Fires damage buildings and claim lives every year, but many are preventable. The District of Columbia Fire and Emergency Medical Services Department inspects buildings and educates residents on fire safety, but it cannot reach every property each year. We partnered with DC Fire & EMS to create a statistical model for which buildings in DC were at the highest risk of having a fire. In a field evaluation, we tested whether fire inspectors found more fire code violations when they were encouraged to inspect the highest risk buildings from our model. We learned that the very highest risk buildings don’t always have more fire code violations, even if they are more likely to catch fire. We also used the scores to help firefighters decide on buildings to target for fire preparedness planning.
A DC Fire & EMS employee installs a smoke detector during August 2023's All Hands on Fire Prevention Outreach. Credit: @dcfireems
Why is this issue important in DC?
Between 2016 and 2022, there were 189,801 calls related to fire incidents in DC, including 15,770 “highest priority” calls 1. Fires in multi-unit buildings and highly-populated neighborhoods can displace dozens of households and businesses at a time. Fire and EMS already uses an evidence-based outreach strategy to install free smoke alarms in neighborhoods2. However, the Department has limited resources and cannot inspect all buildings in DC so it must prioritize which to inspect first. Using predictive modeling may give fire inspectors another tool to prioritize inspections and prevent more building fires.
What did we do?
We worked with publicly available data on Open Data DC and internal data on fire incidents in commercial buildings (i.e. non-residential buildings and residential buildings containing more than four units). We used this data to train a statistical model that ranks individual buildings according to their risk of fire in the future. We then developed a randomized evaluation to test whether using lists of the riskiest buildings helped fire inspectors to identify more fire code violations. We shared lessons from this evaluation with personnel from the Department’s Fire Prevention Division. With their feedback, we have developed a dashboard that will allow them to continue using the data and risk predictions from our models in their ongoing inspections.
What have we learned?
Our model was effective at predicting fire risk. We broke the buildings into 8 groups, from least to most risky, based on what the model predicted. We found that buildings in the top three groups were far more likely to have fires than buildings in lower risk tiers. For the experiment, we gave inspectors lists of buildings from the top tier (about 1500 buildings), where fires were most likely. But giving inspectors lists of the top risk tier buildings did not lead them to find more fire-code violations than they usually do. We think this is because it is relatively common for an inspector to find violations in a commercial building, so they were still likely to find violations in buildings from lower risk tiers. Also, fire inspectors aren’t always allowed to inspect the riskiest spaces in commercial buildings—like apartment kitchens—so they are not able to see the areas where fires are most likely to start.
When we widened the set of buildings to buildings from the top half of risk scores, we observed that inspectors found more violations in those buildings than in those from the bottom half of risk scores. This finding suggests that it might be more important to send inspectors to the top half of risky buildings, instead of just the highest risk tier. You can read more about the model in our predictive modeling report, and about the evaluation in our final report (full version or 2-page version).
What comes next?
We are working with the Office of the Chief Technology Officer to provide DC Fire & EMS with a dashboard that shows risk scores and other information about buildings on a map. Based on our results and multiple rounds of user testing, we are finishing changes that will make the dashboard easier to use. For example, we simplified the risk score, which was previously a number, so it just shows whether a building is in the “high” (top half) or “low” (bottom half) of risk scores. This approach will help inspectors spend more time in the riskier buildings. Second, we are making the dashboard available to more than just fire inspectors. For example, the dashboard will be a new tool for firefighters when they make plans and do dry runs for fighting fires in high-risk buildings.
What happened behind the scenes?
Part of learning how to best use our model meant sitting down with fire inspectors to understand how they schedule inspections and how to best incorporate fire risk score into their current system. This exploration included joining the fire department as they went door-to-door to install smoke alarms and give fire safety tips—and by chance, one of our team members was assigned to visit their own home! During a typical day, the outreach team visits about 500 homes and installs about 20 smoke alarms.
“The mission of the Fire and EMS Department is to save lives. By using state of the art technology to understand where incidents occur and the populations that are most impacted, we can tailor our prevention efforts to target those most in need.”