Discrimination Data Analysis: A Multi-disciplinary Bibliography - Discrimination and Privacy in the Information Society - page 114

the next two subsections. We refer to (Hildebrandt & Gutwirth, 2008) for a cross-
disciplinary perspective of automated profiling.

6.5.1 Racial Profiling

Profiling is an illegal practice as soon as its application results in direct or indirect
discrimination against protected groups. In this section, we concentrate on racial
profiling , defined as “the practice of subjecting citizens to increased surveillance or
scrutiny based on racial or ethnic factors rather than reasonable suspicion” (J. Chan,
2011). Among several possible contexts of racial profiling, vehicle stops have at-
tracted the vast majority of studies 6 . Numerous data collection efforts have been
initiated by law enforcement agencies, often as a result of litigation or of legislation,
for the purpose of understanding the vehicle stop practices of its officers. Attributes
collected concern the stop (time, date, location, reason, duration), driver (race, gen-
der, age), vehicle (make, model), officer (age, gender, race, education, experience),
and the outcome of the stop (e.g., warning, citation, arrest, search, seizure of contra-
band). The objective of data analysis is to identify racial patterns of disparity. One
of the early surveys on racial profiling is due to (Engel et al., 2002). More recent
papers include (Farrell & McDevitt, 2010; Tillyer et al., 2010), reviewing vehicle
stops approaches. The adequacy of statistical analysis of racial profiling in address-
ing legal issues is also discussed in (Tillyer et al., 2008). For a legal comparison of
US and EU laws, see (Baker & Phillipson, 2011).
(Tillyer et al., 2010) categorize existing approaches depending on whether they
deal with the initial decision or with the outcome of a stop.
In initial stop studies, the actual rate of stops by drivers' race is compared with
benchmark data providing the expected rate of stops assuming no police bias. The
outermost difficulty of the approach consists of identifying accurate benchmarks of
the expected driver population at risk of being stopped. (Engel & Calnon, 2004),
and (R. M. Blank et al., 2004, Chapter 9) outline strengths and limitations of six
primary data sources and their use in the design of benchmark data: census data, ob-
servations of roadway usage, official accident data, assessments of traffic violating
behavior, citizen surveys, and internal departmental comparisons. Alternative means
for collecting benchmark data are proposed in (Alpert et al., 2004; Jobard & Levy,
2011; Quintanar, 2009; Ridgeway & MacDonald, 2009; Gelman et al., 2007).
Post-stop outcome studies focus on the identification of racial disparities in a spe-
cific outcome of the stop by taking as reference population the whole set of stops.
An example of post-stop outcome analysis consists of checking whether the search
for drugs among stopped vehicles is biased against the driver's race. In this respect,
starting from the influential paper proposed in (Knowles et al., 2001), several ex-
tensions and critiques have been presented (Antonovics & Knight, 2009; Anwar
& Fang, 2006; Gardner, 2009; Rowe, 2008; Sanga, 2009). We refer to the surveys

6

Other contexts include profiling in airport security (Gabbidon et al., 2011; Persico & Todd,
2005), fraud investigators (Leopold & Meints, 2008), capital sentences (Alesina & Ferrara,
2011), and consumer profiling (Gabbidon et al., 2008; Schreurs et al., 2008).