## Statistics for firefighters!

This is one I’d never thought about . . . Daniel Rubenson writes:

I’m an assistant professor in the Politics Department at Ryerson University in Toronto. I will be teaching an intro statistics course soon and I wanted to ask your advice about it. The course is taught to fire fighters in Ontario as part of a certificate program in public administration that they can take. The group is relatively small (15-20 students) and the course is delivered over an intensive 5 day period. It is not entirely clear yet whether we will have access to computers with any statistical software; the course is taught off campus at a training facility run by the Ontario Fire College.

The students will be highly motivated but they will have no statistics training and whatever math training they have will be from high school, probably more than ten years ago. The main purpose of the course will be to train them in understanding and interpreting the results of quantitative research and reports, not so much in actually performing statistical analysis on their own.

Do you have any advice in terms of topics to cover or avoid, particular approaches to take? As I said, the approach you take in the Teaching Statistics book is excellent and I use many of the demonstrations to great effect in my regular statistics classes. In this case both the students and the compressed nature of the course create some unique challenges.

I have no idea what to say on that, except to retreat to my general thoughts about statistics as comparisons. Start with comparing two averages (or the corresponding graph), then adjust for various potential confounders, consider data collection issues, move to regression and Anova. In 5 days, you can do a couple of examples of this sort, then step back and do a little bit of probability and statistical inference.

1. ZBicyclist says:

Seasonal patterns would be good things to cover. This is something they can probably relate to easily. (broadly using seasonal to refer to hourly, day of week, week of year, patterns).

Also response times relative to target times (when introducing hypothesis testing versus a non-zero null)

2. Ben Hyde says:

The fire insurance industry has lots of data, and the distribution is counter intuitive for most.

3. Michael Braun says:

What about basic notions of probability, random variation? One example might be extending the classic "locations of bombs over London" example to clusters of fire. If one were to segment a city into blocks on a grid, and counted x fires in each grid block, some blocks might have a lot of fires, and others would have none. It could be there is a real cause of the clustering patterns, or it might just be random variation.

Extending this idea (and ZBicyclists's comment), understanding the difference between random noise and a trend might be useful.

4. Steve Sailer says:

I know of a database that all firefighters would find very interesting to analyze using basic statistical techniques: the scores on the 2003 New Haven Fire Department promotion exams:

5. danachandler says:

In Gary Klein's book "Sources of Power", Klein describes a study of his involving how firefighters and other experts make split-second, life or death decisions. It has been a long time since I read the book, but I remember one example with how experienced firefighters were able to tell by the way the floor felt whether it was going to cave in.

You might be able to skim the book and take examples where firefighters or other experts were able to use their priors to make the right decision, perhaps applying Bayes' Law.

6. marcel says:

Do you know, or are you assuming, that tests for US & Canadian fire departments resemble each other? Because if the similarity is low, these firefighters would likely see this as typical US myopia. Further, these are people who are already firefighters, so likely more interested in data having to do with fires and fighting them than in becoming a firefighter.

7. Kevin O'Neil says:

As a firefighter (volunteer) and a demography grad student, I envy Daniel's task. This class will be a LOT of fun.

Basic descriptive statistics are used a lot in the fire service to make important decisions, but not always understood (our department's mean and median response times look quite different, for example, but to most people they're both an "average", which is sometimes helpful). A good working knowledge might help these guys when funding decisions are being made or a consultant comes along with "advice". You might even see if you can get incident and response data from the dept beforehand to play with, it should be readily available. Many firefighters I know both appreciate the power of statistics and suspect that they are "damn lies". Your course might give them the ability to, um, fight fire with fire.

Similarly, the concepts of confidence intervals, outliers, and spurious relationships is an important one. Just because there's been fewer accidents since the new "howler" sirens have been in place, doesn't mean that they're responsible.

Probability theory is something firefighters use every day on the fireground. There are a lot of fun ways to play with this in the distribution of incidents in time and space, as noted above, but one important application is simple cost/benefit analysis. Many depts. spend a lot of time and money preparing for relatively rare events, like a building collapse with firefighters inside, but less on road safety or driver training. There are both cultural and practical reasons for this, but having a more formal framework for assessing risk and measures to address it might help.

Some ability to interpret epidemiological research (what does an odds ratio mean?) might be helpful, give that a lot of research is tying firefighting with cancer risk and that cardiac events (along with road accidents) are the top killer of firefighters.

I suspect the students will have a barrage of topics and applications they are interested in– the hard part will be narrowing it down.

8. Willem says:

What Ben said. First give them a feel for processes that deliver normal distributions. Then give them examples of highly skewed distributions (like the monetary losses in fire). Then you could go on teaching that there is no point in standing next to a building waiting for a fire to happen, but there is a point in distributing firemen over an area as to minimize access times. Some great understandings there.

Teach them a trick or two about inference and why, even with all the best intentions, it's impossible to estimate the effects of an individual departments efforts in fire prevention. Don't talk down prevention, but make them critical of any claim that action X with budget Y causes Z less fires / accidents.

9. Good morning,
First, I gotta say I want to take the class! I'm a U.S. Fire Chief and knowing how to use data is one of my (and the fire services) weakest areas. We accumulate a ton of data through our reporting and use almost none of it for any decision making or planning purposes. Generally, the only statistic that we use is the "average" response time.
There is a push in the fire service to move beyond that, but it is still in its infancy. For example, the National Fire Protection Association (NFPA) has standards for resource deployment. They now require us to evaluate the reponse time at the 90th percentile, instead of the mean. As you can imagine response times (interval from when we're notified of an emergency to when we arrive) are heavily skewed. Most times are in the 3-8 minute range, but we have occasional times out to 10, 15 or even 20 minutes.
Here's a few questions I wish I, or my staff, could answer with our data:
1) Can I justify changing my staffing depending upon time of day, day of week, month of year? So, for the same staffing budget, is there a strong enough (and what does that mean?) correlation that I can short staff on Monday mornings to overstaff on Friday nights (when the bars close and everyone crashes)?
2) When do I need to add a unit? A what point does the workload on an ambulance raise to the point that its no longer available when needed for emergency responses? The follow-on is since I have four ambulances at four stations, should I consider consolidating two units to a busier fire station and none at a slower station, without impacting overall service delivery.
3) How does response time vary between stations/shifts? Does one station/shift do "statistically" better or worse enough to warrant special attention.
4) How do I compare to my different sized neighbor, and what factors influence our service delivery? Does budget, population, demographics affect number of emergencies, number of fires/deaths/crashes?
As I mentioned, there's a growing amount of interest in the use of data that we already collect and sadly unused in a really big database. The U.S. Fire Administration has published "Fire Data Analysis Handbook". The NFPA Standard 1710/1720 (mentioned earlier) has some benchmark specifics detailed. The Center for Public Safety Excellence manages fire department accreditation and has some info and requirements on fire service statistics.
Two other quick thoughts. First, without knowing Ontario's fire service, there is a big gap between big city, suburban and rural departments. Although that seems obvious, it will play a big role in what the agency finds important. For example, my "typical" suburban department serves about 40,000 residents in 50 square miles. We have about 5-10 working structure fires annually and 80% of our workload is ambulance calls. Also, there's a big gap between wildland firefighting agencies (they use bulldozers and shovels) compared to structural departments (we use hose and airpacks). Second, if you can work any sort of geo-statistical components, it would be great. A fair bit our need for stats goes to resource deployment and that's heavily geo based.
Finally, the old saying is very true, the fire service is 200 years of tradition unimpeded by progress (at least in the US).

10. Keith says:

You may wish to clarify what the students are expected to do with their training – if its to make decisions based on other's published or contracted research – you may wish to focus on critical appraisal of literature methods and especially contrast randomized and non-randomized studies – though largely non-statistical.

But if you could sneak in a simple example of calculating the posterior probability of say arson given some test results, perhaps using natural frequencies, you may increase rather than decrease their understanding/appreciation of statistics…

Keith