1 Background

Since at least the 1970s, it has been known that fire-related mortality is unequally distributed in high-income societies [1, 2]. These sociodemographic differences have also increased in recent years [3, 4], alongside general reductions in fire mortality risk [5], leading to ambitious and comprehensive preventative attempts. Several studies have also shown that interventions such as smoke alarms, education, or multi-faceted programs are effective [1, 6,7,8] if they are adjusted to different sociodemographic groups [7, 9]. Despite this, there are still considerable sociodemographic differences in terms of what fire safety devices are present in different homes [10].

A significantly lower use of preventative measures or practices has been seen amongst ethnic minority families [11,12,13,14], single-households, low-income families [15], and in the homes of individuals with a lower educational level [16, 17], as well as those living in socially deprived areas [18, 19]. Whilst such sociodemographic differences in fire prevention habits are well documented from a single factor perspective (for a comprehensive review of the field see [20] and [21]), a previous study [10] proposed that household fire prevention choices could be grouped into five different clusters, differing in terms of sociodemographic factors. Based on age, income, and type of family unit, it was proposed that fire prevention followed a “sociodemographic maturity” curve in which fire protection habits are low in early adulthood, increase with time, and then decrease in old age; ultimately returning to a fire protection strategy primarily reliant on fire alarms. The fact that the curve was an inverted u-curve was surprising, not least given that older age groups are commonly more risk-averse than younger age groups [22], meaning that the expectation would be that older age individuals would have at least the same level of protection as those of middle age, if not higher.

There are several underlying hypothetical reasons for why fire protection habits vary between sociodemographic groups and why the inverted u-curve was observed in the previous study. First, it is important to point out that the choice to disregard safety advice or refrain from investing in fire safety is predominantly an individual choice. In other words, although safety legislation or overarching fire safety policies can be put into place [23], such laws are exceedingly difficult—and morally questionable—to enforce. Rather, it is the individual who decides whether to invest in, or refrain from, safety interventions. Although economic constraints are believed to be an important factor in affecting this choice, some studies have shown that even when smoke alarms were freely distributed—thereby overcoming economic issues—many smoke alarms were not maintained, either through lack of interest or purposeful disabling [24], thereby indicating that deeper norms, values, or behaviours play an important part in the decision-making regarding fire safety. Consequently, it has been proposed that some groups have an erroneous risk perception, leading them to believe that protection is unnecessary [25, 26], i.e., that lower status sociodemographic groups lack an understanding of their risk profile.

A recent study has shown that high-risk individuals (e.g. older adults, and individuals with low income and education) are surprisingly competent in correctly judging their fire risk and seem well aware of their sociodemographic risk profile [27]. Similarly, despite the difficulty in performing correct risk assessments [28,29,30], data suggests that individuals living in communities with poorer societal protection—for example, due to slower response rates—rate their fire morbidity risk higher compared to individuals living in areas with quick response times [31]. Combined, therefore, some evidence would suggest that individuals have somewhat of an awareness—consciously or unconsciously—of their individual vulnerability and societal protection: two factors that are equally important in predicting the risk of fire-related deaths [32].

The aim of this study is thus two-fold. First, it updates the previous paper [10], based on data from 2005, to assess whether sociodemographic differences in fire prevention measures persist in the Swedish population with a particular focus on whether there is still an inverted u-curve in protection habits. Second, the study investigates whether fire protection practices are also influenced by the level of societal protection.

2 Material and Methods

2.1 Material

For this study, a cross-sectional dataset from a national survey was used. The survey was developed by the Swedish Civil Contingencies Agency, and administered, scanned, and entered into a dataset by Statistics Sweden. It was conducted using a combination of web and paper questionnaires. The strategy varied for participants under and over the age of 65. Participants aged 65 and older received two paper questionnaires, while younger participants received one out of a total of four mailings. The survey was sent to a representative, randomised sample of the Swedish adult population aged 18 years and older in 2022. Each respondent received a letter stating the purpose of the survey, that participation was voluntary, and was asked to consent to the collection of complementary register data. Due to the Swedish personal identification number (PIN), a unique identifier that is considered highly reliable as the register covers 99.9% of the Swedish population [33], the survey could be sent to a true representative sample. Also, the PIN enabled Statistics Sweden to link administrative register data to each respondent.

The purpose of the survey was to investigate the prevalence of residential fires and to obtain information regarding whether the household had various types of fire safety equipment, how the equipment was maintained, and whether fire safety education had been completed. The questionnaire also included a variety of sociodemographic questions as well as data related to the individual’s municipality (n.b. Sweden has 290 municipalities and a total population of 10,275,000). Data concerning income was obtained from the Income and Taxation register, and country of birth from the Total Population Register, both via Statistics Sweden [34]. Non-respondents received up to four reminders. An anonymized data file was sent from Statistics Sweden to the Swedish Civil Contingency Agency upon completion. This anonymized data file was then shared with the authors of this paper. The final response rate was 42.9%, yielding a sample size of 9420 individuals. For this study, a subsample of 9109 respondents was used, ensuring that there were no missing values in any of the variables included in the statistical analysis.

To complement this dataset with information on societal protection, freely available data from Statistics Sweden and the Swedish Civil Contingency Agency was collected at the municipal level, that is, all individuals from a certain municipality were attributed to the same data. To measure social protection, the three-year average response time (2020–2022) for rescue services to arrive in the case of a fire was utilized. The three-year average was selected as it provides a more robust measure for smaller municipalities, where the incidence of residential fires per year may be low.

2.2 Latent Class Analysis

As a first step in the statistical analysis, a latent class analysis (LCA) was employed to investigate sub-groups of survey respondents who take different measures to protect themselves from the consequences of residential fires. The poLCA package in R was used for this step. LCA is a statistical method used for identifying unobserved subgroups within a population [35]. These subgroups, or “latent classes”, are not directly observable but can be inferred from patterns in the data. The five questions included in the LCA are summarized in Table 1. Questions 1, 2, 4, and 5 were binary yes/no questions in the survey whereas question 3 was recoded into a binary variable measuring if the respondent checked their smoke alarm every year or more seldom/never.

Table 1 The Five Questions Included in the LCA
Full size table

One of the most critical steps in latent class analysis is the method of deciding the number of latent classes. There are several ways to determine how many latent classes are present in the data [35]. Each latent class must represent a meaningful and interpretable subgroup within the population. Therefore, conceptual and theoretical considerations of the different models were made before examining various goodness-of-fit-measures (see Table 2).

Table 2 Various Goodness of Fit Measures for the LCA
Full size table

The four-class model made the most sense from a theoretical perspective while the goodness of fits measures provided mixed results. The Akaike Information Criterion (AIC) is a measure used to compare models by penalizing complexity to avoid overfitting, while the likelihood ratio (G2) assesses how well the model explains the data compared to a basic model with no patterns. Both the AIC and the G2 improved as the number of classes increased, with the four-class model showing better fit than the three-class model. Although the improvement from the four-class to the five-class model was marginal as indicated by the AIC (a decrease from 37,052.81 to 37,015.58), it is important to consider the principle of parsimony which advocates for the simplest model that sufficiently explains the data.

The entropy value, which indicates the certainty with which cases are classified into latent classes, did indeed show a more substantial increase from the four-class (0.4250959) to the five-class model (0.7080913) compared to the increase from the three-class to the four-class model. However, an excessively high entropy value might also reflect overfitting, especially when not supported by similar improvements in other fit indices.

The Bayesian Information Criterion (BIC), which penalizes model complexity more heavily than the AIC, suggested a preference for fewer classes. The BIC for the four-class model (37,216.67) slightly exceeded that of the three-class model (37,209.38), and the increase to the five-class model (37,222.18) suggested diminishing benefits in terms of parsimony.

Given these considerations, the four-class model emerged as a judicious balance between statistical fit and model complexity. It provided a more nuanced understanding than the three-class model without the potential overfitting suggested by the five-class model. Consequently, the decision was made to proceed with the four-class model.

The four classes in the selected model were named ‘High protection’, ‘High alarm, low extinguish’, ‘Do not check’, and ‘Low protection’. As shown in Table 3, the ‘High protection’ class encompassed 58 percent of the respondents. This class consisted largely of respondents who had a functional fire alarm that they checked regularly and who had both a fire extinguisher and a fire blanket. The ‘High alarm, low extinguish’ class (comprising 25 percent of the population) included respondents who had functional smoke alarms but did not have fire extinguishers or fire blankets. The ‘Do not check’ class (12 percent of the population) comprised respondents who had functional fire alarms, a fire extinguisher, and a fire blanket but refrained from checking if the smoke alarm worked every year. The smallest class, ‘Low protection’, encompassed only five percent of respondents. This class comprised individuals without functional smoke detectors, who did not regularly check them, and where less than a third of the respondents had a fire extinguisher and/or fire blanket.

Table 3 Summary of the Four LCA Classes
Full size table

2.3 Multinomial Logistic Regression

As a second step in the analysis, two multinomial logistic regressions were performed; a statistical analysis technique used when the outcome variable is categorical with more than two categories. As the dependent variable, the results from the LCA with four latent classes was used. In the first regression analysis, the level of urbanisation was used as the main independent variable. This dimension was accounted for by a categorial variable for whether the respondent lived in a ‘rural’ area, a ‘small town (200–9,999)’, a ‘medium town (10,000–99,999)’, or a ‘city (100,000+)’. In the second regression, the average response time of the rescue services, as well as the type of rescue service organisation in the municipality was used. The type of organisation was a categorical variable for whether the municipality had full-time or part-time firefighters or a combination of both full- and part-time. The decision to separate the analysis into two regressions was based on the fact that there was a high correlation between response times and fire department organisation and urbanisation levels.

Several additional control variables were included in the model that are known to be influential in affecting the prevalence of fire safety equipment. Age was accounted for by a categorical variable:’18–30’,’31–64’, ‘65–79’, ‘80+’, and the model included binary variables for the respondents’ sex (‘male’, ‘female’), if the respondent was a parent of a child aged 0–21 (‘not a parent’, ‘parent’), and if the respondent had a foreign background (‘yes’, ‘no’). Moreover, income was accounted for by a categorical variable: ‘low < 25%’, ‘middle 25%–75%’, and ‘high > 75%’, and education was accounted for by a categorical variable: ‘no high school’, ‘high school’, ‘post high school’, and ‘unknown’. Finally, a categorical variable for household characteristics was included, i.e., if the respondent was living with someone (‘cohabiting’), lived alone (‘single person’), or had some other arrangement (‘other’). Table 1 in the appendix shows descriptive statistics for the dependent variables broken down by the four classes.

The multinomial logistic regressions were computed with Class 1 (High protection) as the reference category, that is, the baseline group against which the outcomes of other categories were compared. Class 1 (High protection) was chosen as the reference category since it was the largest class, comprising 58 percent of the respondents. Since multinomial logistic regression models deal with categorical outcomes and non-linear relationships, the traditional R2, designed for linear models with continuous outcomes, is an inappropriate measure. Instead, pseudo-R2 measures, such as Nagelkerke’s R2, are commonly used. Nagelkerke’s R2 adjusts the scale to allow for a maximum value of 1, providing a more interpretable measure of model fit that resembles the traditional R2 in linear regression. The Nagelkerke’s R2 was 0.1997 for the first model and 0.17455 for the second model, indicating that the first model explained a slightly higher proportion of the variance in the outcome compared to the second model. In general, these values suggest that both models had a modest level of explanatory power.

3 Results

As seen in Tables 4 and 5, several statistically significant differences were found between the respondents in Class 1 and the respondents in the other classes. Significance levels of 0.05 or lower are marked in bold in the tables. The coefficients and odds ratios of the control variables were very similar across the two regressions. For the sake of simplicity, the coefficients and odds ratios from the first regression will be referenced when discussing the results related to sociodemographic variables, and the results from the second regression will be referenced when discussing findings related to response time and the organization of rescue services.

Table 4 Results from the First Multinomial Logistic Regression
Full size table
Table 5 Results from the Second Multinomial Logistic Regression
Full size table

3.1 Sociodemographic Analysis

Class 2 (Do not check) differed from Class 1 (High protection) in although respondents generally possessed both smoke alarms and fire extinguishing equipment, they did not regularly check the functionality of their smoke alarms. In terms of age, this class had an overrepresentation of younger individuals (aged 18–30), as evidenced by a coefficient of 0.705 (p = < 0.001). This translates to an odds ratio of 2.024, suggesting that being in the 18–30 age group more than doubled the likelihood of being in Class 2 rather than Class 1. Conversely, individuals aged 65–79 were less common in this class (β = − 0.273, OR = 0.761, p = 0.003). Moreover, people living in cities and medium-sized towns were more common in this class (β = 0.402, OR = 1.494, p = < 0.001, and β = 1.314, OR = 0.097, p = 0.005). It was also more common for women (β = 0.292, OR = 1.339, p = < 0.001), parents (β = 0.353, OR = 1.424, p = < 0.001), and people with higher education to belong to this class.

Class 3 (high alarm, low extinguish) comprised individuals who generally had functioning fire alarms but were less likely to possess fire extinguishers and fire blankets. Like Class 2, this class had a higher prevalence of individuals aged 18–30 (β = 0.649, OR = 1.914, p = < 0.001) and women (β = 0.173, OR = 1.188, p = 0.001). Respondents in this class were more likely to live in small- and medium-sized towns and cities rather than rural areas. Moreover, they were more commonly childless (β = − 0.219, OR = 0.803, p = 0.003 for being a parent), single (β = 0.781, OR = 2.183, p = < 0.001), and had low incomes.

Respondents in Class 4 (Low protection) lacked functioning smoke detectors and did not generally possess fire extinguishing equipment. This class did not differ significantly from Class 1 in terms of gender, suggesting that men might be more likely to belong to this class, as descriptive statistics indicate their overrepresentation in Class 1. Respondents with a foreign background were over 3 times more likely to belong to Class 4 compared to Class 1, as indicated by an odds ratio of 3,562 (β = 1.270, p = < 0.001). Additionally, respondents in Class 4 were also more likely to reside in medium-sized towns and cities (β = 0.666, OR = 1.946, p = < 0.001 and β = 1.314, OR = 3.721, p = < 0.001), have low incomes, and live in single-person households (β = 1.086, OR = 2.961, p = < 0.001).

3.2 Response Time and Organisation of Rescue Services

The data indicates that both response time and the municipal organization of fire departments are significantly related to individual fire protection measures. In comparison with Class 1 (High protection), individuals in all other classes experience significantly shorter average response times in their municipalities. However, the effects are quite small. For instance, the coefficient for respondents in Class 4 is -0.094, with an odds ratio of 0.910, indicating a 9% decrease in the odds of a respondent belonging to Class 4 for each minute increase in response time.

Regarding the organisation of fire departments, the descriptive statistics in Table 6 indicate that it is more common for respondents in Class 1 to live in municipalities with part-time firefighters. Conversely, respondents in Class 3 and Class 4 are more likely to live in municipalities with full-time firefighters or a combination of full- and part-time firefighters. Respondents in Class 2 are more evenly distributed across municipalities. These findings are confirmed by the multinomial regression analysis where the coefficients for the organisation of the fire department are significant for Class 3 and Class 4. For Class 3, there is a negative relationship between belonging to this class and living in a municipality with part-time firefighters (β = − 0.303, OR = 0.739, p = 0.002). For Class 4 there is a negative relationship between belonging to this class and living in a municipality with part-time firefighters and a combination of part- and full-time (β = − 0.465, OR = 0.628, p = 0.021 and β = − 0.306, OR = 0.737, p = 0.021).

Table 6 Percentage of LCA Class for Each Type of Organization of the Fire Department
Full size table

4 Discussion

The aim of this study was two-fold. First, it sought to update the previous paper [10], to assess whether sociodemographic differences in fire prevention measures persist in the Swedish population with a particular focus on whether there is still an inverted u-curve in protection habits. Second, the study sought to investigate whether fire protection practices are also influenced by the level of societal protection.

Regarding sociodemographic differences, the study reveals significant disparities among groups with varying fire prevention practices. In similarity to the previous study [10], the group with the most protection is also the largest, thereby indicating that Swedes in general are consistently good at taking active measures to protect themselves from residential fires. However, the results also show that young age, being female, having an immigrant background, living alone, not having children, having a low income, and having a lower education are still associated with less fire protection measures.

As mentioned in the introduction, the previous study of fire protection practices in Sweden—using data from 2005—demonstrated that the probability of having a high level of fire protection followed an inverted u-curve across the lifespan [10]. Those results were perplexing as safety precautions often increase—not decrease—with age [22]. Therefore, whilst the increase in safety behaviour up to middle age was expected, the decrease in older age was puzzling.

The results of this study show that younger age groups are more likely to be in all classes with less fire protection measures than Class 1, thereby reiterating the safety maturity theory. However, in terms of the decrease in protection amongst older adults observed in the previous paper, there are differences in the data. First, whilst the older age group was overrepresented in Class 2 and Class 4, the odds ratios were much lower compared to the younger age group. Second, in Class 3 there were no significant differences in terms of both 65–79 and 80+, therefore suggesting that the lack of interest in safety measures observed among older adults in the previous study is no longer as pronounced.

The explanation for this change is illustrated in Figure 1. As shown, the lack of fire safety measures does not become apparent until a relatively advanced age (approximately 80 years). In the previous study—using data from 2005—this downturn occurred at around age 60 [10]. Given that the data for this study is almost 20 years newer, the results suggest that the decreased interest has remained within the same age cohort but transgressed with time.

Figure 1
figure 1

Percentage of respondents by age belonging to the four different classes

Full size image

Whilst this indicates that the inverted u-curve still exists, the fact that the downturn occurs among individuals 20 years older reveals a very important aspect: older adults do not appear to reduce their fire safety practices with age. Rather, they seem to retain these practices. In other words, at a population level, it appears that if fire safety practices are established in middle age, they are maintained. From a prevention perspective, this is highly significant knowledge as the results suggest that in 10–20 years, protection practices may no longer resemble an inverted u-curve but instead remain consistently high in the elderly population. Given the overrepresentation of older adults in fire mortality statistics in Sweden [36] and the challenges of introducing new safety measures to an older population [37], the results indicate—although further studies are needed—that ensuring an increased implementation of safety measures in middle age could have a prolonged effect for older adults.

Although the results clearly demonstrate continued sociodemographic differences in fire protection in Sweden, they also indicate that the level of societal protection is associated with the decision to invest in or refrain from safety equipment. Theoretically, living in an urban area could be understood as leading to a reduced need for autonomy regarding fire prevention, as the likelihood of being assisted by neighbours or first responders is considerably higher [38]. To provide further explanation, it could be presumed that living close to others and, therefore, also close to rescue services, may reduce the perceived need to independently detect, extinguish, or evacuate a fire compared to those living in more rural areas.

The study provides evidence supporting this rational choice perspective. Individuals living in rural areas have taken, on average, more fire protection measures than those living in urban areas. Previous research has shown that the most important factor in saving lives is the response times of rescue services [39]. Therefore, it is noteworthy that shorter response times correlate with individuals having fewer protective measures and a lower likelihood of belonging to Class 1 (High protection). Furthermore, given the correlation between response time and organisation type, the more societal protection (short response times and full-time rescue services) a respondent has, the more likely is it that the respondent belongs to Class 3 (High alarm, low extinguish) and Class 4 (Low protection).

This is interesting from the perspective that the prevention of fire mortality can occur by interacting with one of five steps; reducing heat; stopping the ignition of the first object; hindering fire growth; initiating evacuation; or completing evacuation [40]. These actions can be carried out by the individual, neighbours, passers-by, and/or the rescue services. It is known that the likelihood of being saved by neighbours, passers-by, or first responders is considerably higher in urban areas with quicker response times [38]. Consequently, it is not unreasonable for an individual living in an area with many neighbours (who could call for help, extinguish the fire, or assist with evacuation) and within close proximity to full-time rescue services (who could quickly hinder fire growth or extinguish) to focus less on personal protection compared to an individual living more remotely in a rural setting who would need to be more self-reliant.

Simultaneously, such choices present societies with a dilemma. Previous studies have shown that trust in rescue services can both enhance and diminish an individual’s preparedness for crises [41, 42]. While further evidence is required to substantiate this, the current study suggests that higher levels of societal protection are associated with lower levels of individual protection; a development that rescue services undoubtedly want to avoid. However, reducing societal protection to boost individual preparedness or protection is also undesirable, not least given that such a measure would likely be politically unpopular [43].

Despite efforts to ensure methodological soundness, there are some limitations. The study relies on survey data regarding fire protection practices with a response rate of 42.9%. Clearly, a higher response rate would have been desirable. However, given the size of the survey, the dataset is still considerable. In terms of representativity, it is well-known that questionnaires almost always have a certain selection bias in terms of which groups answer, with weaker socio-demographic groups being less likely to respond. Such a selection bias might therefore be correlated with fire prevention practices and, consequently, impact the study’s external validity. While the results should be interpreted with this in mind, it should be noted that the findings of this study in many ways mirror those of the previous study [10], indicating that there is robustness in the data. Also related to the data collection methodology, previous research has pointed to the fact that self-reported data on fire protection practices can overestimate protection levels as individuals often assume their fire alarms are operational even when they are not [44]. To mitigate this issue, respondents were instructed to check if their fire alarms were working while completing the survey. Although it is not possible to ascertain how many participants adhered to this instruction, the requirement likely enhanced the accuracy of the data.

A further limitation arises from the use of municipality-level data for the response times of rescue services. This approach fails to account for the fact that response times can vary greatly within a municipality, especially in larger or more geographically diverse areas. Whilst the large dataset will somewhat automatically adjust for these variations in most municipalities, in the most extreme cases (i.e., the largest or most diverse municipalities), average response times may give a skewed result. The results should be viewed with this in mind. Despite this limitation, the study still provides a valuable contribution to the existing literature by offering insights and a broader understanding of the dynamics between individual fire protection measures and societal protection.

5 Conclusions

This study identifies several important factors related to household fire protection habits. First, sociodemographic differences persist, with the following groups being overrepresented in lacking optimal fire protection measures: younger age groups, women, single households, childfree households, low income, low education, immigrants, and urban dwellers. Second, although a safety maturity curve is still observed, this study indicates that older adults in Sweden today are considerably more protected compared to 15–20 years ago, thereby suggesting that safety practices employed during middle age continue into old age. Third, there are indications of a rational choice perspective in terms of safety measures in relation to the level of societal protection. Specifically, living in areas with more professional and faster rescue services appears to correlate with lower levels of individual protection. Further research is needed to determine whether this correlation has causal underlays and to explore potential interventions.