Testing hot‐spots police patrols against no‐treatment controls: Temporal and spatial deterrence effects in the London Underground experiment

• Date: 01 February 2020
• Author: Lawrence W. Sherman,Mark Newton,Barak Ariel
• Published date: 01 February 2020
• DOI: http://doi.org/10.1111/1745-9125.12231

Received: 27 July 2018 Revised: 31 July 2019 Accepted: 6 August 2019

DOI: 10.1111/1745-9125.12231

ARTICLE

Testing hot-spots police patrols against no-treatment

controls: Temporal and spatial deterrence effects in

the London Underground experiment∗

Barak Ariel1Lawrence W. Sherman2Mark Newton3

1Jerry Lee Centre for Experimental

Criminology; Institute of Criminology,

University of Cambridge, and Institute of

Criminology, Faculty of Law, Hebrew

University of Jerusalem

2Jerry Lee Centre for Experimental

Criminology; Institute of Criminology,

University of Cambridge, and Department of

Criminology and Criminal Justice, University

of Maryland—College Park

3Rail Delivery Group

Correspondence

BarakAr iel, Jerry Lee Centre for Experimental

Criminology,Institute of Criminology, Uni-

versityof Cambridge, Cambridge CB3 9DA,

U.K.

Email:ba285@cam.ac.uk

Fundinginformation

Jerry Lee Centre of Experimental Criminology

∗Wewould like to express our gratitude to the

British Transport Police,wit h special thanks for

their contribution and long-term commitment

toevidence-based policing and experimen-

tal criminology.Generous support for this

researchwas provided by The Jerry Lee Centre

ofExper imental Criminology.We particularly

wishto t hank Jerry Lee, for his commitment

toexperiment al criminology and the global

advancementof evidence-based policing. We

alsot hank DavidWeisburd, Alex Sutherland,

Crispian Strachan,Geoff Bar nes, the Crimi-

nologyeditors, and t he anonymousreviewers

oft his article fort heir useful comments. Weare

indebtedto John MacDonald for his insight-

fulstatistical advice and cr itical read of earlier

versionsof this ar ticle.

Abstract

Our understanding of causality and effect size in random-

ized field experiments is challenged by variations in levels

of baseline treatment dosage in control groups across exper-

iments testing similar treatments. The clearest design is to

compare treated cases with no-treatment controls in a sam-

ple that lacks any prior treatment at baseline. We applied

that strategy in a randomized test of hot-spots police patrols

on the previously never-patrolled, track-level platforms

of the London Underground (LU). In a pretest–posttest,

control-group design, we randomly assigned 57 of the LU’s

115 highest crime platforms to receive foot patrol by offi-

cers in 15-minute doses, 4 times per day, during 8-hour

shifts on 4 days a week for 6 months. The effect of 23,272

police arrivals at the treatment hot spots over 26 weeks was

to reduce public calls for service by 21 percent on treated

platforms relative to controls, primarily when police were

absent (97 percent of the measured effect). This effect was

six times larger than the mean standardized effect size found

in the leading systematic review. This finding provides a

benchmark against the baseline counterfactual of no patrol

in hot spots, with strong evidence of residual deterrence and

no evidence of local displacement.

KEYWORDS

baseline dosage, hot spots, no-treatment controls, randomized experi-

ments, regional deterrence, residual deterrence

Criminology. 2020;58:101–128. wileyonlinelibrary.com/journal/crim © 2019 American Society of Criminology 101

102 ARIEL ET AL.

What is the logic of causality in hot-spots policing experiments? That question has become a prime case

in point for debating what conclusions can and cannot be drawn from randomized field experiments

(Nagin & Sampson, 2019). Although some may see the three-decade history of hot-spots experiments

(Braga, Papachristos, & Hureau, 2012; Sherman & Weisburd, 1995) as providing highly certain con-

clusions, others may see substantial uncertainties in the same evidence. As Nagin and Sampson (2019,

p. 141) concluded, “[I]t is important to be clear with policy makers and our fellow scientists about the

uncertainties in the evidentiary base even where experimental evidence forms the point of departure

for the translation of results to policy.” The uncertainties they identified included the long-term effects

of a whole-city system of hot-spots policing, which has never been compared with cities without such

a system. They also cited the risk of “cognitive narrowing,” by which the logic of causality is confused

with research methods or designs, in which a randomized trial is seen to be not just a “gold standard”

but also a magic bullet that lends certainty to any claim of causality.

In this article, we address a form of uncertainty that has not yet been debated even though the issue is

central to the logic of causality in hot-spots policing. Our concern is that the effect size of redistributing

police patrol presence depends just as much on the “dosage” of patrol in a controlg roup as itdoes on t he

dosage in the treatment group. That logic is independent of whether a test is randomized, of whether

it is short term or long term, or of exactly what police do (or do not do) when they are present in

hot spots. It is also independent of whether “causality” is defined in the ambitious terms proposed

by Nagin and Sampson (2019, p. 124): “[T]he difference between counterfactual worlds that emerge

as a consequence of their being subjected to different universal treatment regimens over a sustained

period of time, what economists would call system-wide equilibrium differences.” Our purpose is not

to contest that definition of causality but to accept it as one of several possible valid definitions, as

well as to use that definition to reassess and enlarge the body of experimental evidence on hot-spots

policing.

The logic of comparing counterfactual worlds must begin with a clear description of both worlds

being compared. In our view, the major uncertainty in hot-spots policing experiments is the general

failure to provide such clarity. The lack of clarity has several dimensions, including the absolute size

and variance of hot-spots boundaries, the frequency and types of crime or calls for service analyzed,

and even the harm level of crimes in each location (Weinborn, Ariel, Sherman, & O’Dwyer, 2017).

None of those characteristics of the sample locations, however, are as theoretically important to the

concept of causality as are the baseline level of the independent variable (police patrol dosage) in the

control (C) group and the corollary ratio of patrol dosage in the T group to the level in the C group.

This ratio between T and C dosage has two dimensions. One is the baseline of dosage in both areas

prior to beginning the experiment, which should ideally start out as equivalent at the point of random

assignment. The other is the size of the gap created between T and C by implementing the experiment’s

random assignments. Of these two dimensions, we suggest that the absolute baseline level is the essen-

tial starting point in developing the counterfactual logic of the experiment. That, in turn, offers more

help to reduce the uncertainties about the estimated effect sizes of adding patrol dosage to hot spots,

at least in the short run, in any community that already provides police patrol.

The logic of causality in hot-spots policing experiments is often vaguely described as testing whether

“adding more” police to high-crime locations prevents more crime compared with business-as-usual

(BAU) levels of patrol in similarly high-crime locations. Systematic efforts to integrate the results

of experiments so described are of great interest (Braga et al., 2012). Yet the uncertainties of any

meta-analysis of the currently available research are massive. These uncertainties begin with the lack

of any measurement, in most trials, of patrol dosage in even the experimental treatment (T) group.

More fundamentally, the uncertainties depend on patrol dosage in the counterfactual:thelevelof

police presence in the control (C) group. Absent this information, any estimates of average effect size