Assessing the counterfactual: The efficacy of drug interdiction absent racial profiling

Duke Law Journal

Volumn 54, Issue 5, 2005, Pages 1089-1141

  Barnes, Katherine Y ^a  

^a UNIVERSITY OF WASHINGTON   (United States)

Author keywords

[No Author keywords available]

Indexed keywords

EID: 33747876338     PISSN: 00127086     EISSN: 00127086     Source Type: Journal    
DOI: None     Document Type: Article

References (170)

1. Jerome H. Skolnick & Abigail Caplovitz, Guns, Drugs, and Profiling: Ways to Target Guns and Minimize Racial Profiling, 43 ARIZ. L. REV. 413, 419 n.36 (2001).
2. RICHARD MIDDLETON, COLONIAL AMERICA: A HISTORY, 1607-1760, at 283-91 (1992).
3. See Exec. Order No. 9066, Feb. 19, 1942, 7 Fed. Reg. 1407 (Feb. 25, 1942) (authorizing the secretary of war to designate "military areas" from which "any and all persons may be excluded"); General DeWitt's Public Proclamation No. 1, March 2, 1942, 7 Fed. Reg. 2320 (March 26, 1942) (designating military areas on the West Coast); see also Civilian Exclusion Order No. 34, May 3, 1942, 7 Fed. Reg. 3967 (May 28, 1942) (excluding persons of Japanese ancestry from Alameda County, California).
4. This belief has little support in historical fact. See Eric L. Muller, Betrayal on Trial: Japanese-American "Treason" in World War II 82 N.C. L. REV. 1759, 1760-62 (2004) (discussing the extremely limited data on disloyalty among the Japanese and Japanese Americans interned).
5. See, e.g., Proclamation No. 2526, Dec. 8, 1941, 6 Fed. Reg. 6323 (Dec. 10, 1941) (referring to Germans); Proclamation No. 2527, Dec. 8, 1941, 6 Fed. Reg. 6324 (Dec. 10, 1941) (referring to Italians). These measures were taken under the auspices of Executive Order 9066, the same order that granted authority for the Japanese American internment. Indeed, on its face, Executive Order 9066 applied to all people, as it did not mention particular ethnicities and only referenced "enemy aliens" in passing. Exec. Order No. 9066, Feb. 19, 1942, 7 Fed. Reg. 1407 (Feb. 25, 1942).
6. A small number of German and Italian Americans, primarily recent immigrants, were also interned during the war. See generally Proclamation No. 2526 (German enemy aliens); Proclamation No. 2527 (Italian enemy aliens).
7. One might suggest, for example, that people of Chinese ancestry would be less likely to engage in espionage and sabotage for the Japanese. This is, of course, speculation, as was the assumption that the Japanese Americans interned in World War II were more likely to be saboteurs, but both of these examples simply suggest that we do not know what the empirical data would demonstrate.
8. Recent scholarship acknowledges that some response may have been prudent, but that the extreme measures used were too drastic. Eric L. Muller, Inference or Impact? Racial Profiling and the Internment's True Legacy, 1 OHIO ST. J. CRIM. L. 103, 116-17 (2003).
9. See Korematsu v. United States, 323 U.S. 214, 220 (1944) ("[W]hen under conditions of modern warfare our shores are threatened by hostile forces, the power to protect must be commensurate with the threatened danger.").
10. In 2001, blacks accounted for 34.5 percent of the drug violation arrests and 44.5 percent of theprison and jail population, but only 13.2 percent of the general population. FED. BUREAU OF INVESTIGATION, CRIME IN THE UNITED STATES 252 t.43 (2001), available at http://www.fbi.gov/ucr/01cius.htm;
11. ALLEN J. BECK ET AL., PRISON AND JAIL INMATES AT MIDYEAR 2001, U.S. DEP'T OF JUSTICE, BUREAU OF JUSTICE STATISTICS BULLETIN 12 t.14 (April 2002), available at http://www.ojp.usdoj.gov/bjs pub/pdf/pjim01.pdf;
12. U.S. CENSUS BUREAU, POPULATION ESTIMATES 1, available at http://www.census.gov/popest/national/asrh/NC-EST2003/NC-EST2003-03.pdf.
13. Self-reported data is based upon a random sample of individuals, rather than a biased selection method, and therefore is not systematically biased to overrepresent black and Hispanic drug use.
14. In 2001, 6.8 percent of whites and 6.9 percent of blacks surveyed reported using illegal drugs in the last month. U.S. DEP'T OF HEALTH & HUMAN SERVS., OFFICE OF APPLIED STUDIES, 2001 NATIONAL SURVEY ON DRUG USE & HEALTH: DETAILED TABLES fig.2.12 (on file with the Duke Law Journal).
15. One extreme example of an inappropriate police response to a witness description involving race is found in Brown v. City of Oneonta, 221 F.3d 329 (2d Cir. 1999), in which the Second Circuit upheld the dismissal of Equal Protection claims against the city. Id at 334. In Brown, the victim, robbed at knifepoint, described her attacker as a young black man and said that his right hand had been cut during the robbery. Id The police responded by stopping over two hundred black individuals and asking to see the individual's right hand. Id The City of Oneonta had fewer than three hundred black residents (not all young men), and approximately one hundred fifty black college students (also not all men) residing within its borders. Id While deeply troubling, the police department's use of race in this case was not racial profiling as I define it. For an argument that removing witness identification from the definition of racial profiling masks the racial justice issues at stake, see generally R. Richard Banks, Race-Based Suspect Selection and Colorblind Equal Protection Doctrine and Discourse, 48 UCLA L. REV. 1075 (2001).
16. Skolnick & Caplovitz, supra note 1, at 419 n.36.
17. Confirmation Hearing on the Nomination of John Ashcroft to Be Attorney General of the United States: Hearing Before the Senate Comm. on the Judiciary, 107th Cong. 492 (2001) (response of Senator Ashcroft, stating that racial profiling is "wrong and unconstitutional no matter what the context").
18. See, e.g., DAVID A. HARRIS, PROFILES IN INJUSTICE: WHY RACIAL PROFILING CANNOT WORK 78-90 (2002) (citing hit rate statistics to demonstrate that "[r]acial profiling is neither an efficient nor an effective tool for fighting crime");
19. William H. Buckman & John Lamberth, Challenging Racial Profiles: Attacking Jim Crow on the Interstate, 10 TEMP. POL. & CIV. RTS. L. REV. 387, 388 (2001) (noting that "as few as one in thirty stops net contraband even as small as a single joint" and that "[t]his figure could be achieved in random stops of all travelers");
20. Deborah A. Ramirez et al., Defining Racial Profiling in a Post-September 11 World, 40 AM. CRIM. L. REV. 1195, 1197-1201 (2003) (reviewing empirical data).
21. See, e.g., Heather Mac Donald, The Racial Profiling Myth Debunked, CITY J., Spring 2002, at 63-64 ("It turns out that the police stop blacks more for speeding because they speed more. Race has nothing to do with it."), available at http://www.cityjournal.org/html/eon_3_ 27_02hm.html;
22. Heather Mac Donald, The Myth of Racial Profiling, CITY J., Spring 2001, at 26 ("Arlington has a 10 percent black population, but robbery victims identify nearly 70 percent of their assailants as black [and a]s long as those numbers remain unchanged, police statistics will also look disproportionate. "), available at http://www.cityjournal.org/html/11_2_the_myth.html;
23. John Derbyshire, In Defense of Racial Profiling: Where Is Our Common Sense?, NAT'L REV. ONLINE, Feb. 19, 2001, at http://www.nationalview.com/ 19feb01/derbyshire021901.shtml ("So long as race is only one factor in a generalized approach to the questioning of suspects, it may be considered.").
24. See, e.g., John Knowles, Nicola Persico, & Petra Todd, Racial Bias in Motor Vehicle Searches: Theory and Evidence, 109 J. POL. ECON. 203, 205 (2001) [hereinafter KPT] (proposing a test to distinguish between racially prejudiced searches and searches that use race as a factor only because it helps to "maximize successful searches"). For a more detailed discussion of the economics literature regarding racial profiling, see infra Part I.B.
25. There is, of course, a huge leap from the descriptive statement "the terrorists were young Arab men" and the speculative statement "young Arab men are more likely to be terrorists." Two points should be made. First, not all terrorists are young Arab men. As has been noted elsewhere, the second deadliest terrorist attack on U.S. soil in recent history was committed by Timothy McVeigh, a European American. Samuel R. Gross & Katherine Y. Barnes, Road Work: Racial Profiling and Drug Interdiction on the Highway, 101 MICH. L. REV. 651, 749(2002). Second, the percentage of terrorists who are Arab men is likely quite different from the percentage of Arab men who are terrorists. The former may be reasonably large-again, a supposition-and may well be a larger percentage than 0.25 percent, the percentage of the U.S. population constituting Arab men, but the latter is almost certainly quite small.
26. See Samuel R. Gross & Debra Livingston, Racial Profiling Under Attack, 102 COLUM. L. REV. 1413, 1414 & n.3 (2002) (gathering examples of the post-September 11 discussion of the benefits of racial profiling).
27. Bernard E. Harcourt, Rethinking Racial Profiling: A Critique of the Economics, Civil Liberties, and Constitutional Literature, and of Criminal Profiling More Generally, 71 U. CHI. L. REV. 1275, 1276-77(2004). While a generalization, this dichotomy provides a useful outline of the recent empirical literature on racial profiling. Perhaps the exception to my dichotomy is Professor Harcourt's recent article itself, which provides a theoretical economic model of racial profiling. See id. at 1291-95, 1315-22 (dividing empirical racial profiling scholarship into "economics" and "civil liberties" strains). Although I do not use the same terminology as Professor Harcourt, the categories are essentially the same.
28. John Lamberth's multiple studies of racial profiling also fall within this category and have the same methodological flaws. For a list of his many studies, most commissioned by police departments themselves, see http://www.lamberthconsulting.com/about.racial-profiling/research-articles.asp.
29. The term "hit rates" is shorthand for the proportion of searched vehicles in which the police found contraband to the total number of searched vehicles.
30. HARRIS, supra note 16, at 78-82.
31. Id. at 80. Professor Harris also argues that racial profiling makes for bad policing-that, as economists would say, it is irrational and inefficient-although he does not engage the theoretical possibility that racial profiling could be rational. Id at 79.
32. Gross & Barnes, supra note 19, at 690-91 (describing a scenario in which equal hit rates would not indicate racial profiling);
33. Harcourt, supra note 21, at 1316-17 (criticizing Harris' focus on hit rates).
34. For an extended discussion of this proposition, see Gross & Barnes, supra note 19, at 690-91.
35. See infra Part III.A and Figure 2 for a graphical example of how selection bias works.
36. See Buckman & Lamberth, supra note 16, at 394 (providing a blueprint for challenging racial profiling in court that relies solely on racial comparisons);
37. Gross & Barnes, supra note 19, at 694 (asserting that "[t]he most likely explanation for this [racial] disparity is the obvious one: Maryland State troopers took race into account in deciding who to stop").
38. Gross & Barnes, supra note 19, at 692.
39. Id at 694. We assume that similar national rates of drug use across races imply similar base rates of drug possession on I-95. Id at 691.
40. Id. at 692.
41. One might think of economists' focus on rationality as a determination of whether racial profiling has the limited benefit of being efficient, rather than based solely on racial animus.
42. KPT, supra note 18, at 227-28.
43. Id. at 205.
44. Id. at 228. KPT's model is flexible enough to allow for different metrics of police success. KPT focus on hit rates-the percentage of searches that find contraband-but also estimate the role of racial animus in police behavior assuming that the police want to maximize the percentage of searches that find not just any quantity of drugs, but a quantity sufficient to constitute a felony. Id. at 224-26. In contrast to their hit rates model, KPT find that felony hit rates for whites and Hispanics are significantly lower than the black felony hit rate, suggesting that rational police behavior would be to search blacks more often. Id. at 226. However, KPT discount this conclusion based on felony hit rates, because it assumes police gain nothing from finding a small amount of drugs. Id.
45. KPT also investigate the profiling of several other groups, including men and individuals driving luxury cars, and find no evidence of discrimination for these groups. Id. at 222-23. Economist David Bjerk criticizes KPT's approach, noting that the model requires troopers to oversample black motorists even though, in equilibrium, black and white motorists have the same hit rates. David Bjerk, Racial Profiling, Statistical Discrimination, and the Effect of a Colorblind Policy on the Crime Rate 14-15 (Mar. 25, 2005) (working paper), available at http://soerserv.socsci.memaster.ca/bjerk/paper.html. According to Bjerk, KPT's model is directly contradictory to a straightforward definition of racial unbiasedness, where troopers search blacks and whites in proportion to their relative likelihood to carry drugs. Id. at 16-37. This criticism, however, ignores both the dynamic equilibrium KPT's model presents, and the difficulty of accurately assessing the likelihood of carrying drugs for blacks and whites.
46. KPT, supra note 18, at 216. This is an earlier version of the search data I use in Parts II and III, although KPT do not use the stop data at all.
47. Id. at 228.
48. Id. at 219-22. KPT's analysis does find that Hispanics have a lower hit rate than whites, which they conclude is evidence of racial animus against Hispanics. Id. at 222.
49. Id. This prediction is a consequence of KPT's model, in which each driver faces the same cost/benefit tradeoff of being caught with drugs versus choosing not to carry drugs. Id. at 211.
50. Police use of profiles, however, is still helpful in deterring criminal behavior.
51. Id. at 205. Economists Kate Antonovics and Brian Knight suggest that this is the legally salient question. Kate L. Antonovics & Brian G. Knight, A New Look at Racial Profiling: Evidence from the Boston Police Department 2-3 (Nat'1 Bureau of Econ. Research, Working Paper No. 10,634, 2004), available at http://papers.nber.org.papers/W10634. On a practical level, Antonovics and Knight may simply be suggesting that causation is difficult to prove in racial profiling cases; a rational use of race might, in fact, be a rational use of some unobserved characteristic of the motorists that is simply correlated with race. Nonetheless, whether the police engage in racial profiling is an important threshold question, as it determines what level of scrutiny is applied in an Equal Protection challenge. See Grutter v. Bollinger, 539 U.S. 306, 326 (2003) (reaffirming that the use of race in government decisionmaking triggers strict scrutiny under the Equal Protection Clause). At best, the KPT model is agnostic with respect to whether race is an explicit factor in the trooper's decision to search.
52. Several economic studies build on KPT's work. See, e.g., Vani K. Borooah, Racial Bias in Police Stops and Searches: An Economic Analysis, 17 EUR. J. OF POL. ECON. 17, 32-33 (2001) (estimating a model similar to KPT's using data on stops and searches in ten separate areas in England);
53. Dhammika Dharmapala & Stephen L. Ross, Racial Bias in Motor Vehicles Searches: Additional Theory and Evidence, 3 CONTRIBUTIONS TO ECON. ANALYSIS & POL'Y, No. 1, art. 12 (2004), at 4-7, available at http://www.bepress.com/ bejeap/conributions/vol.iss1/art12/ (extending KPT to account for the fact that potential offenders can bypass the highway altogether, thus avoiding detection);
54. Harcourt, supra note 21, at 1354-71 (mapping out whether racial profiling is rational, and which population should be profiled, using a series of potential elasticities and offending rate parameters);
55. Rubin Hernandez-Murillo & John Knowles, Racial Profiling or Racist Policing? Bounds Tests in Aggregate Data, 45 INT'L. ECON. REV. 959, 960 (2004) (extending the KPT model to situations in which aggregate, rather than individual-level data, is available);
56. Nicola Persico & Petra Todd, Using Hit Rates to Test/or Racial Bias in Law Enforcement: Vehicle Searches in Wichita 1 (Nat'1 Bureau of Econ. Research, Working Paper No. 10,947, 2004), available at http://papers.nber.org/ papers/W10947 (extending the KPT model to allow for heterogeneous police and motorists).
57. See, e.g., KPT, supra note 18, at 205-06 (assuming that "the police maximize the number of successful searches, [and] net the cost of searching motorists").
58. Antonovics & Knight, supra note 12, at 3;
59. Shamena Anwar & Hanming Fang, An Alternative Test of Racial Prejudice in Motor Vehicle Searches: Theory and Evidence 3-4 (Nat'1 Bureau of Econ. Research, Working Paper No. 11, 264, 2004), available at http:papers.nber.org/ papers/W11264.
60. Antonovics & Knight, supra note 12, at 4;
61. Anwar & Fang, supra note 42, at 6. Because identification of their full model is impossible, Antonovics and Knight assume that black and white troopers are equally prejudiced. That is, they assume that the difference in search cost of black and white motorists is equal, and they compare the white trooper/black motorist search rate to the black trooper/white motorist search rate.
62. Antonovics & Knight, supra note 12, at 12.
63. Technically, both Professors Anwar and Fang and Professors Antonovics and Knight use the search rate for stopped cars, that is, the ratio of the number of vehicles searched to the number of vehicles stopped. Antonovics & Knight, supra note 12, at 15-16, 32;
64. Anwar & Fang, supra note 45, at 25.
65. Antonovics & Knight, supra note 42, at 19-20.
66. Anwar & Fang, supra note 45, at 26.
67. Antonovics and Knight, in particular, identify racial animus on the basis of different search behavior in cross-race interactions between troopers and motorists, Antonovics & Knight, supra note 12, at 12-13, and therefore their results are quite sensitive to this assumption. They address this criticism by providing evidence that the troopers do not have a different amount of information about cross-race motorists. Id. at 23-24.
68. A taste for discrimination, under these models, reduces hit rates because the troopers have a lower threshold of suspicion to search black motorists, and therefore will search more innocent black motorists.
69. KPT do investigate the alternative goal of increasing felony hit rates, but it is not the focus of their article. See supra note 36.
70. See infra Part III.E.
71. Harcourt, supra note 21, at 1303-06.
72. Id. at 1279, 1301-02.
73. Id. at 1279-83, 1329-35.
74. Id. at 1303-05.
75. Id at 1303. While theoretically possible to incorporate the costs of the ratchet effect into the model via the rather amorphous "social cost" of crime, it is clear from Harcourt's exposition thatthis is not his intent. Similarly, Harcourt does not question whether there are more efficient ways to discourage drug trafficking than by searching vehicles on the highway, which deters only drug trafficking on the highway.
76. Id. at 1276-77.
77. For example, commendations and promotions may well be tied to high-profile drug busts. In addition, individual police departments will not necessarily take total social cost into account; they may simply work to keep the criminal activity out of their neighborhood. The neighborhood is, after all, their constituency.
78. As narrowly defined, economic efficiency simply balances the direct cost to searching, in terms of the trooper's time, against the direct benefit of searching, in terms of the amount of drugs seized. But cf. Harcourt, supra note 21, at 1300 (arguing that economists define efficiency too narrowly in the context of the criminal justice system).
79. Data collection has become a focus of antiracial profiling efforts over the past several years. See, e.g., DEBORAH RAMIREZ ET AL., A RESOURCE GUIDE ON RACIAL PROFILING DATA COLLECTION SYSTEMS: PROMISING PRACTICES AND LESSONS LEARNED (2000), available at http:www.ncjjs.org/pdffiles1/bja/184768.pdf. For an up-to-date list of jurisdictions that track at least some information related to racial profiling, see the Racial Profiling Data Collection Resource Center at the Northeastern University website, http:www.racialprofilinganalysis.neu.edu.
80. NICHOLAS LOVRICH ET AL., WSP TRAFFIC STOP DATA ANALYSIS PROJECT 39-111 (June 1, 2003), available at http://www.wsp.wa.gov/reports/wsptraff.pdf. Professors Antonovics and Knight also explicitly model the decision to search using individual-level data. See Antonovics & Knight, supra note 41, at 15-16. Their model, however, does not control for other important variables beyond the race of the trooper and the race of the motorist because they do not have the detailed data necessary to do so. Their model is therefore significantly underspecified.
81. LOVRICH ET AL., supra note 63, at 93.
82. Id. at 107-08.
83. Id. at 109-10. They also control for the level of discretion in the search and find that race has the same effect in low-discretion searches (searches incident to arrest, impound searches, and warrant searches) as in high-discretion searches (consent searches, Terry searches, and K-9 searches). They find this inconsistent with the premise of racial profiling, which would suggest that more discretion allows race to play a greater role in the decisionmaking process. As the authors also believe this to be inconsistent with prior literature on discretionary searches, they posit that they have not controlled for all the factors that are driving the decision to search. Id. It is worth noting that this prior literature is based upon the intuition of many commentators, rather than empirical fact.
84. For further discussion of the Heckman selection model, see infra notes 109 & 111 and accompanying text.
85. I use a probit model to estimate the probability pi that a car with characteristics Xi will be searched. The probit model estimates a model of the form: where is the vector of estimated coefficients that determine the effect of each on and is the inverse cumulative distribution function of the Normal distribution.
86. That is, if one assumes that stops are random events, and further assumes that the relative search rates of stopped vehicles across race are equal, one would underestimate the overall differential search rate for blacks if blacks were profiled at both stages of the process. For a discussion of the salience of the empirical difference between stopped vehicles and all vehicles on the highway, see infra note 110 and accompanying text.
87. The dataset is available at http://law.wustl.edu/Academics/Faculty/ Barnes/index.html.
88. Wilkins v. Md. State Police, No. CCB-93-468 (D. Md. filed 1993) (settled by consent decree in 2003). Its companion case was filed in 1998. NAACPv. Md. Dep't of State Police, No. CCB-98-1098 (D. Md. filed Apr. 10, 1998) (settled by same consent decree in 2003).
89. Wilkins, No. CCB-93-468.
90. Although the search dataset contains information on all searches throughout Maryland, I use only the searches on a portion of 1-95, in order to match the search dataset to the geographically limited stops dataset.
91. See Memorandum in Support of Plaintiffs' Motion for Enforcement of Settlement Agreement and for Further Relief, Wilkins v. Md. State Police (D. Md. 1997) (Civil Action No. CCB-93-468), at Parts III-IV (arguing that the 1-95 corridor and MSP troop numbers 62 (the JFK Barracks) and 55 conduct the vast majority of searches) (on file with the Duke Law Journal);
92. Order, Wilkins v. Md. State Police (D. Md. Apr. 22, 1997) (Civil Action No. CCB-468), 1 (finding a pattern or practice of racial discrimination in vehicle searches on the 1-95 corridor) (on file with the Duke Law Journal);
93. see also Elliott Wolf, ACLU, The Use of Racial Profiling by the Maryland State Police in Drug Interdiction from 1995-2002, at 2-3 (discussing stop and search patterns of the JFK Barracks) (on file with the Duke Law Journal).
94. For the purposes of this study, the primary problem with the stops and searches databases maintained by the MSP is that the two datasets are not linked. Thus, there is no variable in the stops data that indicates whether a search was conducted, or what that search found. One large data project has been the matching of searches to stops based on the time, location, direction of travel, trooper, vehicle make, and race and sex of the driver. In order to match a search to a particular stop, I require that the search occur within two hours of the stop, that the direction of travel match, if available in both datasets, and that the location of the stop be within five miles of the location of the search (to account for variations in data collection). In addition, I match searches based on four additional factors: make of vehicle, name of trooper, race of driver, and sex of driver. I use the order of this list as a tiebreaker. Thus, if two different stops match three criteria for the same search, the stop that matches all but the sex of the driver is preferred to the stop that matches all but the race of the driver, etc. With this rather stringent algorithm, I am able to match 1,248 of 2,583 total searches, or 48.3 percent of the searches on I-95 during the appropriate time period. Stop data from July 1, 2002, through December 31, 2003, did not contain the time of the stop; for this subset of data, stops were matched by date and the above criteria. For the stops and searches before July 1, 2002, when time data was available, 931 of 1486 searches were matched, for a match rate of 62.6 percent. For all of the data, under the assumption that the search data is more accurate, if the data sets diverge on control variables, I use the search data for any value that does not match the stop data.
95. For example, evidence suggests that blacks are slightly more likely to speed than whites, making a white bias less likely. See Gross & Barnes, supra note 19, at 664, 687-88.
96. I do not have data for two particular characteristics of the driver, age and income, and therefore cannot control directly for either. I control for the driver's income indirectly through the make and age of the vehicle driven. I do not have a proxy for age in the data, however, so I cannot control for that variable. As blacks and Hispanics are, on average, younger populations than are whites, if MSP select younger drivers for search more often, this could be masked as a selection on the basis of race.
97. Luxury models include Acura, Audi, Bentley, BMW, Cadillac, Hummer, Infiniti, Jaguar, Land Rover, Lexus, Lincoln, Mercedes, Porsche, Range Rover, Rolls Royce, Saab, Sterling, and Volvo.
98. Large trucks include International, Mack, Peterbilt, and Kenworth trucks.
99. I experimented with different definitions of "older car"; none made a substantive change in the results of the model.
100. There is, for example, a widely held belief that red cars cost more to insure because their drivers are, on average, more reckless. See Carlnsurance.com, FAQs, http://www.carinsurance.com/kb/content10059.aspx (last visited Oct. 17, 2005) (dispelling this myth).
101. In the tables that follow, I denote these categories as Maryland Plates, Northeast Plates, New York Plates, Southeast Plates, DC Plates, and Non-East Coast Plates, respectively.
102. David Kocieniewski, New Jersey Argues That the U.S. Wrote the Book on Race Profiling, N. Y. TIMES, Nov. 29, 2000, at A1; see also United States v. Wilson, 853 F.2d 869, 875 (11th Cir. 1988) (discussing a DEA course that taught Georgia officers that "drug couriers are frequently Hispanics"); United States v. Layman, 730 F. Supp. 332, 334, 337 (D. Colo. 1990) (noting that officers were trained by DEA agents to use drug courier profiles that included race as a factor); TASK FORCE ON GOV'T OVERSIGHT, OPERATION PIPELINE REPORT (1999), available at http://www.aclunc.org/discrimination webb-report.html (noting the California Highway Patrol's use of racial profiling as part of the DEA's "Operation Pipeline" protocol).
103. See Gross & Barnes, supra note 19, at 687-88 (discussing this possibility).
104. Blacks may be more likely to be stopped because of unsafe driving, but unless unsafe driving is positively correlated with drug possession, there is no reason to search unsafe drivers disproportionately more often.
105. Many of the stops include more than one violation.
106. The models I report in Tables 1-2 and 4-8 infra are estimated using the full dataset and drop nighttime as an independent variable. I indicate in footnotes when the alternative model-including nighttime as a variable but excluding stops and searches made after June 30, 2002-provides different results.
107. Out of the 240 troopers who stopped more than 100 vehicles, 20 of the troopers never conducted a search. Thus, variables for these 20 troopers perfectly predict no search (or a probability of searching equal to zero). I therefore drop the stops by these troopers from the model, as they provide no further explanatory power. A total of 14,781 stops were dropped from the full dataset for this reason.
108. But see LOVRICH ET AL., supra note 63, at 107-10 (describing potential means by which "non-discretionary" searches may still allow room for discretionary decisionmaking).
109. A p-value represents the probability that a result as extreme or more extreme would occur when there was, in fact, no relationship between the independent variable and the decision to search. For example, in this case, the p-value of less than 0.0005 represents the probability that after controlling for other relevant variables, the estimated difference in search rates between blacks and whites would be as great or greater if there were no relationship between race and the decision to search. A coefficient with a p-value of 0.05 or less is generally described as statistically significant.
110. If a very powerful independent variable that is highly correlated with race were omitted from the model, the conclusion of racial profiling would be invalid. But this is very unlikely, especially when race is a stronger predictor of searches than variables that are part of the DEA's drug interdiction profile.
111. The relative risk of a search is the additional risk faced by a driver with a specific characteristic, expressed as a ratio, compared to the baseline of a white male motorist driving a newer car with Maryland plates who did not get a ticket. For each variable, the relative risk is given by: Pr (Search given variable (e.g., black))/Pr (Search given baseline)
112. In the standard statistical practice, these variables are not interaction terms, but instead are dummy variables for each demographic group, comparing the group (Hispanic women, for example) against white men. This is equivalent to the standard statistical practice of using "interaction" terms, in which a variable labeled "Hispanic X female" would estimate whether the difference in search rates between Hispanic men and women was the same as the difference in search rates between white men and women. I report dummy variables in the tables to make the demographic differences more transparent. Other relative risks may be found by division: Relative Risk of Variable A to Variable B = Relative Risk of Variable A/ Relative Risk of Variable B. Thus, for example, the relative risk of a Hispanic man being searched as compared to a white woman is: 3.48/0.47 = 7.4.
113. A p-value of 0.000 simply means that the true p-value is less than 0.0005; rounded to three significant digits, this is 0.000.
114. In the alternative model that includes nonconsent searches and controls for these probable cause searches, whether a vehicle was speeding is a statistically significant predictor of a search. Specifically, speeders are less likely to be searched, with a relative risk of 0.40. This implies that nonconsent searches are disproportionately nonspeeders (or, conversely, consent searches are disproportionately performed on speeders) and therefore may suggest that speeding is being used as a pretext for stops.
115. Very few Hispanic motorists were stopped driving luxury vehicles; only 275 Hispanic drivers driving luxury vehicles were stopped, as compared with over 14,000 whites and 7,000 blacks. With such a small amount of data, the coefficient for Hispanic motorists driving luxury vehicles is not statistically significant. Based upon the magnitude and direction of the estimated effectHispanic motorists in luxury cars are over ten times as likely to be searched after being stopped than their white counterparts-the same pattern may hold, but without more data, it is impossible to determine whether this is the case.
116. This is statistically significant, with a p-value of 0.003.
117. The estimates provided in this table hold all other variables constant at their medians (the baseline value).
118. See, e.g., Mac Donald, Myth of Racial Profiling, supra note 17, at 20 (arguing that racial profiling increases the amount of drugs seized).
119. The cut points determining the boundaries between "user" and "courier" are based on government estimates of the average prices and the amounts spent by users of various contraband drugs in 1998. See WILLIAM RHODES ET AL., OFFICE OF NATIONAL DRUG CONTROL POLICY, WHAT AMERICA'S USERS SPEND ON ILLEGAL DRUGS 1988-1998, at 12, 16, 22-23 (2000). The cut points were chosen to balance the cost versus weight for different drugs. For marijuana, the cut point is 455g = $5,120 or five years' supply. For crack and powder cocaine, the cut point is 50g = $7,450 or about nine months' supply. For heroin, the cut point is 10g = $10,290 or about eleven months' supply. Finally, for other drugs, the amounts were receded, very roughly, into dosage units; a cut point of 150 dosage units defines a drug courier.
120. See Gross & Barnes, supra note 19, at 696 n.152, for further details.
121. I recognize that this is not a complete measure of the costs of racial profiling, even on an individual level. Other costs, not easily quantified, include the loss of dignity, the stigma of being searched, and the deterioration of police-citizen relationships in minority communities.
122. See HARRIS, supra note 16, at 84-86 (noting that in "stops and searches," police rarely find evidence of crime);
123. Gross & Barnes, supra note 19, at 750-52.
124. Or, at least, where crack and powder cocaine are. The pattern is similar if one looks at all drugs. Thirty-nine out of forty-nine drivers found with large amounts of drugs are black or Hispanic.
125. See supra p. 1107 for the discussion of selection bias inherent in the dataset.
126. Cf. Kathy Barrett Carter & Ron Marsico, Whitman Fires Chief of State Police, STAR LEDGER (New Jersey), Mar. 1, 1999, at 1A (providing an example of comments by the chief of the New Jersey State Police on the racial stratification of drug trafficking);
127. C.J. Chivers, Ex-Police Leader's Claim of Bias Attacked, N.Y. TIMES, Oct. 4, 1999, at B4 (noting that other documents also describe drug trafficking along racial lines).
128. I define the threshold amount of drugs to be considered a drug courier by reference to government estimates. See supra note 100. These thresholds are based upon a combination of street price and number of dosage units. Gross & Barnes, supra note 19, at 696 n.152.
129. This searching strategy is a form of racial profiling that biases the data. Racial profiling is just one selection mechanism that would bias the data; the model I use controls for whatever selection criteria the MSP actually use.
130. The hypothetical data are simply for illustrative purposes. I created the hypothetical data with two goals: first, in order to make the final graph-Panel C-demonstrate the same offense level across minority status and second, to create a reasonably accurate racial mix of highway drivers.
131. See generally James J. Heckman, Sample Selection as a Specification Error, 47 ECONOMETRICA 153 (1979).
132. The selection model does not control for the potential selection bias inherent in data of all stopped vehicles, rather than all vehicles-i.e., the bias due to the initial selection of which vehicles to stop. This selection bias is almost certainly less salient than the bias introduced by the decision to search; the troopers have much less information when deciding to stop a vehicle than when deciding to search a stopped vehicle. To test this assumption, I re-ran the Heckman models using only stops and searches that occurred at night, when the troopers have even less information and have more difficulty discerning the race of a motorist before stopping the vehicle. The substantive results do not change significantly. Because only slightly less than half the data is used, however, the statistical significance for several variables is lower.
133. This is not to belittle the idea in any way; indeed, James Heckman won the 2000 Nobel Prize in Economics for his insights in this area. See Nobelprize.org, The Bank of Sweden Prize in Economic Sciences in Memory of Alfred Nobel: 2000(noting that Heckman won "for his development of theory and methods for analyzing selective samples"), at http://www.nobel.se. economics/laureates/2000 (last modified Apr. 14, 2005). But Heckman, too, recognizes the limits of his model. See James Heckman & Bo Honore, The Empirical Content of the Roy Model, 58 ECONOMETRICA 1121, 1122 (1990) (noting that the Heckman model can be identified by model assumptions alone, and that care must be taken to make sure that the data are driving the results).
134. As an example, suppose that the selection equation-just like the one estimated in Part II.B-estimates that Troopers A and B search 5 percent of the vehicles they stop, and find, respectively, an average of two grams and four grams of cocaine per stop, after controlling for other factors. Suppose further that the model estimates that Trooper C searches 10 percent of the vehicles stopped, and finds seven grams of cocaine on average, after controlling for other factors. The Heckman selection model then assumes that a motorist's chance of being searched is correlated with the amount of drugs the motorist carries. The Heckman selection model estimates that motorists who have a 10 percent chance of being searched carry about seven grams of cocaine, and motorists who have a 5 percent chance of being searched carry about three grams of cocaine. The model extrapolates to motorists who have even lower (or higher) chances of being searched, and thereby provides a best estimate of what the unsearched motorists stopped on the highway are carrying.
135. Identification, in statistical terms, is an important concept concerning what information is driving the results: information from the data or from the model assumptions. Ideally, one wants to identify the results from the data rather than the model assumptions, which are somewhat arbitrary.
136. This does not mean that racial profiling needs to target blacks or other minorities. It may be that whites carry larger quantities of drugs than blacks or Hispanics do, implying that profiling whites would increase the quantity of drugs seized.
137. Thus, I control for sex of driver, direction of travel, the vehicle's region of origin, whether the vehicle is a luxury model, and the interaction of luxury vehicle with race, age of vehicle, and whether the driver received a ticket. See supra Part II.A.
138. Because crack and powder cocaine represent the bulk of the drugs seized, I ran several alternate models using the total amount of drugs seized, with different composite measures of total drugs seized (equalizing either estimated number of dosages or estimated value of drugs seized). The substantive results were unchanged. For ease of discussion, I concentrate on powder drugs in this portion of the analysis.
139. To estimate the model, I transform the cash value of the drugs seized to the logarithmic scale in order to stabilize the variance. Specifically, I use the transformation y = log(cash value +1) to avoid taking the logarithm of zero. In the tables that follow, however, I transform the logged value back to the standard scale for ease of discussion.
140. To make this estimate, I simply calculate the estimated average value of drugs carried by each type of motorist and sum across all nonsearched vehicles.
141. Given the number of nonsignificant relationships found and the magnitude of the p-value, even this relationship may represent random variation.
142. See supra note 83.
143. See supra Part II.B.
144. Statistically significant variables and their corresponding values are shown in bold.
145. This column represents the average additional (above the baseline) value of drugs that a driver of this description will have, holding all other variables fixed at their medians, which is the baseline value.
146. This is the average value of drugs seized for "baseline" motorists: white men speeding north in their newer, nonluxury cars with Maryland plates.
147. Being nervous when stopped seems ubiquitous; in this context, an individual stopped on the highway is "nervous" if the officer conducting the search lists nervousness as one of the grounds for the search. The results of the model confirm the intuition that after controlling for other factors, nervousness of the driver, as reported by an MSP trooper after the fact, is not correlated with the quantity of drugs carried.
148. As would be true with any model, the data can only determine the profile in use during the time period of data collection. Thus, the data show that the profile used from May 1997 to December 2003 did not mirror the profile that would have yielded the maximum quantity of drugs seized. The data cannot determine whether the profile was appropriate at some prior time, but then became obsolete as drug couriers adapted to the policing strategy. The issue of such dynamic behavior is beyond the scope of this paper.
149. In both cases, the individual would be eligible to be fried as a drug kingpin. MD. CODE ANN., GRIM. LAW §§ 5-612, 5-613 (2002) (setting a threshold of 448 grams, or 16 ounces, of cocaine to be tried and sentenced as a "drug kingpin").
150. Gross & Barnes, supra note 19, at 751-52 (providing a rough estimate that the 33 kilograms seized per year represented about 0.5 percent of the cocaine consumed in the Baltimore/D.C. metropolitan area per year).
151. See supra note 100 (describing the cut points used to discern a drug user from a drug courier). A small subset of the drivers carry more than one drug. I define these individuals to be drug couriers if any one of the drugs they carry meets the appropriate threshold. If, instead, I were to combine the drug amounts by standardizing the amount of each separate drug seized to a percentage of the amount necessary to be deemed a drug courier, no additional drivers would be labeled drug couriers.
152. In this case, I use a variant of the standard Heckman selection model that allows me to estimate a model with a binary dependent variable. See G.S. MADDALA, LIMITED-DEPENDENT VARIABLES AND QUALITATIVE VARIABLES IN ECONOMETRICS 221-23 (1986) (describing the Heckman selection model).
153. Statistically significant variables and their corresponding values are shown in bold.
154. The baseline probability of a trooper arresting a drug courier after stopping the vehicle is 1.0 percent.
155. Again, in the dataset, no Hispanic women-out of 3 searched-carried a large quantity of drugs. Thus, the model's best estimate for the relative risk of being a drug courier for Hispanic women is zero, even though the true relative risk is very likely nonzero, although still small.
156. Similarly, in the dataset, no white men driving luxury vehicles-out of 29 searched-carried a large quantity of drugs. They did, however, carry just under this amount of drugs. Thus, the model's best estimate for the relative risk of being a drug courier for white men driving luxury vehicles is zero, even though the true relative risk is very likely nonzero, although still small.
157. Once again, I use any type of drug found in this analysis, rather than limiting the analysis to powder drugs found.
158. See, e.g., HARRIS, supra note 16, at 78-82 (citing hit rate statistics to argue that "[r]acial profiling is neither an efficient nor an effective tool for fighting crime");
159. Antonovics & Knight, supra note 12, at 6-7 (same);
160. Gross & Barnes, supra note 19, at 689-93 (same);
161. KPT, supra note 18, at 219-24 (modeling whether a motorist carries drugs).
162. Before discussing what the three models I estimate in this Part together imply about the MSP's profiling strategy, let me address one further point. The model predicts that the baseline rate of innocence, for white men from Maryland, is 73 percent, or, equivalently, that 27 percent of stopped vehicles have some contraband. This is a very high percentage. The model also estimates that the MSP's selection criteria do not bias the results; the MSP would not get different results if they randomly searched stopped vehicles. These two facts are consistent: the underlying base rate of offense for stopped vehicles is approximately equal to the offense rate of those vehicles that were selected to be searched. This high offense rate suggests that some selection is happening at earlier stages-in particular, in the trooper's decision to stop a speeding car, or in the motorist's decision to carry drugs while driving on 1-95. Most likely, the decision to stop is not random and is correlated with having drugs. For example, reckless drivers may be overrepresented in both stopped vehicles and vehicles that contain drugs. For this to be a problem with my analysis of the use of race in profiles, however, the overrepresentation of reckless drivers in stops also needs to correlate with race. While I do not have individual-level data on the motorists who were not stopped, and therefore cannot control for selection using a Heckman-style model, I did re-analyze the results using only stops and searches that occurred at night and obtained very similar results. Assuming that it is much more difficult to ascertain the race of motorists at night before stopping them, this suggests that my analysis of the racial element of the MSP's profiling is robust to the selection bias inherent in choosing which vehicles to stop.
163. Profiling black and Hispanic motorists may also have created this problem in the first place, in that black and Hispanic motorists may choose to carry drugs less often because they are searched more often. This is the type of game theoretic argument that economists use in their models. Even if the MSP changed their policy, however, this would not necessarily change the behavior of black and Hispanic motorists: after all, they drive through many jurisdictions quite quickly. Only a wholesale change in the practice of most jurisdictions would lead to a significant change in drug possession.
164. In the alternate model that uses both consent and nonconsent searches, probable cause is the largest determinant of innocent rates. Vehicles searched due to probable cause have innocent drivers 1.5 times more often than vehicles subject to consent searches.
165. Statistically significant variables and their corresponding values are shown in bold.
166. As in Table 1, this value represents the baseline probability that a white male motorist stopped on the northbound highway in a newer inexpensive car with Maryland plates is carrying drugs.
167. This column represents the relative risk of finding and arresting a drug courier, as compared to the baseline value, which holds all variables at their medians.
168. This column represents the relative risk of searching an innocent motorist with the given characteristics compared to the baseline value, which holds all variables at their medians.
169. All cells marked "-" have no statistically significant effect.
170. It is possible that the decision to write a ticket is made after a search is performed, in which case this factor would not be salient.