Jfpe 2013 1 2

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 26

Article

Journal of Fire Protection Engineering


23(1) 51–75
A comparative study ! The Author(s) 2013
Reprints and permissions:

assessing factors that sagepub.co.uk/journalsPermissions.nav


DOI: 10.1177/1042391512469521

influence home fire jfe.sagepub.com

casualties and fatalities


using state fire incident
data
Austin Anderson
Department of Mechanical Engineering, The University of Texas
at Austin, USA

Ofodike A Ezekoye
Department of Mechanical Engineering, The University of Texas
at Austin, USA

Abstract
The major factors that impact residential fire losses and casualties are relatively well
known both on the national level and for any given state. Interestingly, relatively little
has been done in comparing fire loss data between states. If state fire loss data are
compared, one should be able to identify contributing factors that influence differences
in fire losses between states. As an example, it is known that construction standards,
regulatory policy, socio-economic factors, etc. all influence the frequency and severity of
fires. In this study, subsets of National Fire Incident Reporting System (NFIRS) data for
the two largest US states by population are studied in order to identify how areas of
origin, items first ignited, and heat sources contribute to the odds of casualties or
fatalities occurring in fires, as well as to examine differences between the two states
for these three factors. Data on residential home fires from 2006 to 2010 were gath-
ered from the Texas and California NFIRS databases, Texas Fire Incident Reporting
System and California All Incident Reporting System, respectively, for this purpose.
Examination of the datasets separately using logistic regression models emphasized
that fires started in the living room or den, fires in which the item first ignited was a
flammable liquid, piping, or filter, and fires that were initiated from cigarettes, pipes, and
cigars, all have significantly high odds of resulting in both casualties and fatalities for both
states. Additionally, logistic regression modeling with interactions between state and

Corresponding author:
Ofodike A Ezekoye, Department of Mechanical Engineering, The University of Texas at Austin, USA.
Email: dezekoye@mail.utexas.edu
52 Journal of Fire Protection Engineering 23(1)

area of origin, item first ignited, and heat source indicated that for many categories, the
odds of a fire resulting in a casualty in Texas is roughly 1.5 times higher than the same
fire in California.

Keywords
National Fire Incident Reporting System analysis, fire risk, National Fire Incident
Reporting System reliability, logistic regression, USA state comparison

Introduction
Over the past 40 years or so, various organizations, such as the United States Fire
Administration (USFA), National Fire Protection Association (NFPA), and others
have been working together with various stakeholders to develop fire regulations
that mitigate the negative impacts of fire [1]. Thus far, these adopted fire regula-
tions have proven effective in reducing the yearly number of fires and fire-related
casualties and fatalities despite a consensus that fire loads in homes have risen since
the 1970s [2]. Recently however, due to concerns about overall cost control, sus-
tainability, green construction, and the effects of flame-retardant chemicals on the
environment, existing fire safety measures are increasingly requiring greater justi-
fication for their presence and new approaches are being discussed frequently [3].
When considering old and new fire safety measures, studying the viability of their
adoption with an aim to minimizing losses and maximizing safety is paramount.
Statistical analysis is a useful tool in such studies, and may help clarify the impact
of these measures. It has been noted that even descriptive statistics can be useful in
developing an understanding of factors impacting fire [4]. Unfortunately, due to the
complexity of fire phenomena and the multitude of fire safety approaches already
in place, a multitude of confounding effects (e.g. adoption of new technologies or
building practices, etc.) can often obscure examination of any particular issue on
residential fire risk. Development of methodologies that can tease out the threads
of individual issues from the tapestry of available data is important for properly
allocating resources to improve fire safety. While the NFPA provides detailed
analyses of national fire trends, there has been relatively little examination of
state or community level fire impact, perhaps because states differ in the way
they present their annual fire reports. A notable exception is Hall’s work [5] com-
paring unintentional fire death rates by states. Hall examines the impact of popu-
lation demographics on fire deaths for US states and identifies several factors that
seem to be correlated to fire deaths. Other work by Hall [6,7] has used comparative
analysis to examine fire losses in the United States relative to other countries. This
article seeks to examine fire loss data from Texas’ and California’s respective
National Fire Incident Reporting System (NFIRS) databases, TXFIRS (Texas
Fire Incident Reporting System), and CAIRS (California All Incident Reporting
System) with the goals of exposing by comparison differences in the fire problems
Anderson and Ezekoye 53

of Texas and California and opening a dialogue on the differences between the two
states. California was selected for examination because California has often taken a
leading role in the adoption of fire regulations in the USA, and Texas was selected
because it provides a mode of comparison for results, in addition to being of
interest to the researchers. This article first provides an overview of the NFIRS
reporting system and discusses some of the issues involved with using it. Next, a
description of the TXFIRS and CAIRS data is presented so that it is clear what
data were used and how. Next, the analysis, which consisted of using logistic
regression models to examine the two states separately and then compare them
while accounting for various nuisance factors, is discussed in detail and its results
laid forth. Finally, a conclusion summarizes the relevant issues and emphasizes a
few important points to take away from this analysis.

Using the NFIRS database to examine state fire problems


For this study, three issues were considered in using the NFIRS database to ana-
lyze fire incidents. They are summarized as follows:

1. The datasets must be analyzed to determine the importance of fire factors.


2. The integrity of the data must be verified to insure that they provide a reason-
able picture of the fire problem in a state.
3. The data should be sorted to perform the analysis.

Toward the first issue, this article’s approach is based on comparisons between
states, and uses logistic regression on the probability of fire casualties and fatalities
to explore these differences. Such an approach has already been utilized by Hasofer
and Thomas [8] in the past, but that work focused only on national apartment fires
for the year 1993 as a proof of concept. Their article shows how to use generalized
linear models (GLMs), a more robust form of linear regression, to rank order the
importance of fire factors in apartment fires. This study differs in that it expands
their evaluation to include demographic data, enabling comparisons between
states, and also focuses on the broader occupancy of one or two dwelling family
fires. Additionally, a different underlying probability model is used here.
Data verification is difficult, but statistically speaking, NFIRS is the best
resource for fire information in the USA because it captures such a large volume
of data. According to USFA [9], a study performed in 2008 indicated reporting
biases in the sample based on participation of fire departments. USFA also stated
that while there is participation bias, the data are stable from year to year, and that
subsets of the data generally mirror the bulk, indicating reliability. Hall [10] indi-
cates that metropolitan and urban fire departments are more likely than rural fire
departments to participate due to the requirement of documentation systems that
handle high volumes of fires and the necessity of a budget large enough to afford
the equipment required for proper NFIRS reporting. At the national level, NFPA’s
National Fire Experience Survey is typically used in conjunction with the NFIRS
54 Journal of Fire Protection Engineering 23(1)

data to form national estimates of fire. However, the NFPA survey is not publicly
available at the state level, so NFIRS remains the best available source of
information.
The question of data sorting is an important matter that is outlined in detail so
that the reader is adequately informed on what data were analyzed and what
assumptions were used in processing them. The next section is devoted to this
question.

Use of Texas and California NFIRS datasets


Technical designations of fires are defined in US Fire Administration National Fire
Data Center [11] and are supplied largely to clarify the data filtering used in the
analysis. Figure 1 gives a summary of the NFIRS data subset used in the analysis of
Texas and California fires. The analysis focuses on one or two family dwelling
residential fires because they comprise 58% of all structure fires [2]. These dwelling
types have not been closely examined in public literature outside of annual reports
from the USFA or NFPA, both organizations that typically focus on national
trends rather than state level or below. In this study, one or two family dwelling
residential fires refer to the property use associated with fire incidents, but include
any type of fire that occurs on a property of this use, such as an outdoor trashcan

Figure 1. Description of NFIRS data subset used in analysis.


NFIRS: National Fire Incident Reporting System.
Anderson and Ezekoye 55

fire or a vehicle on fire in the driveway of the home. Initially, to further narrow the
scope to fires of interest, the datasets were filtered to include only fire incident types
that involved the structure of the residence. The choice of incident types is noted to
exclude fire types 113–116, and 118, which are confined fire types. The USFA [12],
defines these confined fires as ‘fires in buildings that are confined to noncombus-
tible containers and where there is no flame damage beyond the noncombustible
container’. In order to examine the importance of confined fire incident types, a
second analysis was performed that included the above incident types. Analysis of
fire factors against civilian casualties and fatalities was enabled using the process
outlined in Hasofer and Thomas [8] whereupon casualties and fatalities are
attached to fires by common Incident ID.
A number of studies have shown that fire fatalities and casualties tend to be linked
with a number of risk factors unrelated to fires. Hall [5] and Karter [2] have both
shown that community size can have a strong impact on fatalities. Hall [5] also
examined an interesting approach to lumping ethnicity effects using what he calls
the ‘High risk index,’ defined in percentages of the population as %Black
+ %Indian/Native American%Asian. Federal Emergency Management Agency
likewise compiled a report [13] that detailed differences in risk of fire death for males
and females, and for certain portions of the population within certain age brackets.
To account for these common parameters in the model, fire department identi-
fications associated with the fires were used to determine the county of the fire
department responding to the fire. Estimates of these relevant common parameters
were then obtained on a county level for Texas and California from US census data
and attached to individual fires. The data gathered are summarized as follows:

. percentage of population in a given county whose age is less than 5 or greater


than 55. According to US Fire Administration National Fire Data Center [13],
these are the age brackets for which relative risk of fire death is greater than 1;
. percentage of population in a given county classified as ‘rural;’
. percentage of population in a given county that is male; and
. ‘high risk index’ of the county, as defined above.

All of the above data were available from the 2010 US census data, except the
rural population percentages, which are as yet unavailable. Census 2000 data were
used for rural population proportions instead.
The factor categories available for examination from the Texas and California
datasets come from NFIRS 5.0 Basic, Fire, Structure Fire, and Casualty mod-
ules [11]. They are outlined below, with a brief description for each:

. incident type – the actual situation emergency personnel found on the scene
when they arrived. This article is examining incident types keyed to structure
fires only, defined as incident IDs 111 and 120–123;
. property/contents losses – four categories describing estimates of the USD losses
of contents and property due to the fire;
56 Journal of Fire Protection Engineering 23(1)

. area of fire origin – the room or area in which a fire started;


. heat source – the pilot or source of heat that started the fire;
. item first ignited – the first item ignited in the fire;
. type of material ignited – the composition of the material in the item first
ignited;
. cause of ignition – the general causal factor that resulted in a heat source igniting
a combustible material (e.g. intentional, unintentional, and equipment failure);
. ignition factor – the contributing factors that caused the heat source to ignited
the combustible material (e.g. playing with heat source, discarded cigarette, etc.);
. fire origin – the story where the fire originated in the building;
. fire spread – the extent of fire spread in terms of how far the flame damage
extended;
. item contributing most to flame spread – which item contributed most to flame
spread if different from the item first ignited;
. type of material contributing most to flame spread – the type of material con-
tributing most to flame spread if different from the type of material first ignited;
. detector presence – the existence of fire detection equipment within its designed
range of the fire;
. detector operation – the operation and effectiveness of the detector relative to
the area of fire origin; and
. detector effect – the effectiveness of the fire detection equipment in alerting
occupants.

Of all these categories, area of origin, item first ignited, and heat source are of
primary interest for comparison between the two states. Detector operation is also
included to help account for potential differences in detector operating rates in fires
between the two states.
The contents loss estimates might potentially be useful as another variable to
characterize the effects of fire occurrences, but these estimates can come from both
on-scene firefighter reports and from insurance loss data. It is felt that before using
NFIRS loss estimates, a comparison to a state’s actual insurance payout would be
necessary to ascertain whether the NFIRS loss estimates are adequate for making
observations. The insurance payout data were not available.
Area of origin, item first ignited, and heat source each have a very large number
of potential reporting categories (greater than 30), and many of those categories
have very few reported fires. Thus, to increase stability for the regression model at
the expense of detail, some categories in the factors were collapsed into each other,
either by overarching theme or by common relation, e.g. cigarette, pipe, and cigar
fires were merged into one category. Another problem encountered in the data was
typographical errors in reporting for both the states. To avoid making any assump-
tions about what the error pertained to, without throwing out fires in which only
one reporting factor had a typographical error, incorrect codings or fires for which
the factor was not included were summarily recoded to ‘UU,’ designating undeter-
mined. Tables 1 to 3 provide the strict collapses for each of the factors in terms of
Anderson and Ezekoye 57

Table 1. Category collapses for areas of origin.

Collapsed
Original categories category New definition

01,02,03,04,05,09 01 Means of egress (e.g. hallway)


20,25–28 20 Function areas, other
40–46 40 Storage areas, other
71,73–75 71 Void space within structure, includes attic
10–13,15–17,50–69,80–98 00 Other area of origin

Table 2. Category collapse for items first ignited.

Collapsed
Original categories category New definition

20,23–26 20 Furniture, utensils, other


31–33 31 Mattress, pillow, bedding of some sort, or linens
30,34–38 30 Soft goods, wearing apparel, other
40–47 40 Adornment, recreational material, signs
50–59 50 Storage supplies
60–68 60 Liquids, piping, and filters
70–77 70 Organic materials
81–82 81 Electrical wire, cable insulation, or transformer
83–99 00 Item first ignited, other

Table 3. Category collapse for heat sources.

Collapsed
Original categories category New definition

10,11–12 10 Heat from operating equipment, other


40–43 40 Hot or smoldering object
50–56 50 Explosives, fireworks
61–62 61 Cigarette, pipes, or cigar
64–66 64 Small open flame, e.g. matches
60,63,67–69 60 Heat from open flame or smoking materials, other
70–74,80–82,84,97,00 00 Heat sources, other
58 Journal of Fire Protection Engineering 23(1)

the NFIRS codes, available in US Fire Administration National Fire Data


Center [11], as well as a new definition for the collapsed category.
Overall, the dataset obtained at the end of processing consisted of 46,700 fires
from Texas and 35,100 fires from California, with a total of 2765 fires having at
least one casualty and 606 fires having at least one fatality.

Initial examination of the TXFIRS and CAIRS data


The initial processing of the data consisted of examining the two states separately.
Casualties and fatalities in the Texas and California data were examined separately
using two main effects logistic regression models for each state. This examination
provides a picture of the effects of area of origin, item first ignited, heat source, and
detector operation when controlling for the various aforementioned common par-
ameters. Next, the datasets were aggregated and casualties and fatalities between
the states were examined using a logistic regression model with interaction effects
on area of origin, item first ignited, and heat source. The logistic regression model
is used to determine the odds of either a fatality or a casualty occurring based upon
various underlying factors. This type of model allows for comparison of casualties
and fatalities between the two states by individual categories, allowing significant
differences to be analyzed and quantified. The basic form of a logistic regression
model is
"   !#
P Fjx~
E log   ¼  þ 1 x1 þ 2 x2 þ 3 x3 þ    þ M xM
1  P Fjx~

The model provides the expected value, E[ ], of the log of the odds ratio of F
occurring conditioned on a scenario denoted by the vector of x variables. The
scenario variables, x, are often termed explanatory variables. These explanatory
variables can either be categorical or continuous. In the following section, the types
of explanatory variables used in these models are described. A script for these
models and the relevant tests accompanying them was produced for and run
using the R programming language.

Texas and California main effects models


Two main effects models were produced for each state, one for examining  fires
 with
casualties, the other for examining fires with fatalities. For fatalities, P Fjx~ is the
probability that a fire with given explanatory variables results in a fatality.
The model explanatory variables are as follows

x~ ¼ ðAO, IFI, HS, DO, AP, MP, HRI, RRÞ


Anderson and Ezekoye 59

where the terms in the regression models are defined as follows:

. AO – factor with 12 categories describing various areas of origin;


. IFI – factor with 13 categories describing various items first ignited;
. HS – factor with 10 categories describing various heat sources;
. DO – factor with three levels describing whether a detector operated or not and
undetermined for fires without information;
. AP – the percentage of the population less than age 5 or greater than age 55,
at county resolution;
. MP – the percentage of the population that is male, at county resolution;
. HRI – Hall’s ‘High risk index,’ defined in the previous section, at county reso-
lution; and
. RR – the percentage of the population classified as ‘rural’ by the 2000 US
census, at county resolution.

Before discussing the results of the logistic regression models, it is worthwhile


to describe the nature of the data contributing to them. As mentioned in the
above sections, Texas had 46,720 fires, of which 1659 had at least one reported
casualty and 358 had at least one reported fatality. California had 35,100 fires,
of which 1106 had at least one reported casualty and 248 had at least one
reported fatality. In terms of categories, frequencies for reported areas of origin,
items first ignited, and heat sources for Texas’s and California’s data subsets
are summarized in Tables 4 and 5 along with the frequency of the categories
in fires involving casualties and fatalities. As a reminder, these tables reflect
the frequencies from the NFIRS data subset used in this article, as depicted
in Figure 1.
From examination of Tables 4 and 5, undetermined fires comprise a large por-
tion of the data. This is perhaps expected when dealing with fire data, as it is a
phenomenon that tends to remove evidence. Fires involving fatalities are especially
prone to this lack of detail in reporting, likely due to a combination of fatalities
being an extremely rare and usually severe structure fires. Partial unknown cate-
gories are defined through a set subtraction operation. For example, item first
ignited code 20, corresponding to ‘furniture, utensils other,’ refers to the set of
any furniture or utensils not captured by sets codes 21 or 22, and can be interpreted
as, ‘all furniture or utensils that are not upholstered furniture (21) or non-uphol-
stered chairs and benches (22).’ Likewise, the overall ‘other’ code 0 serves as the
catch-all identifier for fires where the area of origin/item first ignited/heat source
was identified but did not fit into any other code. The ‘undetermined’ code UU
then encompasses all fires for which the area of origin/item first ignited/heat source
was either undetermined, or for which the field was not filled out, or for which a
typographical error was present. Including UU allows the model to control for the
aforementioned effects, but the individual contributions of typographical errors
cannot be quantitatively discussed uncoupled from legitimately undetermined
sources and reporting errors.
60

Table 4. Summary of Texas NFIRS data subset used in logistic regression modeling.

Texas data subset: 46,720 fires, 1659 with casualties, 358 with fatalities

Area of origin Item first ignited Heat source

Frequency Fires Fires with Fires with Fires Fires with Fires with Fires Fires with Fires with
rank (number) casualties fatalities (number) casualties fatalities (number) casualties fatalities

1 UU (11,124) 21 (410) UU (101) UU (21,505) UU (693) UU (234) UU (20,143) UU (610) UU (219)


2 24 (6372) 24 (339) 21 (94) 10 (9631) 70 (163) 10 (22) 10 (8081) 10 (395) 10 (45)
3 21 (5402) 14 (195) 14 (57) 70 (2888) 10 (156) 60 (20) 13 (5264) 64 (179) 13 (20)
4 71 (5064) UU (195) 24 (33) 81 (2494) 60 (132) 31 (20) 0 (4745) 13 (122) 61 (19)
5 20 (4282) 20 (163) 20 (32) 0 (2108) 31 (131) 21 (18) 40 (3411) 0 (102) 0 (17)
6 70 (4023) 47 (91) 0 (10) 20 (2066) 30 (87) 0 (12) 64 (2367) 40 (90) 64 (16)
7 0 (3554) 70 (68) 47 (10) 30 (1658) 20 (77) 81 (12) 60 (1484) 61 (79) 60 (14)
8 47 (2259) 0 (65) 70 (8) 31 (1569) 21 (72) 30 (8) 61 (909) 60 (73) 40 (6)
9 14 (2170) 71 (61) 1 (6) 60 (1244) 0 (70) 20 (7) 50 (210) 50 (7) 50 (1)
10 40 (1434) 40 (40) 71 (6) 21 (885) 81 (51) 70 (3) 83 (106) 83 (2) 83 (1)
11 1 (839) 1 (26) 40 (1) 50 (404) 50 (18) 22 (1) – – –
12 23 (197) 23 (6) 23 (0) 40 (216) 40 (8) 40 (1) – – –
13 – – – 22 (52) 22 (1) 50 (0) – – –
NFIRS: National Fire Incident Reporting System.
Numbers within the parentheses indicate the actual number of fires.
Journal of Fire Protection Engineering 23(1)
Table 5. Summary of California NFIRS data subset used in logistic regression modeling.

California data subset: 35,100 fires, 1106 with casualties, 248 with fatalities
Anderson and Ezekoye

Area of origin Item first ignited Heat source

Frequency Fires Fires with Fires with Fires Fires with Fires with Fires Fires with Fires with
rank (number) casualties fatalities (number) casualties fatalities (number) casualties fatalities

1 24 (4640) 21 (284) 21 (52) UU (12,966) UU (373) UU (142) 10 (8055) UU (341) UU (136)


2 21 (4451) 24 (181) UU (45) 10 (7003) 31 (114) 31 (21) UU (7985) 10 (236) 10 (27)
3 20 (4377) 20 (129) 14 (41) 0 (2645) 60 (95) 10 (19) 0 (5433) 64 (122) 0 (22)
4 71 (3742) 14 (118) 20 (35) 20 (2293) 10 (93) 0 (16) 13 (4017) 0 (105) 61 (16)
5 0 (3645) 47 (112) 24 (25) 81 (2211) 0 (87) 21 (13) 40 (3702) 13 (81) 64 (15)
6 70 (3591) UU (75) 70 (13) 70 (2080) 70 (76) 60 (12) 64 (2324) 61 (78) 60 (14)
7 UU (3504) 70 (53) 0 (8) 30 (1556) 20 (73) 20 (9) 60 (1682) 40 (74) 40 (9)
8 47 (3126) 0 (50) 47 (8) 31 (1472) 30 (62) 30 (7) 61 (876) 60 (62) 13 (8)
9 14 (1843) 71 (41) 71 (8) 60 (1119) 21 (59) 81 (4) 83 (678) 50 (5) 50 (1)
10 40 (1320) 40 (34) 1 (6) 21 (769) 81 (45) 40 (2) 50 (348) 83 (2) 83 (0)
11 1 (760) 1 (24) 40 (5) 50 (652) 50 (20) 50 (2) – – –
12 23 (101) 23 (5) 23 (2) 40 (294) 40 (8) 70 (1) – – –
13 – – – 22 (40) 22 (1) 22 (0) – – –
NFIRS: National Fire Incident Reporting System.
Numbers within the parentheses indicate the actual number of fires.
61
62 Journal of Fire Protection Engineering 23(1)

Four analysis of variance tables testing the significance of predictors in each of


the four models using Wald tests indicated that the four predictors of interest in
this study, area of origin, item first ignited, heat source, and detector operation,
were all significant at the 5% level. Thus, individual examination of the categories
in these factors was pursued to provide some quantification of the effects of fires
with these characteristics on the odds of a fatality or casualty occur. Figures 2
and 3 summarize the proportion of the odds of a fire resulting in a fatality/casualty
for various areas of origin, items first ignited, and heat sources in terms of the odds
of a fire casualty or death for the ‘riskiest’ categories, which for both states and
regardless of examining fatalities or casualties were found to be fires occurring in
the living room (14), with item first ignited being some form of flammable liquid
(60–68), and the heat source being either cigarettes, pipes, or cigars (61–62).
Mathematically, the odds ratio can be represented as follows, using area of
origin (AO ¼ 21) and examining casualties as an example

Oddsð fire resulting in casualty jAO ¼ 21, z~Þ


OR ¼ ¼ eðAO¼21 AO¼14 Þ
Oddsð fire resulting in a casualty jAO ¼ 14, z~Þ

which is the point estimate for the area of origin category (21) when examining
casualties, where the z~ inside the conditioning contains all other explanatory vari-
ables (e.g. item first ignited, heat source, and common factors) that are the same.
Thus, one would say in this instance that for a given fire occurring in Texas, the
odds of it resulting in a casualty, given that it started in the bedroom (AO ¼ 21), are
ð1  0:796Þ  100 ¼ 20:4% lower than the odds of it resulting in a casualty if the
fire had started in the living room or den (AO ¼ 14). Confidence interval bars at
95% are included on both Figures 2 and 3. If a given point estimate lies outside the
error bars of another, then the two estimates are considered significantly different
at the 5% level. Certain categories that had a large standard error compared to
most other estimates, defined as a standard error greater than three times the
average of the standard error across the categories in the factor, have been excluded
from the figures because they are not useful for making statements aside from
‘there were not enough fires in this category to produce a good estimate.’ These
categories can be readily guessed by looking at which categories had very few or no
instances of casualties or fatalities in Tables 4 and 5.
Examining Figure 2, notable risk locations in Texas areas of origin compared to
the worst, living room fires, are bedroom fires (21), and then other function areas
(20,24). There were no particularly large differences between fatalities and casual-
ties. For both casualties and fatalities, there are no significant differences in risk
associated with items first ignited in Texas, except that all are small relative to fires
involving flammable liquids, gas pipes, and filters. Likewise with heat sources in
Texas, fires started with cigarettes, pipes, and cigars generally have significantly
higher odds of fatalities than most other heat sources, and fairly higher odds of
casualties, compared against the other categories, which are generally less than half
as risky. It is notable however that in Texas fires started with matches, lighters, and
Anderson and Ezekoye

Figure 2. Ratio of the odds of a given fire in Texas resulting in a casualty or fatality for areas of origin, items first ignited, and heat sources,
63

as compared against the living room/den (14), flammable liquids, piping, and filters (60), and cigarettes, pipes, or cigars (61).
64

Figure 3. Ratio of the odds of a given fire in California resulting in a casualty or fatality for areas of origin, items first ignited, and heat
Journal of Fire Protection Engineering 23(1)

sources, as compared against the living room/den (14), flammable liquids, piping, and filters (60), and cigarettes, pipes, or cigars (61).
Anderson and Ezekoye 65

candles (64) are worse with regard to causing casualties than they are for causing
fatalities.
Examining Figure 3, for areas of origin in California, fires generally fall into
three separate tiers with respect to their odds ratios for fatalities, living room/den
fires (14), and then other functional areas (20,24,21) which have approximately 0.4
the odds of (14), and finally all other fires (70,71,40,47,0). For casualties, bedroom
(21) and living room fires (14) in California appear equivalent, and while it may
seem that kitchen fires are riskier (24) than the rest of the categories, this assertion
applies only to fires that involved the structure of the home, as confined fires have
been excluded. Examining items first ignited, it seems that there were not enough
fatalities to make any strong statements regarding the effects of items first ignited
on the odds of a fatality, and for casualties it seems possible only to assert that fires
in which flammable liquid, piping, or filters are the item first ignited are far more
likely to cause casualties than any other items. Similarly with heat sources, it seems
that also in California, fires lit from cigarettes, pipes, and cigars have significantly
higher odds of resulting in fatalities or casualties than most other heat sources,
which all had roughly equivalent odds of fatalities/casualties. There were not
enough data however, to confirm that other smoking materials and open flame
ignited fires were any different from cigarette and related fires.

Texas and California interaction effects model


After the initial examination in which the two states were analyzed separately, the
two datasets were aggregated with the addition of a simple dummy variable, ‘ST,’
that specifies in which state a given fire occurred. Six binary logistic regression
models were then formulated with an interaction term included for each factor
of interest, area of origin, item first ignited, and heat source, for both fatalities
and casualties. Shown below is the formulation for the area of origin models:

x~ ¼ ðAO  ST, IFI, HS, DO, AP, MP, HRI, RRl Þ

where x includes both the main effects and simple interaction terms (AO  ST) for
area of origin and the dummy variable state (ST), coded as 0 for California and 1
for Texas. The nuisance terms were once again included to control for potential
demographic differences between the two states.
With the models formulated as above, joint tests for independence between
Texas and California were carried out for each interaction term in each model.
All fatality models but the one containing IFI  ST were found to be significant at
the 5% confidence level. However, when individual odds ratios were examined, it
was found that all three joint tests were largely driven by differences in undeter-
mined fires between Texas and California. The only detectable difference between
the odds ratio of a fire resulting in a fatality between Texas and California was for
fires ignited by heat from operating equipment (10). Joint tests for the models
examining casualty odds between Texas and California were rejected at the 5%
66 Journal of Fire Protection Engineering 23(1)

level for all three factors. Individual differences were explored in more detail for the
casualty models.
Figure 4 summarizes odds ratios of fires resulting in casualties between the states
for a given category, with 95% confidence intervals. Interpretation of the odds
ratios for Figure 4 is different. If the uncertainty in a given category extends to
1.0, that suggests that no difference between the states was detected (the odds are
the same for both). Additionally, values are no longer compared against a reference
category, but are comparisons between the states.
Examining Figure 4, it is apparent that Texas is uniformly worse than California
in many categories. The odds that fires starting in the living room/den (14), other
function areas (20), and bedroom fires (21) resulting in a casualty are roughly 1.5
times higher in Texas than in California. Cooking area/kitchen fires (24) also
appear to be 1.5 times higher, but, as previously noted, this analysis is for fires
directly involving the structure of the home, and confined kitchen fires were
excluded; thus strong statements regarding kitchen fires are discouraged. The
odds that fires in which the item first ignited was a soft material (not mattresses,
bedding, or linens) or wearing apparel (30), a structural component or finish (10),
furniture and utensils not including upholstered furniture or chairs nor non-uphol-
stered chairs and benches (20), and flammable liquids, piping, or filters (60) result-
ing in a casualty are also roughly 1.5 times higher in Texas than in California. The
organic materials (70) category is also significant (notably 1.75 times higher), but
the category includes cooking materials, and thus is subject to the same bias as
kitchen fires. Finally, the odds of fires that were started by heat from operating
equipment (10), matches, lighters, or candles (64), hot or smoldering objects (40),
other smoking materials or flames that are not contained in (64) nor are cigarettes,
pipes, or cigars (61), electrical arcing fires (13), and other fires not distinguished but
recorded (0), resulting in a casualty are roughly 1.5 to 2 times higher in Texas than
in California.
Given the current debate on upholstered furniture fire risk, it is notable that no
difference could be detected in the odds of casualties or fatalities between Texas
and California for upholstered furniture fires for this data set even though living
room fires in Texas appear to be much riskier. It is apparent looking at the indi-
vidual states’ results (Figures 2 and 3) that upholstered furniture first item ignited
fires are major risk factors to injuries and deaths. For both states, the overlap in
odds of fatality associated with these fires overlaps with the worst case scenario of
item first ignited being fuel. The significant spread in the uncertainty is associated
with the amount of data in these categories and this contributes to lack of differ-
entiation in the odds ratio comparison of the two states. Additionally, the inter-
action model that was previously described aggregates the data for the two states
and fits best approximating sensitivity coefficients for the parameters. The model
essentially notes a difference in the two-state data assuming all factors other than
item first ignited have the same behavior. Because the states are in fact different, it
is possible that this relatively simple model form is unable to distinguish the risk
differences between the states. Toward explaining the issues in item first ignited
Anderson and Ezekoye

Figure 4. Ratio of the odds of a given fire resulting in a casualty between Texas and California for various areas of origin, items first ignited,
and heat source. Error bars overlapping with 1.0 mean, there is no detectable difference between the two states for that category.
67
68 Journal of Fire Protection Engineering 23(1)

Figure 5. Simple statistical analysis of casualties and fatalities per capita for California and
Texas based upon first item ignited.

fires, a simpler evaluation of the two-state data was also considered. Fire casualties
and deaths in California and Texas were analyzed using simple statistical analysis.
The NFIRS data for the two states were not corrected for reporting errors and thus
there are limitations to the overall interpretation of the results. Nevertheless,
Figure 5 shows that when looking at upholstered furniture fire injuries and
deaths, the per capita number of fire deaths and/or injuries when the first item
ignited is upholstered furniture is larger in Texas than in California (4.28 per
Anderson and Ezekoye 69

Table 6. Information available from confined fire incident types for Texas and California.
Area of origin Item first ignited Heat source

Identified Undetermined Identified Undetermined Identified Undetermined

Texas (12,817 Number of 3433 9384 2746 10,071 3264 9553


fires, 131 fires
fires with Number 47 84 39 92 43 88
casualties) of fires
with
casualties
California Number 6783 16,039 5895 16,927 6719 16,103
(22,822 of fires
fires, Number 33 47 26 54 34 46
80 fires of fires
with with
casualties) casualties

million versus 1.91 per million for casualties and 0.9 per million versus 0.38 per
million for fatalities). While item first ignited upholstered furniture fires are some-
what lower frequency occurrences than other categories, they rank relatively high
in total number of injuries and deaths. These results are found simply by perform-
ing counts on the incidents. As previously noted, the per capita count differences
between the two states are not identified in the GLM because the GLM state
interaction model aggregates the effects of all other factors between the two
states. This preliminary view of the state to state comparison for item first ignited
suggests that additional data and modeling changes may be required to better
compare the two states.

Effects of confined fire incidents


A second analysis, procedurally identical to the first, was performed on the same
dataset but included the confined fire incident types (113–118). The inclusion of
these incident types resulted in, for California and Texas respectively, an additional
22,822 and 12,817 fires, 80 and 131 fires with casualties, and no new fires with
fatalities. Concerning how much information is actually contributed by these fires,
the USFA [5] noted that ‘a result of abbreviated reporting for confined fires is the
potential for larger unknowns as detailed reporting of fire specifics (e.g. room of
origin) is not required.’ This trend generally shows in the data, as summarized in
Table 6.
While there are indeed many ‘undetermined’ fires brought in by including con-
fined fires in the analysis, a sizable number of identified fires were also added.
Figures 6 to 8 summarize the results of the second analysis that includes con-
fined fires. Including the confined fire incident types did not strongly change the
70

Figure 6. Ratio of the odds of a given fire in Texas resulting in a casualty or fatality for areas of origin, items first ignited, and heat sources,
as compared against the living room/den (14), flammable liquids, piping, and filters (60), and cigarettes, pipes, or cigars (61).
Journal of Fire Protection Engineering 23(1)
Anderson and Ezekoye

Figure 7. Ratio of the odds of a given fire in California resulting in a casualty or fatality for areas of origin, items first ignited, and heat
71

sources, as compared against the living room/den (14), flammable liquids, piping, and filters (60), and cigarettes, pipes, or cigars (61).
72

Figure 8. Ratio of the odds of a given fire resulting in a casualty between Texas and California for various areas of origin, items first ignited,
and heat source. Error bars overlapping with 1.0 mean, there is no detectable difference between the two states for that category.
Journal of Fire Protection Engineering 23(1)
Anderson and Ezekoye 73

conclusions that can be drawn regarding the odds of fires resulting in casualties or
fatalities in both Texas and California. The state to state comparisons do change
(Figures 4 and 8). When including confined cooking fires, the odds of a fire in
which the first item ignited is some form of organic material resulting in a casualty
in Texas is roughly 2.5 times that of California. Additionally, Texas has double the
odds of a kitchen fire resulting in a casualty than California. Also, when Figure 8 is
compared against Figure 4, fires coded as UU which include fires involving casual-
ties, the combination of overall reporting rates/typographical errors/undetermined
fields when confined fire types are included are more hazardous in Texas than in
California. Finally, it is notable that the odds ratio between Texas and California
for fires in which the first item ignited is upholstered furniture moved even closer to
one, consistent with the overall uncertainty in the estimates for such fires and the
limitations of the state interaction GLM.

Conclusions
Analysis of NFIRS data for Texas and California, accounting for various relevant
demographic factors like age, gender, ethnicity, and (roughly) community size at the
county level, as well as operation of fire detecting devices, confirmed the danger of
fires starting in the living room/den, and bedroom of a residence, fires in which the
first item ignited were flammable liquids, pipes, or filters, and fires started by cigar-
ettes, pipes, and cigars. Additionally, logistic regression models including interaction
effects while still controlling for the demographic factors at the county level were
developed to identify differences in civilian risk of death or injury between Texas and
California for various areas of origin, items first ignited, and heat sources. For those
categories that the data could detect significant differences, the odds of a casualty in a
fire is1.5 times greater in Texas than in California. Some categories, such as item first
ignited, that appeared to be different between the two states when looking at casual-
ties and fatalities per capita were not found to differ statistically when analyzed using
the regression model. The differences between these two types of analysis are likely
due to the scarcity of data and limitations of the interaction-based regression model.
This analysis has identified a number of differences between the Texas and California
fire problems, and it is hoped that it will help guide considerations of future exam-
inations between states and development of improved models.

Acknowledgments
The authors appreciate a critical review of an earlier version of this manuscript by an ano-
nymous reviewer. The authors also thank Prof. Elmira Popova for her insight and advice on
aspects of this work. Additionally, the authors would like to thank the California
Department of Forestry and Fire Protection’s NFIRS administrator Kirsti Fong and the
Texas State Fire Marshall’s Office NFIRS administrator Virginia Garza for their help in
querying the data that facilitated this analysis.
74 Journal of Fire Protection Engineering 23(1)

Funding
This research received funding from the American Chemistry Council.

References
1. The National Commission on Fire Prevention and Control. America burning. May 1973.
Washington, DC: The National Commission on Fire Prevention and Control.
2. Karter MJ Jr. Fire loss in the United States during 2010. Report no. FLX10, September
2011. Quincy, MA: NFPA.
3. Krause U, Grosshandler W and Gritzo L. The international FORUM of fire research
directors: a position paper on sustainability and fire safety. Fire Saf J 2012; Vol. 49:
79–81.
4. Richardson LR. What fire statistics tell us about our fire and building codes for housing
and small buildings and fire risk for occupants of those structures. Fire and Mater 2001;
Vol. 25(6): 255–271.
5. Hall JR. U.S. unintentional fire death rates by state. Quincy, MA: NFPA, 2011.
6. Hall JR. Fire in the U.S. and Japan. Quincy, MA: NFPA, 2003.
7. Hall JR. Fire in the U.S. and the United Kingdom. Quincy, MA: NFPA, 2005.
8. Hasofer AM and Thomas I. Analysis of fatalities and injuries in building fire statistics.
Fire Saf J 2006; Vol. 41(1): 2–14.
9. US Fire Administration National Fire Data Center. Fire in the United States, 15th edn.
Emmitsburg, MD: FEMA, 2009.
10. Hall JR and Harwood B. The national estimates approach to US fire statistics. Fire
Technol 1989; Vol. 25(2): 99–113.
11. US Fire Administration National Fire Data Center. National fire incident reporting
system: complete reference guide. Emmitsburg, MD: FEMA, 2010.
12. US Fire Administration National Fire Data Center. Confined structure fires.
Department of Homeland Security. February 2006, p.16.
13. US Fire Administration National Fire Data Center. Fire risk. Topical Fire Research
Series, 4(7), December 2004. Emmitsburg, MD: FEMA.

Appendix
Reference guide for NFIRS codes in the TXCA article
Area of origin
0 ¼ area of origin, other
1 ¼ means of egress
14 ¼ living room, den, common room, family room, etc.
20 ¼ function areas, other
21 ¼ bedrooms
23 ¼ dining rooms
24 ¼ cooking area, kitchen
40 ¼ storage areas, other
47 ¼ garage
70 ¼ exterior structural area
Anderson and Ezekoye 75

71 ¼ attics and void spaces within structure


UU ¼ undetermined

Item first Ignited


0 ¼ item first ignited, other
10 ¼ structural component, finish
20 ¼ furniture, utensils, other
21 ¼ upholstered sofa, chair, vehicle seats
22 ¼ non-upholstered chair, bench
30 ¼ soft goods, wearing apparel other
31 ¼ mattresses, pillows, bedding, blankets, sheets, comforters, and linens like towels
and tablecloth
40 ¼ adornment, recreational material, signs (think Christmas trees)
50 ¼ storage supplies
60 ¼ flammable liquids, piping, and filters
70 ¼ organic materials
81 ¼ electrical wire, cable insulation
UU ¼ undetermined

Heat source
0 ¼ heat source, other
10 ¼ operating equipment
13 ¼ electrical arcing
40 ¼ hot or smoldering object
50 ¼ explosives, fireworks
60 ¼ heat from open flame or other smoking materials
61 ¼ cigarettes, pipes, or cigars
64 ¼ matches, cigarette or cigar lighters, and candles
83 ¼ flying brand, ember, spark
UU ¼ undetermined

You might also like