous facilities, is the best driver of the risk perception differences. The impact of individual differences on risk perception has been extensively discussed in previous studies, but their results remain controversial. In fact, the impact of age, gender, class and even race on risk perception might well be linked to residential segregation – and hazard exposure may well also be a side effect of this residential segregation. This implies that random sample studies as well as location-specific studies might be systematically biased.

38The shift from a location-specific to a space-based framework confirms that ‘optimistic biases’ or ‘risk denial’ are defence mechanisms related to hazard exposure. Residential segregation also leads to the expression of risk ‘captivity’. These defences are counterproductive given that they do not appear to encourage people to take action to reduce risks, at least in this case study on Bucharest, Romania. However, these defences can be opposed through risk communication, on condition that it targets the most exposed populations and is regularly repeated over time. Nevertheless, the seismicity example also confirmed that risk awareness does not favour the memorizing of emergency instructions or adjustment behaviour in Bucharest and that risk communication could simply substitute one concern by another. Bearing in mind the importance of combining risk perception, residential segregation, hazard exposure and vulnerability assessments within a social and spatial framework, it is clear that greater efforts need to be made.