Skip to content
Upcoming courses

Fire statistics: Difference between revisions

Tools
From Fire Safety Academy Wiki
← Older edit
VisualWikitext
Sanne (talk | contribs)
Bots
337 edits
Add diagrams fse-fire-statistics-v1.svg, fse-fire-statistics-v2.svg, fse-fire-statistics-v3.svg (level-1 visual plan)
Mark (talk | contribs)
Bots
284 edits
Revise against the FSE wiki writing guidelines: verified and corrected citations, labelled statistics, See also, links, style (revised by Claude Opus 5.5)
 
Line 1: Line 1:
'''Fire statistics''' are organised data about fires, their causes, development, consequences and the performance of protective measures. They can describe incident frequency, first items ignited, ignition factors, casualties, damage and system operation. Their usefulness depends on definitions, reporting coverage, exposure data and the question being asked.<ref name="Hall-5-65">Hall and Ahrens (2002), pp. 5-65–5-67.</ref><ref name="Hall-5-67">Hall and Ahrens (2002), pp. 5-67–5-69.</ref>
'''Fire statistics''' are numerical data about fires that have occurred, such as how often fires start, what they start in and what harm they cause, together with the summary values calculated from those data. In the SFPE Handbook, Hall defines a statistic as any item of numerical data, or a quantity such as a mean computed from such data. Statistical analysis then condenses large bodies of data into a few summary values from which useful conclusions can be drawn.<ref name="Hall-1-193">Hall (2002c), p. 1-193.</ref> For fires, these calculations start from event frequencies per unit of time, such as how often certain items catch fire or certain types of occupancy have fires.<ref name="HA-5-67">Hall and Ahrens (2002), p. 5-67.</ref>


Statistics support fire risk analysis, design, product evaluation and investigation. They do not replace those essential engineering analyses. A count of incidents cannot by itself establish probability, causation or effectiveness without appropriate exposure and contextual evidence.<ref name="Hall-5-65" /><ref name="Watts-5-2">Watts and Hall (2002b), pp. 5-1–5-3.</ref>
Fire statistics matter because fire risk analysis, performance-based design, product evaluation and the reconstruction of major fires all need data on how likely fires are and how severe they become.<ref name="HA-5-65">Hall and Ahrens (2002), p. 5-65.</ref><ref name="HA-5-66">Hall and Ahrens (2002), p. 5-66.</ref> They are also easy to misread. The national estimates for the United States, for example, cover only fires reported to municipal fire departments, and one term can mean different things in different countries.<ref name="HA-5-70">Hall and Ahrens (2002), p. 5-70.</ref> In fire safety engineering, the quantitative approach relies on measurements and statistical analyses, while the qualitative approach, based on expert judgement, is particularly useful where numerical data are scarce.<ref name="VR-1.4">Van Rossum (2027), §1.4.</ref>


Fire data come from incident reports, field observations, laboratory tests, inspections, simulations and expert judgement. Each source has different strengths and biases. A larger or more detailed dataset is not automatically representative of the population or problem under study.<ref name="Hall-5-70">Hall and Ahrens (2002), pp. 5-69–5-72.</ref>
== Kinds of fire data ==


== Purpose and main data types ==
Hall and Ahrens group the data needed for engineering analysis into a few generic sources: fire incident data and field observations, usage or exposure data, and laboratory test results.<ref name="HA-5-65" /> Incident data describe fires that have occurred. Field observations describe other events, such as faults in protection systems found during routine inspection and testing.<ref name="HA-5-65" /><ref name="HA-5-71">Hall and Ahrens (2002), p. 5-71.</ref> Exposure data describe what is at risk, for example the number of portable electric heaters in use or the floor area of offices. They supply the denominators for frequencies such as fires per year per unit.<ref name="HA-5-65" /> Laboratory tests measure fire size and ignition under planned conditions, but fires and equipment failures are such rare events that measuring their probabilities in the laboratory can be impractical.<ref name="HA-5-65" /><ref name="HA-5-67" />
 
Incident data describe fires that have occurred and their recorded characteristics. They can include occupancy, area of origin, equipment involved, first fuel, ignition factor, casualties and the presence of detection or suppression. The available detail depends on the reporting system and its coding.<ref name="Hall-5-67" />
[[File:fse-fire-statistics-v1.svg|thumb|upright=1.5|alt=Incidents, exposure, laboratory tests and field observations provide distinct evidence.|Fire statistics combine several data types. Based on Hall and Ahrens (2002), pp. 5-65–5-67.]]
[[File:fse-fire-statistics-v1.svg|thumb|upright=1.5|alt=Incidents, exposure, laboratory tests and field observations provide distinct evidence.|Fire statistics combine several data types. Based on Hall and Ahrens (2002), pp. 5-65–5-67.]]


Usage or exposure data describe the population at risk. Examples include the number of products in use, buildings of a type, occupants or floor area. These data provide the denominators needed to express incident frequencies per unit or per period.<ref name="Hall-5-65" />
In the United States, the National Fire Incident Reporting System (NFIRS) records the occupancy, the material first ignited, the equipment involved, the form of heat that caused ignition and the ignition factor. It also records whether automatic detection and suppression systems were present.<ref name="HA-5-67" /> Analysts further use the area of origin and the age, location and activity of victims.<ref name="HA-5-68">Hall and Ahrens (2002), p. 5-68.</ref> Fire incident commanders fill in the coded report after the fire. They can usually name the equipment that started a fire but are unlikely to know which part was involved, and brand and model are not coded.<ref name="HA-5-68" />
 
Laboratory data describe controlled fire behaviour and system response. They can vary fuel, ignition source, geometry or protection to examine particular mechanisms. Their relevance to field conditions must still be evaluated rather than assumed.<ref name="Hall-5-65" /><ref name="Hall-5-72">Hall and Ahrens (2002), pp. 5-71–5-73.</ref>
 
Field observations show how systems and buildings perform in use. They can reveal impairment, degradation, maintenance quality and actual activation. Those observations complement laboratory performance and incident counts.<ref name="Hall-5-72" />


[[Fire risk]] combines these sources to examine likelihood and consequence. The relevant dataset depends on whether the question concerns a product, building, protective system or population. A single source rarely answers every part of that assessment.<ref name="Hall-5-65" /><ref name="Watts-5-2" />
A new product has no fire history, so its likelihood of ignition may have to be estimated from laboratory tests against established products, and subjective judgement will likely be required.<ref name="HA-5-67" /> Every data collection also has a purpose of its own. Hall and Ahrens therefore advise asking why a collection was set up, who paid for it and whether its sponsors had an interest in a particular outcome.<ref name="HA-5-69">Hall and Ahrens (2002), p. 5-69.</ref>


A frequency is a count of occurrences within a defined period. A rate relates that count to an exposure measure such as building-years, product units or floor area. The distinction is necessary before comparing different populations or years.<ref name="Hall-5-65" /><ref name="Watts-5-159">Watts and Hall (2002a), p. 5-159.</ref>
== How fire statistics are collected ==


Probability requires knowledge of relevant opportunities, not just incidents. Watts and Hall illustrate this with fires per appliance use, which also requires the number of times the appliance is used. An annual fire count without use data cannot establish the probability per use.<ref name="Watts-5-159" />
NFIRS, administered by the US Fire Administration, is described as the largest and most detailed fire incident database in the world.<ref name="HA-5-67" /><ref name="HA-5-76">Hall and Ahrens (2002), p. 5-76.</ref> It captures roughly a million fires a year, but only an estimated half of the fires reported to US fire departments. Since 1980, communities in roughly 40 of the 50 states have taken part, although not all of them every year.<ref name="HA-5-67" /> To scale NFIRS up to national estimates, analysts use the annual survey of the National Fire Protection Association ([[NFPA]]), a stratified random sample of roughly one-tenth of all US fire departments. Separate scaling ratios are used for fires, civilian deaths, civilian injuries and direct property damage.<ref name="HA-5-67" />


Exposure can change even when the recorded incident count stays similar. More products, larger buildings or different occupancy can alter the underlying opportunities for fire. Interpretation should therefore include how the exposed population has changed.<ref name="Hall-5-70" />
For 1996, for example, the NFPA survey put civilian deaths in residential structure fires at 4,080, against 1,504 reported to NFIRS. This gave a scaling ratio of 2.71 for that measure and that year.<ref name="HA-5-67" /> Hall and Ahrens judge the sampling uncertainty of NFIRS negligible because of its size, unless a problem involves hundreds of fires a year or fewer. For the NFPA survey they put it at about 10% for deaths and much less for other measures.<ref name="HA-5-68" />
 
Severity measures also require suitable denominators. Deaths per fire, damage per fire and deaths per exposed person describe different quantities. They should not be exchanged when comparing fire consequences across datasets.<ref name="Hall-5-68">Hall and Ahrens (2002), pp. 5-67–5-69.</ref>
 
Individual and societal risk use different consequence presentations. A location-based individual probability and the frequency of multiple fatalities are not the same measure. Statistics must be aggregated at the level appropriate to the question.<ref name="Barry-5-210">Barry (2002), pp. 5-210–5-211.</ref>
 
== Incident reporting and national estimates ==
 
The NFPA chapter uses the US National Fire Incident Reporting System, NFIRS, as a detailed example. It describes data on occupancy, first material ignited, equipment, heat of ignition, ignition factor and protective systems. The discussion is about that US system, not a universal reporting structure.<ref name="Hall-5-67" />
 
Hall and Ahrens explain that NFIRS covered an estimated half of fires reported to US fire departments during the period discussed. Participation varied across states and communities, and not every participant submitted data every year. These facts are specific to the handbook's 2002 account.<ref name="Hall-5-67" />
[[File:fse-fire-statistics-v2.svg|thumb|upright=1.5|alt=NFIRS records and the NFPA survey inform a projection with uncertainty.|National estimates require attention to sample coverage and scaling. Based on Hall and Ahrens (2002), pp. 5-67–5-69.]]
[[File:fse-fire-statistics-v2.svg|thumb|upright=1.5|alt=NFIRS records and the NFPA survey inform a projection with uncertainty.|National estimates require attention to sample coverage and scaling. Based on Hall and Ahrens (2002), pp. 5-67–5-69.]]


NFIRS is a sample whose complete reporting universe cannot be inferred from its own records alone. The NFPA survey provides another data source used to project national estimates. Different severity measures can require different scaling ratios because small and severe fires are represented differently.<ref name="Hall-5-68" />
Statistics of this kind describe reported fires only. The US national estimates exclude fires handled by private brigades, federal fire-fighting organisations or the occupants themselves. Fires in buildings with sprinklers or good detection are more likely to be controlled on site, and no system based on fire service reports can capture fires that are never reported.<ref name="HA-5-70" /> Other countries keep their own records. The United Kingdom has asked fire questions in the British Crime Survey about all fires, not only reported ones, and it records the time before discovery, which NFIRS does not.<ref name="HA-5-76" />


The chapter gives a historical 1996 example of 4,080 civilian deaths in residential structure fires in the NFPA survey compared with 1,504 in NFIRS. The resulting scaling ratio of 2.71 belongs to that measure and year. It must not be reused as a current multiplier.<ref name="Hall-5-68" />
Collections of notable fires are a different kind of source. The NFPA Fire Analysis and Research Division keeps a database of significant incidents found through press clippings. It can serve as a representative, even complete, record of large fires, but for other fires it is only anecdotal.<ref name="HA-5-70" /> Such accounts, like NFPA's reports of [[fire investigation]]s, show the range of fires that can happen. Hall and Ahrens stress, however, that anecdotes cannot be used to estimate incident rates or probabilities.<ref name="HA-5-70" />


Some fires never enter municipal records. Occupants or private brigades may handle small incidents without reporting them, and occupancies with those resources can be underrepresented. Statistics therefore describe reported fires rather than every fire that occurs.<ref name="Hall-5-70" />
== From counts to frequencies and probabilities ==


== Ignition sources and fire causes ==
A count of fires says little until it is related to how much of something is exposed. In a hypothetical example by Meacham, 20 fires a year in toasters as they are switched on is a frequency, not a probability. The probability per switch-on also needs the number of switch-ons, and with 20 million a year it would be one in a million.<ref name="Meacham-5-159">Meacham (2002), p. 5-159.</ref> Exposure can be measured as products in use, establishments, people or floor area. The main challenge is to find exposure data that are statistically representative and structured in the same way as the incident data.<ref name="HA-5-72">Hall and Ahrens (2002), p. 5-72.</ref>


[[Ignition source]] classification requires attention to several fields. The NFIRS discussion separates the form of heat, equipment involved, first material ignited, ignition factor and area of origin. A single equipment label does not capture every part of that sequence.<ref name="Hall-5-67" /><ref name="Hall-5-68" />
A national fire death rate is built in the same way. Notarianni describes the US residential fire death rate as the national number of residential fire deaths divided by the number of occupied residential units. Approximately 3,500 people die each year in US residential fires, but the rate varies with house type, region and community size.<ref name="Notarianni-5-61">Notarianni (2002), p. 5-61.</ref> Data on national practices can also mislead, because they are not weighted by the likelihood of a fire. In 1989 a randomly selected US home had a better than 85% chance of having a smoke detector, but among homes that had fires the chance was below 50%.<ref name="Hall-5-151">Hall (2002b), p. 5-151.</ref>


Coding systems can group products with different properties or uses. The chapter gives an example of one equipment code covering irons, hair dryers and electric blankets. Those devices may be used in different places and ignite different nearby items.<ref name="Hall-5-69">Hall and Ahrens (2002), pp. 5-68–5-70.</ref>
Tables of fire starts per occupancy follow the same idea, and the International Fire Engineering Guidelines (2005 edition) reproduce values from the UK code of practice BS7974, published by BSI in 2001. These give 0.003 fire starts per year for a dwelling and 0.3 for a hospital. That is roughly 3 starts a year for every 1,000 dwellings against 3 for every 10 hospitals.<ref name="BSI-2001">BSI (2001), as cited in National Research Council of Canada et al. (2005), p. 3.2-2.</ref> The guidelines prefer starts per unit of floor area where such data exist, and they caution that the probability of fire starts may have changed with new technologies and materials.<ref name="IFEG-3.2-2">National Research Council of Canada et al. (2005), p. 3.2-2.</ref> Finnish data published in 1998 show why floor area matters. The number of fires per square metre falls as buildings get larger, while the chance of a fire somewhere in the whole building generally rises with floor area.<ref name="Finland-1998">Rahikainen and Keski-Rahkonen (1998), as cited in National Research Council of Canada et al. (2005), p. 3.2-3.</ref>


First-fuel coding can also be ambiguous. A record may identify gasoline rather than the floor covering on which it was spilled, while broad material categories can include several products. Interpretation must consider what the code actually represents.<ref name="Hall-5-69" />
== What fire statistics show ==


Excluding awkward incidents to make a product or design appear safer is unsound. Hall and Ahrens argue that those cases belong to the overall fire problem and should be addressed when evaluating probabilities or acceptability. Selective exclusion would bias the initial estimate.<ref name="Hall-5-70" />
Casualty statistics show where and how people die in fires. Most fatal fire victims die at home. People die in fires because, for example, they are trapped by fire or smoke, are overcome in their sleep or inhale too much toxic smoke.<ref name="VR-7">Van Rossum (2027), §7.</ref> Citing studies published in 1981 and 1989, Gottuk and Lattimer report that over half of all fire fatalities have been attributed to carbon monoxide inhalation.<ref name="GL-CO">Anderson et al. (1981) and Harwood and Hall (1989), as cited in Gottuk and Lattimer (2002), p. 2-54.</ref> Citing a 1994 study, they add that nearly 75% of deaths from [[smoke inhalation]] occur away from the room where the fire started.<ref name="GL-remote">Gann et al. (1994), as cited in Gottuk and Lattimer (2002), p. 2-54.</ref>


[[Fire investigation]] can provide more detail about a particular incident than routine coding allows. That detail can clarify mechanisms and test the meaning of database categories. A detailed case is still not automatically representative of every fire in that category.<ref name="Hall-5-70" />
Statistics also have blind spots. Purser notes that fire statistics cannot show how often smouldering fires occur, because many of them become flaming fires before they are discovered.<ref name="Purser-2-134">Purser (2002), p. 2-134.</ref> UK fire statistics for 1983 nevertheless offer an indication. For victims aged 19 to 49, deaths were 20 times more likely in fires found 30 minutes or more after ignition than in fires found within 5 minutes. Purser considers it likely that such long-burning fires involved smouldering.<ref name="HO-1983">Home Office (1983), as cited in Purser (2002), p. 2-134.</ref>


== Smoke, injury and consequence data ==
Most analyses need the number of fires of a certain type, usually defined by cause, not only the total. Hall and Ahrens list six cause-related characteristics in common use. Two describe the source of heat, or [[ignition source]]: the form of heat of ignition and the equipment involved. Two describe the form and type of material first ignited, and the last two are the ignition factor that brought heat and fuel together and the area of origin.<ref name="HA-5-68" /> The International Fire Engineering Guidelines (2005 edition) note that no quantitative methods exist for predicting the potential for ignition. Fire statistics can instead supply ignition frequencies and the nature of the materials ignited.<ref name="IFEG-2.4-4">National Research Council of Canada et al. (2005), p. 2.4-4.</ref>


[[Smoke inhalation]] and other injury outcomes need clear case definitions and exposure conditions. NFIRS records some victim characteristics, locations and activities, but the coding does not provide every physiological or toxicological detail. Consequence analysis may require other sources and models.<ref name="Hall-5-68" /><ref name="Watts-5-155">Watts and Hall (2002a), pp. 5-154–5-156.</ref>
Statistics also measure the cost of fire. For more than two decades, the World Fire Statistics Centre has compared the total cost of fire in 15 to 20 countries, with losses also shown as percentages of gross domestic product (GDP). In the 2002 SFPE Handbook, the core costs for the United States typically ran at about 1% of GDP. The US fire death rate relative to population consistently ranked among the highest of the countries studied.<ref name="Ram-5-79">Ramachandran (2002b), pp. 5-79–5-80.</ref>


Laboratory toxic effects and fire-effluent data can supplement incident records. They can describe species concentrations, exposure conditions and effects that are not routinely measured at a fire scene. Their relevance to real occupants must be assessed carefully.<ref name="Hall-5-65" /><ref name="Hall-5-72" />
== Use in fire safety engineering ==


Victim location and activity can help examine exposure to smoke and heat. The number of casualties alone does not identify the route or conditions that caused them. Consequence analysis should therefore connect outcome data with the developing environment and occupant actions.<ref name="Hall-5-68" /><ref name="Proulx-3-344">Proulx (2002), pp. 3-344–3-345.</ref>
[[Fire risk]] analysis seeks to answer three questions: what could happen, how bad it would be and how likely it is.<ref name="WH-5-2">Watts and Hall (2002), p. 5-2.</ref> Some assessments look back, inferring probabilities from fire loss experience after an event such as the introduction of home smoke alarms. Others look ahead, for example to the effect of requiring sprinklers in homes.<ref name="WH-5-2" /> The available data differ widely. US homes offer very many documented fires with limited detail, US nuclear power plants far fewer fires with much more detail, and a single building very few fires at all.<ref name="WH-5-3">Watts and Hall (2002), p. 5-3.</ref> In a hazard assessment, historical loss data help engineers avoid focusing on unique hazards while overlooking more common but equally important ones.<ref name="Meacham-5-155">Meacham (2002), p. 5-155.</ref>


Property-loss data can also be influenced by reporting coverage. Small-loss fires may be underrepresented, which can raise the apparent average loss per recorded fire. Separate scaling and careful review are needed for different consequence measures.<ref name="Hall-5-68" />
In Australia, the data sheets that accompany the Fire Safety Verification Method (print version 1.1, 2020) turn incident statistics into characteristic design values, whose use is not mandatory.<ref name="ABCB-B2-1">Australian Building Codes Board (2020a), Data Sheet B2, p. 1.</ref> Their fire growth rates come from an analysis by Holborn and colleagues of fire investigation data from the Greater London area for 1996 to 2000. Of 481 fires in residential buildings, 34% were classed as very slow growing, 50% as slow and 11% as medium.<ref name="Holborn-2004">Holborn et al. (2004), as cited in Australian Building Codes Board (2020a), Data Sheet B2, pp. 4–7.</ref>


Fatalities and injuries should not be treated as interchangeable counts. They have different consequences, coding and uncertainty. The analysis should state which outcome is used and how unknown or missing records are handled.<ref name="Hall-5-68" /><ref name="Hall-5-69" />
The same data sheet estimates the share of fires that would become fully developed if no sprinklers were present. For dwellings it adopts about 18%, based on values of 15.5% for Australia, 18.3% for the USA and Canada and 23.2% for New Zealand. Hotels, hospitals and care homes are given 5% and most other buildings 15%.<ref name="ABCB-B2-6">Australian Building Codes Board (2020a), Data Sheet B2, pp. 6–9.</ref> The data sheet explains the low value by the 24-hour staff and monitored detection that such buildings are likely to have. These bring a higher proportion of their fires to the attention of the fire brigade.<ref name="ABCB-B2-6" /> The accompanying handbook predicts the probability of flashover from the frequency of fires that occupants were unable to manage, taken from fire statistics, together with sprinkler reliability and an analysis of fire brigade intervention.<ref name="ABCB-H-92">Australian Building Codes Board (2020b), p. 92.</ref>


== Reliability and effectiveness ==
In New Zealand, the commentary on the first edition of Verification Method C/VM2 (2013) cites New Zealand Fire Service Emergency Incident Statistics for 2005 to 2010. In these, approximately 3% of all structure fires spread to adjacent property, a risk the commentary calls small but not insignificant.<ref name="MBIE-23">Ministry of Business, Innovation and Employment (2013), p. 23.</ref> In the Netherlands, Van Rossum uses actual outcomes to show that the building regulations (Bbl) were not developed around one quantified goal. Per hour of stay, the chance of dying in a fire is much greater in a residential building than in an assembly building, yet the detection requirements are stricter for assembly buildings.<ref name="VR-1.2">Van Rossum (2027), §1.2.</ref>


[[Reliability of fire protection systems]] concerns whether systems perform when required. Reliability data can come from incident outcomes, inspection, testing, maintenance records and field observations. Each captures a different part of operational performance.<ref name="Hall-5-72" />
== Reliability of protection systems ==


Inspection programmes can reveal failures that do not draw attention, but they may miss transient impairments. They can also overrepresent properties with advanced maintenance and trained staff. Reliability from those programmes may not apply to typical buildings.<ref name="Hall-5-71">Hall and Ahrens (2002), pp. 5-71–5-72.</ref>
Incident data are one of five kinds of source that Hall and Ahrens identify for the [[reliability of fire protection systems]], for which data are very scarce.<ref name="HA-5-70" /><ref name="HA-5-71" /> The Australian data sheets define reliability as the probability that a system performs to its specification. Efficacy is the degree to which the system then achieves its design objective, and effectiveness combines the two.<ref name="ABCB-C3-87">Australian Building Codes Board (2020a), Data Sheet C3, p. 87.</ref> For smoke alarms, reliability counts only fires large and close enough to activate an alarm, and effectiveness is efficacy multiplied by reliability.<ref name="ABCB-C3-91">Australian Building Codes Board (2020a), Data Sheet C3, p. 91.</ref> A battery-operated alarm system with limited coverage, a typical reliability of 85% and an efficacy of 93% thus has a typical effectiveness of 79%.<ref name="ABCB-C3-92">Australian Building Codes Board (2020a), Data Sheet C3, p. 92.</ref>


Incident-based effectiveness requires a clear outcome definition. ABCB distinguishes smoke-alarm efficacy, meaning achievement of a specified objective when the system performs to specification, from reliability under conditions expected to cause activation. Those are different measures.<ref name="ABCB-89">Australian Building Codes Board (2020), pp. 88–90.</ref>
UK Home Office statistics for dwelling fires from 2010/11 to 2017/18, analysed in the same data sheet, show how much definitions matter. Mains-powered smoke alarms failed to operate in 21% of the fires where they were present. Fires whose smoke never reached a detector, or that started outside the covered area, do not count as failures; excluding them left about 4%, a reliability of 96%. For battery-powered alarms the same steps turned 39% into 13%, a reliability of 87%.<ref name="HO-2018">Home Office (2018), as cited in Australian Building Codes Board (2020a), Data Sheet C3, pp. 105–109.</ref>


ABCB also warns that its general estimate of waking effectiveness spans all times of day and residential occupancies. It should not be interpreted as the ability to wake every occupant type. Age, alcohol or drugs, background noise and alarm characteristics can alter response.<ref name="ABCB-90">Australian Building Codes Board (2020), pp. 89–90.</ref>
For US home fires from 2009 to 2013, an NFPA analysis by Ahrens (2015) found that battery-operated alarms operated in 80% of fires large enough to activate them. Hard-wired alarms with battery backup operated in 95%.<ref name="Ahrens-2015">Ahrens (2015), as cited in Australian Building Codes Board (2020a), Data Sheet C3, pp. 109–111.</ref> Measured as alerting the occupants once an alarm operates, efficacy in the US data was about 97%. Measured as the reduction in deaths, US and UK data gave only 50% to 55%, so the data sheet stresses that the definition must be stated whenever an efficacy is quoted.<ref name="ABCB-C3-111">Australian Building Codes Board (2020a), Data Sheet C3, p. 111.</ref>


An operating system does not guarantee the desired human outcome. Detection, audible alerting, occupant response and reduced injury are successive possible criteria. Statistics used for one should not be presented as evidence for all of them.<ref name="ABCB-89" /><ref name="Proulx-3-350">Proulx (2002), p. 3-350.</ref>
Reported fires can make protection look less reliable than it is, because successful systems often keep fires too small to be reported. In the United States, reported home fires consistently show roughly one-third of smoke alarms not operating, against one-fifth in a study of home installations in general.<ref name="HA-5-72" /> For sprinklers, reported fires suggest failure to operate in one-sixth to one-fifth of the fires large enough to activate them, while real reliability could easily be well over 90%.<ref name="HA-5-72" /> Inspection and testing data lean the other way, because they tend to come from properties with very advanced maintenance programmes. This is one reason why estimates of sprinkler reliability in Australia are much higher than in the United States.<ref name="HA-5-71" /> In Australia, Marryatt's 1988 compilation found control in over 99% of fires in sprinklered buildings, but it left out fires in which the sprinklers failed to operate or had no water.<ref name="Marryatt-1988">Marryatt (1988), as cited in National Research Council of Canada et al. (2005), pp. 3.7-2–3.7-3.</ref>


== Unknown data and reporting bias ==
== Limits and sources of bias ==


Unknown or blank fields can materially change category patterns. If the proportion of unknown causes changes over time, ignoring those records can make every known cause appear to increase. Analysis must address the unknown records rather than treating them as absent events.<ref name="Hall-5-68" />
Missing entries can distort patterns. If the number of fires stayed the same while the share with an unreported cause steadily fell, ignoring the unknowns would make every specific cause appear to increase.<ref name="HA-5-68" /> Most national estimates therefore allocate unknown entries in the same proportions as the known ones. This assumes that the unknown fires would show the same pattern if their details were known.<ref name="HA-5-68" /> The Australian data sheets apply the same method, for example to unknown fire sizes in New Zealand data and to unknown reasons why smoke alarms did not operate.<ref name="ABCB-B2-6" /><ref name="ABCB-C3-106">Australian Building Codes Board (2020a), Data Sheet C3, pp. 106, 108.</ref>
[[File:fse-fire-statistics-v3.svg|thumb|upright=1.5|alt=Known and unknown incident fields lead to different apparent category shares and uncertainty.|Unknown data and reporting bias must be addressed explicitly. Based on Hall and Ahrens (2002), pp. 5-67–5-70.]]
[[File:fse-fire-statistics-v3.svg|thumb|upright=1.5|alt=Known and unknown incident fields lead to different apparent category shares and uncertainty.|Unknown data and reporting bias must be addressed explicitly. Based on Hall and Ahrens (2002), pp. 5-67–5-70.]]


Proportional allocation assigns unknown records in the same shares as known records. That provides a transparent method but assumes that unknown and known cases have the same distribution. The assumption may not hold, so its uncertainty should be examined.<ref name="Hall-5-68" /><ref name="Hall-5-69" />
Codes can also lump different things together. One NFIRS equipment code covers irons, hair dryers and electric blankets, which are used in very different places and tend to ignite different items.<ref name="HA-5-69" /> The floor covering code includes carpet, tile, rugs, hardwood, vinyl flooring and even stairs.<ref name="HA-5-68" /> One of the materials most often reported as first ignited in floor covering fires is petrol, which refers to a liquid spilled on the floor rather than to the covering itself. Hall and Ahrens consider it unsound to exclude such fires as unfair tests, because all fires involving a product belong to its fire problem.<ref name="HA-5-69" />
 
Coding inconsistencies and factual errors create additional uncertainty. A second source can help in some cases, but it is not necessarily more reliable and is often unavailable. Data cleaning must not erase unusual but genuine fires.<ref name="Hall-5-69" />
 
Opportunity samples arise when cases are collected because they are easy to obtain or newsworthy. They can overrepresent severe, urban or unusual events. Such data can show what is possible without establishing how often it occurs.<ref name="Hall-5-70" />
 
The purpose of data collection also matters. Hall and Ahrens recommend examining why the system was created, who funded it and which questions it was designed to answer. Those factors can affect definitions, coverage and availability.<ref name="Hall-5-70" />
 
Laboratory and simulation studies can investigate conditions that are rare or unsafe in real fires. They allow controlled variation but may simplify geometry, fuel, ventilation or occupant behaviour. Their representativeness must be assessed before applying results to a building.<ref name="Hall-5-72" />
 
Evacuation drills are a form of simulation. They can reveal human response and route performance but may avoid dangerous conditions or include trained participants. Drill data should not be treated as exact predictions of every actual emergency.<ref name="Hall-5-72" /><ref name="Proulx-3-349">Proulx (2002), pp. 3-348–3-350.</ref>
 
Expert judgement is used where systematic observations are limited. Experience can supply useful estimates, but recollection and halo or horns effects can bias assessments towards higher or lower reliability. Group methods may reduce some bias without eliminating it.<ref name="Hall-5-71" />
 
New products create a particular problem because they lack incident history. Laboratory comparisons may be needed, but rare ignition makes direct probability measurement difficult. Subjective estimates and explicit assumptions can remain part of the analysis.<ref name="Hall-5-67" />
 
Fire models can supplement missing observations by representing physical processes. Their outputs still depend on scenario selection and input data. A model cannot create representative fire statistics merely by producing a detailed result.<ref name="Notarianni-5-43">Notarianni (2002), pp. 5-43–5-44.</ref><ref name="Hall-5-72" />
 
== Organisational sources and roles ==
 
[[NFPA]] publishes fire-analysis research and statistical resources in addition to codes and standards. Hall and Ahrens are identified as NFPA fire-analysis specialists in the 2002 handbook. That role is separate from adopting NFPA documents as legal requirements.<ref name="Hall-5-65" /><ref name="SFPE-iii">DiNenno (2002), pp. iii–iv.</ref>
 
NFIRS and the NFPA survey serve different functions in the cited US method. NFIRS supplies detailed incident records, while the NFPA survey provides a structured basis for national projection. Their strengths are used together rather than treated as interchangeable datasets.<ref name="Hall-5-67" /><ref name="Hall-5-68" />
 
Other databases may contain detailed fire-investigation or insurance records, but access and representativeness can be limited. Proprietary data may not be available, and older records may not exist in accessible digital form. Data availability must be confirmed before planning an analysis.<ref name="Hall-5-70" />
 
Statistical definitions can vary between countries and collection systems. Similar terms can cover different equipment or events, making direct comparison unsafe. Definitions and coding guidance should be reviewed before merging sources.<ref name="Hall-5-70" />
 
[[Fire risk assessment]] must document these choices of source, definition and scope. The report should make clear which figures are observed, projected, modelled or judged. This allows a reviewer to distinguish statistical evidence from the assumptions used to interpret it.<ref name="Hall-5-65" /><ref name="Watts-5-160">Watts and Hall (2002a), p. 5-160.</ref>
 
== Use and review of statistics ==
 
A useful analysis starts with the decision and the population to which it applies. It then identifies the required incident, exposure and consequence data. Collecting a convenient dataset first can lead to detailed but irrelevant results.<ref name="Hall-5-65" /><ref name="Hall-5-70" />
 
Each statistic should retain its source year and collection conditions. The NFIRS examples in the 2002 handbook are not current national estimates. Likewise, an effectiveness value from a particular occupancy or study should not be treated as universal.<ref name="Hall-5-68" /><ref name="ABCB-90" />


Sensitivity checks can show how unknown allocation, sampling assumptions and model inputs affect the result. Where those changes alter a decision, the uncertainty should be reported rather than hidden by one preferred estimate. This is part of responsible statistical interpretation.<ref name="Hall-5-69" /><ref name="Notarianni-5-43" />
Definitions also differ between countries and data systems. In fire incident databases a space heater heats a relatively small area, whereas the US Department of Energy uses the term for all heating systems.<ref name="HA-5-70" /> Ramachandran notes that a "room" in fire brigade reports is not necessarily a fire compartment. Fires recorded as spreading beyond the room include fires that passed through an open door or window. Such counts therefore do not give a valid estimate of how often the walls, floors and ceilings of a room fail.<ref name="Ram-3-388">Ramachandran (2002c), p. 3-388.</ref>


Data should be revisited when the building or population changes. Altered fuel, occupancy, detection or maintenance can make old evidence less relevant. A previous statistical conclusion is not a permanent property of the system.<ref name="IFEG-0-3-8">Australian Building Codes Board et al. (2005), p. 0.3-8.</ref><ref name="Hall-5-72" />
Past data may not suit the future. Meacham asks whether fire loss data for cellulosic materials before 1960 could indicate losses involving synthetic materials after 1960.<ref name="Meacham-5-154">Meacham (2002), p. 5-154.</ref> Because fires are rare and people change buildings and their contents over time, Meacham calls predicting future risks from past loss experience challenging at best.<ref name="Meacham-5-159" /> The International Fire Engineering Guidelines (2005 edition) likewise warn that data may lose relevance through new technologies, materials, regulations and construction practices.<ref name="IFEG-3.1-2">National Research Council of Canada et al. (2005), p. 3.1-2.</ref> Data sometimes have to be borrowed from abroad: the Australian data sheets quote a Senate committee finding that Australian fire statistics are hard to locate and compare. For smoke alarms they therefore draw mainly on data from the US, the UK and New Zealand.<ref name="ABCB-C3-91" />


The aim is not to find a single impressive number. It is to assemble representative evidence that answers a defined question while exposing uncertainty. Good fire statistics make both the observed patterns and their limitations available for engineering decisions.<ref name="Hall-5-65" /><ref name="Hall-5-70" />
Fire losses also vary at random from year to year, since several large-loss fires may occur in one year and few or none the next. A US cost-benefit study of residential sprinklers described by Notarianni therefore used mean values for the five years from 1989 to 1993.<ref name="Notarianni-5-61" /><ref name="Notarianni-5-62">Notarianni (2002), p. 5-62.</ref> Where systematic data are lacking, estimates from experienced people can be used. Their recollections, however, tend to make reliable systems seem even more reliable and unreliable ones even less reliable.<ref name="HA-5-71" />


== See also ==
== See also ==


* [[Fire investigation]]
* [[Fire risk assessment]]
* [[Fire risk]]
* [[Quantitative risk assessment]]
* [[Ignition source]]
* [[Event tree analysis]]
* [[Uncertainty analysis]]
* [[Societal risk]]


== Notes ==
== Notes ==
Line 136: Line 88:
== References ==
== References ==


* Australian Building Codes Board. (2020). ''Fire safety verification method data sheets'' (Version 1.1). Author.
* Australian Building Codes Board. (2020a). ''Data sheets: Fire Safety Verification Method'' (Handbook annex, Print version 1.1).
* Australian Building Codes Board, National Research Council of Canada, International Code Council, & Department of Building and Housing New Zealand. (2005). ''International fire engineering guidelines'' (2005 ed.). Australian Building Codes Board.
* Australian Building Codes Board. (2020b). ''Handbook: Fire Safety Verification Method'' (Print version 1.1).
* Barry, T. F. (2002). Quantitative risk assessment in chemical process industries. In P. J. DiNenno (Ed.), ''SFPE handbook of fire protection engineering'' (3rd ed.). National Fire Protection Association.
* Gottuk, D. T., & Lattimer, B. Y. (2002). Effect of combustion conditions on species production. In P. J. DiNenno, D. Drysdale, C. L. Beyler, W. D. Walton, R. L. P. Custer, J. R. Hall Jr., & J. M. Watts Jr. (Eds.), ''SFPE handbook of fire protection engineering'' (3rd ed., pp. 2-54–2-82). National Fire Protection Association; Society of Fire Protection Engineers.
* DiNenno, P. J. (Ed.). (2002). ''SFPE handbook of fire protection engineering'' (3rd ed.). National Fire Protection Association.
* Hall, J. R., Jr. (2002b). Product fire risk. In P. J. DiNenno, D. Drysdale, C. L. Beyler, W. D. Walton, R. L. P. Custer, J. R. Hall Jr., & J. M. Watts Jr. (Eds.), ''SFPE handbook of fire protection engineering'' (3rd ed., pp. 5-143–5-152). National Fire Protection Association; Society of Fire Protection Engineers.
* Hall, J. R., Jr., & Ahrens, M. J. (2002). Data for engineering analysis. In P. J. DiNenno (Ed.), ''SFPE handbook of fire protection engineering'' (3rd ed.). National Fire Protection Association.
* Hall, J. R., Jr. (2002c). Statistics. In P. J. DiNenno, D. Drysdale, C. L. Beyler, W. D. Walton, R. L. P. Custer, J. R. Hall Jr., & J. M. Watts Jr. (Eds.), ''SFPE handbook of fire protection engineering'' (3rd ed., pp. 1-193–1-202). National Fire Protection Association; Society of Fire Protection Engineers.
* Notarianni, K. A. (2002). Uncertainty. In P. J. DiNenno (Ed.), ''SFPE handbook of fire protection engineering'' (3rd ed.). National Fire Protection Association.
* Hall, J. R., Jr., & Ahrens, M. J. (2002). Data for engineering analysis. In P. J. DiNenno, D. Drysdale, C. L. Beyler, W. D. Walton, R. L. P. Custer, J. R. Hall Jr., & J. M. Watts Jr. (Eds.), ''SFPE handbook of fire protection engineering'' (3rd ed., pp. 5-65–5-78). National Fire Protection Association; Society of Fire Protection Engineers.
* Proulx, G. (2002). Movement of people: The evacuation timing. In P. J. DiNenno (Ed.), ''SFPE handbook of fire protection engineering'' (3rd ed.). National Fire Protection Association.
* Meacham, B. J. (2002). Building fire risk analysis. In P. J. DiNenno, D. Drysdale, C. L. Beyler, W. D. Walton, R. L. P. Custer, J. R. Hall Jr., & J. M. Watts Jr. (Eds.), ''SFPE handbook of fire protection engineering'' (3rd ed., pp. 5-153–5-175). National Fire Protection Association; Society of Fire Protection Engineers.
* Watts, J. M., Jr., & Hall, J. R., Jr. (2002a). Building fire risk analysis. In P. J. DiNenno (Ed.), ''SFPE handbook of fire protection engineering'' (3rd ed.). National Fire Protection Association.
* Ministry of Business, Innovation and Employment. (2013). ''Commentary for Building Code clauses C1–C6 and Verification Method C/VM2''.
* Watts, J. M., Jr., & Hall, J. R., Jr. (2002b). Introduction to fire risk analysis. In P. J. DiNenno (Ed.), ''SFPE handbook of fire protection engineering'' (3rd ed.). National Fire Protection Association.
* National Research Council of Canada, International Code Council, Department of Building and Housing, & Australian Building Codes Board. (2005). ''International fire engineering guidelines'' (2005 ed.). Australian Building Codes Board.
* Notarianni, K. A. (2002). Uncertainty. In P. J. DiNenno, D. Drysdale, C. L. Beyler, W. D. Walton, R. L. P. Custer, J. R. Hall Jr., & J. M. Watts Jr. (Eds.), ''SFPE handbook of fire protection engineering'' (3rd ed., pp. 5-40–5-64). National Fire Protection Association; Society of Fire Protection Engineers.
* Purser, D. A. (2002). Toxicity assessment of combustion products. In P. J. DiNenno, D. Drysdale, C. L. Beyler, W. D. Walton, R. L. P. Custer, J. R. Hall Jr., & J. M. Watts Jr. (Eds.), ''SFPE handbook of fire protection engineering'' (3rd ed., pp. 2-83–2-171). National Fire Protection Association; Society of Fire Protection Engineers.
* Ramachandran, G. (2002b). Measuring fire consequences in economic terms (J. R. Hall Jr., Rev.). In P. J. DiNenno, D. Drysdale, C. L. Beyler, W. D. Walton, R. L. P. Custer, J. R. Hall Jr., & J. M. Watts Jr. (Eds.), ''SFPE handbook of fire protection engineering'' (3rd ed., pp. 5-79–5-92). National Fire Protection Association; Society of Fire Protection Engineers.
* Ramachandran, G. (2002c). Stochastic models of fire growth. In P. J. DiNenno, D. Drysdale, C. L. Beyler, W. D. Walton, R. L. P. Custer, J. R. Hall Jr., & J. M. Watts Jr. (Eds.), ''SFPE handbook of fire protection engineering'' (3rd ed., pp. 3-381–3-401). National Fire Protection Association; Society of Fire Protection Engineers.
* Van Rossum, E. (2027). ''Introduction to Fire Safety Engineering''. Fire Safety Academy.
* Watts, J. M., Jr., & Hall, J. R., Jr. (2002). Introduction to fire risk analysis. In P. J. DiNenno, D. Drysdale, C. L. Beyler, W. D. Walton, R. L. P. Custer, J. R. Hall Jr., & J. M. Watts Jr. (Eds.), ''SFPE handbook of fire protection engineering'' (3rd ed., pp. 5-1–5-7). National Fire Protection Association; Society of Fire Protection Engineers.


[[Category:Fire Risk Analysis]]
[[Category:Fire Risk Analysis]]

Latest revision as of 18:53, 2 November 2026

Fire statistics are numerical data about fires that have occurred, such as how often fires start, what they start in and what harm they cause, together with the summary values calculated from those data. In the SFPE Handbook, Hall defines a statistic as any item of numerical data, or a quantity such as a mean computed from such data. Statistical analysis then condenses large bodies of data into a few summary values from which useful conclusions can be drawn.[1] For fires, these calculations start from event frequencies per unit of time, such as how often certain items catch fire or certain types of occupancy have fires.[2]

Fire statistics matter because fire risk analysis, performance-based design, product evaluation and the reconstruction of major fires all need data on how likely fires are and how severe they become.[3][4] They are also easy to misread. The national estimates for the United States, for example, cover only fires reported to municipal fire departments, and one term can mean different things in different countries.[5] In fire safety engineering, the quantitative approach relies on measurements and statistical analyses, while the qualitative approach, based on expert judgement, is particularly useful where numerical data are scarce.[6]

Kinds of fire data

Hall and Ahrens group the data needed for engineering analysis into a few generic sources: fire incident data and field observations, usage or exposure data, and laboratory test results.[3] Incident data describe fires that have occurred. Field observations describe other events, such as faults in protection systems found during routine inspection and testing.[3][7] Exposure data describe what is at risk, for example the number of portable electric heaters in use or the floor area of offices. They supply the denominators for frequencies such as fires per year per unit.[3] Laboratory tests measure fire size and ignition under planned conditions, but fires and equipment failures are such rare events that measuring their probabilities in the laboratory can be impractical.[3][2]

Incidents, exposure, laboratory tests and field observations provide distinct evidence.
Fire statistics combine several data types. Based on Hall and Ahrens (2002), pp. 5-65–5-67.

In the United States, the National Fire Incident Reporting System (NFIRS) records the occupancy, the material first ignited, the equipment involved, the form of heat that caused ignition and the ignition factor. It also records whether automatic detection and suppression systems were present.[2] Analysts further use the area of origin and the age, location and activity of victims.[8] Fire incident commanders fill in the coded report after the fire. They can usually name the equipment that started a fire but are unlikely to know which part was involved, and brand and model are not coded.[8]

A new product has no fire history, so its likelihood of ignition may have to be estimated from laboratory tests against established products, and subjective judgement will likely be required.[2] Every data collection also has a purpose of its own. Hall and Ahrens therefore advise asking why a collection was set up, who paid for it and whether its sponsors had an interest in a particular outcome.[9]

How fire statistics are collected

NFIRS, administered by the US Fire Administration, is described as the largest and most detailed fire incident database in the world.[2][10] It captures roughly a million fires a year, but only an estimated half of the fires reported to US fire departments. Since 1980, communities in roughly 40 of the 50 states have taken part, although not all of them every year.[2] To scale NFIRS up to national estimates, analysts use the annual survey of the National Fire Protection Association (NFPA), a stratified random sample of roughly one-tenth of all US fire departments. Separate scaling ratios are used for fires, civilian deaths, civilian injuries and direct property damage.[2]

For 1996, for example, the NFPA survey put civilian deaths in residential structure fires at 4,080, against 1,504 reported to NFIRS. This gave a scaling ratio of 2.71 for that measure and that year.[2] Hall and Ahrens judge the sampling uncertainty of NFIRS negligible because of its size, unless a problem involves hundreds of fires a year or fewer. For the NFPA survey they put it at about 10% for deaths and much less for other measures.[8]

NFIRS records and the NFPA survey inform a projection with uncertainty.
National estimates require attention to sample coverage and scaling. Based on Hall and Ahrens (2002), pp. 5-67–5-69.

Statistics of this kind describe reported fires only. The US national estimates exclude fires handled by private brigades, federal fire-fighting organisations or the occupants themselves. Fires in buildings with sprinklers or good detection are more likely to be controlled on site, and no system based on fire service reports can capture fires that are never reported.[5] Other countries keep their own records. The United Kingdom has asked fire questions in the British Crime Survey about all fires, not only reported ones, and it records the time before discovery, which NFIRS does not.[10]

Collections of notable fires are a different kind of source. The NFPA Fire Analysis and Research Division keeps a database of significant incidents found through press clippings. It can serve as a representative, even complete, record of large fires, but for other fires it is only anecdotal.[5] Such accounts, like NFPA's reports of fire investigations, show the range of fires that can happen. Hall and Ahrens stress, however, that anecdotes cannot be used to estimate incident rates or probabilities.[5]

From counts to frequencies and probabilities

A count of fires says little until it is related to how much of something is exposed. In a hypothetical example by Meacham, 20 fires a year in toasters as they are switched on is a frequency, not a probability. The probability per switch-on also needs the number of switch-ons, and with 20 million a year it would be one in a million.[11] Exposure can be measured as products in use, establishments, people or floor area. The main challenge is to find exposure data that are statistically representative and structured in the same way as the incident data.[12]

A national fire death rate is built in the same way. Notarianni describes the US residential fire death rate as the national number of residential fire deaths divided by the number of occupied residential units. Approximately 3,500 people die each year in US residential fires, but the rate varies with house type, region and community size.[13] Data on national practices can also mislead, because they are not weighted by the likelihood of a fire. In 1989 a randomly selected US home had a better than 85% chance of having a smoke detector, but among homes that had fires the chance was below 50%.[14]

Tables of fire starts per occupancy follow the same idea, and the International Fire Engineering Guidelines (2005 edition) reproduce values from the UK code of practice BS7974, published by BSI in 2001. These give 0.003 fire starts per year for a dwelling and 0.3 for a hospital. That is roughly 3 starts a year for every 1,000 dwellings against 3 for every 10 hospitals.[15] The guidelines prefer starts per unit of floor area where such data exist, and they caution that the probability of fire starts may have changed with new technologies and materials.[16] Finnish data published in 1998 show why floor area matters. The number of fires per square metre falls as buildings get larger, while the chance of a fire somewhere in the whole building generally rises with floor area.[17]

What fire statistics show

Casualty statistics show where and how people die in fires. Most fatal fire victims die at home. People die in fires because, for example, they are trapped by fire or smoke, are overcome in their sleep or inhale too much toxic smoke.[18] Citing studies published in 1981 and 1989, Gottuk and Lattimer report that over half of all fire fatalities have been attributed to carbon monoxide inhalation.[19] Citing a 1994 study, they add that nearly 75% of deaths from smoke inhalation occur away from the room where the fire started.[20]

Statistics also have blind spots. Purser notes that fire statistics cannot show how often smouldering fires occur, because many of them become flaming fires before they are discovered.[21] UK fire statistics for 1983 nevertheless offer an indication. For victims aged 19 to 49, deaths were 20 times more likely in fires found 30 minutes or more after ignition than in fires found within 5 minutes. Purser considers it likely that such long-burning fires involved smouldering.[22]

Most analyses need the number of fires of a certain type, usually defined by cause, not only the total. Hall and Ahrens list six cause-related characteristics in common use. Two describe the source of heat, or ignition source: the form of heat of ignition and the equipment involved. Two describe the form and type of material first ignited, and the last two are the ignition factor that brought heat and fuel together and the area of origin.[8] The International Fire Engineering Guidelines (2005 edition) note that no quantitative methods exist for predicting the potential for ignition. Fire statistics can instead supply ignition frequencies and the nature of the materials ignited.[23]

Statistics also measure the cost of fire. For more than two decades, the World Fire Statistics Centre has compared the total cost of fire in 15 to 20 countries, with losses also shown as percentages of gross domestic product (GDP). In the 2002 SFPE Handbook, the core costs for the United States typically ran at about 1% of GDP. The US fire death rate relative to population consistently ranked among the highest of the countries studied.[24]

Use in fire safety engineering

Fire risk analysis seeks to answer three questions: what could happen, how bad it would be and how likely it is.[25] Some assessments look back, inferring probabilities from fire loss experience after an event such as the introduction of home smoke alarms. Others look ahead, for example to the effect of requiring sprinklers in homes.[25] The available data differ widely. US homes offer very many documented fires with limited detail, US nuclear power plants far fewer fires with much more detail, and a single building very few fires at all.[26] In a hazard assessment, historical loss data help engineers avoid focusing on unique hazards while overlooking more common but equally important ones.[27]

In Australia, the data sheets that accompany the Fire Safety Verification Method (print version 1.1, 2020) turn incident statistics into characteristic design values, whose use is not mandatory.[28] Their fire growth rates come from an analysis by Holborn and colleagues of fire investigation data from the Greater London area for 1996 to 2000. Of 481 fires in residential buildings, 34% were classed as very slow growing, 50% as slow and 11% as medium.[29]

The same data sheet estimates the share of fires that would become fully developed if no sprinklers were present. For dwellings it adopts about 18%, based on values of 15.5% for Australia, 18.3% for the USA and Canada and 23.2% for New Zealand. Hotels, hospitals and care homes are given 5% and most other buildings 15%.[30] The data sheet explains the low value by the 24-hour staff and monitored detection that such buildings are likely to have. These bring a higher proportion of their fires to the attention of the fire brigade.[30] The accompanying handbook predicts the probability of flashover from the frequency of fires that occupants were unable to manage, taken from fire statistics, together with sprinkler reliability and an analysis of fire brigade intervention.[31]

In New Zealand, the commentary on the first edition of Verification Method C/VM2 (2013) cites New Zealand Fire Service Emergency Incident Statistics for 2005 to 2010. In these, approximately 3% of all structure fires spread to adjacent property, a risk the commentary calls small but not insignificant.[32] In the Netherlands, Van Rossum uses actual outcomes to show that the building regulations (Bbl) were not developed around one quantified goal. Per hour of stay, the chance of dying in a fire is much greater in a residential building than in an assembly building, yet the detection requirements are stricter for assembly buildings.[33]

Reliability of protection systems

Incident data are one of five kinds of source that Hall and Ahrens identify for the reliability of fire protection systems, for which data are very scarce.[5][7] The Australian data sheets define reliability as the probability that a system performs to its specification. Efficacy is the degree to which the system then achieves its design objective, and effectiveness combines the two.[34] For smoke alarms, reliability counts only fires large and close enough to activate an alarm, and effectiveness is efficacy multiplied by reliability.[35] A battery-operated alarm system with limited coverage, a typical reliability of 85% and an efficacy of 93% thus has a typical effectiveness of 79%.[36]

UK Home Office statistics for dwelling fires from 2010/11 to 2017/18, analysed in the same data sheet, show how much definitions matter. Mains-powered smoke alarms failed to operate in 21% of the fires where they were present. Fires whose smoke never reached a detector, or that started outside the covered area, do not count as failures; excluding them left about 4%, a reliability of 96%. For battery-powered alarms the same steps turned 39% into 13%, a reliability of 87%.[37]

For US home fires from 2009 to 2013, an NFPA analysis by Ahrens (2015) found that battery-operated alarms operated in 80% of fires large enough to activate them. Hard-wired alarms with battery backup operated in 95%.[38] Measured as alerting the occupants once an alarm operates, efficacy in the US data was about 97%. Measured as the reduction in deaths, US and UK data gave only 50% to 55%, so the data sheet stresses that the definition must be stated whenever an efficacy is quoted.[39]

Reported fires can make protection look less reliable than it is, because successful systems often keep fires too small to be reported. In the United States, reported home fires consistently show roughly one-third of smoke alarms not operating, against one-fifth in a study of home installations in general.[12] For sprinklers, reported fires suggest failure to operate in one-sixth to one-fifth of the fires large enough to activate them, while real reliability could easily be well over 90%.[12] Inspection and testing data lean the other way, because they tend to come from properties with very advanced maintenance programmes. This is one reason why estimates of sprinkler reliability in Australia are much higher than in the United States.[7] In Australia, Marryatt's 1988 compilation found control in over 99% of fires in sprinklered buildings, but it left out fires in which the sprinklers failed to operate or had no water.[40]

Limits and sources of bias

Missing entries can distort patterns. If the number of fires stayed the same while the share with an unreported cause steadily fell, ignoring the unknowns would make every specific cause appear to increase.[8] Most national estimates therefore allocate unknown entries in the same proportions as the known ones. This assumes that the unknown fires would show the same pattern if their details were known.[8] The Australian data sheets apply the same method, for example to unknown fire sizes in New Zealand data and to unknown reasons why smoke alarms did not operate.[30][41]

Known and unknown incident fields lead to different apparent category shares and uncertainty.
Unknown data and reporting bias must be addressed explicitly. Based on Hall and Ahrens (2002), pp. 5-67–5-70.

Codes can also lump different things together. One NFIRS equipment code covers irons, hair dryers and electric blankets, which are used in very different places and tend to ignite different items.[9] The floor covering code includes carpet, tile, rugs, hardwood, vinyl flooring and even stairs.[8] One of the materials most often reported as first ignited in floor covering fires is petrol, which refers to a liquid spilled on the floor rather than to the covering itself. Hall and Ahrens consider it unsound to exclude such fires as unfair tests, because all fires involving a product belong to its fire problem.[9]

Definitions also differ between countries and data systems. In fire incident databases a space heater heats a relatively small area, whereas the US Department of Energy uses the term for all heating systems.[5] Ramachandran notes that a "room" in fire brigade reports is not necessarily a fire compartment. Fires recorded as spreading beyond the room include fires that passed through an open door or window. Such counts therefore do not give a valid estimate of how often the walls, floors and ceilings of a room fail.[42]

Past data may not suit the future. Meacham asks whether fire loss data for cellulosic materials before 1960 could indicate losses involving synthetic materials after 1960.[43] Because fires are rare and people change buildings and their contents over time, Meacham calls predicting future risks from past loss experience challenging at best.[11] The International Fire Engineering Guidelines (2005 edition) likewise warn that data may lose relevance through new technologies, materials, regulations and construction practices.[44] Data sometimes have to be borrowed from abroad: the Australian data sheets quote a Senate committee finding that Australian fire statistics are hard to locate and compare. For smoke alarms they therefore draw mainly on data from the US, the UK and New Zealand.[35]

Fire losses also vary at random from year to year, since several large-loss fires may occur in one year and few or none the next. A US cost-benefit study of residential sprinklers described by Notarianni therefore used mean values for the five years from 1989 to 1993.[13][45] Where systematic data are lacking, estimates from experienced people can be used. Their recollections, however, tend to make reliable systems seem even more reliable and unreliable ones even less reliable.[7]

See also

Notes

  1. ↑ Hall (2002c), p. 1-193.
  2. ↑ 2.0 2.1 2.2 2.3 2.4 2.5 2.6 2.7 Hall and Ahrens (2002), p. 5-67.
  3. ↑ 3.0 3.1 3.2 3.3 3.4 Hall and Ahrens (2002), p. 5-65.
  4. ↑ Hall and Ahrens (2002), p. 5-66.
  5. ↑ 5.0 5.1 5.2 5.3 5.4 5.5 Hall and Ahrens (2002), p. 5-70.
  6. ↑ Van Rossum (2027), §1.4.
  7. ↑ 7.0 7.1 7.2 7.3 Hall and Ahrens (2002), p. 5-71.
  8. ↑ 8.0 8.1 8.2 8.3 8.4 8.5 8.6 Hall and Ahrens (2002), p. 5-68.
  9. ↑ 9.0 9.1 9.2 Hall and Ahrens (2002), p. 5-69.
  10. ↑ 10.0 10.1 Hall and Ahrens (2002), p. 5-76.
  11. ↑ 11.0 11.1 Meacham (2002), p. 5-159.
  12. ↑ 12.0 12.1 12.2 Hall and Ahrens (2002), p. 5-72.
  13. ↑ 13.0 13.1 Notarianni (2002), p. 5-61.
  14. ↑ Hall (2002b), p. 5-151.
  15. ↑ BSI (2001), as cited in National Research Council of Canada et al. (2005), p. 3.2-2.
  16. ↑ National Research Council of Canada et al. (2005), p. 3.2-2.
  17. ↑ Rahikainen and Keski-Rahkonen (1998), as cited in National Research Council of Canada et al. (2005), p. 3.2-3.
  18. ↑ Van Rossum (2027), §7.
  19. ↑ Anderson et al. (1981) and Harwood and Hall (1989), as cited in Gottuk and Lattimer (2002), p. 2-54.
  20. ↑ Gann et al. (1994), as cited in Gottuk and Lattimer (2002), p. 2-54.
  21. ↑ Purser (2002), p. 2-134.
  22. ↑ Home Office (1983), as cited in Purser (2002), p. 2-134.
  23. ↑ National Research Council of Canada et al. (2005), p. 2.4-4.
  24. ↑ Ramachandran (2002b), pp. 5-79–5-80.
  25. ↑ 25.0 25.1 Watts and Hall (2002), p. 5-2.
  26. ↑ Watts and Hall (2002), p. 5-3.
  27. ↑ Meacham (2002), p. 5-155.
  28. ↑ Australian Building Codes Board (2020a), Data Sheet B2, p. 1.
  29. ↑ Holborn et al. (2004), as cited in Australian Building Codes Board (2020a), Data Sheet B2, pp. 4–7.
  30. ↑ 30.0 30.1 30.2 Australian Building Codes Board (2020a), Data Sheet B2, pp. 6–9.
  31. ↑ Australian Building Codes Board (2020b), p. 92.
  32. ↑ Ministry of Business, Innovation and Employment (2013), p. 23.
  33. ↑ Van Rossum (2027), §1.2.
  34. ↑ Australian Building Codes Board (2020a), Data Sheet C3, p. 87.
  35. ↑ 35.0 35.1 Australian Building Codes Board (2020a), Data Sheet C3, p. 91.
  36. ↑ Australian Building Codes Board (2020a), Data Sheet C3, p. 92.
  37. ↑ Home Office (2018), as cited in Australian Building Codes Board (2020a), Data Sheet C3, pp. 105–109.
  38. ↑ Ahrens (2015), as cited in Australian Building Codes Board (2020a), Data Sheet C3, pp. 109–111.
  39. ↑ Australian Building Codes Board (2020a), Data Sheet C3, p. 111.
  40. ↑ Marryatt (1988), as cited in National Research Council of Canada et al. (2005), pp. 3.7-2–3.7-3.
  41. ↑ Australian Building Codes Board (2020a), Data Sheet C3, pp. 106, 108.
  42. ↑ Ramachandran (2002c), p. 3-388.
  43. ↑ Meacham (2002), p. 5-154.
  44. ↑ National Research Council of Canada et al. (2005), p. 3.1-2.
  45. ↑ Notarianni (2002), p. 5-62.

References

  • Australian Building Codes Board. (2020a). Data sheets: Fire Safety Verification Method (Handbook annex, Print version 1.1).
  • Australian Building Codes Board. (2020b). Handbook: Fire Safety Verification Method (Print version 1.1).
  • Gottuk, D. T., & Lattimer, B. Y. (2002). Effect of combustion conditions on species production. In P. J. DiNenno, D. Drysdale, C. L. Beyler, W. D. Walton, R. L. P. Custer, J. R. Hall Jr., & J. M. Watts Jr. (Eds.), SFPE handbook of fire protection engineering (3rd ed., pp. 2-54–2-82). National Fire Protection Association; Society of Fire Protection Engineers.
  • Hall, J. R., Jr. (2002b). Product fire risk. In P. J. DiNenno, D. Drysdale, C. L. Beyler, W. D. Walton, R. L. P. Custer, J. R. Hall Jr., & J. M. Watts Jr. (Eds.), SFPE handbook of fire protection engineering (3rd ed., pp. 5-143–5-152). National Fire Protection Association; Society of Fire Protection Engineers.
  • Hall, J. R., Jr. (2002c). Statistics. In P. J. DiNenno, D. Drysdale, C. L. Beyler, W. D. Walton, R. L. P. Custer, J. R. Hall Jr., & J. M. Watts Jr. (Eds.), SFPE handbook of fire protection engineering (3rd ed., pp. 1-193–1-202). National Fire Protection Association; Society of Fire Protection Engineers.
  • Hall, J. R., Jr., & Ahrens, M. J. (2002). Data for engineering analysis. In P. J. DiNenno, D. Drysdale, C. L. Beyler, W. D. Walton, R. L. P. Custer, J. R. Hall Jr., & J. M. Watts Jr. (Eds.), SFPE handbook of fire protection engineering (3rd ed., pp. 5-65–5-78). National Fire Protection Association; Society of Fire Protection Engineers.
  • Meacham, B. J. (2002). Building fire risk analysis. In P. J. DiNenno, D. Drysdale, C. L. Beyler, W. D. Walton, R. L. P. Custer, J. R. Hall Jr., & J. M. Watts Jr. (Eds.), SFPE handbook of fire protection engineering (3rd ed., pp. 5-153–5-175). National Fire Protection Association; Society of Fire Protection Engineers.
  • Ministry of Business, Innovation and Employment. (2013). Commentary for Building Code clauses C1–C6 and Verification Method C/VM2.
  • National Research Council of Canada, International Code Council, Department of Building and Housing, & Australian Building Codes Board. (2005). International fire engineering guidelines (2005 ed.). Australian Building Codes Board.
  • Notarianni, K. A. (2002). Uncertainty. In P. J. DiNenno, D. Drysdale, C. L. Beyler, W. D. Walton, R. L. P. Custer, J. R. Hall Jr., & J. M. Watts Jr. (Eds.), SFPE handbook of fire protection engineering (3rd ed., pp. 5-40–5-64). National Fire Protection Association; Society of Fire Protection Engineers.
  • Purser, D. A. (2002). Toxicity assessment of combustion products. In P. J. DiNenno, D. Drysdale, C. L. Beyler, W. D. Walton, R. L. P. Custer, J. R. Hall Jr., & J. M. Watts Jr. (Eds.), SFPE handbook of fire protection engineering (3rd ed., pp. 2-83–2-171). National Fire Protection Association; Society of Fire Protection Engineers.
  • Ramachandran, G. (2002b). Measuring fire consequences in economic terms (J. R. Hall Jr., Rev.). In P. J. DiNenno, D. Drysdale, C. L. Beyler, W. D. Walton, R. L. P. Custer, J. R. Hall Jr., & J. M. Watts Jr. (Eds.), SFPE handbook of fire protection engineering (3rd ed., pp. 5-79–5-92). National Fire Protection Association; Society of Fire Protection Engineers.
  • Ramachandran, G. (2002c). Stochastic models of fire growth. In P. J. DiNenno, D. Drysdale, C. L. Beyler, W. D. Walton, R. L. P. Custer, J. R. Hall Jr., & J. M. Watts Jr. (Eds.), SFPE handbook of fire protection engineering (3rd ed., pp. 3-381–3-401). National Fire Protection Association; Society of Fire Protection Engineers.
  • Van Rossum, E. (2027). Introduction to Fire Safety Engineering. Fire Safety Academy.
  • Watts, J. M., Jr., & Hall, J. R., Jr. (2002). Introduction to fire risk analysis. In P. J. DiNenno, D. Drysdale, C. L. Beyler, W. D. Walton, R. L. P. Custer, J. R. Hall Jr., & J. M. Watts Jr. (Eds.), SFPE handbook of fire protection engineering (3rd ed., pp. 5-1–5-7). National Fire Protection Association; Society of Fire Protection Engineers.
Retrieved from "https://fire-safety.academy/w/index.php?title=Fire_statistics&oldid=2696"