racial disparities in taxicab tipping

7/31/2019 Racial Disparities in Taxicab Tipping

1/52

Social Science Research Network Electronic Paper Collection at:

http://papers.ssrn.com/abstract=401201

Center for Law, Economics and Public Policy

Research Paper No. 276

Public Law and Legal Theory

To Insure Prejudice: Racial Disparities in Taxicab Tipping

This paper can be downloaded without charge from:

Yale Law School

Ian Ayres, Fredrick E. Vars and Nasser Zakariya

Research Paper No. 50
http://papers.ssrn.com/abstract=401201http://papers.ssrn.com/abstract=401201http://papers.ssrn.com/abstract=394842http://papers.ssrn.com/abstract=394282http://papers.ssrn.com/abstract=384595http://papers.ssrn.com/abstract=373604http://papers.ssrn.com/abstract=362080http://papers.ssrn.com/abstract=349500http://papers.ssrn.com/abstract=325640http://papers.ssrn.com/abstract=329680http://papers.ssrn.com/abstract=316120http://papers.ssrn.com/abstract=283822


2/52

1

To Insure Prejudice:Racial Disparities in Taxicab Tipping

Ian Ayres, Fredrick E. Vars and Nasser Zakariya*

Abstract:

We collected data on over 1000 taxicab rides in New Haven, CT in 2001. Aftercontrolling for a host of other variables, we find two potential racial disparities in tipping:(1) African-American cab drivers were tipped approximately one-third less than whitecab drivers; and (2) African-American passengers tipped approximately one-half theamount of white passengers (African-American passengers are 3.7 times more likely thanwhite passengers to leave no tip).

Many studies have documented seller discrimination against consumers, but this study

tests and finds that consumers discriminate based on the sellers race. African-Americanpassengers also participated in the racial discrimination. While African-Americanpassengers generally tipped less, they also tipped black drivers approximately one-thirdless than they tipped white drivers.

The finding that African-American passengers tend to tip less may not be robust toincluding better controls for passenger social class. But it is still possible to test for theracialized inference that cab drivers (who also could not directly observe passengerincome) might make. Regressions suggest that a rational statistical discriminatorwould expect African Americans to tip 56.5% less than white passengers.

These findings suggest that government-mandated tipping (via a tip included decal)might reduce two different types of disparate treatment. First, mandated tipping woulddirectly reduce the passenger discrimination against black drivers documented in thisstudy. Second, mandated tipping might indirectly reduce the widely-documentedtendency of drivers to refuse to pick up black passengers.

*Ian Ayres is the Townsend Professor of Law at Yale Law School. Fredrick Vars is an associate at theChicago law firm of Miller Shakman & Hamilton. Nasser Zakariya is a fellow at the Yale Law SchoolCenter for the Study of Corporate Law. Please send comments to: [email protected]. This article isdedicated to Underhill Moore and Suzanne Perry. Underhill Moore took to the streets of New Havenduring the 1930s to see whether people observed parking meter regulations. See John Henry Schlegel,

American Legal Realism and Empirical Social Science: The Singular Case of Underhill Moore, 29 BUFF.L. REV. 195 (1980). Nearly seventy years later, Perry conducted a pilot study of taxi and pizza deliverytipping that was the inspiration and foundation for the present effort. The authors thank Aditi Bagchi,Caroline Harada, Lee Harris, and Ian Slotin for their heroic efforts as auditors. Jennifer Brown, EmmaColeman, Neil Katyal, and seminar participants at Georgetown Law School provided helpful comments.


3/52

2

Table of Contents

Introduction......................................................................................................................... 1I. Race and the History of Tipping................................................................................ 4II. Description of Data .................................................................................................... 7

III. Results...................................................................................................................... 10A. Lower Tips for Minority Drivers .......................................................................... 11B. Lower Tips By Minority Passengers..................................................................... 11C. Driver and Passenger Racial Intersections............................................................ 13D. Regression Analysis.............................................................................................. 15

IV. Alternative (Non-Racial) Hypotheses...................................................................... 23A. Censored Data? ..................................................................................................... 23B. Lower Tips for Minority Drivers .......................................................................... 25C. Lower Tips by Minority Passengers ..................................................................... 27

V. Why Are Consumers Discriminating?..................................................................... 31VI. Normative Implications ........................................................................................... 34

A. Adding Insult to Injury?........................................................................................ 34B. Service Compris.................................................................................................... 361. Implementation ................................................................................................. 362. Reducing Driver Discrimination....................................................................... 373. Countervailing Effects ...................................................................................... 46

Conclusion ........................................................................................................................ 48


4/52

1

Introduction

It has become increasingly common to test whether sellers in retail markets discriminateagainst buyers.1 But this paper, to our knowledge, is the first to test the other side of the

market.2

We test whether retail consumers discriminate against sellers on the basis of thesellers race. Even though Gary Becker, long ago understood that consumers taste fordiscrimination could cause sellers to discriminate against other customers3 forexample, leading restaurant owners to maintain racially segregated lunch counters noone has tested whether consumers taste for discrimination might be directed at a sellersrace itself (or the race of a sellers employees).

This failure to test should not be surprising. Consumer-side race discrimination tests arehard to implement. Consumer refusals to deal are difficult to document.4 In manycontexts, it would be infeasible to use an audit procedure where two different-race sellersmade identical offers to an individual buyer. And the ability of consumers to

discriminate in the terms of contracting is usually severely constrained. Oftenconsumers only promise to pay a non-discretionary price. In contrast, tipping is anunusual and natural place to test for consumer-side discrimination, because it is a

1 Peter Siegelman, Race Discrimination in Everyday Commercial Transactions: What Do We Know,What Do We Need to Know, and How Can We Find Out, in A NATIONAL REPORT CARD ONDISCRIMINATION IN AMERICA:THE ROLE OF TESTING (Michael Fix & Margery Austin Turner eds., 1998);

see also John Yinger, Evidence of Discrimination in Consumer Markets, 12 J. ECON PERSPECTIVES 23(1998); IAN AYRES, PERVASIVE PREJUDICE? UNCONVENTIONAL EVIDENCE OF RACE AND GENDERDISCRIMINATION (2001).2 Some studies have indirectly inferred the presence of consumer discrimination. See, e.g., Clark Nardinelli& Curtis Simon, Customer Racial Discrimination in the Market for Memorabilia: The Case of Baseball,

105 QUARTERLY J.ECON. 575, 576 (1990) (The appeal of sports for the study of discrimination is that itis possible to separate consumer discrimination from the ability to do the work.); Lawrence Kahn & PeterSherer, Racial Differences in Professional Basketball Players Compensation, 6 J. LABOR ECON. 40, 42(1988) ([A]ll else equal, white representation on a team contributes to home attendance, providingevidence consistent with the idea of consumer discrimination.). Employment audits are non-retail tests ofwhether consumers of labor (i.e., employers) discriminate on the basis of seller race. Keith R. Ihlanfeldt,Madelyn V. Young,Intrametropolitan Variation in Wage Rates: The Case of Atlanta Fast-Food RestaurantWorkers, 76 REV.ECON. & STAT.425,425 (1994) (Evidence on discrimination suggests that consumer

prejudice affects the wages paid to black workers); see also John Yinger, Measuring RacialDiscrimination with Fair Housing Audits: Caught in the Act, 76 AM. ECON. REV. 881, 881 (1986)(Housing agents cater to the racial prejudice of current or potential white customers.).3 GARY S.BECKER,THE ECONOMICS OF DISCRIMINATION (1971).4 There have been some important sociological studies analyzing consumer preferences for dealing with

sellers of particular ethnic and/or racial groups. See Jennifer Lee,From Civil Relations to Racial Conflict:Merchant-Customer Interactions in Urban America, 67 AM. SOC. REV., 77 (2002); ST. CLAIR DRAKE &HORACE R. CAYTON, BLACK METROPOLIS ([1945] 1993); A.E. McCormick & G.C. Kinloch, InterracialContact in the Customer-Clerk Situation, 126 J.SOC.PSYCH. 551 (1986). More recently, there has beendiscussion of the rise of FUBU (for us, buy us) consumerism which at heart is a movement of race-contingent consumer choice. Jerre B. Swann, Sr. et al., Trademarks and Marketing,91 Trademark Rep.787, 802 (2001) (FUBU (For Us, By Us) brand clothing became popular in the African-Americancommunity in part by tapping into the sense of cultural unity and authenticity that wearing the brandfostered.). But these studies tend to be qualitative, failing to measure the degree of preference orstatistical tests of its significance.


5/52

2

dimension of consumer economic behavior that is both discretionary and potentiallyobservable.

We collected data on over 1000 tips to taxicab drivers in New Haven, Connecticut in2001. After controlling for a host of other variables, we find two potential racial effects:

(1) African-American cab drivers were tipped approximately one-third less than whitecab drivers; and (2) African-American passengers tipped approximately one-half theamount of white passengers.5

African-American passengers also seemed to participate in the racial discriminationagainst African-American drivers. While African-American passengers generally tippedless, they also tipped black drivers approximately one-third less than they tipped whitedrivers.

The propensity to stiff by which we mean to leave no tip was particularlyracialized. African-American drivers were 80% more likely to be stiffed than white

drivers (28.3 vs. 15.7%). And African-American passengers were almost 4 times morelikely than white passengers to leave no tip at all (39.2 vs. 10.6%).

Several caveats, however, are in order before accepting these racial interpretations of thedata. First, the data were based on cab drivers reports. Cab drivers racial stereotypes orpreconceptions may have led them to systematically under-report black passenger (orover-report white passenger) tips. Second, we could only crudely control for passengersocial class. What we attribute to passenger race may be caused at least in part by atendency of poorer people to tip less. And third, we do not have strong controls fordriver quality. Lower tips by African-American passengers might be explained by ageneral tendency of passengers to give lower tips for poorer service, coupled with driversproviding inferior service to African-American passengers. And if the white drivers inour study had cleaner cars or for some other reason provided better service, that couldexplain the race-of-driver disparity.

We will respond to each of these alternative hypotheses below by bringing to bearadditional pieces of evidence. Audit testing of the participating drivers provides limitedevidence that the drivers were providing equal service to passengers of different races.And recent surveys of consumers suggest that even after controlling for consumer incomeand quality of service, African Americans tend to tip substantially less than whites. Inthe end, we believe the study provides strong prima facie evidence of the two highlightedracial effects that is, of passenger disparate treatment against African-American driversand a lower African-American propensity to tip generally. But this study is far from thefinal word.

5 In the main body of the Article, we will discuss the tipping behavior with regard to other passenger anddriver races. The results with regard to Hispanic passengers are similar to those of African-American

passengers. But both for the sake of brevity and because no other races were as well represented in ourdata, we limit the discussion of the results in the introduction to just whites and blacks.


6/52

3

If we tentatively accept the finding of a race-of-driver disparity, a natural question to askis Why? The data are not well suited to answer this question, but they do contain someclues. First, the higher propensity of passengers to stiff black drivers seems moreconsistent with conscious decision making than the less visible tendency of non-stiffingpassengers to tip black drivers a lower percentage. Because driver allocation was more

or less random, it does not appear that some passengers simply do not tip anybody rather, there seem to be passengers who are more likely to decide not to tip African-American drivers.

In contrast, another portion of the disparity may resonate more with unconsciousdisparate treatment. Passengers tend to round up their tips (to the nearest dollar abovetheir target level) more often when tipping white drivers than when tipping black drivers(32.3% vs. 24.6%).6 When confronted with a last-second decision (based on the finalfare) about whether to round up or round down, even passengers who believe they arehard-wired 15% tippers may in practice unconsciously allow the drivers race to impacttheir rounding decision.7 When we decompose the overall racial disparity in tips received,

we find that racial disparities in stiffing and rounding account respectively for about 27%and 36% of the overall disparity.

It is less clear that we should accept the evidence that African-American passengers tendto tip less. But it turns out that this seemingly racial result may nonetheless haveimportant policy implications. While passenger poverty instead of race may really bedriving our finding that African-American passengers tend to tip less, it is important toemphasize that, like us, cab drivers also cannot directly observe passenger wealth. Butthey can observe passenger race and a variety of non-racial measures (such as pickuplocation, dress, etc.). While our limited data do not allow us to conclusively addresswhether poverty is the true cause of the result, our data do allow us to estimate whatkind of statistical inferences a cab driver would make about the size of the likely tip giventhe observable characteristics of the passengers. Our statistical discriminationregressions suggest that rational drivers might expect to earn a 56.5% lower tip from anAfrican-American passenger than from a white passenger (after controlling for a host ofnon-racial observable characteristics). Overall, in our data a driver should expect about13.8% lower revenue when stopping to pick up an African-American passenger (relativeto a white passenger).

This result has policy relevance because such driver inferences may play a role in thewell-documented refusal to deal with minority passengers.8 The data suggest that at leasta portion of driver-side discrimination may be caused, not by animus or by (rational orirrational) statistical inferences about crime, but instead by inferences about how much

6 These percentages are taken over all fares, so that 24.6 percent of all black driver fares are rounded up tothe nearest dollar above the passengers target level.7

This finding parallels the results for the Implicit Association Tests (IATs) which analogously suggest

that unconscious racial influences affect timed sorting decisions.Anthony G. Greenwald et al.,MeasuringIndividual Difference in Implicit Cognition: The Implicit Association Test, 74 J. PERSONALITY & SOC.PSYCH. 1464 (1998); AYRES,supra note 1, at 184.8 See infra at Part VI.B.2.


7/52

4

passengers of different races are likely to tip. Indeed, we will show that this revenueeffect is orders of magnitude greater than any rational inferences that might be madeabout the propensities of passengers of different races to rob cab drivers.

Our two core racial findings are relevant to a single normative implication. These

findings suggest that government-mandated tipping (via a tip included decalprominently posted in cabs) might reduce two different types of disparate racialtreatment. First, mandated tipping would directly reduce the passenger discriminationagainst black drivers documented in this study. Second, mandated tipping mightindirectly reduce the tendency of drivers to refuse to pick up black passengers at least tothe extent that this driver discrimination is caused by statistical inferences aboutdifferences in tipping.

There are, however, at least two reasons to pause before eliminating discretionarytipping. First, although research suggests that tips are not strongly correlated with qualityof service, tipping (at least in theory) may induce better service. Second, poorer

individuals, whose rides are currently subsidized by passengers who do tip, will be lessable to afford the increased fare under a mandatory tipping regime. (To the extentminority individuals tend to be less wealthy, this shift would have a disparate racialimpact.)

The remainder body of the paper is divided into six parts. Part I briefly reviews the roleof race in the history of tipping in the United States. Part II describes the data collectedfor this study. Part III presents the core results highlighting both the racial and non-racial determinants of cab driver tipping. Part IV considers alternative, non-racialhypotheses. Part V explores what might be causing the customer discrimination. Andfinally, Part VI discusses normative implications of the findings.

I. Race and the History of Tipping

Tipping is a substantial component of our economy. More than thirty service professionsare regularly tipped.9 Restaurant tips alone in the United States have been estimated at$26 billion a year.10 The tipping norm is now broadly accepted both as a matter of equity to increase the wages of workers in the service industry and as a matter of efficiency to increase the quality of service.11 People tend not to know what percentage of income

9 Michael Lynn et al., Consumer Tipping: A Cross-Country Study, 20 J.CONSUMERRES. 478 (1993).10 Ofer H. Azar, The Social Norm of Tipping: A Review (unpublished manuscript, 2003), available athttp://www.papers.ssrn.com.11Uri Ben Zion & Edi Karni, Tip Payments and the Quality of Service, in O.C. Ashenfelter & W.E. Oates,ESSAYS IN LABOR MARKET ANALYSIS 37 (1977), explicitly modeled a repeated interaction between acustomer who chooses how much to tip and a service agent who chooses how much effort to provide toshow how a marginal reward for effort could induce the service agent to provide more than the minimaleffort level.


8/52

5

their parents give to charity, but they know their parents tipping percentage. 12 Indeed,parents often explicitly tell their children how much they should tip in various settings.13

But what is less well known is that the social practice of tipping was much morecontroversial 100 years ago. Critics referred to the practice as un-American and

incompatible with democracy.

14

Former Yale Law Professor, William Howard Taft wasthe patron saint of the anti-tip crusade15 and Ralph Waldo Emerson roundly condemnedthe practice: I sometimes succumb and give the dollar, yet it is a wicked dollar which byand by I shall have the manhood to withhold.16 Tipping was attacked as bribery and astraining school for graft.17

In the early twentieth century 7 states and the District of Columbia passed anti-tippingstatutes that to varying degrees outlawed the practice.18 Today many patrons and workersin the service industry look upon the tipping practice, not only as non-stigmatizing, butindeed as a workers entitlement for work well done. At the turn of the last century, incontrast, tipping was often viewed as a marker of degradation. Both the giving and the

receiving of tips were perceived as an acceptance of the recipients inferiority.

19

In The

12 BARRYNALEBUFF &IAN AYRES,WHYNOT?SIMPLE METHODS OF EVERYDAY INGENUITY (forthcoming2003). Researchers have explored a variety of server strategies that can enhance restaurant tipping:

In one well-known 1984 experiment, researchers found that a waitress who touched her customers,whether male or female, on the hand or shoulder when asking if the meal was all right, raised hertips to 14 per cent, from 11 per cent. . . . So does crouching at the table when taking an order or, ifthe server is a women, putting a smiley face on the bill. For male waiters, the smiley face cuts tips.

William Grimes, The Tip: A Reward, But for Whom?, N.Y.TIMES, at A16 (Feb. 24, 1999).13Id.14 Segrave,supra note 11, at 5-6 (What, may I ask, is more un-American than tipping? It doesnt belong inAmerican society; it doesnt belong in a democracy. It is the product of lands where for centuries there has

been a servile class.). Scott, supra note 11, at 43 (In the American democracy to be servile isincompatible with citizenship. Every tip given in the United States is a blow at our experiment indemocracy.)15 Taft An Anti-Tipper, N.Y. TIMES 2 (June 20, 1908) (his barber observed Never a tip did he give. Iunderstand that he thinks he has paid for the work when he gives the regular price and I guess he is right.).16Regulating Tips, 45 SCRIBNERS MAGAZINE 252 (Feb. 1909).17 In 1920 William Rufus Scott launched the Commercial Bribery and Tipping Review. Scott argued thattipping was not only a form of bribery whereby one customer sought unfair advantage over another but a

breeding ground for social corruption more generally. Scott wondered whether a messenger who thoughtthe public owed him gratuities would develop into a man with sound morals. Segrave, supra note 11, at43:

There is a direct connection between corruption in elections and the custom of tipping. The manwho lives upon tips will not see the dishonesty of selling his vote.

Scott,supra note 10, 148.18 Stephanie Cook,A History of 'Handing it Over', CHRISTIAN SCIENCE MONITOR (Oct. 23, 2000) ([M]anyAmericans loathed the custom, branding it un-American and undemocratic. The Anti-Tipping Society ofAmerica, an alliance of 100,000 traveling salesmen, managed to have tipping abolished in seven statesfrom 1905 to 1919.);see also KERRY SEGRAVE,TIPPING 29 (1998).19 The Oxford English Dictionary in 1933 defined tip as a small present of money given to an inferior. 18OXFORD ENGLISH DICTIONARY 134 (1933). [Tipping] makes the daily income of the worker dependentupon his subservience to the passing humor of the customer. It promotes fawning and sycophancy, andkills dignity and independence. SEGRAVE,supra 18, at 35.


9/52

6

Itching Palm, a 1916 manifesto against the practice, William R. Scott said that tipping is[the] willingness to be servile for a consideration.

This degradation conception of tipping was intimately tied to race. Kerry Segrave hastaken the lead in excavating this history:

[A] Guntons Magazine article in 1896 . . . remarked that in the United States wehave been comparatively free from this offensive semi-mendicancy. However,tipping was a growing custom among a certain class of laborers such as domesticservants, coachmen, barbers, waiters, and railroad porters: It will be observedthat these occupations are nearly all filled by foreigners and negroes who for themost part have been reared under the patronizing and semi-feudal influences ofpaternal or ante-wage condition. Centuries of slavery had left blacks in menialjobs while Europeans were menial workers due to the aristocratic, patronizingconditions of Europe.20

For some, the practice of tipping was intimately connected to the perceived inferiority ofAfrican Americans. In 1902, for example, a Southern journalist named John Speedremarked:

I have never known any but Negro servants. Negroes take tips, of course; oneexpects that of them -- it is a token of their inferiority. But to give money to awhite man was embarrassing to me. I felt defiled by his debasement andservility. Indeed, I do not know how any native-born American could consent totake a tip. Tips go with servility, and no man who is a voter in his country bybirthright is in the least justified in being in service.21

The modern tipping norm was incubated in a history rife with explicit racism -- as can beseen in the public prominence given to a seemingly insignificant vignette, again reportedby Segrave:

In 1907 Senator Tillman of South Carolina tipped black porter George Hollister25 cents as he departed Omaha, Nebraska's Paxton Hotel. Tillman was wellknown for maintaining that he never tips a nigger. Familiar with the senator'sviews, Hollister joked that he would have the quarter made into a watch charm.So newsworthy was the event that theNew York Times editorialized that Tillman'sprevious position was not meanness. We must assume that the Senator abstainsfrom tipping as a matter of principle. Tipping blacks was not unusual then, andpeople on long journeys, said the editor, frequently were forced to convertthemselves into fountains playing quarters upon the circumambient Africans.22

The practice of tipping far from being perceived as a way of increasing the pay ofservice workers -- was frequently seen as an employer strategy for exploiting its workers,

20 SEGRAVE,supra note 12, at 6.21 John Gilmer Speed, Tips and Commissions, 69 LIPPINCOTTS MAGAZINE748 (June 1902).22 SEGRAVE,supra note 12, at 11.


10/52

7

particularly black workers. The Pullman Company in particular was repeatedly singledout for fostering the tipping norm for its all black workforce as a way of economizing onits wage bill. In 1914 when the Railroad Commission of California asked why thecompany only hired blacks from the south, an executive explained, the southern Negrois more pleasing to the traveling public. He is more adapted to wait on people with a

smile.

23

The St. Louis Republic newspaper concluded:

It was the Pullman Company which fastened the tipping habit on the Americanpeople and they used the negro as the instrument to do it . . .24

Pullman made public the fact that its African-American porters were poorly paid so thepublic would pay them instead.

When the Pullman porters organized into the Brotherhood of Sleeping Car Porters in1925, one of the first things they did was to petition the Interstate CommerceCommission asking for an orderprohibitingtips. The petition in part said:

Only Negroes, many of them ex-slaves, were employed as porters. This causedthe work to be looked on as menial and servile and led to the giving and taking ofgratuities.25

That porters would ask the ICC to prohibit a form of their compensation is remarkable.True, a tipping prohibition would put pressure on the Pullman Company to pay higherwages. But it is harder to imagine why the increase in wages would more than offset theloss of tipping income.26 To our minds, the petition might have reflected a non-economicmotive. Even if the prohibition would not increase their net incomes, the black portersunion might have wanted it nonetheless possibly to increase dignitary dimensions oftheir employment. They too may have seen the receiving of tips as a token of theirinferiority and wanted to move away from an equilibrium where they had to scrape andbow for their living.

This brief detour does not begin to serve as a full history of tipping practices or itsintersections with race. But it may somewhat destabilize and problematize current normsabout the inherent desert of service workers to tips. We will pick up this history againboth when we discuss the tendency of minority passengers to tip less than whites andwhen we discuss our proposal to discourage tipping.

II. Description of Data

In April and May 2001, we collected tipping data from 1066 surveys completed by 12different New Haven medallion taxicab drivers (six black men, four white men, two

23 WILLIAM R.SCOTT,THE ITCHING PALM:ASTUDY OF THE HABIT OF TIPPING IN AMERICA 105-07 (1916).24Id. at 111-12.25Porters Assail Tipping, N.Y.TIMES, 10, at 3 (Nov. 24, 1927).26 One possibility is that making transparent the low net income of the porters would induce government toincrease their pay.


11/52

8

other minority27 men). Like most other localities, New Haven regulates both thenumber of taxies on the road and the price the taxis can charge.28 The taxis are commoncarriers who have a duty to service all customers.29

There are about 140 medallions in New Haven, predominantly owned by the MetroCab

Company. With only a few exceptions, the cabs in our study were leased from MetroCabon a fixed cost basis meaning that the drivers were the residual claimants of allrevenues including tips.30

The drivers were instructed to complete the surveys immediately after dropping off theirpassengers and were paid one dollar per survey. The amount tipped was calculated as thedifference between the amount due (the fare) and the total amount paid by thepassenger to the driver. Drivers reported information on passenger and driver profiles,including sex, race, age, passenger dress (as a proxy for wealth), and driver experience.Drivers were also asked to indicate if they knew the passengers, if the passengers wereregular clients, if conversation took place between them, if the pick-up was in response to

a call and if the passenger paid cash. Other data included pick-up and drop-offneighborhoods, travel times, day of week, time of day, temperature, and weather.

Table 1 reports summary statistics for this survey dataset.

27 In this other minority racial category, one of the cab drivers self-reported his race as being Arab(Franco-Moroccan) and the other reported his race as being Asian (Indian). The racial composition of

New Haven cab drivers (who are predominantly white or black) differs markedly from that of New York or

Washington D.C. J. M. Shenoy,African Americans Flag Down NY Cabbies (Nov. 6, 1999), available atwww.rediff.com/news/1999/nov/06us1.htm (About 40 per cent of New Yorks 25,000 drivers of yellowand livery cabs are from the Indian subcontinent.).28 See Lee A. Harris, Taxicab Regulation: The Ever-Elusive Freedom to Contract for a Ride 11(unpublished manuscript 2001):

[T]he main constraints in the taxicab industry can be schematized, loosely, into two categoriesentry controls and price restrictions. First, regulation in the taxicab industry often takes the formof entry constraints, via a medallion system, like the New York model. [Second,] regulations takethe form of price controls. In New York, London, and Washington DC, the amount charged for agiven trip is determined by local regulatory agencies. In Indianapolis, by contrast, only amaximum amount is prescribed by law.

The artificial restriction of supply has caused New York medallions to be worth well in excess of $150,000.Marcus Cole has suggested that the medallion system reduces competition and allows drivers more leeway

in refusing to pickup minority passengers. Marcus Cole,Medallion Monopoly Drives Taxicab Racism, 9Liberty & Law 4 (Feb. 2000).29 In many jurisdictions, the service all rule is enforced by prohibiting taxicab drivers from asking thedestination of passengers until they have entered the vehicle. Id. at 22. In New Haven, dispatchers inquireas to destination, but do not pass that information along to the driver. Some drivers in the study, however,circumvent the dispatchers by distributing personal business cards with cell phone numbers. During ourtesting process, one such driver declined to pick up a tester after inquiring as to drop-off location.30 Drivers tend to be paid either via a fixed cost system, where drivers are paid an amount over a certain

payment to the taxicab firm, or (to a far lesser extent) a commission system, where drivers are paid apercentage of the profits. Id. at 5 n.16.


12/52

9

Table 1. Summary Statistics

Continuous Variables Obs Mean Std. Dev Min Max

Tip ($) 1059 1.22 2.25 -5 27.5

Tip as % of Fare 1059 0.16 0.27 -0.14 5.22

Stiffing Rate 1059 0.24 0.43 0 1

Travel Time (min) 952 9.94 11.89 1.5 200Travel Distance (mil) 943 4.59 7.73 0.2 90

Average Speed (mil/hr) 937 27.12 20.38 3.20 180

Amount Due ($) 1064 9.26 11.51 2.5 150

Amount Paid ($) 1059 10.48 12.86 3 170

Temperature (F) 1044 52.25 10.38 28 80

Passenger Age* 1039 32.77 13.45 5 85

Driver Age 1016 39.78 7.95 24 51

Driver Exp (wks) 1066 62.69 53.47 2 192

Indicator Variables** Obs Percent*** Mean Tip Mean Tip%

Passenger Sex:

Female 506 47.47 0.97 15.5%

Male 443 41.56 1.46 16.6%

Passenger Race:

Asian 95 9.06 1.02 16.2%

Black 319 30.41 0.60 9.2%

Hispanic 138 13.16 0.81 12.0%

Other 14 1.33 0.84 10.7%

White 483 46.04 1.82 21.6%

Driver Race:

Black 517 48.5 1.02 12.6%

Other 99 9.29 0.76 12.4%

White 450 42.21 1.54 20.3%

Passenger Dress:

Below Average 146 13.7 0.79 12.8%

Average 858 80.49 1.16 15.1%

Above Average 41 3.85 1.74 21.8%Respond to Call 648 68.9 1.24 15.9%

Luggage 196 18.39 1.61 15.3%

Regular Customer 188 17.64 1.87 25.0%

Acquaintance 254 23.83 1.85 24.6%

Conversation 720 67.54 1.37 17.3%

Rain or Snow 27 2.53 2.28 37.8%

Cash 935 98.1 1.21 16.5%

***Variable percents may not add to 100 if variable is undefined for any observations

**Among the variables not listed here are indicators for neighborhood and addresses,

days of the week, times of day, fare types (on the dollar, 25 50 and 75 above the

dollar), individual drivers, and categorical variables derived from continous variables

*For multiple passenger rides, drivers were instructed to record information only for the

passenger who paid the fare

Overall, the average tip was $1.22, and the average tip as a percentage of fare was 15.8%.23.8% of the passengers left no tip (the mean stiffing rate). The data containsubstantial numbers of both male and female passengers and is also well balanced with


13/52

10

regard to black and white driver observations (N = 517 and 450, respectively), 31 whichaids greatly in developing statistically reliable tests of whether consumers discriminate intipping.

Unlike New York or Washington cab drivers who sometimes obtain a substantial portion

of their fares from passengers who hail them from the street, New Haven cab driversobtain fares predominantly by responding to a radio call or by waiting at a designatedstand (for example, at the New Haven train station or airport). More than two thirds ofour observations come from responses to a dispatchers call and the remainder are mostlyfrom drivers waiting their turn at cab stands. Because both the dispatcher and cab standcalls are distributed on a queued basis and because New Haven drivers do not have asmuch discretion (as hailed cab drivers) in turning down fares, the structure of serviceprovision tends to randomize driver-customer pairings. The tendency towardrandomization increases the power of our test of consumer discrimination. But weemphasize that this randomization effect is not perfect. Some fares are generated bydirect (cellphone) requests. Moreover, Table 2 shows that passenger races were not in

fact randomly distributed across driver races:Table 2. Driver and Passenger Race Frequency

Driver Race White Black Hispanic Asian Other

Number of

Fares

White 50.7% 25.6% 13.0% 9.4% 1.3% 446

Black 44.2% 35.1% 12.5% 7.1% 1.0% 504

Other 34.3% 28.3% 17.2% 17.2% 3.0% 99

Number of

Fares 483 319 138 95 14 1049

Pearson Test of Independence: chi2(8) = 25.9082 Pr = 0.001

Passenger Race

Black drivers were more likely than white drivers to have black passengers (35.1 vs.

25.6%) and white drivers were more likely than black drivers to have white passengers(50.7 vs. 44.2%). But drivers of each race were still exposed to substantial numbers ofblack, Hispanic and white customers (which again aids in testing for the existence ofcustomer discrimination).32

III. Results

While we will ultimately rely on multivariate regression analysis, our central results aresuggested by simply calculating the race-specific means for the tip amount.

31 This did not occur by chance drivers were recruited with the goal of achieving balance between blackand white drivers. Accordingly, the data are not necessarily representative of the racial mix of cab driversin New Haven. In this regard, it should be noted that more than half of drivers approached (while queuedat the train station) declined to participate in the study.32 There was a substantially higher percentage of African Americans and Asians among our passengers(30.4% and 9.1% respectively) than is found in the Greater New Haven population (11.2% and 2.4%respectively) and there was a substantially lower percentage of whites (46% in our sample vs. 74.8% inthe general population). See Census Table DT_DEC_2000_SF3_U_DATA1 (2003).


14/52

11

A. Lower Tips for Minority Drivers

Table 3 summarizes the central evidence of consumer discrimination:

Driver

Race Avg. Tip Avg. Tip % tip tip%

White $1.54 20.3%

Black $1.02 12.6% 0.66 0.62

Other $0.76 12.4% 0.50 0.61

*Disparity here is defined as the given statistic divided by the white

statistic

Race Disparity Ratios*:

Table 3. Average Tips & Tipping Percentage by Driver Race

White drivers are tipped substantially more than black (or other) drivers whethermeasured either in terms of the average tip amount or the average tip percentage. Whitedrivers were tipped 61% more than black drivers (20.3 vs. 12.6%) and 64% more thanour Other Minority drivers (20.3 vs. 12.4%). Put simply, passengers systematicallytipped white drivers substantially more than non-white drivers.

Disparate treatment in tipping can also be seen in the disparate propensities to stiffdrivers of different races.

Driver

Race

Stiffing

Rate

Race

Disparity

Ratios*:

White 15.7%

Black 28.3% 1.80

Other 36.4% 2.31

Table 4. Stiffing Rate by Driver Race

*Disparity here is defined as the given

statistic divided by the white statistic

Table 4 shows that the rate at which white drivers were stiffed (15.7%) was far less thanthat of non-white drivers (28.3 and 36.4% for black and Other Minority drivers,

respectively). Black drivers were 80% more likely to be stiffed than white drivers (andour Other Minority drivers were 131% more likely).

B. Lower Tips By Minority Passengers

The second racial effect that can be seen in the aggregate data concerns the propensity ofdifferent racial groups to tip different amounts.


15/52

12

Avg. Tip Avg. Tip % tip tip%

White $1.82 21.6%

Black $0.60 9.2% 0.33 0.42Hispanic $0.81 12.0% 0.44 0.55

Asian $1.02 16.2% 0.56 0.75

Other $0.84 10.7% 0.46 0.49

*Disparity here is defined as the given statistic divided by the white

statistic

Race Disparity Ratios*:

Table 5. Average Tips & Tipping Percent by Passenger Race

Passenger

Race

Table 5, for example, shows that the average tipping percentage of African-Americanpassengers was only 42% of the average tipping percent of white passengers (9.2 vs.21.6%). The tipping percentage of 138 Hispanic passengers was only slightly more thanhalf of the white passenger tipping percentage (12.0 vs. 21.6%). And Asian passengers

tipped only 75% of the white passenger tipping percentage (16.2 vs. 21.6%).

And, as before, the racial disparity in the average tip amount or tip percentage can also beseen in the different propensities of passengers to stiff drivers.

Passenger

Race

Stiffing

Rate

Race

Disparity

Ratios*

White 10.6%

Black 39.2% 3.69Hispanic 34.3% 3.23

Asian 15.8% 1.49

Other 35.7% 3.36

Table 6. Stiffing Rate by Passenger

Race

*Disparity here is defined as the given

statistic divided by the white statistic

Table 6 shows that African-American passengers are 3.7 times more likely to stiff (andHispanic passengers are 3.2 times more likely) than white passengers (39.2% and 34.3%stiffing rates for blacks and Hispanics respectively vs. only a 10.6% stiffing rate for whitepassengers). It is important to emphasize, however, that the racial disparities uncovered

in Tables 5 and 6 (as well as the racial disparities uncovered in Tables 3 and 4) are justfirst cuts at the data. These tables do not control for other variables such as the socio-economic status of the passenger or the service quality of the driver that might bedriving the result. These issues and others will be explicitly discussed below inconsidering alternative hypotheses (see infra Part IV).

But, as emphasized in the introduction, these passenger race results may have policyrelevance even without controlling for any non-racial factors. An irrational statistical


16/52


17/52

14

The tendency of minority passengers to join in the larger pattern of passengerdiscrimination against minority drivers can also be seen in an analysis of stiffing rates.

Table 9. Stiffing Rate by Passenger and Driver Race

Driver Race Rate Observations

Race Disparity

Ratios*

White 6.3% 224

Black 12.2% 222 1.95

Other 29.4% 34 4.71

White 33.0% 112

Black 43.2% 176 1.31

Other 39.3% 28 1.19

White 22.8% 57

Black 42.9% 63 1.88

Other 41.2% 17 1.81

White 9.5% 42

Black 13.9% 36 1.46

Other 35.3% 17 3.71

White 33.3% 6

Black 20.0% 5 0.60

Other 66.7% 3 2.00

Asian

Other

*Disparity here is defined as the given (Black or Other) statistic

divided by the white statistic

White

Black

Hispanic

Table 9 shows that white passengers are almost twice as likely to stiff black drivers aswhite drivers (12.2 vs. 6.3%). And again, the behavior of minority passengersqualitatively mirrors this disparity. From Table 6, we already know that African-

American passengers are more likely to stiff drivers than white passengers, but Table 9shows that African-American passengers are almost a third more likely to stiff black

Table 8. Average Tips by Passenger and Driver Race

Passenger

Race Driver Race Avg. Tip Observations

Race Disparity

Ratios*

White $2.21 224

Black $1.55 222 0.70

Other $0.97 34 0.44

White $0.72 112

Black $0.49 176 0.68

Other $0.75 28 1.04

White $1.11 57

Black $0.59 63 0.53Other $0.57 17 0.51

White $0.84 42

Black $1.43 36 1.70

Other $0.63 17 0.75

White $1.04 6

Black $0.90 5 0.86

Other $0.33 3 0.32

White

* Disparity is defined as the given (Black or Other ) statistic divided

by the white statistic

Black

Hispanic

Asian

Other

The results of Tables 7 and 8 are qualitatively the same. Again, black and Hispanic passengers, like whites,tipped black cab drivers less.


18/52

15

drivers as white drivers (43.2 v. 33.0%) and Hispanic passengers are 88% more likely tostiff black drivers than white drivers (42.9 v. 22.8%).

D. Regression Analysis

The disparities reported in the previous tables, however, are provided only for heuristicpurposes. Without more, we would not know whether the results would be statisticallysignificant, or whether the racial results would hold up once we controlled for a host ofnon-racial factors that might influence the amount that passengers tip.

This section corrects for these deficiencies by offering a series of regressions testingwhether the foregoing racial effects persist in a more nuanced analysis. Table 10 reportsthe result of four nested regressions that relate the tipping percentage to the passenger anddriver race as well as an increasing number of non-racial right-hand side variables.


19/52

16

Table 10. Regressions with Tipping Percentage as Dependent Variable

1 2 3 4*

Constant** 0.252 0.379 0.289 0.410

Racial Effects:

Driver Black -0.067 -0.095 -0.092 -0.091

Driver Other -0.067 -0.166 -0.158 -0.134

Passenger Black -0.117 -0.093 -0.090 -0.090

Passenger Hispanic -0.093 -0.061 -0.065 -0.042

Passenger Asian -0.053 -0.044 -0.048 -0.047

Passenger Other -0.107 -0.058 -0.051 -0.022

Other Variables:*** Mean Std. Dev

Passenger Female -0.004 -0.004 -0.010

Passenger Age 32.77 13.45 0.041 0.042 0.034

Below Average Dress -0.034 -0.035 -0.020

Above Average Dress 0.032 0.033 0.036

Weeks Driving Cab 0.000 0.000 0.001

Survey Experience

-0.015 -0.009 -0.016

Driver Age 39.778 7.947 -0.035 -0.040 -0.045

Conversation (1=yes) 0.029 0.023 0.019

Repeat Passenger (1=yes) 0.039 0.047 0.052

Acquaintance (1=yes) 0.110 0.112 0.100

Multiple Passengers (1=yes) 0.038 0.037 0.045

Dispatched Pick-up (1=yes) 0.010 0.017 0.000

Amount Due 9.262 11.506 -0.108 -0.105 -0.103

Amount Due Squared 218.034 1123.087 0.062 0.062 0.054

Fare 25 0.037 0.036 0.057

Fare 50 0.051 0.053 0.052

Fare 75 0.001 0.000 0.017Cash (1=yes) 0.002 0.007 0.002

Travel Time 9.940 11.890 -0.026 -0.023 -0.010

Travel Distance 4.587 7.727 0.050 0.045 0.033

Average Speed (Travel Time/Travel Distance) 0.452 0.340 -0.020 -0.015 -0.007

Night (Between 7PM and 7AM; 1=yes) 0.017 0.025 0.025

Late (Between 11PM and 5 AM; 1=yes) 0.012 0.010 -0.025

Temperature 52.25 10.38 -0.012 -0.008 -0.005

Rain/Snow (1=yes) 0.206 0.196 0.136

Luggage (1=yes) 0.020 0.017 0.008

Pick-up Nghbd with Below Average 911 Calls

-0.012 -0.074

Pick-up Nghbd with Above Average 911 Calls

0.047 0.025

Drop-off Nghbd with Below Average 911 Calls

0.044 0.167

Drop-off Nghbd with Above Average 911 Cal ls

0.026 -0.019

Train Pick-up -0.010

Train Drop-off -0.021

1059 841 841 841

0.061 0.196 0.203 0.298

N Y Y Y

NOT SIG NOT SIG NOT SIG

NOT SIG NOT SIG NOT SIG

*Neighborhood dummies were included in this r egression -- the coefficients are not r eported.

***For continous variables, the effects of a one standard-deviation change are reported.

Categories are based on Total year 2000 911 calls divided by neighborhood population, with extrapolations to missing data/suburbs.

**The omitted categories for the indicator variables ar e Driver White, Passenger White, Average Dress, Pick-up And Drop-off

Neighborhood variables with Average 911 Calls. To avoid losing observations and to keep the omitted category pure, indicator variables

equal to one for missing data were also included but are not reported.

Survey Experience is defined on a scale of 1-3, depending on whether the driver was filli ng out his first, second or third set of surveys

Underlined coefficients are significant at the 10% level, coefficients in bold are significant at the 5% level, and coefficients underlined and

in bold are significant at the 1% level.

Number of Observations

Hausmann Test

Variance Test

Random Effects

R-Squared

Observations fell out of models for two reasons: (1) incomplete dr iver surveys, and (2) a mid-study change in survey design replacing type

of building with neighborhood for the pick-up and drop-off information

The first regression specification (reported in column 1) simply regresses the tipping

percentage on passenger and driver racial indicator variables. Because the white driverand white passenger variables are omitted, the constant term (25.2%) equals the predictedtipping percentage for a white passenger tipping a white driver. The coefficients on theremaining variables represent the incremental difference for the specified driver orpassenger racial type. For example, the first specification suggests that African-American driver tips would be 6.7 percentage points lower than the tips given to a whitedriver (and the bolded and underlined font indicates that this shortfall is statisticallysignificant at the 1% level).


20/52

17

The results of the first specification are consistent with the preceding analysis. Thenegative coefficients on the driver variables indicate that customers tip a lowerpercentage to minority drivers than to white drivers and the negative coefficients on thepassenger variables indicate that minority customers tip a lower percentage than white

customers. For example, the regression suggests that African-American passengers tip11.7 percentage points less than white passengers. Moreover, the regression lets us seefor the first time that both of these types of racial disparities are statistically significant.

But the first specification still does not control for a host non-racial factors that might beinfluencing the tipping percentage. Regression specifications 2 through 4 addsuccessively more right-hand side control variables to test whether the racial effectsuncovered in the previous tables (and in specification 1) are merely the byproduct of whateconometricians call omitted variable bias. Regression 2 adds twenty-six variablesrelated to non-racial demographic characteristics of passenger and driver (such as age andgender) and to characteristics of the ride itself. Regression 3 then adds four more

variables related to the crime rate (measured by number of 911 calls per resident) foundin the pickup or drop off neighborhoods. Finally, specification 4 adds individualindicator controls for 48 pick-up and drop-up neighborhoods.35

Regressions 2 through 4 also employ a random effects estimation method. In thiscontext, a random effects model tries to take into account that different drivers may haveidiosyncratic propensities to be tipped a particular percentage. Imagine for example thatsome random process makes some people better or worse drivers (and hence more or lesslikely to receive a good tip). A random effects model simultaneously estimates (i) thesize of these individual driver random effects and (ii) whether (after controlling for theindividual driver random effects) there are still statistically significant driver raceeffects.36

The variance test reported for these regressions suggest that estimated variance of theindividual driver random effects is not statistically different than zero. This suggests thatthe individual driver effects are not dominant in this dataset. The real action in the data isbetween races and not idiosyncratic differences within race.37

After controlling for random driver effects and a host of time, manner and place effects,these specifications suggest that the customer discrimination result is quite robust.Adding additional variables to the regression does not materially impact the size orstatistical significance of the driver race variables. For the most complete regression(specification 4), we find that black drivers are tipped 9.1 percentage points less thanwhite drivers (and that this result is statistically significant at the 1% level). Thisregression predicts that an African-American driver would be tipped 43.6% less than a

35 There are 61 neighborhood dummies, but many of these were dropped as a result of multicollinearity. Tosave space, neighborhood-dummy effects are not reported.36 The Hausmann tests that are reported for these regressions cannot reject the hypothesis that the randomeffects regression is appropriate (relative to a fixed-effects regression).37 However, later in the discussion of alternative hypotheses (part IV), we will return to this issue again.


21/52

18

similarly-situated white driver,38 and that this tipping shortfall causes the overall revenueper fare for African-American drivers to be 7.0% less than that of white drivers.39 Thepassenger discrimination imposes the economic equivalent of a 7% tax on the income ofblack cab drivers.40

The finding that minorities tip systematically less, in contrast, is not as robust to theinclusion of additional right-hand side controls. The size of the coefficients becomesmaller as additional controls are added and the Hispanic disparity becomes lessstatistically significant. For the most complete regression (specification 4), the size of theblack passenger effect is diminished (from 11.7 percentage points in specification 1) to a9 percentage point shortfall.

While 9 percentage points is still substantial, the large point here is that evenspecification 4 contains only poor controls for the socio-economic class of the passenger(above or below average dress, and characteristics of pickup or drop off locations). Withbetter controls, the seeming tendency of minorities to tip less might simply become a

tendency of poor people to tip less. We will explore this hypothesis below in thealternative hypotheses discussion. But for now it is important to note that the cab driver,like the researcher, will not be able to observe many of these additional variables and, aswe will stress below (in the Implications section), thus might be moved to makeinferences including racial inferences about whom to pick up on the bases of similarlyrestricted data.

The regressions also suggest that customers discriminate against older drivers.Specification 4 shows that a driver whose age is one standard deviation (about 8 years)above the mean driver age (about 40) should expect to receive tips that are 4.5 percentagepoints less than average. And this disparity is statistically significant at the 10% level (p.< .1).41

For those interested in norms of tipping more generally, the non-racial controls provide apotpourri of interesting results. In contrast to driver age, we learn that older passengerstip a systematically higher percentage. In our data, for example, a passenger whose agewas one standard deviation above the mean passenger age was predicted in regression 4to tip 3.4% percentage points more.

38 The tipping percentage shortfall of 9.0% divided by the predicted white driver tip percentage (evaluatedat the means of the non-driver race variables) of 20.9% equals 43.6%.39 The revenue shortfall of $.76 divided by the predicted white driver revenue (evaluated at the means ofthe non-driver race variables) of $10.82 equals 7.0%.40

In simple economic models, the lower revenues available for minority drivers would tend to causeminorities to substitute away from driving cabs where they would not have to pay this discrimination tax.But our conversations with both minority and non-minority drivers suggest that the drivers are not wellinformed about the discrepancies in customers willingness to tip minority drivers. And even if minoritydrivers learned of their ill-treatment, they may also face restricted opportunities in finding and beingcompensated for other forms of employment.41 It is important to note that the model also includes a variable for weeks of driver experience, whichyielded a positive but not statistically significant coefficient. Thus, the results suggest that although theremay be a slight advantage to experience (all else held equal), the more significant effect is that older driversat a given experience level receive lower tips.


22/52

19

We also learn in specification 4 that the tipping percentage is statistically higher:

during inclement whether (13.6 percentage points);for acquaintances (10 percentage points); and,

for lower fairs (10.3%).

42

In contrast, we found no passenger gender effects. Men and women seem to tip roughlythe same percentage in all of our specifications.

A variant of the foregoing regression specifications can also be used to test whether ourearlier finding -- that minority passengers seem to participate in the discriminationagainst minority cab drivers is statistically significant and robust to inclusion of otherright-hand side controls. The previous regressions do not allow for such a test becausethe passenger and driver race effects were forced by the specification to enterindependently. But it is possible in a less constrained specification to have the

regressions estimate effects for all 15 of the possible driver/passenger race permutations.By including these interaction effects it becomes possible to test, for example, whetherblack passengers tend to tip black drivers less than white drivers.

Passenger

Race

Driver

Race

Regression

1*

Regression

2*

Regression

3*

Regression

4*

Black -0.088 -0.104 -0.100 -0.101

Other -0.135 -0.212 -0.208 -0.166

Black -0.036 -0.079 -0.075 -0.064Other 0.021 -0.104 -0.095 -0.071

Black -0.104 -0.092 -0.093 -0.112

Other -0.062 -0.161 -0.152 -0.149

Black 0.020 -0.062 -0.065 -0.055

Other -0.038 -0.186 -0.176 -0.158

Black -0.038 -0.037 -0.049 -0.094

Other -0.130 -0.058 -0.078 -0.070*These effects are based on regressions of the same form as 1, 2, 3 and 4 where

the passenger and driver racial variables are replaced with interacted variables.

Underlined coefficients are significant at the 10% level, coefficients in bold are

significant at the 5% level, and coefficients underlined and in bold are

significant at the 1% level.

Table 11. Tests of Consumer/Passenger Discrimination against Different

Driver Race, by Passenger Race

Minority Driver Disparity (Relative to White Drivers) --

Difference in Driver Effects:

White

Black

Hispanic

Asian

Other

Table 11 reports the tipping percentage shortfalls for minority drivers (relative to whitedrivers) from particular passenger races in regression specifications that are otherwiseidentical to specifications 1 through 4 in Table 10. For example, we see in specification

42 The regression suggests that the tipping percentage increases with amount due squared. But we foundthat the squared term only began to dominate the linear term for fare amounts due that were out of sample(approximately $186).


23/52

20

2* that white passengers tip black drivers 10.4 percentage points less than white drivers.The black driver shortfall is large and statistically significant in all of the specifications.

Again, there is some evidence that minorities discriminate against black drivers, but thedifferences are less significant than in the case of white passengers. The reported black

disparities are all negative indicating an estimated shortfall for black drivers, but for themost controlled specification, the black driver shortfalls are insignificant. Most of theother passenger/driver race permutations were statistically insignificant. However, itshould also be noted that we may be simply running into a small numbers (or whateconometricians sometimes call a degrees of freedom) problem. When you ask aregression to estimate 15 separate driver/passenger race effects as well as nearly 100other right-hand control variable effects it can become increasingly difficult to identifystatistically significant effects.

To further analyze stiffing behavior, we regressed the stiffing indicator against the sameset of independent variables in the tipping percentage specifications. Table 12 reports the

results of these regressions, now relating the probability of stiffing to passenger anddriver race and the nested set of non-racial right-hand side variables.


24/52

21

Table 12. Regressions with Stiffing as Dependent Variable

1 2 3 4*

Racial Effects:**

Driver Black 0.107 0.100 0.096 0.098

Driver Other 0.223 0.194 0.176 0.239

Passenger Black 0.307 0.279 0.273 0.192

Passenger Hispanic 0.286 0.198 0.205 0.131

Passenger Asian 0.061 0.087 0.088 0.102

Passenger Other 0.306 0.188 0.180 0.181Other Variables: Mean Std. Dev

Passenger Female 0.006 0.006 0.004

Passenger Age 32.77 13.45 -0.032 -0.035 -0.050

Below Average Dress 0.069 0.069 0.062

Above Average Dress -0.093 -0.087 -0.065

Weeks Driving Cab 0.001 0.001 0.000

Survey Experience*** 0.001 -0.007 -0.005

Driver Age 39.778 7.947 0.017 0.018 0.044

Conversation (1=yes) -0.016 -0.005 0.015

Repeat Passenger (1=yes) 0.013 -0.003 -0.014

Acquaintance (1=yes) -0.068 -0.067 -0.035

Multiple Passengers (1=yes) 0.013 0.013 0.008

Dispatched Pick-up (1=yes) 0.017 0.006 0.010

Amount Due 9.262 11.506 0.034 -0.051 -0.154

Amount Due Squared 218.034 1123.087 -0.134 -0.010 0.088

Fare 25 -0.138 -0.127 -0.104

Fare 50 -0.161 -0.154 -0.133

Fare 75 -0.185 -0.177 -0.151

Cash (1=yes) -0.360 -0.395 -0.285Travel Time 9.940 11.890 0.029 0.058 0.110

Travel Distance 4.587 7.727 -0.084 -0.114 -0.146

Average Speed (Travel Time/Travel Distance) 0.452 0.340 0.019 0.024 0.039

Night (Between 7PM and 7AM; 1=yes) 0.073 0.066 0.045

Late (Between 11PM and 5 AM; 1=yes) -0.047 -0.034 0.008

Temperature 52.25 10.38 0.030 0.029 0.023

Rain/Snow (1=yes) -0.030 -0.021 0.028

Luggage (1=yes) -0.104 -0.099 -0.077

Pick-up Nghbd with Below Average 911 Calls

0.500 0.890

Pick-up Nghbd with Above Average 911 Calls

-0.038 0.642

Drop-off Nghbd with Below Average 911 Calls

0.059 -0.062

Drop-off Nghbd with Above Average 911 Calls

-0.022 0.116

Train Pick-up 0.105

Train Drop-off 0.073

1059 837 837 837

0.119 0.313 0.332 0.416

*Neighborhood dummies were included in this regression -- the coefficients are not reported.

Categories are based on Total year 2000 911 calls divided by neighborhood population, with extrapolations to missing data/suburbs.

Number of Observations

**Coefficients reported here are the changes in the probability of stiffing resulting from infinitesimal changes in the continuous variablesand discrete changes in the indicator variables from 0 to 1. For continous variables, the effects of a one standard-deviation change are

reported. The omitted categories for the indicator variables are Driver White, Passenger White, Average Dress, Pick-up And Drop-off

Neighborhood variables with Average 911 Calls. To avoid losing observations and to keep the omitted category pure, indicator variables

equal to one for missing data were also included but are not reported.***Survey Experience is defined on a scale of 1-3, depending on whether the driver was filling out his first, second or third set of

surveys.

Underlined coefficients are significant at the 10% level, coefficients in bold are significant at the 5% level, and coefficients underlined

and in bold are significant at the 1% level.

Pseudo R-Squared

Observations fell out of models for two reasons: (1) incomplete driver surveys, and (2) a mid-study change in survey design replacing

type of building with neighborhood for the pick-up and drop-off information

Since stiffing is a dummy variable, we ran the analogous stiffing regressions using probitmodels so that the output in Table 12 indicates the change in the likelihood of a stiff on

the basis of a discrete 0-1 change in the dummy variables or a one-standard-deviationchange in the continuous variables. Thus, in the first specification, the black drivercoefficient of 10.7 indicates that a black driver is about 11 percentage points more likelyto be stiffed than a white driver.

Once again, the results of all specifications are quite consistent with our initialtabulations. Minority passengers are more likely to stiff and minority drivers are morelikely to be stiffed. For example, in all specifications, a Black passenger is significantly


25/52

22

more likely to stiff than a white passenger, though with decreasing magnitude as we addmore controls. In addition, we now see that the majority of these effects are highlysignificant (prob. < .01).

Again, the consumer discrimination effects, now in terms of stiffing, remain robust to the

addition of racial and non-racial controls. Although the second and third regressionsyield black driver effects that are significant only at the 10% level, the magnitude of theeffects actually remains roughly in the area of 10 percentage points. The most controlledspecification is significant at the 1% level, indicating that black drivers are almost 10percentage points more likely to be stiffed than white drivers (while, according to thesame specification, Other Minority drivers are almost 24 percentage points more likely tobe stiffed).

As is in the case of the tipping regressions, the passenger effects tend to diminish with theaddition of more controls. Comparing the first and fourth regressions, we see that theblack passenger effects diminish by over 11 percentage points and the Hispanic passenger

effects diminish by over 15 percentage points -- although the broad levels of significanceremain the same.43

Not surprisingly, many of the non-racial controls that were significant in the tippingpercentage regressions remain significant in the stiffing regressions.44 Passenger anddriver age, above average dress, and acquaintance of driver with passenger yield largelysignificant effects with the same tendencies as in the tipping regressions.45 So, forexample, in the fourth regression, we see that a passenger one standard deviation abovethe mean age is 5% less likely to stiff while a driver one standard deviation above themean age is 4.4% more likely to be stiffed. Again, this suggests that passengers mayhave engaged in discrimination against older cab drivers as well.

All in all, the regression analysis suggests that there is strong evidence of customerdiscrimination against minority drivers measured by tipping and stiffing differences thatpersists and is statistically significant after controlling for a variety of non-racial factors.And there is some evidence that black and Hispanic passengers participate in thisdiscrimination, although the estimates of minority driver shortfalls in the most controlledregression are less significant. The second finding that minority passengers tend to tip

43 However, it is again important to keep in mind that none of our specifications contain good controls forthe socio-economic class of the passenger so that effects the stiffing regressions are attributing to race mayin fact be more dependent on the wealth of the passenger. This and its implications to cab driver inferenceswill be considered further in Part IV.C.44

Note, however, that the fare dummies are all negative and highly significant. This indicates thatpassengers are more likely to stiff with integer fares than with any other type, perhaps in part becauseletting the driver keep the change is an option for non-integer fares. We will discuss this furtherinfranote 68. Also, somewhat surprisingly to our minds, a passenger paying cash is less likely to stiff. Somedrivers may have neglected or refused to record a non-cash tip.45 Though rain and snow are not significant disincentives to stiffing in these regressions, the temperatureeffects show that passengers are about 2.3% less likely to stiff with every 10% drop in temperature,suggesting that cold/inclement weather does have some prohibitive effect on stiffing. Additionally,

passengers are more likely to stiff at night under the cover of darkness (4.5% more likely in the thirdregression, and significant at the 5% level).


26/52

23

a lower percentage and stiff more frequently is confirmed as statistically significant forAfrican Americans in all specifications (but the absolute size of these effectssystematically decreases as more non-racial controls are introduced).

IV. Alternative (Non-Racial) Hypotheses

Although the prior regressions control for a host of non-racial variables, there are stillmany aspects of the transaction that we do not observe and hence cannot include in theanalysis. As is often the case, the omission of right-hand side control variables creates apossibility that the racial effects reported above may in fact be caused by non-racialfactors for which we did not control. This section outlines the major alternativehypotheses and assesses with the best data available the extent to which they qualify thetwo racial results of the last section.

But before proceeding to consider the particular omitted variables that may be driving theminority driver and the minority passenger effects, we first take on more globalconcerns about the quality of the data reported by the cab drivers.

A. Censored Data?

It is imported to remember that all the results reported above are based upon surveysfilled out by individual cab drivers. Either misreported or censored data wouldimportantly reduce our confidence in the results.

A weak indication of survey reliability can be found in the non-significance of theSurvey Experience variable reported in both the Table 10 and 12 regressions. Thecoefficients on this variable are both very small and not statistically significant. Thisindicates that reported tips of the drivers did not vary as they filled out more surveys. Ifthe drivers were misreporting fares, they at least seem to be consistently misreportingthem over time.

The earlier discussed rejection of individual driver random effects (as evinced byrejecting the hypothesis that the random effects variance was different from zero) alsoprovides some small measure of assurance that drivers were accurately reporting fairs.The random effects regressions suggest that drivers of the same race were treated

similarly: If white drivers were misreporting fare data, they seemed to be doing itconsistently as a group.46 The possibility that drivers explicitly colluded to misreport is

46 In sharp contrast, an earlier pilot study conducted by Suzanne Perry found that one of the driverssurveys had markedly different (and implausible) survey answers. This driver reported that virtually all ofhis passengers failed to tip. Our impression from interacting with the participating drivers was that eachtook the study seriously. Most asked questions about the study and several expressed interest in obtaining acopy of the results. One driver returned a subset of his 50 surveys, apologizing that he could not completethe set because he was going to be unable that month to make the lease payment on his cab.


27/52


28/52

25

audited. (That every driver reported some substantial tips, however, argues against thistroubling hypothesis.) If the recorded fares are not representative of the larger universe,then the prior results may not be indicative of the broader tipping experience. However,as one of the first studies of taxicab tipping practices and the first quantitative study ofconsumer-side discrimination, even qualified results raise important concerns about the

possibility of disparate treatment (and at a minimum suggest the appropriateness offurther testing).

Putting aside for the moment these important censoring concerns, we next turn to thepossibility of omitted variable problems with the two core racial results.

B. Lower Tips for Minority Drivers

Our finding of customer discrimination against minority drivers was both quite stable androbustly significant in Tables 10 and 12s series of nested regressions. But it is alwaysimportant to consider whether omitted variables may be driving this disparate-treatment

result. Here we pause to consider three possibilities: individual driver effects, disparatedriver quality, and disparate customers.

Individual Driver Effects. To begin, it is useful to assess whether what we reported asdriver race effects might instead be caused by idiosyncratic characteristics of the twelveindividual drivers in the data. As an initial matter, we found that our random effectsmodels, controlling for an increasing range of non-racial factors, rejected the presence ofindividual driver effects (and after attempting to control for them nonetheless foundpronounced evidence of customer discrimination). Second, if we simply calculate meantipping percentage for each of 4 white and 6 black drivers, we find that white driversgarnered three of the four highest tipping percentages, while black drivers garnered three

of the four lowest tipping percentages.49

But it is possible in non-random effectspecifications to alternatively control for individual driver effects by asking theregression to cluster the data by individual driver.50 When we rerun the regressionsclustering by individual cab drivers, we still find evidence of customer discrimination butthe results are not as statistically significant.51 In the clustering regressions that parallelthe first specification in Tables 10 and 12, we still find that black and other minoritydrivers receive lower percentage tips and are more likely to be stiffed, but the results areonly marginally statistically significant.52

49 One of the white drivers average tipping percentage was particularly high (30.5%) and one of the black

drivers average tipping percentage was particularly low (5.8%).50See STATA USERS GUIDE 256 (1999).51 The clustering procedure, by its nature, generates the same coefficients as reported in Tables 10 and 12,

but lets the data test whether clustering increases or decreases the level of statistical significance.52 The Table 10 analog suggests that the Black Driver tipping percentage result is only significant at the11.9% level, while the Other Minority Driver result was no longer statistically significant. The Table 12analog suggests that the Black Driver stiffing result is only significant at the 6.9% level, while the OtherMinority Driver result is only significant at the 9.5% level. The much smaller pilot study of Suzanne Perry,discussed supra note 46, also was not able to identify statistically significant customer discriminationagainst minority drivers.


29/52

26

On net, we still have some lingering concerns about individual driver effects. But using avariety of reasonable approaches that alternatively control for these effects, we still findwhat seems to be an independent and statistically significant disparity in the tips receivedby minority and non-minority drivers.

Disparate Driver Quality. The driver race disparity might alternatively be explained bypotential differences in the quality of service that minority and non-minority driversprovided. If minority drivers provided systematically poorer service than white drivers,then minority drivers may have received lower tips not because of their race, but becausepassengers may give lower tips for poorer service.

We do not have much information to respond to this theoretical possibility. The speedvariable and the indicator variable for whether the driver conversed with the passengercrudely control for two dimensions of quality. Also, other studies of tipping generallyhave found that variation in service quality does not explain a very large percentage ofdifferences in the amounts that people tip.53 These studies would at least suggest that the

degree of the racial disparity is not likely to be caused by differences in quality.

We also undertook a modest amount of auditing of the drivers in our study to see if therewere gross differences in the quality of service they provided.54 Based on a total of onlyten audit rides with participating drivers (six rides with white drivers, four rides withblack drivers), we did not find support for the hypothesis that minority drivers providedlower quality service. Indeed, our testers subjectively rated the quality of service higherfor black drivers than for white drivers (average 4.5 out of 5 for black drivers vs. anaverage rating of 3.3 for white drivers). This miniscule sample does not allow us toconfidently test for quality differences. But when combined with the authors ownexperience of taking numerous cabs in New Haven, we are fairly confident that the driverrace effect is not well explained by racial disparities in driver quality.

Disparate Customers. Finally, we considered whether the driver-race disparity might becaused by minority drivers serving different types of customers than non-minoritycustomers. Under this hypothesis, minority drivers would receive lower tips than whitedrivers not because customers discriminate but because minority drivers tend to provideservice to low-tipping customers while non-minority drivers tend to provide service tohigh-tipping customers.

As discussed above, there are some structural variables that tend to push New Havendrivers toward a more random selection of customers. Both dispatchers and cab standspurport to operate on a queued basis allocating the next customer to the next driver inline.

53 Michael Lynn & Michael McCall, Gratitude and Gratuity: A meta-analysis of Research on the Service-Tipping Relationship, 29 JOURNAL OF SOCIO-ECONOMICS 203, 212 (2000) (Although the averagerelationship between tip size and service evaluations was statistically significant in this review, it was alsoquite smallaccounting for less than two percentage of the variability in tip percentages.).54See supra note 48 (discussing the driver audits).


30/52

27

But there are several dimensions on which non-random allocations of passengers canoccur. Dispatchers may, contrary to stated policy, give poorer fares to minority drivers.Drivers may engage in different strategies as to which cab stand they wait at or whetherthey queue for the next dispatchers call. Waiting at the airport may expose drivers to adifferent mix of passengers than choosing to wait at the train station or at the Shubert

theater.

55

And passengers may directly call a driver to schedule service.

56

Table 2 already showed one dimension of non-random allocation of passengers.Minority drivers were more likely to service minority passengers and white drivers weremore likely to be paired with white passengers. We also explored a few other dimensionsof non-randomness. We found, for example, that a Pierson chi-squared test of statisticalindependence rejected the hypothesis that African-American and white drivers pickuppassengers from the same neighborhoods (p = .01). But we found no statistical differencein the average fare of black, white or other minority drivers.

While the evidence of non-randomized allocations makes it more difficult to test for

customer discrimination, our regressions controlled for a host of non-racial differencesand still found robust statistical evidence that minority drivers were tipped less evenafter controlling for the heightened chance that minority drivers have of serving minoritycustomers and even after controlling for their non-random allocation of neighborhoods.Accordingly, the disparate customer hypothesis does not, in the end, present a strongchallenge to our earlier results.

C. Lower Tips by Minority Passengers

It is superficially inviting to categorize our earlier results as that the minority status ofeither the driver or the passengercause the expected tipping percentage to be lower. But

these two racial effects stand on a very different theoretical and empirical footing. As atheoretical matter, it is straightforward to test whether one persons race influenced thedecisions of another person for example, X refused to sell to Ybecause Ywas black.But it is much harder to test whether the behavior of a decision-maker herself wasaffected by the decision makers own race. Such attempts run the risk of crudelyessentializing racial characteristics. At the extreme, such interpretations suggest farfetched genetic predispositions. But, as is well understood in the identity literature,57 theother extreme of emptying race of all experiential content is a social scientistsnightmare. Of course, cultural differences in behavior such as a different propensity totip are likely at some level to be the byproduct of differential cultural experiences. Butit is literally impossible to independently control for the myriad of present and past

55 Steve Salop helpfully suggested that we reanalyze the data to try to better control for more aggregateddriver strategies over the course of a shift. In this more aggregated analysis, we would have tested whetherdriver racial disparities persist at the shift level when we take into account that waiting at the airport takeslonger but is expected to generate a larger fare. However, the problem of incomplete shift data, discussed

supra Section IV.A, unfortunately precludes us from analyzing shift data in a systematic way.56 But this last possibility may still be an example of customer discrimination, if customers systematically

prefer to schedule with white drivers.57See, e.g., Kenji Yoshino, Covering, 111 YALE L.J.769 (2002).


31/52

28

disparities in cultural experiences. So to demand an experiential based quantitative testof culture differences is in effect a request for statisticians to put away their regressions.

The prior results, as a theoretical matter, exist in the intermediate range between theessentializing and non-essentializing extremes. The passenger race coefficients should be

interpreted (at best) as crude measures of current cultural differences taking the(historical and present) ambient experiential differences as given. The regressions askwhether, after controlling for aspects of the observed transactions, individuals ofparticular racial cultures (exposed to different experiences) tend to tip differently.58

As an empirical matter, it is important to remember that passenger race effects were notas robust as driver race effects. While the black and Hispanic passengers were found totip systematically less than whites in all the regressions of Table 10 and to stiffsystematically more in all the regressions of Table 12, the degree of the disparitynarrowed as better neighborhood controls were added to the regressions.

Nonetheless, analysis of data from other tipping studies makes us fairly confident thatpassenger race in the limited sense just described is likely to play an independent rolein expected tip amount. It turns out that an independent literature exists examiningwhether customer race affects the tipping size particularly (but not exclusively) inrestaurants. The least persuasive studies merely survey service providers about theirgeneral impressions. For example, a Houston survey of 51 waiters and waitresses foundthat 94% of the servers classified African Americans as poor tippers whereas none ofthe servers classified whites as poor tippers.59

There are also numerous racialized incidents concerning server perceptions of AfricanAmericans as being poor tippers.60 For example:

On October 23, 1999, Charles Thompson and Theresa White went out for dinnerat Thai Toni Restaurant in Miami Beach, Florida. When they got their bill, theyfound that a 15 percent gratuity had been added even though no similar chargeswere added to the bill of a nearby couple. Mr. Thompson asked the restaurants

58 Most residents of New Haven (96.3%) are native United States citizens and of the foreign born residentsmost (86.5%) are from Puerto Rico. Census Table,supra note 32. But two or three of our six African-American drivers were born abroad. This raises the possibility that some of the tipping disparity that we

attribute to race discrimination may instead be attributable to national origin discrimination. Unfortunately,at the time of collecting the data we did not ask the national origin of the drivers and hence cannotseparately control for this affect.59 Michael Lynn, Servers Perceptions of Who are Good and Poor Tippers (unpublished manuscript 2000).60 For example, an anonymous posting from a discussion board at www.tipping.org expressed the following

perception, which it described as common among servers:When I worked at T.G.I Fridays, I noticed that many tips from African American parties were not

based upon a percentage of the check, but were typically an arbitrary amount and were usually inthe two-to-five-dollar range. This is not to say that all African Americans left small tips, but asignificant number did.


32/52

29

owner/manager for an explanation and was told: You black people dont tipwell.61

The Miami city counsel reacted by passing an ordinance forbidding discriminatorytipping practices in eating establishments.62

More authoritative studies take one of two forms. Some of the studies, like ours, haverestaurant servers record information about their customers.63 Other studies interviewedcustomers as they depart the restaurant.64 Indeed, Michael Lynn and Clorice Thomas-Haysbert conducted something akin to a meta-analysis in which they pooled therestaurant data from five different studies of restaurant tipping in Houston.65 The authorsfound that African-American customers were expected to tip 3.59 percentage points lessthan white customers (p < .0001).66

While the restaurant studies are suggestive, they too may be infected by the problem ofomitted variable bias. Next we examine two dimensions in which our prior tipping

regressions are incomplete and may misattribute effects to passenger race that are notreally

racial disparities in taxicab tipping

Documents