It’s time for Compstat to change

If we are to promote more thoughtful and evidence-based policing, then Compstat has to change. The Compstat-type crime management meeting has its origins in Bill Bratton’s need to extract greater accountability from NYPD precinct commanders in late 1990s New York. It was definitely innovative for policing at the time, and instigated many initiatives that are hugely beneficial to modern policing (such as the growth of crime mapping). And arguably it has been successful in promoting greater reflexivity from middle managers; however these days the flaws are increasingly apparent.

Over my years of watching Compstat-type meetings in a number of departments, I’ve observed everyone settle into their Compstat role relatively comfortably. Well almost. The mid-level local area commander who has to field questions is often a little uneasy, but these days few careers are destroyed in Compstat. A little preparation, some confidence, and a handful of quick statistics or case details to bullshit through the tough parts will all see a shrewd commander escape unscathed.

In turn, the executives know their role. They stare intently at the map, ask about a crime hot spot or two, perhaps interrogate a little on a case just to check the commander has some specifics on hand, and then volunteer thoughts on a strategy the commander should try—just to demonstrate their experience. It’s an easy role because it doesn’t require any preparation. In turn, the area commander pledges to increase patrols in the neighborhood and everyone commits to reviewing progress next month, safe in the knowledge that little review will actually take place because by then new dots will have appeared on the map to absorb everyone’s attention. It’s a one-trick pony and everyone is comfortable with the trick.

There are some glaring problems with Compstat. The first is that the analysis is weak and often just based on a map of dots or, if the department is adventurous, crime hot spots. Unfortunately, a map of crime hot spots should be the start of an analysis, not the conclusion. It’s great for telling us what is going on, but this sort of map can’t really tell us why. We need more information and intelligence to get to why. And why is vital if we are to implement a successful crime reduction strategy.

We never get beyond this basic map because of the second problem: the frequent push to make an operational decision immediately. When command staff have to magic up a response on the spot, the result is often a superficial operational choice. Nobody wants to appear indecisive, but with crime control it can be disastrous. Too few commanders ever request more time to do more analysis, or time to consider the evidence base for their operational strategies. It’s as if asking to think more about a complex problem would be seen as weak or too ‘clever’. I concede that tackling an emerging crime spike might be valuable (though they often regress to the mean, or as Sir Francis Galton called it in 1886, regression towards mediocrity). Many Compstat issues however, revolve around chronic, long-term problems where a few days isn’t going to make much difference. We should adopt the attitude that it’s better to have a thoughtfully considered successful strategy next week than a failing one this week.

Because of the pressure to miracle a working strategy out of thin air, area commanders usually default to a limited set of standard approaches, saturation patrol with uniform resources being the one that I see at least 90 percent of the time. And it’s applied to everything, regardless of whether there is any likelihood that it will impact the problem. It is suggested by executives and embraced by local area commanders because it is how we’ve always escaped from Compstat. Few question saturation patrols, there is some evidence it works in the short term, and it’s a non-threatening traditional policing approach that everyone understands. Saturation patrol is like a favorite winter coat, except that we like to wear it all year round.

Third, in the absence of a more thoughtful and evidence-based process, too many decisions and views lack any evidential support and instead are driven by personal views. There is a scene in the movie Moneyball where all the old baseball scouts are giving their thoughts on which players the team should buy, based only on the scouts’ experience, opinion and personal judgment. They ignore the nerd in the corner who has real data and figures … and some insight. They even question if he has to be in the room. In the movie, the data analyst is disparaged, even though he doesn’t bring an opinion or intuition to the table. He brings data analysis, and the data don’t care how long you have been in the business.

Too many Compstat meetings are reminiscent of this scene. The centerpiece of many Compstat meetings is a map of crime that many are viewing for the first time. A room full of people wax lyrical on the crime problem based on their intuitive interpretation of a map of crime on the wall, and then they promote solutions for our beleaguered commander, based too often on opinion and personal judgement and too little on knowledge of the supporting evidence of the tactic’s effectiveness. Because everyone knows they have to come back in a month the strategies are inevitably short-term in nature and never evaluated. And without being evaluated, they are never discredited, so they become the go-to tactical choice ad infinitum.

So the problems with Compstat are weak analysis, rushed decision-making, and opinion-driven strategies. What might the solutions be?

The U.K.’s National Intelligence Model is a good starting point for consideration. It has a strategic and a tactical cycle. The strategic meeting attendees determine the main strategic aims and goals for the district. At a recent meeting a senior commander told me “We are usually too busy putting out fires to care about who is throwing matches around.” Any process that has some strategic direction to focus the tactical day-to-day management of a district has the capacity to keep at least one eye on the match thrower. A monthly meeting, focused on chronic district problems, can generate two or three strategic priorities.

A more regular tactical meeting is then tasked with implementing these strategic priorities. This might be a weekly meeting that can both deal with the dramas of the day as well as supervise implementation of the goals set at the strategic meeting. It is important that the tactical meeting should spend some time on the implementation of the larger strategic goals. In this way, the strategic goals are not subsumed by day-to-day dramas that often comprise the tyranny of the moment. And the tactical meeting shouldn’t set strategic goals—that is the role of the strategic working group.

I’ve previously written that Compstat has become a game of “whack-a-mole” policing with no long-term value. Dots appear, and we move the troops to the dots to try and quell the problem. Next month new dots appear somewhere else, and we do the whole thing all over again. If we don’t retain a strategic eye on long-term goals, it’s not effective policing. It’s Groundhog Day policing.

Year-to-date comparisons and why we should stop doing them

Year-to-date comparisons are common in both policing and the media. They involve comparing the cumulative crime count for the current year up to a certain date and comparing to the same point in the preceding year. For a Philadelphia example from April of this year, NBC reported that homicides were up 20 percent in 2017 compared to 2016. You can also find these types of comparison in the Compstat meetings of many police departments.

To gauge how reliable these mid-year estimates of doom-and-gloom are, I downloaded nine years (2007-2015) of monthly homicide counts from the Philadelphia Police Department. These are all open data available here. I calculated the overall year change as well as the cumulative change monthly from year to year. In the table below you can see a row of annual totals in grey near the bottom, below which is the target prediction as a percentage of the previous year (white text, blue background). For example, the 332 homicides in 2008 were 14.7% lower than the previous year, expressed in 2007 terms.

Let’s determine that we can tolerate our prediction to be within 5 percent plus or minus the eventual difference between this year and the preceding year. That stipulates a fairly generous 10% range as indicated by the Low and High rows in blue.

Each month you can see the percentage difference between the indicated year-to-date at the end of the month, and the calendar year-to-date (YTD) for the same period in the previous year. So for example, at the end of January 2008 we had 21.9% fewer homicides than at the end of January 2007. By the time we get to December, we obviously have all the homicides for the year, so the December percentage change exactly matches the target percentage difference.

Cells highlighted with a green background have a difference on the previous year that is within our +/- 5 percent tolerance. By the end of each January, we only had one year (2012) with a percentage difference that was within 5 percent of how the city ended the year. The 57% increase in January 2011 was considerably different that the eventual 6% increase over 2010 at the end of December. When Philadelphia Magazine dramatically posted “Philly’s Murder Rate is Skyrocketing Again in 2014” on January 14th of that year, the month did indeed end up nearly 37 percent over 2013. But by year’s end, the city had recorded just one homicide more than the preceding year – a less dramatic increase of 0.4%.

In fact, if we seek out a month where the difference is within our 10% range and later months will remain consistently accurate through to the end of the year, then we have to wait until the months shown with a border. 2009 performed well, however while 2010 was fairly accurate throughout the summer, the cumulative totals in September and October were more than 5% higher than the previous year when the year ended only 0.3% higher.

To use calendar YTD comparisons with any confidence, we have to wait until the end of October before we can be more than 50% confident that the year-to-date is indicative of how we will enter the New Year. And even then we still have to be cautious. There was a chance at the end of November 2010 that we would end the year with fewer homicides, though the eventual count crept into increase territory.

The bottom line is that with crimes such as homicide, we need not necessarily worry about crime panics at the beginning of the year. This isn’t to say we should ever get complacent and of course every homicide is one too many; however the likely trend will only become clear by the autumn.

Alternatives exist. Moving averages seem to work okay, but another alternative I like is to compare full (annual) YTDs to the prior annual (i.e. full 12 month) YTD. So instead of (for example) comparing January-April 2010 to January to April 2009, you could compare the 12-month change May 2009-April 2010 against the May 2008-April 2009 total. I’ve done that in the red graph below. The first available point is December 2008 and as we know from the previous table, the preceding 12 months had outperformed the annual year 2007 by 14.7%. But then each subsequent month measures not just the calendar YTD but the 12-month YTD.

The result is a graph that shows the trend changing over time from negative (good) territory to positive (bad for homicides because it show an increase). Not only do you get a more realistic comparison that is useful throughout the year, you can see changing trend. Anything below the horizontal axis is good news – you are doing well. Above it means that your recent 12 months (measured at any point) was worse than the preceding 12 months.

You can have overlapping comparison periods. The graph in blue below compares 24 months of accumulated counts with the 24 month totals for the previous year. For example, the first point available is December 2009. This -11.7% value represents the change in total homicides from the 24 months January 2008 to December 2009 and compares it to the 24 month total from a year previous to this (January 2007 to December 2008). For comparison purposes, I have retained the same vertical scale but note the change in horizontal axis.

You can see there is more smoothing, but the general trend over time is still visible. Lots of variations available and you might want to play with different options for your crime type and crime volume.