Article Text

Comparison of the characteristics of fire and non-fire households in the 2004–2005 survey of fire department-attended and unattended fires
1. Michael A Greene
1. US Consumer Product Safety Commission (retired), 4340 East West Highway, Bethesda, MD, USA
1. Correspondence to Dr Michael A Greene, 9811 Dairyton Court, Montgomery Village, MD 20886-1121, USA; michael.greene{at}comcast.net

## Abstract

Objective Comparison of characteristics of fire with non-fire households to determine factors differentially associated with fire households (fire risk factors).

Design, setting and subjects National household telephone survey in 2004–2005 by the US Consumer Product Safety Commission with 916 fire households and a comparison sample of 2161 non-fire households. There were an estimated 7.4 million fires (96.6% not reported to fire departments) with 130 000 injuries.

Main outcome measure Bivariate analysis and multivariate logistic regression analyses to assess differences in household characteristics.

Results Significant factors associated with fire households were renting vs. owning (OR 1.988 p<0.0001); household members under 18 year of age (OR 1.277 p<0.0001); lack of residents over 64 years old (OR 0.552 p=0.0007); and college or higher education (some college OR 1.444 p=0.0360, college graduate OR 1.873, p<0.0001, postgraduate OR 2.156 p<0.0001). Not significant were age of house; race; ethnicity; and income. Number of smokers was borderline significant (OR 1.132 p=0.1019) but was significant in the subset of fire households with non-cooking fires (OR 1.383 p=0.0011). Single family houses were associated with non-fire households in the bivariate analysis but not in the multivariate analyses.

Conclusion Renting, household members under 18 years old and smokers are risk factors for unattended fires, similar to the literature for fatal and injury fires. Differences included household members over 65 years old (associated with non-fire households), college/postgraduate education (associated with fire households) and lack of significance of income. Preventing cooking fires (64% of survey incidents), smoking prevention efforts and fire prevention education for families with young children have the potential for reducing unattended fires and injuries.

• Epidemiology
• statistical issues
• surveys

## Introduction

The objective of this research is to add to the understanding of fire risk factors with data from the US Consumer Product Safety Commissionb's (CPSC) 2004–2005 National Sample Survey of Unreported Residential Fires.1 Most literature on fire risk factors is associated with the risk of death because of the importance of and availability of information about fatal fires and fire victims. Less information is available on non-fatal fires, and aside from household surveys in 1974,2 19843 and this survey, there is no information on risk factors for unwanted household fires in the USA that were not attended by fire departments. Risk factors for unwanted household fires in England and Wales have been collected in surveys since the late 1990s.4–6

The 1974 and 1984 surveys revealed that there are a large number of unwanted home fires that were not reported to or attended by fire departments. The 1984 survey estimated 23.7 million unwanted residential fires (96.6% unreported to fire departments) with 1.5 million injures. The 2004–2005 survey estimated an annual average of 7.4 million unwanted residential fires, 130 000 injuries and also 96.6% of fires not reported. Although differing from fire department-attended fires in the severity of injuries and losses, the large number of fires and injuries makes these fires a public health issue and the associated risk factors important.

This research builds on the 1984 survey and other literature by identifying risk factors from contrasting characteristics of fire households with a sample of non-fire households. Telephone interviews were conducted between June 2004 and September 2005. Details about the sample and methods of the 2004–2005 survey are found later in this paper.

## Background: fire risk factor literature

Almost all the literature on fire risk factors focuses on the risk of fire injury or death in fire department-attended fires. Housing-related risk factors include manufactured homes, older homes, renting (compared with owning) and rural location. Social and demographic factors include young (especially under 5 years of age), older people, minority race, low income, unemployment, lower education levels, physical and cognitive disabilities, drug and alcohol impairment, and presence of smokers. Some literature on risk factors is summarised in table 1.

Table 1

Selected literature on risk factors of fire injury or death in fire department-attended fires

There is less information on risk factors associated with fires that were not reported to or attended by fire departments. The 1984 US survey was the first to compare factors differentiating fire and non-fire households.3 That survey was a national telephone survey of 1708 fire households that reported to interviewers experiencing at least one unwanted fire during the previous 3 months. A comparison sample of 769 non-fire households was also included in the survey. Comparisons were presented for one factor at a time (bivariate analyses), and statistical significance was assessed using unweighted χ2 tests of association.

The 1984 survey found the following characteristics statistically significant: fire households had more family members, more members under 18 years old and more family members that smoked. Also, heads of fire households completed more grades of school. Not significant were type of residence (eg, single family, mobile home), owning compared with renting, age of residence, race of head of household and household income.

Household fire surveys were conducted in England in 2004/2005 in the Survey of English Housing,4 and in England and Wales in 2001/2002 and 2002/2003 in the British Crime Survey.5 6 All surveys involved face-to-face interviews. Sample sizes were 18 000 households in 2004/2005 (272 fire households) and 35 000 households in the earlier surveys (485 fire households in 2001/2002 and 537 in 2002/2003). Respondents were asked about fires in the previous 12 months, including fires not attended by the fire service.

Logistic regression analysis was used to estimate and test risk factors comparing fire and non-fire households. Significant risk factors in the 2004/2005 survey included candle use, living in an economically depressed location, dissatisfaction with housing accommodations or location, space heaters, household head employed or unemployed (contrasted with retired) and household head an employer or in higher managerial/professional occupations (contrasted with lower supervisory and technical occupations). In the bivariate analysis, low annual income (<£5000 annually) and highest annual income (more than £50 000) were identified as risk factors. One would expect income to be monotonic, that is, risk either would increase with increasing income or would decrease with increasing income. The lack of monotonicity may have been the reason for excluding income from the multivariate analysis.

The 2002/2003 survey identified the following multivariate fire risk factors: victim of crime, renters, two or more adults with or without children compared with a single adult or head of household 60 years or older, disabled head of household, smokers, fair or poor condition of house and household head having below A-level qualification or A-level and above qualification as compared with no qualifications. (A-levels are granted by examination in the UK and are required for college admission.) The bivariate analysis also found race and financially unstable households as risk factors, but those variables were not significant in the multivariate analysis. The multivariate analyses in the 2001/2002 survey identified the same factors as significant as the 2002/2003 survey with the following exceptions: head of household disability and renting were not significant risk factors, while household financial instability and geographic area with a high degree of physical disorder were risk factors.

## Methods

In the 2004–2005 survey similar to the 1984 survey, fires were defined as including ‘… any incident, large or small, that occurred in or around the home, resulted in unwanted flames or smoke, and that could have caused damage to life and property if left unchecked.’ Further survey prompts elaborated that fires included cooking fires and other types of fire incidents that required action to extinguish, but excluded ‘friendly fires,’ such as barbecues and bonfires, unless they got out of control and spread to the home. Motor vehicle fires were included only if they spread to the home.

The survey frame was all US residential (landline) telephone numbers divided into 11 strata defined by region of the country and urban/rural composition. Random digit dialling was used. The qualifying event was at least one unwanted residential structure fire during the previous 90 days. The survey sample was 916 households that experienced 961 fires. A comparison sample of 2161 households was selected randomly from 1/40th of households that responded that they did not have a fire in the previous 90 days. Approximately 580 000 numbers were dialled between 4 June 2004 and 7 September 2005. Using definitions from the American Association for Public Opinion Research,20 response rate (RR) 1 was 22.5% and RR3 was 31.6%. The formulae for calculating RR3 are below:RR3=Completed Screening InterviewsCompleted Screening Interviews+Partial Interviews+Eligible Non-Interviews+eUnknown Eligibilitywheree=Completed Screening Interviews+Partial InterviewsCompleted Screening Interviews+Partial Interviews+Eligible Non-Interviews

RR1 sets e to 1, conservatively assuming unknown eligibles were non-responses. RR2 and RR4 are similar but include partial interviews in the numerator (the number of partial interviews in the survey was negligible). Further information about the survey design can be found in the full report.1

The literature on recall of adverse events suggests that recall drops off with increasing time from the interview and that more severe events tend to be recalled longer than less severe events (for a summary of the literature, see Warner et al21 and Greene and Andres22). Like the 1984 survey, the longer the time from the interview, the fewer number of fires reported by survey respondents; that could only be explained by respondents' failure to recall fire events. To estimate fire incidence, fires were stratified by severity, missing fire dates were imputed and various hypothetical recall periods of <90 days were compared. The optimum recall periods for estimating fire incidence were determined to be up to 2 weeks from the interview for lower severity fire events and up to 3 weeks for incidents with higher severity. The fire incidence estimation sample included 257 of the original 916 fire households. With this subsample, the annual estimates of 7.4 million unwanted residential fires (7.2 million or 96.6% unattended) and 130 000 injuries were obtained. These results are described in Greene and Andres.1 22

The analysis of fire risk factors in this paper instead used the entire sample from the full 3-month recall period of 916 fire households and 2161 non-fire households. Two considerations motivated this choice: (1) loss of sample size resulting in lowered statistical power and (2) comparability with the 1984 survey that also used a 3-month recall period in the risk factor analysis (and also used a shorter period for incidence estimates). As a result of the longer period, it is likely that some non-fire households may have had fires during the 90-day period that they did not recall. The resulting contamination of the non-fire households sample with some fire households complicates identifying characteristics that differ between fire and non-fire households. Note also that the fire household sample is biased towards a higher proportion of more easily recalled higher severity incidents. As an example, 7.7% of fires in the full 3-month recall period were fire department attended, more than twice the percentage of attended fires in the fire estimation sample.

Bivariate statistics contrasting fire and non-fire households were computed using the SAS (V.9.1) software system.23 Counts were weighted by the sampling weight using methods appropriate for stratified samples. Tables were prepared using Proc Surveyfreq; averages, with Proc Surveymeans; and differences, between averages with Proc Surveyreg. Two-way tables were tested for independence using the Rao-Scott likelihood ratio F statistic, which is corrected for the survey design.24 Proc Surveylogistic was used for multivariate logistic regression models. Missing data, responses of ‘don't know,’ or refusals to respond were excluded before computation of percentages—a procedure that allocates non-responses in proportion to responses but excludes non-responses from the sample size for computing standard errors and test statistics.

## Results

### Bivariate analysis

The variables considered for the bivariate analysis were suggested by the literature and the 1984 survey. These included housing characteristics (urban vs non-urban region, dwelling type and dwelling age), number and ages of household members, various other characteristics (income, education levels, race and ethnicity) and the presence of smokers.

Bivariate analyses are presented in table 2.

Table 2

Bivariate analyses

To summarise table 2, note that fire households were significantly less likely to be in single-family houses than non-fire households, and less likely to own. Fire households had significantly more members than non-fire households (3.27 vs 2.83 on average). Fire households averaged significantly more people under 18 years of age and, in particular, under age 5; more people between 18 and 64 years old; and fewer 65 years and older. Also, heads of fire households tended to have higher education levels than heads of non-fire households.

The difference in the age of the residence was not statistically significant. There was a slightly but not significantly larger proportion of fire households with heads of Hispanic or Latino descent. Also, there were more fire households than non-fire households with income under $35 000, more non-fire households with incomes between$35 000 and $75 000, and about the same proportion of fire and non-fire in the$75 000 and over category. The distribution of household incomes between fire and non-fire households was not significantly different and was not monotonic as would be expected. Note that there were fewer respondents willing to provide income information than any other variable.

### Multivariate analyses

Candidate variables were selected from the significant bivariate variables and also variables that were significant in the literature. The analysis was based on 2895 cases; the loss of cases from the bivariate analyses was due to listwise case deletion in Proc Surveylogistic.

In table 3, the response variable modelled was the probability that a household was a fire household. Positive signs indicate increases in that probability; negative signs indicate decreases relative to the reference level. The numbers given in the column ‘Estimate’ are the natural logarithms of the ORs. ORs and associated 95% confidence limits are shown in table 4.

Table 3

Multivariate analysis of the probability a response is a fire household

Table 4

Multivariate analysis ORs

Note that most variables significant in the bivariate analysis were significant in the multivariate analysis. Renters (compared with owners), household members under 18 years of age and household heads with more than a high school education were associated significantly with fire households. Also, fire households were less likely to have members over 64 years old. Number of smokers was borderline significant in the bivariate analysis and remained borderline significant in the multivariate analysis.

In a second analysis, the number of people under 18 years of age was replaced with the number of people under age 5. This age variable was statistically significant (OR: 1.400, 95% CI 1.200 to 1.633, p<0.0001). There were no other interesting changes in the estimates, except that the number of smokers became significant (OR: 1.159, 95% CI 1.007 to 1.335, p=0.0401).

Tables 5 and 6 used the same model as tables 3 and 4, but the fire household samples were restricted to the 314 households with non-cooking fires only. The reason for the restriction was to determine if the unusual results on education and smoking might have been associated with households that had cooking fires.

Table 5

Multivariate analysis of the probability a response is a non-cooking fire household

Table 6

Multivariate analysis ORs

In comparing the non-cooking fire households' subset in table 5 with the results from table 3, almost everything has remained the same except for the larger and significant parameter estimate associated with the number of smokers. This was not unexpected, as one would not expect smoking to play a role in cooking fires as much as other types of fires.

ORs are shown in table 6.

## Discussion

This study analysed factors that differed between fire and non-fire households in the survey conducted in 2004–2005. Fire households reported at least one unwanted fire incident during the previous 90 days. Non-fire households were a 1/40th sample of respondents that did not have a fire in the previous 90 days. The survey projected to 113.3 million households, of which 1.29 million households (1.13%) were fire households and 112.05 million (98.16%) were non-fire households. Sample sizes were 916 fire households and 2161 non-fire households.

By focusing on fires, most of which were not attended by fire departments, this study presents a different view of fire risk factors. Risk factors in this present study and the 1984 study that agree with the fatal/injury fire risk factor literature include the number of young household members, renters (vs owners) and smokers (borderline significant for all fires and significant for non-cooking fires only). Also, the 2001/2002 and 2002/2003 British studies of unreported fires found smokers as a significant risk factor, but the 2004/2005 study did not. Renting was a statistically significant risk factor in the 2002/2003 British study, but not in the other studies. Race and ethnicity, which seem to appear in almost every fatal/injury fire-based risk factor analysis, were not significant risk factors in the present study or the 1984 US study.

The finding that some factors were not statistically significant does not mean that they are not risk factors. The lack of statistical significance may be due to lack of statistical power associated with sample size or contamination of non-fire households with fire households because of unrecalled fires. Also, significant bivariate associations may disappear in multivariate associations as a result of multicollinearity. That may explain why housing type was not significant in the multivariate analyses.

The results for education and income are problematic in the three British and two US studies but are too persistent to ignore. In contrast to the fatal and injury fire risk factor literature, the results for education identify a higher level of education as a risk factor. In the 2004–2005 US study, college education and postgraduate-level education are factors associated with the probability of being a fire household compared with high school or less education. This was also found in the 1984 survey. In the 2001/2002 and 2002/2003 British surveys, but not the 2004/2005 survey, fire risk was associated with respondents having A-level qualifications or below as compared with not having any qualifications; there was also an inverse relationship between risk and education.

Income was not monotonic with risk, with the 2004–2005 survey bivariate relationship showing a higher percentage of fire households in the lowest income category (<$35 000), a higher percentage of non-fire households in the middle category ($35 000–\$74 999) and about an equal amount in the highest category. The 1984 survey found that fire households had higher incomes than non-fire households, but the difference was not statistically significant. Both US surveys had a large number of respondents not specifying income. The British survey identified low income as a significant risk factor only in the 2001/2002 study, and like the present 2004–2005 US study, income was not monotonic with risk in the 2004/2005 British study.

Limitations of the US and British surveys are that fire and demographic information collected in the survey is self-reported. Assignment in the survey to fire or non-fire groups relies on respondents' recall of fire incidents. Although there is no supporting evidence, one might speculate that that there may be a correlation between education and the ability to recall fire incidents or the patience to participate in surveys. If so, such a relationship might explain the inverse relationship between educational levels and fire risk found in the two US surveys and the British survey findings about A-levels. Also, as memory decays with increasing age, failure to recall fire incidents may also explain why the presence of household members 65 years and older, when respondents, seems to be associated with non-fire households.

Another limitation of this survey is that hospitalised or displaced fire victims cannot be selected for telephone surveys. This could result in an underestimate of the number and severity of fires and injuries.

Comparing survey estimates with official statistics for fire department-attended fires, for 2004–2005, there was an average of 380 600 fire department-attended residential fires with 13 075 injuries or approximately one injury in every 29.1 fires.25 The 2004–2005 survey data of mostly unreported fires estimated 130 000 injuries from 7.4 million fires or one injury in every 56.9 fires. Although there were fewer injuries per fire, unreported fires had almost 10 times as many injuries as fire department-attended fires.

Cooking fires play a larger role in unreported fires than fire department-attended fires. About 37% of reported fires involve cooking equipment25 compared with 64% in the survey, almost all unreported. Preventing cooking fires would reduce the number of fire department-attended fires and the large number of unattended fires with associated injuries and property damage. Programs to prevent smoking would be helpful in reducing non-cooking fires, as smoking is a risk factor. With the presence of children a risk factor for fires in this study, and for death and injury in other studies, fire prevention efforts directed at families with young children also have the potential for reducing reported and unreported fires and fire injuries.

### What is already known on the subject

• Fire risk factors for fire department attended fires from the literature mainly on fatal fires include the following: young and elderly household members, lower socioeconomic status, disability, smokers, renting, older housing.

• For fires not attended by fire departments, in the USA there is only the 1984 household telephone survey. This study estimated that there were 23.7 million unwanted residential fires (96.6% unattended) with 1.5 million fire injuries.

• The 1984 study found the following statistically significant fire risk factors: larger household, smokers, household members under 18. Type and age of residence, owning versus renting and household income were not significant.

• This study updates the 1984 survey to 2004–2005, with 961 fire households and 2161 non-fire households associating renting, household members under 18, no household members 65 and smokers with risk factors. Type of housing, race, ethnicity and income were not identified as risk factors.

• The estimated number of fires has declined from 1984 totals to 7.4 million and estimated injuries to 130 000. The number of injuries is about 10 times the number in fire department attended fires.

• This study uses logistic regression to identify risk factors in contrast to most studies that use bivariate analyses. Logistic regression helps sort out variables that are highly correlated and eliminates the multiple comparison problem of bivariate analyses.

• Prevention efforts aimed at prevented reported cooking fires, smoking related fires and fire prevention efforts with families with children can also be helpful in reducing the number of unreported fires.

View Abstract

## Footnotes

• Funding In addition to funding from the Consumer Product Safety Commission, funding was also provided by the Division of Unintentional Injury Prevention of the National Center for Injury Prevention and Control in the Centers for Disease Control and Prevention, the US Department of Health and Human Services, and the US Fire Administration, Department of Homeland Security.

• Competing interests None.

• Provenance and peer review Not commissioned; externally peer reviewed.

## Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.