This story was originally published by Data-Smart City Solutions.
About a year ago, Cincinnati, Ohio, felt the first major shock in its opioid epidemic. While the surrounding county was already one of Ohio’s hardest hit, Cincinnati’s late-summer 2016 surge was “unprecedented.” It started August 23rd; by August 28th, 174 people had overdosed. The city, where paramedics typically made about four overdose runs per day, averaged almost 20 each day for a week.
It was only the start. After a comparatively slow week, overdoses spiked even higher the second week of September. Rates gradually declined, bottomed out over the winter, and then rose again in the spring.
These counts represent individuals’ struggles with addiction, the strain on their families, and the dangers of synthetic opioids like fentanyl and carfentanil. They also measure the daily toll on frontline workers like paramedics, police officers and social workers. First responders accumulate compassion fatigue after resuscitating parents in front of their young children and reviving the same patients repeatedly. Saving lives with naloxone, the opioid reversal drug, is usually thankless, since the medication sends overdose victims directly into withdrawal.
Meanwhile, financial costs stack up for city governments, at around $1,000 for each emergency medical service (EMS) run, usually several per day. Other city services can take a hit: Cincinnati Mayor John Cranley even argued that serious crimes rose in 2016 because overdoses demanded so much attention from police officers.
With these problems mounting, city agencies must devise their own solutions. In Part I of this two-part series, I explored interagency collaboration in New York City, grounded in data sharing, and how medical examiners’ reports can help cities understand who is overdosing and on what. In this installment, I dive into police and EMS dispatch data revealing when and where overdoses occur. I show how cities can use these data to create an effective early warning system for overdose spikes, identify chronic hotspots to target interventions, and assess performance across units.
These analyses are possible only because Cincinnati has made its overdose dispatch data openly available—a model for data-driven policy that I examine as well.
For a city like Cincinnati, struggling with an overdose crisis, opening overdose records to public scrutiny might seem like an act of radical transparency. Cincinnati, after all, was the first city in Ohio with an open data policy and is one of the relatively few U.S. city governments that also routinely publish records of police-involved shootings.
But for Brandon Crowley, the city’s Chief Data Officer, the objective was simply to help curb the epidemic. “What I’ve learned is that ultimately, access to data is the greatest contributing factor to any success…whether it’s potholes or overdoses,” Crowley told me.
After last summer’s overdose surge, Crowley’s office decided to begin publishing records from all fire incident responses—including EMS calls—with data fields indicating whether dispatchers had coded an incident as a suspected overdose. The records include the request’s time, date and location, both by street address and geographical coordinates. To protect residents’ privacy, the office partially redacts street numbers and randomly skews the coordinates.
The result is a big raw dataset with thousands of suspected overdoses embedded among 180,000 incident records. To help users interpret it, Crowley created an interactive dashboard for the city’s public-facing analytics site, visualizing temporal and geographical trends in overdose incidents. He said the dashboard has become a resource both internally, for city health officials and first responders, and externally, for partners like Talbert House, a local nonprofit that uses the data to guide resource allocation.
Cincinnati’s dataset is valuable, in part, because information about overdoses is scarce—even taboo. For Crowley, the overdose crisis requires a culture change, specifically a “collective willingness to share painful facts” inside and outside government. Cincinnati’s decision highlights the opportunity, so far mostly missed, for cities with robust open data programs to address overdoses head-on, even though the topic is sensitive. A handful of cities, including Seattle, WA, Germantown, MD, and Northampton, MA, have released maps or geocoded data for overdose incidents, but the paucity of these releases is striking. Many more cities publish the locations of street trees.
And yet there is a clear demand for information about overdoses. The data I analyzed in Part I came from Allegheny County, PA (via the Western Pennsylvania Regional Data Center) and the State of Connecticut. Connecticut Chief Data Officer Tyler Kleykamp told me accidental drug overdoses has been the most popular dataset on the state’s open data portal, with twice as many downloads as the next most popular dataset since it was uploaded in 2015. One user was TrendCT, an online news outlet, which published an in-depth analysis last year.
With city agencies responding to overdoses every day, their own records can provide in-depth information about the nature of the problem. Even when a city has not been intentionally tracking overdoses, first responders have identified workable methods for deriving overdose data from computer-aided dispatch records.
Using dispatch data, Cincinnati’s interactive dashboard allows users to see patterns at a glance and to dig deeper according to variables like time period and neighborhood. The dataset is timely and reliable enough that the Cincinnati Enquirer has consulted the site in its reporting on overdose outbreaks.
Yet U.S. city governments are underutilizing data on local trends, like those visualized on Cincinnati’s dashboard, to guide their overdose responses. For example, many jurisdictions in Ohio, including Cincinnati, have established quick response teams, which pair cops and paramedics with addiction counselors. The teams visit recent overdose victims and offer them services, especially referrals to addiction treatment; officials in nearby Colerain Township say the strategy helped reduce overdose rates there. But while response teams require data, such as contact information for overdose victims, the strategy involves little or no data analysis.
One reason for sparse data analysis may be that few trends are immediately evident among overdoses. Some types of injuries cluster at specific times, such as car accidents, which tend to occur during rush hour and overnight on weekends; overdoses, however, are relentlessly regular. Other types of injuries, like homicides, cluster in specific locations; as Crowley told me, a remarkable aspect of Cincinnati’s overdoses is their geographical dispersion across many neighborhoods.
Nonetheless, data analysis can uncover useful patterns could make cities’ responses even quicker and smarter. I illustrate three examples below.
Cincinnati’s 2016 rise in overdoses was hardly unique among American cities. As the Washington Post recently reported, urban overdose deaths exploded between 2014 and 2016, including a 600% increase in deaths involving fentanyl, a synthetic opioid often manufactured and distributed illicitly. Intense short-term spikes, lasting a few days or weeks, have been recorded in Louisville, KY, Huntington, WV and elsewhere.
Cincinnati’s late summer overdose surge was linked to carfentanil, a synthetic opioid 100 times stronger than fentanyl and licensed only for tranquilizing large mammals. The effects became unmistakable on August 23rd, when cases hit 25—a level never seen before. Local media was reporting on the story by the 25th. But in fact, the signal had emerged several days earlier.
A good early warning system would have triggered an alert by August 20th. Epidemiologists have developed such “syndromic surveillance” methods to monitor for possible disease outbreaks. Typically, they apply algorithms to detect anomalies in daily case rates. One simple algorithm, using what’s called a C1 statistic, looks at the past seven days’ rates, calculating the average case count and the amount of variation around that average. It is run each day, and if a given day’s rates deviate enough from the past week’s numbers, it triggers an alert.
Such a method, using EMS overdose runs, would have triggered an alert after cases jumped on August 19th. Eleven overdose runs was not unheard of, but it was a significant enough change to suggest something was going on. Whatever caused the late-summer surge may already have been introduced on August 19th. With this advance warning, could health and safety agencies have prevented the spike? Likely not, but perhaps they could have mitigated its effects, identified the cause earlier, or put more resources in place to respond.
Here is how the C1 method, using a standard threshold, would have performed on recent Cincinnati data:
While the method generates a few false alarms, and a few trivial positives (when an outbreak was already unmistakable), it also flags the initial bumps that appear to have been associated with major increases last August and October, an upward turn in January, and a large spike in March.
The ability to map overdose locations is an obvious and important advantage of geocoded dispatch data. While it is tempting to look for emergent spatial patterns, I found little evidence that today’s overdoses predict the locations of tomorrow’s. Although New York City monitors timely geographical patterns to look for emerging overdose hotspots, the city is a poor comparator: its smallest borough, Staten Island, has twice the residents of Cincinnati. Moreover, the “pizza delivery” inspired model of drug retailing described in Sam Quinones’s Dreamland, along with rapid police response to each overdose, might tend to disperse cases. (Indeed, with the SaTScan algorithm New York City uses, I could not find a single meaningful, statistically significant spatiotemporal cluster in Cincinnati’s 2016 overdose incidents.)
Over time, however, the same areas experience enough overdoses to warrant a targeted response. I mapped every Cincinnati EMS run in 2016 onto a grid with squares of 250 x 250 meters—about the size of a city block. While most grid squares had zero overdoses or just a few, there were 15 outliers with 11 overdoses or more. Here the results are visualized, using the free spatial analysis software GeoDa:
Cincinnati suspected overdoses, 2016 (excluding false alarms). Data source: Cincinnati open data portal
This quick mapping can help understand outliers. In Cincinnati, the single-highest incidence square covers the downtown Hamilton County Public Library facility, where local news reports have noted a major problem with overdoses and other emergencies. Overdose issues are not unique to the Cincinnati library, and other cities have taken action: Philadelphia, for example, implemented overdose drills in its libraries. The city trained librarians to use Narcan; one has administered it at least four times to revive library visitors.
Next, we can inspect for broader spatial patterns, including hot spots. To be confident we are observing real spatial patterns, we conduct a cluster analysis that imposes a rigorous process for assessing statistical significance. For this map of the local Gi* statistic, an algorithm picks out zones, in red, with higher-than-average incident counts, given their size; the blue ones have lower-than-average counts.
Cincinnati Gi* overdose clusters at p < 0.05, 2016
Cluster maps can help pinpoint problem areas, including ones that extend across neighborhood boundaries. For example, this map shows a higher-incidence swath extending south and northeast of the city’s medical area, home to many of Cincinnati’s health services providers, including the county’s only needle exchange.
Within this cluster, the highest-incidence square is a dense commercial block where fast food chain restaurants—whose publicly-accessible bathrooms have been called “ground zero” for heroin use elsewhere—intersperse with health service agencies. Perhaps partnerships between health providers and local businesses could help address unsafe drug use; the proximity to health providers could offer favorable conditions for a supervised injection facility, should Cincinnati consider the option.
Trouble in public spaces has been characteristic of the opioid crisis nationally. The EMS director in Berkeley County, WV, told The New Yorker that opioid users may seek out public places so that someone will find them if they overdose. “They’re struggling with using but not wanting to die,” said Brian Costello. In addition, RTI International epidemiologist Alex Kral told me, overdoses might be more likely in public places where users, cramped or hurrying to conceal their use, could be prone to dosing mistakes or other errors: his study of clients at an unsanctioned supervised injection facility found that almost 85% rushed when they injected outside the facility.
In response to public overdoses, some cities have considered enlisting Good Samaritans, including with lockboxes containing Narcan, which strangers could access after calling 911 and before paramedics arrive. The health commissioner in Baltimore, MD, issued a standing prescription allowing any city resident to obtain Narcan.
Mayor Cranley’s assertion that overdoses have hampered policing was striking, and initially difficult to square with the data. Suspected overdoses accounted for less than one percent of all Cincinnati police dispatches in 2016. How could this tiny proportion affect performance?
To test the idea, I consulted the police department’s 2016 dispatch records and used them to calculate response times—an indicator of departmental performance. Consistent with Mayor Cranley’s concern, overdoses and average response times both increased over the summer. But after adjusting for these seasonal trends, there was no correlation citywide.
Cincinnati police districts & overdose hotspots (Gi* high clusters), 2016
The effects, however, vary by police district. Among Cincinnati’s five districts, the west side’s District 3 is the busiest. While District 3’s response time rises on its busier days, overdoses do not appear to have much effect. In contrast, Districts 4 and 5 both show an apparent correlation between overdoses and response time, irrespective of the number of non-overdose calls:
In District 5, this effect turns out to be large and statistically significant. After controlling for seasonal trends and non-overdose calls, my regression model found that every overdose in District 5 increased the expected response time for other incidents by approximately one minute. For residents of District 5, this estimated delay could cause considerable concern; for city officials, it could justify investigating causes and potentially shifting resources to address the problem. Notably, two 2016 overdose hotspots, around Camp Washington and southeast Winton Place, lie on the periphery of District 5: this geography might draw responding units away from where they are needed for other incidents.
Dispatch records are a powerful tool for data-smart opioid responses. While some of the same information can be pieced together from hospital emergency department reports, dispatch data offers several advantages. Not every overdose requires an emergency room visit, so dispatch data may capture more incidents, generating more statistical power. Dispatch data is timely and often geocoded, allowing more precise analysis of the locations of recent overdoses. First response agencies will likely also find their own dispatch data to be more accessible and easier to analyze than hospital data. This advantage may be especially salient in small and medium-sized cities, where data analytic resources might be most concentrated in police departments.
As I showed, EMS dispatch data appear to be a viable data source for early warning systems. Analytically, there are methods that tend to outperform the C1 statistic I used above, but the C1 method is simple enough to implement as a spreadsheet formula and appears well suited to overdose counts, where seasonal and day-of-week trends are unclear. Moreover, a city can use the C1 method just a week after beginning to track overdose counts. After a few months of accumulating the data, it will become possible to assess geographical trends and performance like in sections (2) and (3).
Finally, cities should work to communicate this information to residents effectively. An open data approach helps engage the public, particularly with an interface like Cincinnati’s data visualization tool. While Part I explained the value of bringing city agencies together around data, open data creates opportunities for other community members to contribute. Many of the same steps are necessary for interagency collaboration and for public release, so it may be sensible to pair the New York City and Cincinnati models. With overdose rates continuing to rise, cities should put as many resources as possible behind data-driven local strategies.