Household Fire Protection Practices in Relation to Socio-demographic Characteristics: Evidence from a Swedish National Survey

The sociodemographic inequalities in the ownership of residential fire safety equipment, fire prevention practices and fire protection knowledge was studied using an inductive and data-driven approach based on the responses to a national Swedish survey containing individual-level data on several dimensions of home fire safety practices (n = 7507). Cluster analysis was used to summarise home fire safety data and sociodemographic characteristics of the sample were then regressed on the data ordinal regression analysis. The results showed significant correlations between the level of fire protection and a range of factors (sex, age, family composition, income, housing type and country of birth), suggesting a positive effect of socioeconomic success. Further, the results imply that having experienced a residential fire has a positive impact on future fire protection practices, and that higher levels of fire protection interest increases the probability of having a functional smoke detector.


Introduction
Although large risk reductions in fire-related deaths have been observed in most high-income countries during the last 50-60 years [1], household fires are still a considerable societal problem [2,3]. Specifically, although fire mortality has decreased from a general perspective, these reductions seem to have been disproportionate in terms of different socio-demographic groups, as well as there being a levelling-off of the decreasing trend. For example, whilst large decreases have been * Correspondence should be addressed to: Finn Nilson, E-mail: finn.nilson@kau.se Fire Technology, 56, 1077-1098, 2020 Ó 2019 The Author(s) Manufactured in The United States https://doi.org/10.1007/s10694-019-00921-w seen amongst adults and children, only minor rate reductions have been observed amongst older adults [4]. Also, in regards to older adults, several studies have suggested that the changing demographics, in which the number of older people are increasing substantially [5] will lead to increases in the number of deaths in countries such as Japan [6] and Spain [7].
In terms of prevention, fire-related deaths can be hindered at five points in the fire process; reduce heat; stop ignition of first object; hinder fire growth; initiate evacuation; and complete evacuation [24]. Starting with the first two steps, i.e. the development of an unwanted fire, previous studies have shown that the risk of fire, regardless of result, is higher amongst socio-demographically ''strong'' groups (well educated, high income households) compared to the rest of the population [25,26]. Therefore, it would seem that it is not that vulnerable socio-demographic groups have a higher risk of fire but rather a reduced ability to hinder fire growth and/or evacuate. Previous studies on child injuries in general have found that sociodemographic differences exist in the possession of safety equipment and the perception of safety. Specifically, they have found that safety equipment is significantly less prevalent in the homes of ethnic minorities [27,28], single-households [29], low income families [29] and families in rented accommodation [30]. Similar socio-demographic differences have been seen with regards to older people and their fire prevention equipment and evacuation preparedness [31]. If a similar pattern exists regarding the possession and knowledge of fire safety equipment in the general population, this could serve as a potential explanation for the socio-demographic differences in mortality and aid in the identification of prevention measures. It could also help clarify the conflicting results between epidemiological studies of the social determinants of residential fires and studies of fire mortality [32].

Method and Materials
For this study, cross-sectional data from a national survey, that was sent to a random sample of the Swedish adult population aged 18-79 years in 2005, was used. The purpose of the survey was to investigate the prevalence of residential fires and to obtain information regarding if the household had various types of fire safety equipment, how the equipment was maintained, and if fire safety education had

Statistical Analysis
To effectively explore the socio-demographic differences in residential fire protection practices, the different components of the questionnaire relating to these were summarised. Since the variables available were mainly categorical, multiple correspondence analysis with agglomerative hierarchical clustering was used [35], which is a cluster analysis method that allows for the summary of a larger set of categorical (e.g., nominal or ordinal) variables into a smaller number of clusters [36][37][38]. The FactoMineR package for R was used for this part of the analysis.
In the cluster analysis, variables entered to contribute to the clustering procedure are called active variables whilst supplementary variables are used to aid in the interpretation of the clusters, even though they do not actively create the clusters. Table 1 details the role of each variable in the cluster analysis. The available variables that capture safety equipment use, education, information and practices were entered as active variables in order to capture clustering around latent factors related to safety attitudes and behaviours. The goal was to identify a set of clusters that clearly show a variation in the degree to which an individual is interested in, or practice, fire-related safety in their home. Categorical respondent and household characteristics were entered as supplementary variables to analyse how these were distributed between different fire safety clusters (Table 1).
After this, the optimal number of clusters (Q) can be selected in two different ways. The first approach is based on subjective input after a graphical analysis of a hierarchical tree plot (or dendogram) and prior theoretical beliefs regarding the principal components in the data. The second approach is data-driven, and applies an algorithm that automatically selects the optimal Q based on the inertia gain for each additional partitioning [35]. Since there were no prior hypothesis regarding the optimal number of clusters, the latter approach was chosen. The identified clusters were then interpreted using multivariate v-tests to study the statistically significant differences to the sample averages (see [39] for details).
To test other hypotheses (where appropriate), Pearson's v 2 -test (to test bivariate differences between groups), ordinal logistic regression (to estimate the effects of multiple variables on categorical outcome variables), and log-binomial regression models (for binary outcome variables) were used. These analyses were conducted in Stata version 15.1.

Results
Six different fire protection clusters were identified in the analysis. Of these, the smallest cluster (n = 82) was mainly clustered around a large number of non-responses regarding safety equipment and other key variables. For this reason, this cluster was omitted from further analysis. Quantitative data from the remaining five clusters can be found in Table 2, where they are compared to the sample average on a range of fire safety behaviours and equipment use.
The results are interpreted in that the clusters represent five distinct levels of safety interest and behaviours related to residential fire safety: (1) Uninterested in fire safety, with negative responses to almost all questions regarding safety equipment in the home; (2) Minimal fire safety, where individuals belonging to this cluster have smoke detectors, but do not test their functionality; (3) Reliance on detection, which is similar to the previous cluster, but with regular testing of the smoke detector's functionality; (4) Formally educated in fire safety, which is characterised by individuals who are safety conscious, have extinguishing equipment in their home, and have obtained their knowledge through formal fire safety education, and (5) Informally educated in fire safety, which exhibit similar fire safety practices to individuals in the previous cluster, but who have obtained their safety information elsewhere (e.g. through leaflets or newspapers), meaning that compared to cluster 4, knowledge and information has more actively been searched for.
While the rank order of the clusters in terms of fire safety interest is clear, the exact distinction between cluster 1 and 2 and between 4 and 5 is less pronounced. For example, the clusters Uninterested in fire safety and Minimal fire safety, i.e. clusters 1 and 2, mainly differ in whether or not a smoke detector is installed. In Sweden, the owner of a property is responsible for maintaining a reasonable level of fire protection and therefore, if the property is a rental property, the fire protection responsibility is not with the resident, but with the landlord [40], which could serve as an underlying cause for the observed difference in smoke detector use. Unfortunately, this could not be tested using the available data.
Both cluster 4 and 5 exhibit a high level of safety consciousness and therefore rank higher than the other three. As Table 2 suggests, almost all individuals in the formally educated cluster (cluster 4) have obtained formal fire safety training (n = 1875, 98%), while only half of the informally educated cluster (cluster 5) has taken part in such training (n = 500, 48%). To explore this further, the differences in the context in which individuals in the two clusters generally obtained their fire training was studied using Pearson's v 2 -test ( Table 3). The results imply that individuals in the formally educated cluster who had received fire training Table 2 Characteristics of the Clusters Compared to the Sample Average on Observed Fire Safety Practices  Table 2 continued In essence, the formally educated cluster appears more likely to have held jobs where fire training is provided, while individuals in the informally educated cluster are more likely to have actively sought out information on their own (even after obtaining formal fire safety training).

Socio-demographic Differences Between the Clusters
Several statistically significant differences emerged when supplementary, socio-demographic variables were used to characterise the clusters. The quantitative results are presented in Table 4, and an interpretation of the cluster analysis, from a socio-demographic perspective, is presented in Table 5. The socio-demographic variables that are highlighted are those that are over-represented in the clusters compared to the sample average (according to the multivariate v-tests seen in Table 4), and thus represent how the clusters distinguishes themselves from the sample norm.
As can be seen in Tables 4 and 5, considerable socio-demographic differences exist between the five clusters. As mentioned previously, Uninterested in fire safety and Minimal fire safety merely differed in whether a smoke detector was installed. However, with the addition of supplementary variables, socio-demographic differences appeared between these clusters that could explain the differences in protec- Notes The data presented above is a subset of the sample that answered yes to having received formal fire safety education. Hence, the cluster sizes (n) do not correspond to the actual cluster size reported in the main tables tion. Specifically, although being unmarried was more common in both clusters, in the Minimal fire safety cluster, female respondents were more prevalent compared to the Uninterested in fire safety cluster where men were more common. Gender differences in fire protection has previously been well established [41] and could Notes (+) = significantly greater than the sample average (at the 0.05-level) according to a multivariate v-test, (-) = significantly lower than the sample average (tests were not performed for continuous variables). The values in each cell represent the percentage of individuals in the cluster belonging to each variable category unless otherwise stated. The sum of observations from all clusters does not correspond to the sample total due to omission of 82 individuals who formed an uninterpretable, ''unknowns'' cluster therefore serve as a partial explanation for the difference. Likewise, in the Minimal fire safety cluster, having children was more common, a factor that has previously been shown to increase worry and risk perception [42], and therefore likely to increase the motivation to protect.
Socio-demographic differences were also observed between the two other similar clusters; Formally educated in fire safety and Informally educated in fire safety. It would seem that differences exist regarding income, age and whether children live at home (Informally educated in fire safety earn more, are older and are less likely to have children living at home). Therefore, although job type is not available in the dataset, given the sociodemographic differences, it may be that the individuals in the informal education group more often have jobs where formal fire training is less likely to be required.
The Reliance on detection cluster differs considerably from other clusters, in that older adults and women are more prevalent in this group. Given the prevalence of testing smoke detectors in various ways, this group seems to be fire safety conscious, while heavily reliant on detection rather than extinguishing or escaping the fire. This could potentially be an artefact of a perceived (or actual) ability to cope with a fire by other means than escape or by the help of the rescue services. Specifically, old age has considerable effects on the physical and cognitive abilities of an individual [43] meaning that evacuation or more complex fire extinguishing can be difficult or impossible. Therefore, an early detection becomes the only reasonable preventative measure for older adults with reduced capabilities.    (17) 654.0*** 641.2*** n 7425 7425 Notes The four level scale is coded as follows: (1) Uninterested in fire safety, (2) Minimal fire safety, (3) Reliance on detection, and (4) Formally educated in fire safety + Informally educated in fire safety. The three level scale merges (1) and (2) into one category. The odds ratios (OR) can be interpreted as the change in odds for a belonging to a higher level on the fire protection scale associated with a change in predictor category compared to its reference value (indicated by ''reference'' in the table), keeping all other variables in the model constant ***p < 0.001

Regression Results
Many of the socio-demographic variables presented in Table 4 co-vary (e.g. age and income), and it is therefore also important to consider how each variable independently affects fire safety behaviour. To identify which variables still appeared to modify safety practices, while keeping the others constant, a fourlevel fire protection scale (from 1 to 4, where 1 low and 4 is high) was coded using the obtained clusters, merging the formally and informally educated clusters into one due to their similarities in exhibited fire safety behaviour. The results from this can be found in Table 6. The robustness of the results was also tested to a three-level version of the scale, merging the Uninterested in fire safety and Minimal fire safety clusters as well. As can be seen, the inferences and effect sizes are largely invariant to coding scheme. They were also robust to using the full range of the clusters in a five-level scale, where switching the rank order of the two educated clusters does not affect the results (available from the authors upon request). Running ordinal logistic regression models on these scales shows that men score higher on the fire safety scale than women, and that young respondents score significantly lower than older respondents. Marital status does not appear to affect these behaviours when adjusted for the other covariates. Rather, it appears that family type is the dominant variable, where single adult households score much lower than households with children or adult-only households with more than one adult. Individuals with lower income are on average less likely to exhibit fire safety behaviours than respondents in the middle-or high-income groups, and immigrants from non-Scandinavian countries also score significantly lower than native Swedes or immigrants from other Scandinavian countries. Finally, respon- dents living in single family homes tend to score higher than those living in multifamily homes (Table 6).

Correlation with Fires in the Past 5 Years
While the survey was not designed to test the causal effects of different safety behaviours (which would require an experimental or quasi-experimental setting), correlations were tested with self-reported residential fires in the past five years using a log-binomial regression model (residential fires reported in the sample = 273). For this, the four-level fire safety scale derived above was used (the inferences were invariant to using the alternative scales). The results produced a positive coefficient, which if taken at face value would suggest that higher fire safety scores are associated with a higher risk of fires (Risk Ratio [RR] 1.17, 95% confidence interval [CI]: 1.02-1.32). However, since the questions regarding safety practices refer to the individual's current state, while the residential fire question encompasses a five-year span, this could be an artefact of reverse causality. This notion is supported by the fact that omitting individuals who reported having changed their fire safety practices due to a past fire (n = 100) from the model yields a non-significant coefficient (RR 1.02, 95% CI: 0.88-1. 19). The differences in past fire prevalence by cluster, and the effect of removing the individuals who have changed their safety practices since, are shown in the left panel of Fig. 1. This result is consistent with previous research on individual disaster preparedness and fires [44], but cannot explain a large part of the variation in fire safety behaviours due to the low prevalence of residential fires in the sample (2 percent).

Correlation with Smoke Detector Functionality
During the survey, respondents were also asked to check the functionality of their smoke detector and report the results. In total, 82.1 percent of the sample reported having at least one functional smoke detector in their home. Testing the correlation between the four-level fire safety scale and functionality in the same manner as above, the probability of having a functional smoke detector increases, on average, by 19 percent for each step in the scale (RR 1.19, 95% CI: 1.18-1.20). As can be seen in Fig. 1, the cluster that does not frequently test the functionality (Minimal fire safety) clearly has a lower probability of having a functional smoke detector as compared to those that do, despite the fact that they are just as likely to have a smoke detector in their home ( Table 2). Removing the individuals who reported not being able to test their alarms functionality at the time of the survey, these differences were smaller, but still remained statistically significant (RR 1.14, 95% CI: 1.13-1.16). In fact, even when ignoring the Uninterested in fire safety cluster while accounting for ability to test, the prevalence of functional smoke detectors is still significantly greater in the three clusters that regularly perform functionality tests compared to the Minimal fire safety cluster (89.7 vs. 96.1%, v 2 (1) = 81.6, p < .001).

Discussion
The aim of this study was to investigate socio-demographic differences in the ownership of residential fire safety equipment, fire prevention practices and knowledge of fire prevention. The results show clearly that, in Sweden, considerable differences exist in household fire protection practices between different socio-demographic groups. These results are consistent with previous studies that have found a significantly lower use of preventative measures or practices amongst ethnic minority families [27,28,45,46], single-households and low income families [29], individuals with a lower educational level [47,48] as well as those living in socially deprived areas [49,50], thereby indicating that the level of protection is a highly plausible cause of the socio-demographic differentiation in fire-related mortality.
The results in this study also show that there seems to be a certain ''socio-demographic maturity'' in the probability of belonging to a high fire protection cluster that takes the form of an inverted u-curve across the lifespan, as shown in Fig. 2. Specifically, younger individuals living in single households with low income tend to exhibit low levels of fire protection. The level of protection then increases with sociodemographic development, to peak during middle-age when individuals have higher income and live in single-family homes with children and to then decrease again with old age, a pattern also seen in a UK government study [51]. Whether this is true from an individual perspective, i.e. that the level of protection varies throughout an individual's life, cannot be tested without access to longitudinal data, although previous studies have shown that adding a child to a household greatly increases the probability of the household having an existing fire escape plan and the probability decreases with old age [52,53]. This is particularly interesting given the fact that the curve in Fig. 2 does not mimic cross-sectional evidence of changes in positive attitudes towards risk-taking across the life span, which are consistently negative in most risk-taking domains [54]. Likewise, given that experiences of fires or similar emergency situations have been shown to increase precautionary behaviour [44] and that logically more older people would have experienced emergencies, it could be expected that a linear, increasing, fire protection curve could be seen.
Hypothetically, the regression in protective behaviours in old age compared to middle-age may be indicative of a change in the ability to perform active protective behaviours rather than an effect of changes in attitudes and perceptions of fire risks. If this is the case, i.e. that the reduced protection amongst older adults is the result of physical and mental aspects rather than attitude or risk perception, the interventions required to increase the resilience towards residential fires will likely differ between younger and older age groups as well as requiring more innovative solutions for older adults [55].
With regards to the groups with low levels of protection, a number of studies have shown effective interventions such as smoke alarm installations, education or multi-facetted programs [10,[56][57][58]. Also, a recent Cochrane review found little evidence that effective interventions to promote home fire safety practices differed in effectiveness by social group [57] meaning that it would seem that the socio-demographic differences in fire protection are not carved in stone. For the oldest age groups, given that it would seem as traditional preventative efforts are somewhat abandoned with increased age as a result of decreased physical and mental capabilities, other prevention efforts with different approaches need to be developed. As highlighted by both Jennings [59] and Corcoran et al. [60] in their respective theoretical models, differences in fire risk and fire protection are most likely the results of complex interactions of individual, societal and structural factors. For older adults this may be particularly important, especially in regards to societal factors such as loneliness, social exclusion and financial difficulties. Such aspects have been highlighted as important to include in prevention programmes [55] given that they have also been shown to increase risk behaviour [61,62]. Therefore, whilst holistic, multi-facetted programs are required for all groups with low levels of prevention, it would seem unreasonable to suggest that the same interventions are suitable for all groups.

Limitations
Firstly, data was used from a previously conducted survey and therefore no influence was had on the definition and scope of the variables collected. However, the survey captured many important aspects of residential fire safety behaviours and thus sufficiently served the purposes of this study. Still, since the procedure surrounding the creation and interpretation of the clusters, and the subsequent fire safety scales, is inductive and data-driven, it should be noted that the results could be affected by the addition of more variables relating to fire safety (e.g. explicit questions regarding safety attitudes, knowledge tests, and the presence of passive interventions such as sprinkler systems). Another issue with the data is that some of the safety questions were answered on the behalf of the household, whilst the register data was linked to the respondent, which may introduce some bias into the observed correlations between the affected variables (e.g., age and smoke detector functionality testing). Secondly, the survey response rates might be nonrandomly conditional on sociodemographic factors in a manner that is correlated with fire safety practices. If true, this could affect the external validity of the study in the sense that, for instance, respondents with low socioeconomic status are not necessarily representative of non-respondents from the same strata. Thirdly, while we hope that the results are generalisable to other contexts, they may not be comparable to countries in which cultures, fire protection laws and socioeconomic conditions differ greatly from that of Sweden.

Conclusion
Considerable socio-demographic differences exist in the level of residential fire protection. This study suggests that socio-demographic factors associated with fire protection are similar to those associated with fire mortality but not with the risk of fire regardless of outcome. Therefore, from a preventative perspective, it would seem important to focus on increasing the fire protection capabilities amongst individuals with lower socio-demographic levels. In particular, in terms of access to information, training and extinguishing equipment.