Electronic copy available at: http://ssrn.com/abstract=1615149
The Visible Hand:
Race and Online Market Outcomes
Jennifer L. Doleac Luke C.D. Stein∗
May 2010†
Preliminary
Abstract
We examine the effect of race on market outcomes by selling iPods through local online
classified advertisements throughout the United States in a year-long field experiment. Each
ad features a photograph of the product being held by a dark- or light-skinned (“black” or
“white”) hand. To provide context, we also consider a group of sellers against whom buyers
might statistically discriminate for similar reasons: white sellers with wrist tattoos. Black
sellers do worse than white sellers on a variety of market outcome measures: They receive
13% fewer responses and 17% fewer offers. These effects are strongest in the Northeast, and
are similar in magnitude to those associated with the display of a wrist tattoo. Conditional
on receiving at least one offer, black sellers also receive 2–4% lower offers, despite the selfselected—and
presumably less biased—pool of buyers. In addition, buyers corresponding
with black sellers exhibit lower trust: They are 17% less likely to include their name in e-mails,
44% less likely to accept delivery by mail, and 56% more likely to express concern about
making a long-distance payment. We find evidence that black sellers suffer particularly
poor outcomes in thin markets; it appears that discrimination may not “survive” in the
presence of significant competition among buyers. Furthermore, black sellers do worst in
the most racially isolated markets and markets with high property crime rates, consistent
with channels through which we might expect statistical discrimination to operate.
JEL Classifications: J15, J71, C93, L86.
We are grateful to B. Douglas Bernheim, Nicholas Bloom and Caroline Hoxby for extensive advice and guidance,
and have also benefited from conversations with participants in several Stanford seminars, the San Francisco
Federal Reserve Bank’s Applied Micro Summer Conference, and the University of Chicago Experimental Economics
Lunch. Brandon Wall made important contributions to our experimental design and piloting. We appreciate the
generous support of the George P. Shultz Dissertation Support Fund.
∗
Department of Economics, Stanford University; Stanford, CA 94305 (both authors). E-mail: jdoleac@stanford.edu
and lstein@stanford.edu.
†
This version updated March 4, 2011.
Electronic copy available at: http://ssrn.com/abstract=1615149
1 Introduction
1 Introduction
Economic outcomes in the United States are highly correlated with race, but the causal mechanisms
underlying these correlations are not well understood. In particular, it is unclear how
much of the correlation is due to discrimination and how much is due to other characteristics
that are correlated with race, such as education.
There is an extensive literature on the effect of race on market outcomes, focusing on both
labor and goods (including housing) markets. However, “discrimination” is not a monolithic
phenomenon, and it is important to investigate its extent and causes in a variety of different
contexts (across which they may differ). Our paper asks the following question: When the typical
person engages in a consumer transaction (usually as a buyer), does he or she try to avoid dealing
with minority sellers, and does she treat minority sellers differently?
This is an important question for at least two reasons. First, a large amount of commerce
takes place through these kinds of transactions. Second, discrimination by consumers may
in fact underlie other forms of discrimination. For example, if white customers prefer not to
deal with black sellers, retailers might avoid hiring black salespeople. Becker (1971) identifies
discrimination by consumers, employers, and fellow workers as the three potential sources of
the racial disparity in labor market outcomes, and Nardinelli and Simon (1990) note that in a
relatively competitive labor market like the United States, consumer discrimination is the most
likely cause of the persistent disparity.1
However, the resulting labor market discrimination is
difficult in practice to distinguish from lower ability because both affect observed productivity.
Our study contributes to the literature by examining the effect of consumer discrimination
on sellers’ market outcomes using a field experiment that rigorously isolates the effect of sellers’
skin color. This complements existing evidence of discrimination in labor and housing
markets which—though striking—does not imply the existence of discrimination in common
consumer transactions (though the latter could be one explanation for labor market discrimination).
Indeed, we might expect that individuals are more likely to indulge in their taste for
anti-black animus when renting a house or hiring an employee than when engaging in the
limited interaction of a consumer transaction.
We posted classified advertisements offering an iPod Nano portable digital music player for
sale on several hundred locally focused websites throughout the United States, and analyze
here the effect of the seller’s skin color on several outcomes of interest. Notably, we signal skin
1
Altonji and Blank (1999) summarize the theory and evidence regarding race and the labor market in their Handbook
of Labor Economics chapter, and document the persistent black-white gap in earnings, labor participation,
and education.
2
1 Introduction
color using photographs instead of one of the more problematic treatments used in previous
discrimination studies.2
By including a photograph of a dark-skinned (black) or light-skinned
(white) hand holding the item, we were able to randomly vary the apparent race of the seller
while fixing other advertisement and market characteristics.3
We also compare the effect of
race with that of a social signal that can be communicated through the appearance of a seller’s
hand: a wrist tattoo. Tattooed sellers are likely statistically discriminated against for many of the
same reasons as black sellers, so—in addition to providing general context for interpreting the
magnitude of the black-white differences we observe—this third group of sellers can serve as a
“suspicious” white control group.
The market in which we run our experiment facilitates our investigation of consumer discrimination
in multiple ways: Buyers have no reason to make offers that they do not anticipate ending
in a transaction. They anticipate having to meet a seller in order to complete the transaction—
perhaps on the seller’s terms—with the non-trivial possibility of deception or theft. Thus, trust
plays a key role in the interactions we observe. These are characteristics of many (labor and
goods) market transactions that may be less present in the decision to call back a job applicant,
bid in an online auction, or make a purchase guaranteed by a third party (such as from a
store where the salesperson is merely an employee). Furthermore, the local focus of the online
classified advertisements allows us to analyze regional differences as well as variation by local
economic and demographic characteristics.
Given the challenges of adequately controlling for unobservable characteristics that may be
correlated with race, we are, unsurprisingly, not the first to use an experimental approach. Riach
and Rich (2002) offer a useful survey of field experiments designed to assess discrimination. Actorbased
audit studies (e.g., Ayres and Siegelman, 1995; Neumark, Bank, and Van Nort, 1996)—in
which actors apply for jobs, consider housing, or negotiate sales—attempt to match differentrace
candidates on as many dimensions as possible, but the match quality will never be perfect.
Additionally, these studies are typically not double-blind, and actors’ awareness of the object of
study and experimental design may bias the results.
One similar study, of discrimination by buyers and sellers in a goods market, is List (2004). As
in more typical audit studies, List uses black and white agents who directly signal their race—in
this case to dealers at a baseball card show. The agents are drawn from the general population
2
Ravina (2008) and Pope and Sydnor (Forthcoming) have also studied race through photographs in online markets,
although in non-experimental settings where non-random samples and unobservable characteristics are a
concern.
3
Skin color is clearly highly correlated with race. We believe that discrimination based on skin color is of primary
interest when people discuss racism, but there are surely many other relevant components of race that our study
ignores.
3
1 Introduction
of show attendees, so they inevitably differ along characteristics other than race. These nonrace
differences may help explain a disparity between offers made to black and white agents
somewhat larger than those we find (average offers 3–30% worse vs. our 2%), although there
are many other differences between our market environments. List also offers evidence from
complementary experiments that statistical discrimination is the main force explaining blackwhite
gaps; given the large number of card dealers competing for business and their expertise
(which could prepare them to effectively discriminate statistically), it is perhaps unsurprising
that animus plays less of a role. The online classified advertising market we consider is perhaps
more typical of interpersonal negotiation between non-experts where outside options (i.e.,
what a seller might do with an unsold iPod) are less clear. Although our experimental design
does not provide similar opportunities to tease out mechanisms, we do find discrimination is
heterogeneous across markets in patterns consistent with some expected channels through
which statistical discrimination might operate.
A number of studies have avoided the issues (and costs) associated with hiring actors by
signaling race through the use of racially distinctive names. Perhaps the best known application
of this approach was conducted by Bertrand and Mullainathan (2004), who responded to job
postings in Boston and Chicago using fictitious resumés randomly assigned either a distinctively
black or white name. The authors measured whether employers followed up with each application,
and found that those with black names had a callback rate one-third lower than those
with white names. This difference was remarkably consistent across industries, and persisted
for “higher-quality” (i.e., better educated and more experienced) applicants as well as for those
randomly assigned mailing addresses in more affluent neighborhoods. Similar studies have been
conducted in several other countries considering (for example) Middle Eastern names in Sweden
(Carlsson and Rooth, 2007) and Turkish names in Germany (Kass and Manger, 2010). A number
of other authors have used racially distinctive names to experimentally investigate the impact of
race in online markets including apartment rentals (e.g., Carpusor and Loges, 2006; Ahmed and
Hammarstedt, 2008; Ewens, Tomlin, and Wang, 2009; Bosch, Carnero, and Farré, 2010; Hanson
and Hawley, 2010) and low-value auctions (Nunley, Owens, and Howard, 2010).
Two primary criticisms of the Bertrand and Mullainathan design have been raised. The first
results from the use of names as a proxy for race, rather than a more direct signal. Market
actors may have viewed stereotypically black names as signals of socioeconomic status or family
background, and responded in a way that they might not have to a more typical black candidate.4
4
Bertrand and Mullainathan are forthright in recognizing this concern, suggesting even in the paper’s title that
their conclusions are fundamentally about applicants with names that are much more common among either
blacks or whites (like Lakisha, Jamal, Emily, and Greg).
4
2 Procedure
This concern applies for most “correspondence” studies, since names are typically the most
appropriate way of signaling race to a potential buyer, seller, or employer in an experimental
setting. Second, the measured outcome—callbacks, for instance—is often not the ultimate
outcome of interest. While the number of callbacks is interesting, it does not tell us how many of
those applicants might have been offered a job or apartment rental, or what wage or price they
might have received.
Our experimental design attempts to address these concerns in two ways. First, by signaling
race through the inclusion of a photograph, we can vary race while holding constant all other
signals sent about attributes of the seller.5
Second, the fact that online transactions are brought
near completion without face-to-face contact makes it possible to consider outcome measures
that are relatively “close” to the true outcomes of economic interest. One attractive feature of the
classified advertising market we consider relative to online auctions is that buyers expect that
completing their transaction will involve face-to-face interaction with the black or white seller;
as this is typical of many non-online transactions, we expect that our results will be informative
about discrimination offline.
The paper proceeds as follows: We begin with an overview of our experimental procedure, including
discussion of local classified advertising markets, the contents and timing of our ads, and
the manner in which we negotiated with respondents (Section 2). Next, we compare the response
to white, black, and tattooed sellers’ advertisements along a number of dimensions (Section 3).
Black (and tattooed) sellers receive fewer, less trusting responses and fewer, less valuable offers.
Section 4 considers whether the outcome differences vary systematically across markets. Black
sellers seem to fare particularly poorly in less competitive markets, in the Northeast, and in
markets with highest degrees of racial isolation or property crime. Section 5 concludes.
2 Procedure
The goal of this paper is to rigorously isolate the effect of race on market outcomes via a carefully
constructed field experiment. We have developed a procedure that avoids several confounding
factors present in other studies, and which is replicable in a variety of settings. We posted online
classified advertisements on locally focused websites throughout the United States over the
5
Of course, our treatment does signal that the seller has chosen to show her hand—and thus reveal her race.
Approximately 16% of the iPod Nano ads on the websites we consider with personal photos include the seller’s
hand, suggesting the practice is not uncommon (though not typical). To the extent that buyers interpret this
as a signal of the seller’s confidence or naïveté, for instance, this could be a confounding factor, though we do
not believe it has a large effect on our results. We would also note that it is impossible to hide one’s race in most
market transactions.
5
2 Procedure
course of one year, with variation along three key dimensions: race or social group of the seller
(as indicated by a photograph), asking price, and the “quality” of the advertisement text. The
photographs used are shown in Figure 1. Table 14 (in Appendix A) tabulates these advertisement
characteristics which—along with the markets in which the ads are posted and our posting and
negotiation procedures—are discussed in greater detail below.
Figure 1: Advertisement photographs
Note: These photographs have been slightly scaled down from the size included in our advertisements.
6
2 Procedure
2.1 Overview: Online classified advertisements
We posted advertisements for new Apple iPod Nano 8 GB Silver portable digital music players6
on local classified advertising websites in approximately 300 geographical markets. The sites
together compose a network that is a major national source of online classified advertising. All
sites are publicly accessible and fee-free for those looking to buy or sell items. We used all sites
available in the network as of March 2009.
Potential buyers responded via advertisement-specific, anonymized e-mail addresses. We
then followed typical practice in these markets, where a seller replies to individual buyers to
negotiate a final price and—if they reach agreement—arrange to complete the transaction. An
ad might receive zero responses or a dozen, depending on the market demand for a particular
good, the contents of the advertisement, and any number of idiosyncratic factors. The ensuing
negotiations are, in general, ad hoc; that is, there is no formal bidding mechanism, and either
party can cease communication at any time without facing any consequences.7
Among the experimental advantages of considering classified advertising in this setting are
the local focus and the lack of information each potential buyer has about other buyers and
their offers. Given the local focus of the sites on which we posted, buyers generally assume that
sellers are local. In addition, the network of sites provides no facility for viewers to browse or
search for advertisements across multiple markets, further encouraging local use. This is in
contrast to online auctions like eBay, where it is normal to do national searches. The local focus
allows us to analyze regional differences as well as variation by local economic and demographic
characteristics; it also made it feasible to post multiple advertisements in a limited time frame
while minimizing the risk that our analysis of any given advertisement is contaminated by our
other postings. Clearly, potential buyers’ bids are affected not only by how much the iPod is worth
to them, but their assumptions about the seller and the other buyers who might be bidding.
2.2 Local markets
Over the course of the experiment we posted at least three advertisements in each market
available in the network,8
which collectively covers the geography of all fifty states and Washington,
D.C. There are over 300 local sites, which include a wide variety of locations—from small
6
Apple released an updated iPod Nano model in the midst of our experiment. Our advertisements offer the current
model—the “4th generation” (model MB598LL/A) before September 9, 2009, and the “5th generation” (model
MC027LL/A) after that date. The two models appear almost identical in their packaging.
7
Indeed, our experience confirms anecdotal evidence that potential buyers regularly cease communicating in the
midst of discussing a potential sale.
8
In fact, we attempt at least three postings, not all of which were successful (as discussed below).
7
2 Procedure
towns in rural areas to the centers and suburbs of large cities. More information on the specific
markets is included in Appendix A. Within a single market, sellers choose a category in which to
list their advertisement; we posted in the “electronics for sale” category (as do the vast majority
of other advertisers offering iPods for sale).
We collected data on the number of iPod Nano advertisements in each market at the time of
our listing, and denote as “large” markets those with at least 20 advertisements posted between
one and seven days prior.9
At the time we posted our advertisements, the average market had
15.7 other advertisements for iPod Nanos that had been listed in the previous week, and 18% of
our ads were posted in large markets. Markets with more advertisements posted presumably
get more traffic from potential buyers. Thus, advertisements in thicker markets may get more
responses on average. On the other hand, in markets with more sellers our advertisements face
greater competition for prospective buyers’ attention and dollars.
Table 1 shows summary statistics for several market characteristics. In addition, Table 13 (in
Appendix A) shows average values for these characteristics broken down by advertisement type.
2.3 Advertisement contents
The contents of our advertisements varied along three dimensions: photograph (including skin
color), advertisement text, and asking price.
Photograph
Each advertisement included a photograph of a new, unopened iPod held in a black hand, a
white hand, or a white hand with a wrist tattoo. All the photographs were of a man’s hand so,
strictly speaking, our experimental design will only allow us to assess the discrimination faced
by black men. Three styles of photographs were used for the black and white hands; the need to
display the tattoo prevented us from matching all three hand positions perfectly in the last series
of photos. Nevertheless, the pictures (reproduced in Figure 1) are very similar in all ways other
than the apparent race or social group of the seller.
Photographs are very common in online classified advertisements, and are included in approximately
60% of the other iPod Nano advertisements we observe. Typically these are either
9
This count is based on a search for other advertisements in the same market that include the phrase “iPod Nano”
(regardless of capitalization) in their title. (This count therefore includes both new and used items, and some
non-iPod items, such as accessories.) Note that we have data on the stock of advertisements listed on the site
when we post, but not on any flow of advertisements posted. Sellers can remove their listings, so the number of
listings of vintage less than one week gives only a lower bound on the number of sellers to which a potential buyer
may have been exposed during that week.
8
2 Procedure
Table 1: Market characteristics—Summary statistics
Mean Std. Dev. 25% 50% 75%
Mkt. ads week 15.7 33.2 1 3 11
20+ weekly ads 0.18
Northeast 0.13
Midwest 0.24
South 0.36
West 0.27
% pop. White 77.0 16.1 67.1 81.5 90.1
% pop. Black 12.8 14.6 2.4 7.2 16.9
% pop. Hispanic 13.5 16.6 3.2 6.9 16.7
% pop. Asian 3.3 4.1 1.3 2.0 3.6
Poverty rate 15.7 6.3 11.7 14.7 19.1
Med. HH income ($K) 46.3 10.9 39.4 44.5 51.1
Black isolation index 0.21 0.17 0.06 0.19 0.34
Property crime rate 358 125 276 338 412
Observations 1200
Note: All observations equally weighted. Local racial composition, poverty rate, and household income are
from the 2007 American Community Survey. Black isolation index is degree to which “the average black
resident lives in a census tract in which the black share of the population exceeds the overall metropolitan
average” in 2000 (from Glaeser and Vigdor, 2001). 2008 property crimes are per ten thousand people (from
United States Department of Justice and Federal Bureau of Investigations, 2009). Isolation index and crime
data are not available for all markets.
9
2 Procedure
stock/marketing images or personal photographs of the item for sale; our photos are similar in
style to the personal photos many others use.
Advertisement text
Our advertisements (and the ensuing e-mail correspondence discussed in Section 2.5) randomized
over six different texts: three types, each with a “high quality” and a “low quality” variant.
We used multiple text types to create within-market variety that minimizes the apparent suspiciousness
associated with repeatedly posting ads in the same market. (We were concerned here
both with the websites’ users and with spam filters present on the sites themselves.) All six texts
are included in Appendix B.
The three high-quality texts use proper capitalization, punctuation and grammar, and were
generally well-written. Our low-quality advertisements had the same content, but with less
sophisticated wording and incorrect spelling, grammar, and capitalization. Our aim was to
provide a signal of the seller’s socioeconomic status, proxied by his education level and writing
ability.
Asking price
Each advertisement also included an asking price (both in a searchable price field and in the
text of the listing) of either $90, $110, or $130.10
The iPod we advertised were popular and widely
available through electronics retailers, mass market stores, online vendors, and Apple Stores. It
had a list price of $149.99 (plus local sales tax) and was available for sale prices of approximately
$135 throughout our experimental period, so all three asking prices were below the amount
buyers would have paid in a store. This asking price represented the “first offer” in the sale
negotiation, and we expect to see buyers’ responses depend on it. In addition to producing
an anchoring effect (as in Tversky and Kahneman, 1974), the specific asking price also sent
prospective buyers a signal about market conditions, the seller, and the quality of the product.
2.4 Timing
Our experimental period covered March 16, 2009 through March 15, 2010, excluding the periods
around major holidays (and various other times at which we suffered technical difficulties). Advertisements
were posted in the morning and evening (at approximately 9:30 A.M. and 9:30 P.M.,
10
We limited asking prices to $90 and $110 beginning in December, 2009 due to $130 advertisements’ very low
response rates.
10
2 Procedure
Pacific Time), with no more than four online at any given time. A tabulation of advertisement
timing by advertisement type is provided in Table 15 in Appendix A.
We removed our advertisements approximately twelve hours after they were posted; no new potential
buyers would view or respond to an ad after that point, though ongoing e-mail exchanges
could and did continue well beyond the twelve-hour mark. During a pilot of the experiment
in which we posted advertisements for longer durations, we found that the vast majority of
responses were received within twelve hours, and it was common practice to complete transactions
within a day or two after posting. Thus, our twelve-hour window gives us sufficient time to
receive responses from most likely buyers.
We added the non-race social signal dimension of this experiment after we had already begun
posting ads with black and white photos. Thus, a larger share of the later ads include tattoos.
The results reported below are robust to the inclusion of a quadratic time trend to control for
this correlation between ad type and timing, as well as several alternate strategies for controlling
for advertisement timing such as including in regression specifications the order in which ads
were posted within each market.
The weeks around two particular gift-giving holidays, Christmas and Valentine’s Day, saw a
large increase in responses to our ads and the offers received. Our analyses therefore include
controls for these two periods.11
2.5 Negotiation with respondents
Beginning approximately two hours after each advertisement is posted, we sent a response via
e-mail to each respondent saying that we had received many e-mails and asking for her best
offer (or to confirm that an offer made in an initial e-mail was indeed her best). The text of all
interactions was scripted and is included in Appendix C, together with additional details about
our negotiation procedure.
In the course of our correspondence with potential buyers, we received a large number of
“scam” offers (both as initial responses to our advertisements, and following our first e-mails).
These scams generally comprised offers to pay high prices to have the item shipped overseas;
several samples are included in Appendix D.12
We coded all requests for shipping or non-cash
11
We define the Christmas period as the Monday after Thanksgiving (November 30) through December 21. We did
not post advertisements from December 22 through January 5. The Valentine’s Day period runs from two weeks
before the holiday to one week after (January 31–February 21). We include the days after the holiday because
some buyers reported looking for gifts to reciprocate gifts they had unexpectedly received.
12
The associated fraud appears to operate in at least two ways. First, the “buyer’s” payment—whether by online
payment service, check, or money order—is counterfeit, allowing her to acquire the item at no cost. The second
technique is more insidious. The seller receives an e-mail purporting to be from her bank or an online payment
11
3 Results: Average effects
payments (and other similar responses) as scams, and ceased correspondence with these re-
spondents.13
Approximately 48 hours after removing each advertisement, we offered to sell the iPod, by
postal mail, to the respondents who made and confirmed the highest offer and (when available)
the second-highest offer. We apologized for being out of town, and told the respondent we were
willing to mail her the iPod in exchange for payment via PayPal, an electronic payment system
widely used for online person-to-person transactions. The time delay was intended to make our
shipment proposal less suspicious; buyers might think we were local but had to leave town after
posting the ad. We sent the iPods to those who agree to this, and replied to all other confirmed
bidders that the iPod was no longer available.
The reasons we chose to offer shipment rather than in-person delivery were principally logistical,
but we also sought to avoid introducing unobservable (and uncontrollable) variation. Given
the local nature of our advertisements and the sites we posted them on, most high-bidders are
understandably wary of a long-distance transactions; those who agreed to trade this way are
unlikely to compose a representative sample of potential buyers. Nonetheless, we completed as
many transactions as possible in the spirit of honestly following through on our advertised offer
to sell.
3 Results: Average effects
We consider six different outcome measures: whether our advertisements were prematurely
removed by website users, the number of responses received, qualitative characteristics of the
responses’ contents, the dollar amounts offered, high bidders’ reactions to our stated inability to
deliver the iPod in person, and the probability that an ad resulted in a successful sale. Average
values for these measures by advertisement type are reported in Table 2.
The following subsections present findings on the effects of race on these outcomes, controlling
for a variety of advertisement, timing, and market characteristics. Given that seller race was
experimentally varied independent of these other characteristics, the race effects are consistently
estimated by the difference or ratio of means reported in Table 2; including additional controls
merely allows us to increase the estimates’ precision.
service, confirming that a payment has been received. The web links in this e-mail lead to sites controlled by the
scammer, who hopes that the seller will enter her bank account or online payments password.
13
After several months of reading and responding to potential buyers’ e-mails, it became increasingly obvious
which e-mails were attempted scams. Since not all of these e-mails result in follow-ups that would confirm our
suspicions, we coded such responses as “probable scams” to distinguish them from genuine offers. We code
responses as probable scams if the text of the e-mail or e-mail address is identical to those from a confirmed scam
12
3 Results: Average effects
Table 2: Key outcome averages by advertisement type
White Black Tattoo Total
Prematurely removed 0.028 0.056 0.041 0.041
Number of responses
Number of non-scams 2.49 2.11 2.11 2.25
Number of offers 1.71 1.39 1.47 1.53
Response characteristics (given ≥ 1 non-scam response)
Incl. name 0.388 0.307 0.318 0.341
Polite 0.408 0.373 0.351 0.381
Incl. personal story 0.038 0.048 0.048 0.044
Offer amount (given ≥ 1 offer)
Mean offer 86.02 83.89 83.53 84.63
Best offer 94.05 90.25 90.41 91.79
Reaction to delivery proposal (given delivery proposed)
Scam/payment concern 0.076 0.113 0.078 0.089
No response 0.375 0.421 0.401 0.397
Other 0.193 0.127 0.204 0.175
Prefer to wait 0.299 0.315 0.267 0.295
Willing to ship 0.056 0.025 0.051 0.045
iPod shipped 0.038 0.016 0.030 0.028
Mean values are reported.
Note: Observations are weighted by state population/number of ads posted in each state. All rows except “prematurely
removed” exclude prematurely removed advertisements.
13
3 Results: Average effects
Table 3: Key outcome regressions
(1) (2) (3) (4) (5) (6)
Prem. rem. Nonscams Offers Mean offer Best offer Shipped
Probit Neg. Bin. Neg. Bin. OLS OLS Probit
Black 0.0270∗
0.875∗∗
0.830∗∗
-1.872+
-3.558∗∗
-0.0141∗
(0.0149) (0.0525) (0.0631) (1.250) (1.404) (0.00766)
Tattoo 0.0136 0.827∗∗∗
0.841∗∗
-2.676∗∗
-4.149∗∗∗
-0.00513
(0.0141) (0.0525) (0.0603) (1.328) (1.554) (0.00705)
High quality -0.0158+
0.975 1.005 0.325 0.760 -0.00559
(0.0103) (0.0526) (0.0603) (1.152) (1.229) (0.00680)
Price $110 0.00607 0.422∗∗∗
0.388∗∗∗
11.38∗∗∗
6.269∗∗∗
-0.0101+
(0.0116) (0.0264) (0.0278) (1.211) (1.323) (0.00615)
Price $130 -0.0242∗∗
0.223∗∗∗
0.182∗∗∗
20.32∗∗∗
12.83∗∗∗
-0.0278∗∗∗
(0.0109) (0.0180) (0.0190) (2.104) (2.145) (0.00735)
Christmas 0.0209 1.951∗∗∗
2.119∗∗∗
5.531∗∗∗
11.80∗∗∗
0.0248
(0.0307) (0.180) (0.242) (1.733) (2.106) (0.0262)
Valentine’s Day -0.00724 1.290∗∗∗
1.329∗∗∗
2.230 1.895 0.00608
(0.0195) (0.110) (0.120) (2.516) (2.439) (0.0143)
Night -0.00836 0.696∗∗∗
0.665∗∗∗
-1.118 -2.938∗∗
0.000771
(0.00972) (0.0427) (0.0422) (1.130) (1.273) (0.00641)
20+ weekly ads -0.00482 2.021∗∗∗
2.133∗∗∗
0.863 6.156∗∗∗
-0.00619
(0.0187) (0.174) (0.199) (1.255) (1.874) (0.00892)
Med. HH inc. (log) -0.0615 2.608∗∗∗
3.209∗∗∗
12.67∗∗
10.87+
0.0719∗∗
(0.0447) (0.805) (1.089) (5.326) (6.717) (0.0352)
Poverty rate -0.00198 1.026∗∗
1.039∗∗∗
0.267 0.251 0.00253∗
(0.00155) (0.0122) (0.0138) (0.221) (0.243) (0.00133)
% pop. White 0.000108 0.996∗
0.997 -0.00465 -0.0574 -0.000122
(0.000455) (0.00235) (0.00255) (0.0442) (0.0469) (0.000294)
Northeast -0.0111 0.978 1.090 -4.116∗∗
-2.468 -0.00516
(0.0144) (0.0971) (0.115) (1.953) (2.301) (0.0128)
Midwest -0.0107 0.980 0.974 -1.000 0.0924 0.00558
(0.0136) (0.0817) (0.0911) (1.475) (1.739) (0.0116)
South -0.0154 0.879 0.924 -2.015 -2.486 0.00619
(0.0152) (0.0820) (0.0975) (1.506) (1.801) (0.0103)
Observations 1200 1147 1147 663 663 1147
White mean 0.0278 2.493 1.707 86.02 94.05 0.0378
Marginal effects from probit estimation are reported in (1) and (6); (2)–(3) report incidence rate ratios from negative binomial estimation;
(4)–(5) report OLS coefficients. Standard errors clustered by market are reported in parentheses.
+ p < 0.15, ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
Note: Observations are weighted by state population/number of ads posted in each state. (2)–(6) exclude prematurely removed advertisements,
and (4)–(5) limited to advertisements receiving at least one offer.
14
3 Results: Average effects
3.1 Premature advertisement removal
The sites on which we posted provide tools for users to mark advertisements as inappropriate or
unwelcome. If enough users protest a particular ad in this way, it is removed from the website.
4.1% of our advertisements were removed in this manner. In addition to legitimate use of this
feature, other sellers may disingenuously mark competing ads as inappropriate in order to
reduce competition.14
Column 1 of Table 3 provides estimated marginal effects associated with probit estimation
assessing which advertisements are most likely to be prematurely removed. The regression
controls for a variety of advertisement, market, and timing characteristics that explain a substantial
amount of the variation in our dependent measures. Appendix E reports results for a
specification without controls and another that replaces the market controls with market fixed
effects.
On average, black sellers’ ads are prematurely removed 2.7% more of the time than white
sellers’; the likelihood is thus almost twice as high that a black sellers’ ad will be removed.
Clearly if an advertisement is removed, it limits the seller’s opportunity to receive responses
and bids from potential buyers. Although this is an economically relevant outcome for a seller,
she can also repost her advertisement. In the name of conservatism, we therefore exclude
prematurely removed advertisements from our subsequent analyses; Appendix E replicates
several of the paper’s results including these ads.
3.2 Number of responses
Approximately 82% of our advertisements received some response, and on average they received
2.7 responses. We identify a number of our responses as disingenuous “scams,” and partition
the remainder based on whether or not they result in a specific dollar offer. Table 4 provides
summary statistics on the number of responses received broken down by response type.
Given that the number of responses (of each type) received are count variables, we estimate
e-mail we received earlier. Our results are robust to this alternative coding procedure.
14
In addition, the websites implement filters (based on unknown algorithms that appear to change frequently)
to identify unwelcome advertisements. On several occasions, all of the ads we posted on a given morning or
evening were immediately removed from the site. This universal, simultaneous premature removal suggested
that our ads were caught in the websites’ filters. Similarly, some ads did not show up in search results despite
appearing to have posted successfully; this is also due to the websites’ screening for unwelcome ads. All of these
advertisements are entirely excluded from our analyses. Stratifying the advertisements we posted simultaneously
by advertisement quality and market size greatly reduced this automatic removal.
15
3 Results: Average effects
Table 4: Number of responses—Summary statistics
Mean Std. Dev. 25% 50% 75% 95% Max. Frac. > 0
Responses 2.70 2.78 1 2 4 8 17 0.82
Scams 0.45 0.79 0 0 1 2 10 0.33
Non-scams 2.25 2.75 0 1 3 8 17 0.71
Offers 1.53 2.07 0 1 2 6 15 0.60
Observations 1147
Note: Observations are weighted by state population/number of ads posted in each state. Excludes prematurely removed
advertisements.
the impact of race and other covariates using models of the form
responsesi ∼ Poisson νi exp(xi β) (1a)
νi ∼ Gamma(1/α,α), (1b)
where i indexes advertisements and xi is the ith
row of the data matrix X, containing the covariates
for advertisement i. This yields a Negative Binomial distribution for the outcome
of interest (conditional on covariates)15
Note that this Negative Binomial distribution has
E[responsesi ] = exp(xi β); thus the reported exponentiated coefficient estimates (corresponding
to exp(βj ) in Equation 1) should be interpreted as incidence rate ratios. A covariate has a
positive effect on the outcome measure precisely when its corresponding exponentiated coefficient
is greater than one; to determine the combined effect of several covariates, multiply the
exponentiated coefficients together.
Responding to an advertisement requires no commitment and limited time, so it is cheap,
but it is not free. There is no incentive for anyone to respond to an ad in which he is completely
uninterested. Also, the number of responses received is unaffected by our subsequent e-mail
correspondence, which may send additional signals about the seller and the local market. In
particular, our first scripted e-mail response suggests that there is a lot of interest in our iPod (i.e.,
that the market is competitive) and that the seller is fairly savvy and organized in his approach to
selling the item. Thus, the number of responses may best reflect local buyers’ prior assumptions
about black and white sellers, as well as the demand to purchase from them. To the extent that
our correspondence provided additional information that contradicts these assumptions, some
buyers might have ceased communication because they were no longer interested in purchasing
15
In these Negative Binomial models, α 0 parameterizes over/underdispersion relative to the Poisson distribution,
since responsesi → Poisson exp(xi β) as α → 0.
16
3 Results: Average effects
from us, not because they were not serious to begin with.
Column 2 of Table 3 reports the results of a maximum likelihood estimation of Equation 1
for the number of non-scam responses received. While our average advertisement received 2.3
non-scam responses, black sellers receive 13% fewer responses than white sellers. Tattooed
sellers appear to suffer even more discrimination than blacks along this margin, receiving 17%
fewer responses than white sellers.
Several other covariates seem to have the expected effects: high asking prices depress response,
and advertisements posted at night or in markets with few other ads fare poorly. Perhaps surprisingly,
advertisement quality appears to have no effect on the number of responses received.
The number of dollar-valued offers16
received may be a more reliable measure of serious
interest, especially if we think that some buyers were searching for a good deal by indiscriminately
responding to many sellers’ ads. We record the dollar amount of an offer whether it comes in
the initial inquiry or in response to our reply. Approximately two thirds of non-scam responses
resulted in an offer, and the average advertisement received 1.5 offers.
We report Negative Binomial regression results for the number of offers in column 3 of Table 3.
Black sellers receive 17% fewer offers than white sellers, while tattooed sellers receive 16% fewer.
3.3 Response characteristics
The manner in which buyers respond to advertisements may indicate their underlying level of
respect or trust. We analyze the text of the first e-mail each buyer sends, identifying whether
1. The buyer included or signed her name (34% of responses);
2. The words “please,” “thank you,” or variations such as “pls,” “thx,” or “thanks” appeared
anywhere in the e-mail text (38%); and/or
3. The buyer included a personal story, presumably to appeal to the seller’s sentiments and
get a lower price (4%).
Note these characteristics are neither mutually exclusive nor collectively exhaustive; examples of
responses exhibiting each are included in Appendix D.
Table 5 reports probit regression results for these three attributes of buyers’ responses. Obviously,
this analysis is restricted to advertisements which received at least one non-scam response,
which may introduce some selection effect. Overall, buyers are more likely to act respectfully
when communicating with white sellers. Approximately 7% fewer buyers sign their names when
16
Throughout the paper, we refer as “offers” only to cash offers. Approximately 4% of non-scam respondents offered
to trade various goods and services—from live snakes to auto detailing—for our iPod. Several examples are
included in Appendix D.
17
3 Results: Average effects
responding to black rather than white sellers; thus the average response received by a black seller
is 17% less likely to include the buyer’s name. This is similar to the effect observed for tattooed
sellers.
Buyers are slightly less likely to use polite language when responding to black or tattooed
sellers’ advertisements, although these results do not rise to the level of statistical significance.
Table 5: Probit regression of response characteristics
(1) (2) (3)
Name Polite Personal
Black -0.0652∗∗
-0.0135 0.0134
(0.0290) (0.0299) (0.0124)
Tattoo -0.0740∗∗
-0.0364 0.00614
(0.0311) (0.0321) (0.0139)
Standard ctls.
Observations 2488 2488 2488
White mean 0.388 0.408 0.038
Probit marginal effects are reported. Standard errors clustered by advertisement are reported
in parentheses.
+ p < 0.15, ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
Note: Observations are weighted by reciprocal of number of responses per advertisement. Excludes
responses to prematurely removed advertisements. “Standard controls” are: high advertisement
quality, asking price ($130 and $110 dummies; $90 excluded), holidays (Christmas
and Valentine’s day dummies), night, 20+ iPod Nano ads in market over previous week,
median household income, poverty rate, non-Hispanic white fraction of local population,
and region (Northeast, Midwest, and South dummies; West excluded).
3.4 Offer amount
The ultimate reason that the number of responses is economically important to a seller is that it
increases the probability of receiving a good offer and of completing a sale. We thus look at both
the mean and maximal offers made in response to each advertisement, limiting our analysis
to the 60% of advertisements that received at least one dollar-valued offer. To the extent that a
seller is able to successfully complete a sale with the highest bidder at that bidder’s offered price,
the “best offer” received is the outcome of primary economic importance to the seller.
Conditional on receiving an offer, the mean offer received averaged $86.02. We present
Ordinary Least Squares results assessing the effect of advertisement type and other covariates on
the average offer value in column 4 of Table 3. Compared with white sellers, black sellers receive
offers $1.87 lower, and tattooed sellers $2.68 lower.
18
3 Results: Average effects
Although higher asking prices lead to higher offers (conditional on receiving at least one offer),
this is clearly offset by the reduced number of offers received. As we found for the number of
responses received, advertisement quality appears to have no effect on offer value. Finally, we
note that average offers are approximately $1–4 higher in the West (the excluded category) than
in the rest of the country, with the lowest average offers in the Northeast.
Column 5 of Table 3 reports Ordinary Least Squares regression results for the best offer received.
Conditional on receiving at least one offer, the average maximal offer was $94.05. Given our
earlier findings that black sellers receive fewer and lower offers, it is unsurprising that the their
best offers are $3.56 lower than whites’. Tattooed sellers also suffer, with best offers $4.15 lower
than white sellers. These differences are statistically significant.17
3.5 Reactions to delivery proposal
After we took an advertisement down, we contacted the highest bidder to say that we would mail
the iPod to her if she would pay us using PayPal. Because the websites include warnings about
the risks of non-local transactions, we did not expect many buyers to accept this offer. However,
the manner in which they declined can tell us something about their inclination to trust the
seller. Buyers’ initial responses to our delivery proposal fall into one of five mutually exclusive
categories, listed here in order of most to least positive:
5. Suggesting an openness to receiving the iPod by mail (6% of proposed deliveries);
4. Offering to wait and meet when we get back into town (30%);
3. Declining for some other reason (19%);
2. No response (37%), which we interpret as a signal of some distrust; or
1. Explicitly accusing us of trying to scam them, or saying they don’t want to use PayPal,
which we interpret as a concern about being scammed (8%).
Examples of each type of reaction are included in Appendix D.
In Table 6, we report the results of ordered probit regressions of buyers’ reactions to our
delivery proposal on advertisement type. These regression specifications allow us to test whether
each seller type received “more positive” reactions as measured by the ordinal ranking above.
Although the coefficient estimates are difficult to interpret, the statistically significant negative
coefficients on black support the hypothesis that black sellers receive worse reactions to their
17
We were curious about the extent to which the lower best offers received by black and tattooed sellers are merely
mechanically driven by the fact that they receive fewer offers: taking fewer draws from even identical offer
distributions would produce lower maxima. Table 8 shows the regression separately for advertisements receiving
one and more than one offer. The difference in best offers received by black and white sellers seems mainly to be
driven by those advertisements that only receive a single offer.
19
3 Results: Average effects
delivery offers, suggesting an underlying distrust of black sellers. Note that this is the case even
though the sample consists only of the (presumably less biased) potential buyers who chose to
respond to those sellers’ ads.
Table 6: Ordered probit regression of “positivity” of reaction to delivery pro-
posal
(1) (2)
React. to delivery proposal React. to delivery proposal
Black -0.181∗
-0.184∗
(0.104) (0.105)
Tattoo -0.145 -0.145
(0.108) (0.109)
Standard ctls.
Observations 604 604
Ordered probit coefficients are reported. Standard errors clustered by advertisement are reported in paren-
theses.
+ p < 0.15, ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
Note: Observations are weighted by reciprocal of number of delivery proposals per advertisement. Outcomes
ranked from most to least positive are: willing to ship, prefer to wait, other, no response, and scam/payment
concern. Excludes deliveries proposed to respondents to prematurely removed advertisements. “Standard
controls” are: high advertisement quality, asking price ($130 and $110 dummies; $90 excluded), holidays
(Christmas and Valentine’s day dummies), night, 20+ iPod Nano ads in market over previous week, median
household income, poverty rate, non-Hispanic white fraction of local population, and region (Northeast, Midwest,
and South dummies; West excluded).
Table 23 in Appendix E reports probit estimates of the frequency of receiving each individual
reaction type. For example, a buyer is 44% less likely to accept delivery by mail (an absolute
difference of 2.5 percentage points) and 56% more likely to express concern about making a
long-distance payment (an absolute difference of 4.2 percentage points) when responding to a
black rather than white seller; both these differences are significant at the 15% level.
3.6 Shipment
After offering to ship the iPod to the highest bidder, our procedure becomes more ad hoc out of
necessity (we must respond to questions, and work out the logistics of shipment and payment),
but remains blind to the seller’s type. Column 6 of Table 3 reports the effect of seller type on the
probability that advertisement results in a successful transaction. While the number of successes
is small (since delivery by mail is not typical in this market), the race of the seller still has an
effect: ads posted by black sellers ultimately result in sales almost 40% less often than ads posted
by white sellers.
20
4 Understanding observed discrimination
4 Understanding observed discrimination
In the previous section, we analyzed the differences in a number of outcomes faced by white,
black, and tattooed sellers. We now investigate whether these differences vary systematically
across markets, focusing on three key outcome measures: the number of offers received, the
mean offer and the best offer. It is important to emphasize that unlike race, which we vary experimentally,
the evidence in this section relies on variation across markets—competitive, regional,
and demographic—that we do not control. While evidence that discrimination is particularly
prevalent in particular types of markets can help shed light on its underlying mechanisms, we
are mindful that a variety of mechanisms might be consistent with this evidence.
4.1 Degree of market competition
In theory (Becker, 1971), discrimination against black sellers should be less present in more
competitive markets.18
We use an indicator of market thickness—the presence of more than 20
ads per week19
—to test whether this is the case in our data; that is, we test the hypothesis that
the coefficient β20+×Black is greater than one in the Negative Binomial and greater than zero in
the Ordinary Least Squares regressions.
Table 7 suggests that black sellers indeed face less discrimination in more competitive markets.
In markets with more than 20 weekly advertisements, black sellers receive approximately the
same number of offers as whites ( ˆβBlack · ˆβ20+×Black −1 = −1%), while in less competitive markets
black sellers receive 24% fewer offers (= 1− ˆβBlack). Similarly, in more competitive markets black
sellers’ best offers are about the same as whites’ ( ˆβBlack + ˆβ20+×Black = −$0.22), rather than being
$4.64 lower (= ˆβBlack).
Similarly, if discrimination is competed away, then black sellers should do relatively better
when their advertisements are likely to receive more offers (whether due to the market, or nonrace
characteristics such as asking price). Table 8 shows regression results of offer value separately
for advertisements that received one vs. multiple offers. When there is less competition for their
product, the offers received by black sellers are more than $6 lower than those received by whites;
when there is more competition (indicated by at least two offers), the difference in offer values
18
We are unable to observe the number of potential buyers in each market, but do know the number of other iPod
Nano ads online at the same time as our ad. We use this measure of market thickness as a proxy for the number of
potential buyers, but we do not know how highly the two are correlated. If the number of competing ads is a good
proxy for the number of buyers, we should see less discrimination in thicker markets. If the number of competing
ads is only an indicator of the number of sellers, discrimination could be greater in thicker markets, since buyers
have more power.
19
The results are similar for a variety of threshold values above 14 ads per week.
21
4 Understanding observed discrimination
Table 7: Heterogeneous results by number of weekly market ads
(1) (2) (3)
Number of offers Mean offer Best offer
Black 0.765∗∗∗
(0.0673) -3.350∗∗
(1.566) -4.642∗∗∗
(1.746)
Tattoo 0.817∗∗
(0.0767) -4.044∗∗
(1.753) -5.513∗∗∗
(1.893)
20+ weekly ads 1.899∗∗∗
(0.224) -2.661+
(1.836) 3.197 (2.732)
20+ × Black 1.298∗
(0.203) 6.030∗∗
(2.380) 4.422+
(2.793)
20+ × Tattoo 1.108 (0.164) 4.932∗∗
(2.430) 4.788 (3.353)
Standard ctls.
Observations 1147 663 663
All black=0 p. 0.00961 0.0363 0.0298
Incidence rate ratios from negative binomial estimation are reported in (1); OLS coefficients are reported in (2)–(3).
Standard errors clustered by market are reported in parentheses.
+ p < 0.15, ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
Note: Observations are weighted by state population/number of ads posted in each state. Excludes prematurely removed
advertisements, and (2)–(3) limited to advertisements receiving at least one offer. “Standard controls” are:
high advertisement quality, asking price ($130 and $110 dummies; $90 excluded), holidays (Christmas and Valentine’s
day dummies), night, 20+ iPod Nano ads in market over previous week, median household income, poverty rate, nonHispanic
white fraction of local population, and region (Northeast, Midwest, and South dummies; West excluded).
nearly disappears. The results are similar for tattooed sellers.
An alternate interpretation of these results is that black sellers face less discrimination in cities,
where thicker online markets are more likely to be found. Residents of cities tend to be more
racially diverse and younger (per the 2000 Census), and may be more accustomed to interacting
with people of other races and ethnicities. Because these market characteristics (thickness and
urbanity) are highly correlated, we are unable to distinguish whether market competition has
“crowded out” discrimination, or whether buyers inclined against discrimination are merely
more likely to live in thick competitive markets.
4.2 Regional cultural norms
The results reported in previous sections have included region fixed effects to control for geographic
variation in factors such as the popularity of online classified advertising, as well as
cultural norms and incomes. However, these local norms should be an important determinant of
individuals’ racial biases, and culture differs greatly across regions. We thus expect that the effect
of race on our economic outcomes of interest may be heterogeneous with respect to region. In
Table 9 we report regression results that allow the differences between seller types to vary by
region. (The West is the excluded region.)
We originally hypothesized that, due to its history of slavery and troubled race relations, the
22
4 Understanding observed discrimination
Table 8: Offer value by number of offers
(1) (2) (3)
1 offer Mean of 2+ Best of 2+
Black -6.214∗∗
(2.666) 0.648 (1.107) -1.058 (1.332)
Tattoo -5.739∗∗
(2.773) -0.721 (1.231) -1.979 (1.619)
Standard ctls.
Observations 277 386 386
OLS coefficients are reported. Standard errors clustered by market are reported in parentheses.
+ p < 0.15, ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
Note: Observations weighted by state population/number of ads posted in each state. Excludes prematurely
removed advertisements. “Standard controls” are: high advertisement quality, asking price ($130 and $110
dummies; $90 excluded), holidays (Christmas and Valentine’s day dummies), night, 20+ iPod Nano ads in
market over previous week, median household income, poverty rate, non-Hispanic white fraction of local
population, and region (Northeast, Midwest, and South dummies; West excluded).
Table 9: Heterogeneous results by region
(1) (2) (3)
Number of offers Mean offer Best offer
Black 1.010 (0.145) -2.388 (1.845) -3.495∗
(2.041)
Tattoo 0.953 (0.121) -2.025 (2.444) -1.822 (3.141)
Northeast 1.360∗∗
(0.207) -3.377 (3.017) -1.108 (3.503)
Midwest 1.130 (0.161) -1.303 (2.270) 0.710 (2.822)
South 0.993 (0.137) -2.317 (2.003) -1.836 (2.611)
Northeast × Black 0.672∗
(0.149) -0.803 (4.416) -2.295 (4.704)
Northeast × Tattoo 0.691∗
(0.151) -2.023 (4.407) -2.229 (4.984)
Midwest × Black 0.761 (0.172) 0.377 (3.150) -0.154 (3.721)
Midwest × Tattoo 0.816 (0.174) 0.737 (3.351) -2.050 (4.152)
South × Black 0.840 (0.159) 1.835 (2.826) 1.312 (3.225)
South × Tattoo 0.950 (0.171) -1.078 (3.294) -3.853 (4.049)
Standard ctls.
Observations 1147 663 663
All black=0 p. 0.0524 0.526 0.116
Incidence rate ratios from negative binomial estimation are reported in (1); OLS coefficients are reported in (2)–(3).
Standard errors clustered by market are reported in parentheses.
+ p < 0.15, ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
Note: Observations are weighted by state population/number of ads posted in each state. Excludes prematurely
removed advertisements, and (2)–(3) limited to advertisements receiving at least one offer. “Standard controls”
are: high advertisement quality, asking price ($130 and $110 dummies; $90 excluded), holidays (Christmas and
Valentine’s day dummies), night, 20+ iPod Nano ads in market over previous week, median household income,
poverty rate, and non-Hispanic white fraction of local population.
23
4 Understanding observed discrimination
South would be the worst region for black sellers. However, black sellers were at the greatest
disadvantage in the Northeast, where they received 32% fewer offers than whites. In contrast,
these gaps were 23% in the Midwest and 15% in the South; black sellers received 1% more offers
than white sellers in the West.
4.3 Other heterogeneous effects consistent with statistical
discrimination
The presence of larger black-white outcome gaps in some settings than others could suggest
the presence of statistical discrimination as distinct from animus. The former generally refers
to discrimination where race is used as a proxy for other characteristics that buyers cannot
observe directly, but wish to avoid (e.g., low socioeconomic status). Animus, or taste-based
discrimination, is a negative reaction to race itself, independent of other characteristics.
We expect that buyers might statistically discriminate in this market to avoid one or more
of the following: (1) buying fake or stolen goods,20
(2) sellers they would need to meet in an
inconvenient/dangerous neighborhood, and (3) sellers who would not complete the transaction
because they are unreliable.
In general, statistical discrimination should decrease when more information about the seller’s
other (more relevant) characteristics becomes available. Animus would not be affected by
additional information.
Advertisement text
Each of our advertisements was randomly assigned either a high- or low-quality text to provide a
signal of the seller’s socioeconomic status, proxied by his education level and writing ability. If
low socioeconomic status is highly correlated with the characteristics that buyers are trying to
avoid, discrimination should decrease in the presence of a high-quality ad. That is, if statistical
discrimination against black sellers is operative, it should be smaller when advertisements are
high quality; it might therefore manifest itself as coefficients on the interaction between black
and high advertisement quality being greater than one in the Negative Binomial and greater than
zero in the Ordinary Least Squares regressions reported in Table 10.
In terms of the number of offers received, black sellers do benefit slightly more than white
sellers by posting high-quality ads, but the effect is fairly small, and statistically insignificant.
However, given the limited importance we find for our quality measure even in our average effects
20
Indeed, the iPod Nano we sell is probably more likely to be fake or stolen than many other goods, so it is a
particularly good test for this form of statistical discrimination.
24
4 Understanding observed discrimination
Table 10: Heterogeneous results by advertisement quality
(1) (2) (3)
Number of offers Mean offer Best offer
Black 0.795∗∗
(0.0798) -1.293 (1.806) -2.529 (2.155)
Tattoo 0.824∗
(0.0916) -2.414 (1.852) -4.210∗
(2.190)
High quality 0.968 (0.0935) 0.819 (1.677) 1.379 (1.951)
HQ × Black 1.087 (0.154) -1.098 (2.744) -1.971 (3.246)
HQ × Tattoo 1.039 (0.160) -0.473 (2.650) 0.128 (3.004)
Standard ctls.
Observations 1147 663 663
All black=0 p. 0.0356 0.318 0.0372
Incidence rate ratios from negative binomial estimation are reported in (1); OLS coefficients are reported in
(2)–(3). Standard errors clustered by market are reported in parentheses.
+ p < 0.15, ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
Note: Observations are weighted by state population/number of ads posted in each state. Excludes prematurely
removed advertisements, and (2)–(3) limited to advertisements receiving at least one offer. “Standard
controls” are: asking price ($130 and $110 dummies; $90 excluded), holidays (Christmas and Valentine’s
day dummies), night, 20+ iPod Nano ads in market over previous week, median household income, poverty
rate, non-Hispanic white fraction of local population, and region (Northeast, Midwest, and South dummies;
West excluded).
analysis, it seems likely that our low- and high-quality advertisements are simply insufficiently
different to affect response.21
Racial isolation
Buyers might statistically discriminate against black sellers if they assume it would be inconvenient
to travel to meet those sellers. (If the seller is the one traveling, buyers might assume he
is less reliable because of the inconvenience.) This is more likely when local black and white
populations are more geographically isolated from one another. Glaeser and Vigdor (2001)
created an “isolation” index to measure segregation in metropolitan areas across the country;
their data map to approximately 82% of our markets. The index increases from zero to one with
greater isolation, and indicates the degree to which “the average black resident lives in a census
tract in which the black share of the population exceeds the overall metropolitan average.” That
is, it measures how geographically segregated the local black population is from the local white
population.
We denote markets in the top quartile of racial isolation scores as exhibiting “high isolation,”
21
Buyers might also have interpreted the low-quality text as signaling youth or “hipness” instead of low socioeconomic
status. To the extent that different buyers had opposing reactions to ad quality, the effects might have
canceled out.
25
4 Understanding observed discrimination
and consider the differential effect of race in those markets. If statistical discrimination is
operative in this market due to buyers assuming that buying from a black seller is likely to involve
greater inconvenience, black sellers should have worse outcomes in high-isolation markets.22
This would result in coefficients on the interactions between black and high isolation that are
less than one in the Negative Binomial and less than zero in the OLS regressions.
Table 11: Heterogeneous results by black isolation index
(1) (2) (3)
Number of offers Mean offer Best offer
Black 0.783 (0.137) -6.610∗∗∗
(2.474) -7.321∗∗∗
(2.789)
Tattoo 0.827 (0.150) -4.298 (3.513) -4.372 (4.120)
High isolation 1.351∗∗
(0.181) 0.836 (2.201) 3.617 (2.759)
High iso. × Black 0.544∗∗∗
(0.112) -3.504 (3.526) -9.368∗∗
(4.373)
High iso. × Tattoo 0.789 (0.142) 1.474 (4.050) -4.428 (4.424)
Regions × Black/Tattoo
20+ × Black/Tattoo
Standard ctls.
Observations 924 577 577
All black=0 p. 0.00813 0.0935 0.0636
Incidence rate ratios from negative binomial estimation are reported in (1); OLS coefficients are reported in (2)–(3). Standard errors
clustered by market are reported in parentheses.
+ p < 0.15, ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
Note: Observations are weighted by state population/number of ads posted in each state. Excludes prematurely removed advertisements,
and (2)–(3) limited to advertisements receiving at least one offer. “High isolation” markets are top 25% as measured by
degree to which “the average black resident lives in a census tract in which the black share of the population exceeds the overall
metropolitan average” in 2000 (from Glaeser and Vigdor, 2001). “Standard controls” are: high advertisement quality, asking
price ($130 and $110 dummies; $90 excluded), holidays (Christmas and Valentine’s day dummies), night, 20+ iPod Nano ads in
market over previous week, median household income, poverty rate, non-Hispanic white fraction of local population, and region
(Northeast, Midwest, and South dummies; West excluded).
Indeed, this is what we find. To account for the fact that racial isolation varies systematically
with other important market characteristics, we exploit only variation in isolation between
similarly thick markets in the same region. For example, among Western markets with fewer than
twenty weekly ads (the excluded categories), black sellers receive 57% fewer offers than white
sellers in high isolation markets, compared with 12% fewer in low isolation markets. The best
offers black sellers receive are nearly $17 lower than those white sellers receive in high isolation
markets, compared with a $7 gap in markets with less isolation. We do not expect the effect of
22
Of course, the extent to which black residents are geographically isolated from white residents in a particular
market may also be positively correlated with animus for a variety of reasons. (Either one could cause the other,
or both might be attributable to some third factor.) To the extent that this is the case, ours is an imperfect test for
statistical discrimination.
26
4 Understanding observed discrimination
having a tattoo to vary as much with the degree of racial isolation, and indeed it does not.
Property crime rate
Buyers might also statistically discriminate against black sellers if they think those sellers are
more likely to sell stolen goods or that it is more dangerous to meet those sellers in person
(because the sellers live in high-crime markets or are criminals themselves). Using data from the
Uniform Crime Statistics (United States Department of Justice and Federal Bureau of Investigations,
2009), which map to 88% of our markets, we designate markets with 2008 property crime
rates in the top quartile of our sample as “high-crime” areas. We then test the hypothesis that
buyers are more likely to discriminate against black sellers in areas with high property crime
rates than they are in areas with less crime, as might be expected if transacting with a black seller
is perceived as posing a disproportionate risk of exposure to criminality.23
Indeed, we do find that black sellers have the worst outcomes in high-crime areas: They
receive 21% fewer offers and $12 lower best offers than white sellers, compared with 2% fewer
offers and $4 lower best offers in low-crime (thin, western) markets. The effect is directionally
similar, albeit somewhat smaller, for tattooed sellers.
23
As with our racial isolation test, this is an imperfect test of statistical discrimination since property crime rates are
likely positively correlated with racial animus. We are testing a hypothesis consistent with a particular channel
through which statistical discrimination might be expected to operate; we are not testing for the existence of this
form of statistical discrimination. We might also be concerned about whether race is actually serving as a proxy
for crime, or whether the latter is merely correlated with some other factor that respondents are (statistically)
discriminating against. Appendix E shows that our results for racial isolation and property crime are robust to
joint inclusion in one regression.
27
5 Conclusions
Table 12: Heterogeneous results by property crime rate
(1) (2) (3)
Number of offers Mean offer Best offer
Black 0.980 (0.168) -4.748∗
(2.436) -4.198+
(2.648)
Tattoo 0.970 (0.171) -3.887 (3.274) -3.547 (3.855)
High prop. crime rate 0.983 (0.140) 3.376∗
(1.907) 4.060∗
(2.324)
High crime × Black 0.803 (0.165) -5.820∗∗
(2.742) -7.984∗∗
(3.363)
High crime × Tattoo 0.913 (0.170) -5.315+
(3.274) -5.414+
(3.726)
Regions × Black/Tattoo
20+ × Black/Tattoo
Standard ctls.
Observations 1007 604 604
All black=0 p. 0.0799 0.0242 0.0230
Incidence rate ratios from negative binomial estimation are reported in (1); OLS coefficients are reported in (2)–(3). Standard
errors clustered by market are reported in parentheses.
+ p < 0.15, ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
Note: Observations are weighted by state population/number of ads posted in each state. Excludes prematurely removed advertisements,
and (2)–(3) limited to advertisements receiving at least one offer. “High crime” markets are top 25% as measured by
2008 property crimes per capita (from United States Department of Justice and Federal Bureau of Investigations, 2009). “Standard
controls” are: high advertisement quality, asking price ($130 and $110 dummies; $90 excluded), holidays (Christmas and
Valentine’s day dummies), night, 20+ iPod Nano ads in market over previous week, median household income, poverty rate,
non-Hispanic white fraction of local population, and region (Northeast, Midwest, and South dummies; West excluded).
5 Conclusions
In this paper, we present strong evidence that black sellers suffer worse market outcomes than
their white counterparts in the environment we consider. In particular, their advertisements
receive 13% fewer responses, and 17% fewer offers. These effects are strongest in the Northeast,
and are similar in magnitude to those associated with a seller’s display of a wrist tattoo. Conditional
on receiving at least one offer, a black seller’s average offer is approximately $1.87 lower
than a white seller’s, with an even greater difference in the highest offers: the best offer received
by a black seller is typically $3.56 lower. These represent gaps of 2.2% and 3.8%, respectively,
below white sellers’ offers.
Respondents to advertisements with black photographs also exhibit lower trust. Compared
with correspondents with white sellers, they are 17% less likely to include their name in their
initial e-mail to the seller. Furthermore, the high bidders on black sellers’ advertisements—
presumably among the least biased of potential buyers—are 44% less likely to accept delivery by
mail and are 56% more likely to express concern about making a long-distance payment.
We test (albeit with limited ability) various explanations for this observed discrimination. The
28
5 Conclusions
disadvantage faced by black sellers is greatly reduced in more competitive markets; this provides
evidence in favor of Becker’s hypothesis that discrimination can be competed away. Our results
do not vary significantly by advertisement quality, but discrimination is greater in markets in
which black and white residents are geographically isolated from one another, and in markets
with high property crime rates. This is consistent with statistical discrimination used to avoid
a fraudulent, inconvenient, or dangerous sale (although it is also possible that animus against
black sellers is higher in high-crime or high-isolation markets).
29
References
References
Ahmed, A. M., and M. Hammarstedt (2008): “Discrimination in the rental housing market: a field
experiment on the internet,” Journal of Urban Economics, 64(2), 362–372.
Altonji, J. G., and R. M. Blank (1999): “Race and gender in the labor market,” in Handbook of Labor
Economics, ed. by O. C. Ashenfelter, and D. C. Card, vol. 3 of Handbook of Labor Economics,
chap. 48, pp. 3143–3259. Elsevier.
American Community Survey (2007): “2007 1-year estimates,” Online database.
Ayres, I., and P. Siegelman (1995): “Race and gender discrimination in bargaining for a new car,”
The American Economic Review, 85(3), 304–321.
Becker, G. S. (1971): The economics of discrimination. University of Chicago Press.
Bertrand, M., and S. Mullainathan (2004): “Are Emily and Greg more employable than Lakisha
and Jamal? A field experiment on labor market discrimination,” American Economic Review,
pp. 991–1013.
Bosch, M., M. A. Carnero, and L. Farré (2010): “Information and discrimination in the rental
housing market: evidence from a field experiment,” Regional Science and Urban Economics,
40(1), 11–19.
Carlsson, M., and D.-O. Rooth (2007): “Evidence of ethnic discrimination in the Swedish labor
market using experimental data,” Labour Economics, 14(4), 716–729.
Carpusor, A. G., and W. E. Loges (2006): “Rental discrimination and ethnicity in names,” Journal
of Applied Social Psychology, 36(4), 934–952.
Ewens, M., B. Tomlin, and C. Wang (2009): “Statistical discrimination in the U.S. apartment
rental market: a large sample field study,” http://dss.ucsd.edu/~btomlin/pdfs/
ewenstomlinwang.pdf.
Glaeser, E. L., and J. L. Vigdor (2001): “Racial segregation in the 2000 census: promising news,”
Center on Urban and Metropolitan Policy, The Brookings Institution.
Hanson, A., and Z. Hawley (2010): “Do landlords discriminate in the rental housing market?
Evidence from an internet field experiment in U.S. cities,” Unpublished Working Paper.
Kass, L., and C. Manger (2010): “Ethnic discrimination in Germany’s labour market: a field
experiment,” IZA Discussion Paper 4741.
List, J. A. (2004): “The nature and extent of discrimination in the marketplace: evidence from the
field,” The Quarterly Journal of Economics, 119(1), 49–89.
Nardinelli, C., and C. Simon (1990): “Customer racial discrimination in the market for memorabilia:
the case of baseball,” The Quarterly Journal of Economics, 105(3), 575–595.
30
References
Neumark, D., R. J. Bank, and K. D. Van Nort (1996): “Sex discrimination in restaurant hiring: an
audit study,” The Quarterly Journal of Economics, 111(3), 915–941.
Nunley, J. M., M. F. Owens, and R. S. Howard (2010): “The effects of competition and information
on racial discrimination: evidence from a field experiment,” Middle Tennessee State University
Department of Economics and Finance Working Paper 201007.
Pope, D. G., and J. R. Sydnor (Forthcoming): “What’s in a picture? Evidence of discrimination
from Prosper.com,” Journal of Human Resources.
Ravina, E. (2008): “Love & loans: the effect of beauty and personal characteristics in credit
markets,” SSRN Working Paper 1101647.
Riach, P. A., and J. Rich (2002): “Field experiments of discrimination in the market place,” The
Economic Journal, 112(483), F480–F518.
Tversky, A., and D. Kahneman (1974): “Judgment under uncertainty: heuristics and biases,”
Science, 185(4157), 1124–1131.
United States Department of Justice and Federal Bureau of Investigations (2009): “Crime in the
United States, 2008,” Online database.
31
A Market, advertisement and timing characteristics
A Market, advertisement and timing characteristics
The classified network through which we posted included sites for 326 geographic markets as of
March 2009. When posting an advertisement, a seller is asked to write in her specific location. For
those markets where the site provides a list of submarkets (e.g., the New York metropolitan area),
we posted separate advertisements in each submarket and entered the name of that submarket
in the location field. For most remaining markets (e.g., Boise), we merely entered the name of the
market. A few markets (e.g., South Dakota) cover such broad areas that it would be unrealistic
for a single seller to conduct a sale everywhere within the market; in these cases, we entered the
name of the highest-population city in the market.
Table 13 shows average values for a number of market characteristics broken down by advertisement
types. An F-test of the joint hypothesis that all fourteen means are equal across the
three advertisement types has a p-value in excess of 99% (which is unsurprising given that we
aimed to post every type in each market).
Table 13: Market characteristic averages by advertisement type
White Black Tattoo Total
Mkt. ads week 14.86 15.32 17.25 15.68
20+ weekly ads 0.171 0.174 0.210 0.183
Northeast 0.125 0.134 0.119 0.127
Midwest 0.241 0.243 0.243 0.242
South 0.369 0.356 0.362 0.362
West 0.264 0.266 0.277 0.268
% pop. White 77.26 76.78 77.09 77.04
% pop. Black 12.93 12.88 12.52 12.80
% pop. Hispanic 13.02 13.74 13.91 13.53
% pop. Asian 3.160 3.301 3.305 3.250
Poverty rate 15.62 15.92 15.55 15.71
Med. HH inc. ($K) 46.30 45.85 46.79 46.27
Black isolation index 0.211 0.218 0.215 0.215
Property crime rate 359.6 359.6 352.1 357.6
Observations 1200
Note: All observations equally weighted. Local racial composition, poverty rate, and household income are
from the 2007 American Community Survey. Black isolation index is degree to which “the average black
resident lives in a census tract in which the black share of the population exceeds the overall metropolitan
average” in 2000 (from Glaeser and Vigdor, 2001). 2008 property crimes are per ten thousand people (from
United States Department of Justice and Federal Bureau of Investigations, 2009). Isolation index and crime
data are not available for all markets.
We described in Sections 2.3 and 2.4 several key dimensions along which our advertisement
32
A Market, advertisement and timing characteristics
contents and post timing vary. Tables 14–15 show our stratification of advertisement types across
these attributes.
The tables in this appendix include data on prematurely removed advertisements.
Table 14: Advertisement characteristics
Ad type
White Black Tattoo Total
No. % No. % No. % No. %
Ad. quality
Low 215 49.0 217 50.2 163 49.5 595 49.6
High 224 51.0 215 49.8 166 50.5 605 50.4
Total 439 100.0 432 100.0 329 100.0 1200 100.0
Asking price
90 164 37.4 161 37.3 115 35.0 440 36.7
110 169 38.5 153 35.4 148 45.0 470 39.2
130 106 24.1 118 27.3 66 20.1 290 24.2
Total 439 100.0 432 100.0 329 100.0 1200 100.0
33
B Advertisement text
Table 15: Advertisement timing
Ad type
White Black Tattoo Total
No. % No. % No. % No. %
Day of week
Weekday 335 76.3 325 75.2 249 75.7 909 75.8
Saturday 45 10.3 53 12.3 39 11.9 137 11.4
Sunday 59 13.4 54 12.5 41 12.5 154 12.8
Total 439 100.0 432 100.0 329 100.0 1200 100.0
Post time
Day 249 56.7 247 57.2 180 54.7 676 56.3
Night 190 43.3 185 42.8 149 45.3 524 43.7
Total 439 100.0 432 100.0 329 100.0 1200 100.0
Holiday
Christmas 16 3.6 16 3.7 18 5.5 50 4.2
Valentine’s Day 28 6.4 24 5.6 29 8.8 81 6.8
Other 395 90.0 392 90.7 282 85.7 1069 89.1
Total 439 100.0 432 100.0 329 100.0 1200 100.0
B Advertisement text
The text used in our advertisements is shown in Tables 16–18.24
Our advertisements were posted
in the “electronics for sale” category (as were the vast majority of other advertisers offering iPods
for sale).
24
Prior to Apple’s introduction of the fifth generation iPod on September 9, 2009, advertisements offered fourth
generation iPods (substituting “4th” for “5th,” throughout).
34
B Advertisement text
Table 16: Advertisement text: “A” text
High-quality advertisement Low-quality advertisement
Ad title iPod Nano 8GB 5th Gen silver - *New in
Box*
iPod Nano 8GB 5th Gen silver - *New in
Box*
Ad text I recently received an iPod Nano as a
gift, but I already have one. It is a
BRAND NEW silver 8GB iPod Nano 5th
generation (with video). Still in the box never
opened. Retails for $149 but I’ll
sell it for $[price] or the best offer I
receive.
* Holds up to 2000 songs
* Holds up to 8 hours video
* View photos/video in portrait or
landscape
* Up to 24-hr. battery life
I can meet wherever is convenient for
you.
i recently recieved a ipod nano as a gift,
and i already have one. its BRAND NEW
silver 8GB iPod Nano 5th generation
(with video)!! still in the box never
opened. retail for $149 but ill sell for
$[price] or best offer
* Holds up to 2000 songs
* Holds up to 8 hours video
* View photos/video in portrait or
landscape
* Up to 24-hr. battery life
can meet wherever works for you
35
B Advertisement text
Table 17: Advertisement text: “B” text
High-quality advertisement Low-quality advertisement
Ad title BRAND NEW silver 8gb IPOD NANO!!!!
Never opened!
BRAND NEW silver 8gb IPOD NANO!!!!
Never opened!
Ad text NEW IN BOX - Never been opened, still
sealed. Exactly what you’d get if you
bought from Apple.
5th generation, Silver, with Video - 8GB
(holds up to 2000 songs or 8 hours
video) - Battery life up to 24 hours
I won it, but I don’t need it; I need the
cash instead.
Help me out and make me an offer. The
retail price $149, but I will sell for
$[price] or the best offer I receive.
We can meet wherever is convenient for
you.
NEW IN BOX!!! never even opened.
exatcly what youd get from apple.
5th generation, silver, with video - 8gb holds
up to 2000 songs/8 hours video battery
lasts up to 24 hours
i won it but dont need it. need the cash
instead!
help me out!!! make me an offer. retail
price $149–will sell for $[price] or the
best offer i get.
meet wherever works for you
36
B Advertisement text
Table 18: Advertisement text: “C” text
High-quality advertisement Low-quality advertisement
Ad title NEW 5th gen silver 8gb IPOD NANO –
NEVER OPENED!
NEW 5th gen silver 8gb IPOD NANO –
NEVER OPENED!
Ad text I have a brand new iPod Nano that I
don’t need (I already have one). Silver,
8GB, 5th generation (with video), never
opened!
Holds up to 8 hours of video or 2000
songs. 24-hour battery life. This is the
latest version!
It sells for $149 plus tax in the store, but
I’ll let it go for $[price] or the best offer I
receive. So, make me an offer! I’ll meet
you wherever is convenient.
We can meet wherever is convenient for
you.
ive got a brand new ipod nano that i
dont need (already have one). silver,
8GB, 5th generation (with video), never
opened at all!!
holds up to 8 hours of video or 2000
songs. 24-hour batttery life. this is the
newestt version!!
it sells for $149 plus tax in teh store, but
i’ll let it go for $[price] or the best offer i
get. so make me an offer!!! i’ll meet you
wherevers good for you.
37
C E-mail correspondence with potential buyers
C E-mail correspondence with potential buyers
Potential buyers initially contacted us via advertisement-specific anonymized e-mail addresses
assigned by the websites. Our responses and all subsequent correspondence were conducted
using the address rmitchell203@gmail.com. “Mitchell” is an extremely common last name
among both black and white Americans, we do not anticipate the number “203” sending any
particular signal to correspondents, and Gmail is a popular, freely available e-mail service.
All of our messages were sent in plain text (i.e., without any HTML coding) and maintained
the subject of the potential buyer’s initial e-mail response, prepending “Re:.” Our software was
configured so that potential buyers’ e-mail programs would display the e-mail address either
alone or together with the name “R. Mitchell.”
Table 19: Correspondence text
High-quality advertisement Low-quality advertisement
e-mail 1
(offer):
“A” text
Thank you for your interest in my iPod
Nano. I’ve received a lot of responses,
and would like to sell this quickly to the
person who makes me the best offer.
CASH ONLY, no trades. Is $[offer] your
best offer? Thanks.
[link to ad]
[text of ad]
thank you for your interest in my ipod
nano. i got a lot of responses, and
would like to sell this quickly to the
person who makes me the best offer.
CASH ONLY, no trades. is $[offer] your
best offer? thanks.
[link to ad]
[text of ad]
e-mail 1
(no
offer):
“A” text
Thank you for your interest in my iPod
Nano. I’ve received a lot of responses,
and would like to sell this quickly to the
person who makes me the best offer.
CASH ONLY, no trades. Please send me
your best offer, and I’ll let you know if
you get it. Thanks.
[link to ad]
[text of ad]
thank you for your interest in my ipod
nano. i got a lot of responses, and
would like to sell this quickly to the
person who makes me the best offer.
CASH ONLY, no trades. please send me
your best offer, and i’ll let you know if
you get it. thanks.
[link to ad]
[text of ad]
Continued
38
C E-mail correspondence with potential buyers
High-quality advertisement Low-quality advertisement
e-mail 1
(offer):
“B”/“C”
text
Hi –
Thanks for your email. I had a lot of
responses and I’m sure you understand
that I want to get the most that I can for
this iPod. Am I correct that $[offer] is
your best offer? Cash only/no trades,
please.
Thanks.
[link to ad]
[text of ad]
hi
thanks for your email. i got lots of
responses and im sure you understand i
wanna get the best price i can for the
ipod. am i right $[offer] is ur best offer?
cash only no trades please!
thanks
[link to ad]
[text of ad]
e-mail 1
(no
offer):
“B”/“C”
text
Hi –
Thanks for your email. I had a lot of
responses and I’m sure you understand
that I want to get the most that I can for
this iPod. Send me your best offer, and
I’ll email you if you get it. Cash only/no
trades, please.
Thanks.
[link to ad]
[text of ad]
hi
thanks for your email. i got lots of
responses and im sure you understand i
wanna get the best price i can for the
ipod. send me ur best offer and ill email
you if you get it. cash only no trades
please!
thanks
[link to ad]
[text of ad]
e-mail 2
(high
offer)
Hello. Congratulations, you made the
best offer so the new iPod Nano is yours
for $[offer]. Unfortunately, I am out of
town at the moment so can’t meet in
person for a while. I’m happy to send it
to you if you can pay me using Paypal.
Sorry about that. Let me know if you
want to do this.
hello. congratulations, you made the
best offer so the new ipod nano is yours
for $[offer]. unfortunatley, im out of
town at the moment so can’t meet in
person for a while. i can sent it to you if
you can pay me using paypal. sorry
about that. let me know if you want to
do this
Continued
39
D Sample responses from buyers
High-quality advertisement Low-quality advertisement
e-mail 2
(low
offer)
The iPod Nano is no longer available.
Thank you for your time, and good luck
finding one elsewhere.
the ipod nano is gone. thanks for your
time and good luck finding another one
D Sample responses from buyers
D.1 Response types
The e-mails below are representative of each type of response received from potential buyers.
Some are initial responses to our advertisement, and some followed our first e-mail (“e-mail 1”
above).
Scams
thanks for getting back to me ,i am a medical doctor soo i av to perform an operation outside the
country in west africa and i dont have the time to pick the item up so i want you to help me ship
the the item to this country i will be responsible for the shipping and handling fee so get back to
e on how you wish to get your money....
regards Dr [name]
Nice hearing back from you,i want you to know that i am so serious in buying your item, and i
will not be able to come for the pickup.i intended to come to your door to come and give you
the money and use the opportunity to check the condition of the item,but due to the fact that
am very busy at work nowaday"s,i wouldn"t have the opportunity to do so.I will be glad if you
can be so nice to help me ship it down to my daughter whom schools oversea,and i will cover
the shipping cost for the item to be sent to her...So i am ready to pay you $800 for the item
cost including the shipping cost , and i will prefer paying you via paypal,because very fast and
secure,so i want you to get back with your PayPal email address and your full name.so that the
instant payment will be made..
Or if you prefer a bank transfer or western union.
Best regards,
[name]..
40
D Sample responses from buyers
You don’t have to sell this.
I used to sell stuff on [website] just like you were doing to pay my bills. I was really getting ripped
off by selling stuff so cheap. My buddy told me about this site and I made about $4k from it last
month. Its great and the has helped me and friends so much. I’m doing better than my current
fulltime job and this is just extra money! To give it a try go to [URL]
If you need help getting started please let me know.
Non-offers (non-trade)
You still have the ipod?
[link to ad]
I am interested in buying your Ipod, therefore what is the best cash price you will take for this. I
am local in [city] and could meet you tomorrow.
Thank you,
[name]
Non-offers (trade)
I was just wondering if you would be interested in trading for a full blooded boxer pup if so give
me or my wife a call [phone number] [name] or [phone number] [name] thanks
[link to ad]
may I know if you will consider a silver edition game cube with wireless controller and 3 games
as trade for your Ipod nano?
I will not be available until 3:00 pm.
would you be interested in a trade for a 6 ft red tail boa with custom enclosure and a laptop that
works great just needs a hard drive it comes with a wifi card and a bb gun co2 semi auto hand
gun let me know
Offers
Hi, $75 is all i can afford right now. This will be given as a gift. I don’t even get to keep it for
myself.
Cash in hand and ready to pick up tomorrow if you accept my offer. Br,
[name]
take 60
Hey, just wondering if its still available. Had one of my friends get his nano jacked so if $100
would work that would be very appreciated. Thanks
41
D Sample responses from buyers
D.2 Response characteristics
The e-mails below illustrate the three characteristics (inclusion of the buyer’s name, politeness,
and inclusion of a a personal story to invoke sympathy) discussed in section 3.3.
Name
If you still have the nano iPod for sale i would like to get it from you.
Thanks
[Name]
[link to ad]
Hi is this still available and if so would u take $70 dollars..[Name] at [phone number].
If you will take $75 cash let me know I am off today and my name is [Name]. Let me know Thanks,
my cell is [phone number].
No Name
$75 for your iPod nano?
Would you go down to a hundred.. I’ll pick it up today. I do work at 1 tho so asap would be best.
IS THIS STILL AVAILABLE FOR SALE?GET BACK TO ME ASAP.
Polite
Hello,
Would you take $50 for the ipod?
Thanks
HI , MY NAME IS [NAME] , I AM INTERESTED IN YOUR IPOD , CAN YOU GIVE ME YOUR
PHONE NUMBER PLEASE?OR IN CASE THIS IS MY PHONE NUMBER [PHONE NUMBER] ,
THANKS.
Hi!! Do you still have your ipod? Would you be able to sell it for 100? Thanks for your time!!!
-[Name]
Not Polite
i can offer you $70 for your ipod ...let me know
Are you willing to trade for anything like gps for the car or a gas car. if so text/call [Name] at
[phone number]
Im coming down from [city] tomorrow would you be able to meet in [city] sometime tomorrow
night?
42
D Sample responses from buyers
Personal
Hi,
I’m interested in the Nano and can meet you in [neighborhood] today. I’d really like to spend $75.
I know it’s not the $110 you’re asking for, but I’m a grad student and have to stay on budget! I
need it because I promised my little cousin I’d get her Nano for her birthday but I’m short on the
cash. In this case, the Nano will be good as new!
Please consider the offer–it’s $75 bucks in your pocket that you weren’t expecting before you
got the gift, and it would make my 14 year old cousin a very happy birthday girl! Call me at the
number below and let’s make it happen!
Thanks so much,
[Name] [phone number]
[link to ad]
DO YOU STILL HAVE THE IPOD?
IF SO I AM WILLING TO OFFER $75 FOR IT. I WOULD OFFER MORE BUT IT IS ALL THAT I HAVE
LEFT TO SPEND.JUST TRYING TO GET AN EXTRA VALETINES GIFT FOR MY DAUGHTER.I AM
IN [CITY] AND I CAN MEET YOU ANY TIME.
LET ME KNOW
MAHALO
[NAME]
Hi, saw your ad on [website]. I know that you’re asking 90.00 for the ipod, but I only have 50.00.
It’s fore my son’s birthday in a few weeks and that’s all I can afford...would you consider it? email
me back and let me know.
[link to ad]
Not Personal
hi 70 for ipod 8gb,call me please [name]„„„„„„„„„„„„„„„
[phone number]——
Hello. How about 70usd? thanks
hello,
i saw your post for the ipod and am very interested!
i am available all day tomorrow and can meet you where/whenever.. i live in [neighborhood].
please do not hesitate to call me.
thanks,
[name]
[phone number]
43
D Sample responses from buyers
D.3 Reactions to delivery proposal
The e-mails below illustrate the types of reactions received to our proposal to ship the iPod in
exchange for PayPal payment. These reactions are discussed in section 3.5.
Open to Shipment
Great. Any idea when you will be back or can you send ro me and I will pay then? Where does
payment need to be sent ?
Thanks
paypal would be fine, do you live near [city]? could pick it up. Your call.
I do have a paypal accout that I use when using ebay but Im not sure how to use this outside of
ebay. That would be great if this is possible. I will wait to hear back from you. Thanks’
[Name]
Prefer to Wait
I’ll wait till you are back in town and can get it that way.
Thanks for letting me know, thats great news! I will wait until you are back in town to meet up
and give4 you cash. Send me a message when you get back, and when it is good for you to meet.
All the best.
[Name]
let me know when u get to town and we‘ll hook up ok??
Other or Too Late
already bought another one, i needed it the same day i e-mailed you. sorry.
Sorry bought one already.
[Name]
No thanks
Payment Concern
Hsha [expletive] you spammer
Sorry, I DON’T DEAL WITH SCAMMERS!!!! Try someone else.
i don’t use paypal buddy. face to face ok.
44
E Additional results
E Additional results
E.1 Regressions with alternate controls
We also consider the results reported in Table 3 with alternate controls for advertisement, market,
and timing characteristics. Table 20 reports results for a specification without controls (other
than advertisement type). Table 21 replaces the market controls with market fixed effects (where
the effects on likelihood of premature removal and shipment are now estimated using OLS to
avoid an incidental parameter problem).
Table 20: Key outcome regressions (without controls)
(1) (2) (3) (4) (5) (6)
Prem. rem. Nonscams Offers Mean offer Best offer Shipped
Probit Neg. Bin. Neg. Bin. OLS OLS Probit
Black 0.0297∗
0.845∗
0.811∗
-2.122+
-3.802∗∗
-0.0210∗∗
(0.0165) (0.0728) (0.0885) (1.417) (1.585) (0.0107)
Tattoo 0.0157 0.846∗
0.859+
-2.488+
-3.643∗∗
-0.00597
(0.0159) (0.0737) (0.0843) (1.634) (1.712) (0.0111)
Observations 1200 1147 1147 663 663 1147
Marginal effects from probit estimation are reported in (1) and (6); (2)–(3) report incidence rate ratios from negative binomial estimation;
(4)–(5) report OLS coefficients. Standard errors clustered by market are reported in parentheses.
+ p < 0.15, ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
Note: Observations are weighted by state population/number of ads posted in each state. (2)–(6) exclude prematurely removed advertisements,
and (4)–(5) limited to advertisements receiving at least one offer.
E.2 Robustness to including prematurely removed ads
We would be concerned if our findings were driven by a large extent by the higher likelihood
of black sellers’ advertisements being prematurely removed discussed in Section 3.1. However,
our average effects results are affected only slightly by the exclusion of prematurely removed
advertisements; this is illustrated in Table 22.
E.3 Reactions to delivery proposal regressions
Table 23 reports probit estimates of the effect on the frequency of receiving each reaction
associated with the non-white seller types. Although these separate results are not generally
statistically significant, the fact that the “black” and “tattooed” coefficients are generally negative
in the left-hand columns and positive in the right-hand ones offers further evidence that black
45
E Additional results
Table 21: Key outcome regressions (with market fixed effects)
(1) (2) (3) (4) (5) (6)
Prem. rem. Nonscams Offers Mean offer Best offer Shipped
OLS Neg. Bin. Neg. Bin. OLS OLS OLS
Black 0.0198+
0.923 0.854∗∗
-0.500 -2.167 -0.0195+
(0.0135) (0.0518) (0.0580) (1.323) (1.558) (0.0119)
Tattoo 0.00995 0.824∗∗∗
0.796∗∗∗
-1.571 -3.750∗∗
-0.00972
(0.0146) (0.0498) (0.0574) (1.412) (1.664) (0.0129)
High quality -0.0178 0.969 0.973 0.162 0.994 -0.00851
(0.0127) (0.0501) (0.0600) (1.250) (1.472) (0.0112)
Price $110 0.00492 0.427∗∗∗
0.416∗∗∗
10.59∗∗∗
5.118∗∗∗
-0.0130
(0.0149) (0.0254) (0.0293) (1.409) (1.659) (0.0132)
Price $130 -0.0136 0.225∗∗∗
0.191∗∗∗
23.08∗∗∗
14.60∗∗∗
-0.0446∗∗∗
(0.0171) (0.0197) (0.0209) (1.863) (2.195) (0.0151)
Christmas 0.0539∗
1.791∗∗∗
1.878∗∗∗
5.430∗
9.836∗∗∗
0.0422+
(0.0319) (0.174) (0.214) (2.861) (3.371) (0.0287)
Valentine’s Day 0.00515 1.222∗∗
1.374∗∗∗
0.0399 1.035 0.0147
(0.0254) (0.108) (0.142) (2.165) (2.551) (0.0224)
Night -0.0246∗
0.689∗∗∗
0.664∗∗∗
0.347 -2.323+
-0.00480
(0.0132) (0.0385) (0.0443) (1.357) (1.598) (0.0117)
Observations 1195 1093 1024 578 578 1139
Market FE 319 297 274 192 192 315
OLS coefficients are reported, except columns (2)–(3) report incidence rate ratios from negative binomial estimation. Standard errors
clustered by market are reported in parentheses.
+ p < 0.15, ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
Note: Samples are as follows: (1) includes all advertisements in markets with more than one advertisement posted; (2) includes advertisements
not prematurely removed, in markets where more than one advertisement was not prematurely removed and some ad received at
least one nonscam; (3) includes advertisements not prematurely removed, in markets where more than one advertisement was not prematurely
removed and some ad received at least one offer; (4)–(5) include advertisements not prematurely removed and receiving at least one
offer, in markets where more than one advertisement was not prematurely removed; (6) includes advertisements not prematurely removed,
in markets where more than one advertisement was not prematurely removed.
46
E Additional results
Table 22: Key outcome regressions (including prematurely removed ads)
(1) (2) (3) (4) (5)
Nonscams Offers Mean offer Best offer Shipped
Neg. Bin. Neg. Bin. OLS OLS Probit
Black 0.869∗∗
0.822∗∗
-1.835+
-3.506∗∗
-0.0127+
(0.0532) (0.0643) (1.225) (1.374) (0.00776)
Tattoo 0.826∗∗∗
0.836∗∗
-2.541∗
-4.067∗∗∗
-0.00381
(0.0523) (0.0589) (1.315) (1.556) (0.00731)
High quality 0.988 1.022 0.397 0.858 -0.00560
(0.0535) (0.0614) (1.151) (1.227) (0.00672)
Price $110 0.422∗∗∗
0.388∗∗∗
11.96∗∗∗
7.012∗∗∗
-0.0119∗
(0.0266) (0.0281) (1.197) (1.292) (0.00613)
Price $130 0.224∗∗∗
0.183∗∗∗
20.67∗∗∗
13.35∗∗∗
-0.0289∗∗∗
(0.0180) (0.0190) (2.102) (2.123) (0.00717)
Christmas 1.873∗∗∗
2.027∗∗∗
5.534∗∗∗
11.69∗∗∗
0.0224
(0.181) (0.239) (1.664) (2.028) (0.0246)
Valentine’s Day 1.312∗∗∗
1.361∗∗∗
1.983 1.675 0.00448
(0.110) (0.120) (2.437) (2.400) (0.0136)
Night 0.698∗∗∗
0.663∗∗∗
-1.557 -3.418∗∗∗
0.00152
(0.0432) (0.0426) (1.149) (1.285) (0.00642)
20+ weekly ads 2.002∗∗∗
2.103∗∗∗
0.412 5.414∗∗∗
-0.00754
(0.179) (0.203) (1.295) (1.888) (0.00837)
Med. HH inc. (log) 2.665∗∗∗
3.249∗∗∗
12.65∗∗
11.05+
0.0698∗∗
(0.833) (1.121) (5.388) (6.754) (0.0353)
Poverty rate 1.026∗∗
1.038∗∗∗
0.256 0.232 0.00238∗
(0.0122) (0.0137) (0.221) (0.242) (0.00133)
% pop. White 0.995∗
0.997 -0.0109 -0.0619 -0.000154
(0.00244) (0.00261) (0.0437) (0.0464) (0.000282)
Northeast 0.952 1.061 -3.687∗
-2.234 -0.00669
(0.0926) (0.109) (1.967) (2.285) (0.0119)
Midwest 0.981 0.977 -0.836 0.341 0.00295
(0.0829) (0.0929) (1.516) (1.752) (0.0106)
South 0.887 0.937 -1.698 -2.098 0.00527
(0.0828) (0.0996) (1.525) (1.797) (0.00988)
Observations 1200 1200 681 681 1200
Incidence rate ratios from negative binomial estimation are reported in (1)–(2); (3)–(4) report OLS coefficients; (6)
reports marginal effects from probit estimation. Standard errors clustered by market are reported in parentheses.
+ p < 0.15, ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
Note: Observations are weighted by state population/number of ads posted in each state. (3)–(4) limited to advertisements
receiving at least one offer.
47
E Additional results
and tattooed sellers are more likely to receive “bad” and less likely to receive “good” reactions.
Dividing the estimated “black” coefficients by the white means reported in Table 2, for example,
suggests that a buyer is 44% less likely to accept delivery by mail and 56% more likely to express
concern about making a long-distance payment when responding to a black rather than white
seller.
Table 23: Probit regression of reactions to delivery proposal
(1) (2) (3) (4) (5)
Ship Wait Other None Concern
Black -0.0248+
0.000723 -0.0494 0.0280 0.0424+
(0.0159) (0.0452) (0.0355) (0.0484) (0.0272)
Tattoo -0.000782 -0.0590 0.00678 0.0376 0.0165
(0.0182) (0.0461) (0.0391) (0.0513) (0.0279)
Standard ctls.
Observations 572 604 604 604 604
Probit marginal effects are reported. Standard errors clustered by advertisement are reported in parentheses.
+ p < 0.15, ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
Note: Observations are weighted by reciprocal of number of delivery proposals per advertisement. Excludes deliveries proposed to respondents
to prematurely removed advertisements. Sample size is slightly reduced for the “willing to accept shipment” outcome since no buyers
had this response to the 32 delivery proposals made during the Christmas period (and these observations were dropped). “Standard
controls” are: high advertisement quality, asking price ($130 and $110 dummies; $90 excluded), holidays (Christmas and Valentine’s day
dummies), night, 20+ iPod Nano ads in market over previous week, median household income, poverty rate, non-Hispanic white fraction of
local population, and region (Northeast, Midwest, and South dummies; West excluded).
E.4 Racial isolation and property crime joint regression
The degree of racial isolation is somewhat positively correlated with the property crime rate,
but the results presented in the main text are robust to inclusion of both measures in a single
regression.
48
E Additional results
Table 24: Heterogeneous results by racial isolation and property crime rate
(1) (2) (3)
Number of offers Mean offer Best offer
Black 0.819 (0.153) -6.191∗∗
(2.696) -6.241∗∗
(2.984)
Tattoo 0.851 (0.167) -3.003 (3.808) -3.488 (4.418)
High isolation 1.380∗∗
(0.192) 0.147 (2.388) 1.828 (2.919)
High iso. × Black 0.540∗∗∗
(0.108) -3.408 (3.692) -8.739∗
(4.526)
High iso. × Tattoo 0.765+
(0.134) 1.908 (4.199) -2.319 (4.649)
High prop. crime rate 0.974 (0.130) 3.102+
(1.957) 3.636+
(2.441)
High crime × Black 0.860 (0.166) -4.827∗
(2.805) -7.184∗∗
(3.385)
High crime × Tattoo 0.910 (0.169) -5.324+
(3.438) -5.375 (3.961)
Regions × Black/Tattoo
20+ × Black/Tattoo
Standard ctls.
Observations 858 546 546
All black=0 p. 0.0137 0.0449 0.0176
Incidence rate ratios from negative binomial estimation are reported in (1); OLS coefficients are reported in (2)–(3). Standard
errors clustered by market are reported in parentheses.
+ p < 0.15, ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
Note: Observations are weighted by state population/number of ads posted in each state. Excludes prematurely removed advertisements,
and (2)–(3) limited to advertisements receiving at least one offer. “High crime” markets are top 25% as measured by
2008 property crimes per capita (from United States Department of Justice and Federal Bureau of Investigations, 2009). “Standard
controls” are: high advertisement quality, asking price ($130 and $110 dummies; $90 excluded), holidays (Christmas and
Valentine’s day dummies), night, 20+ iPod Nano ads in market over previous week, median household income, poverty rate,
non-Hispanic white fraction of local population, and region (Northeast, Midwest, and South dummies; West excluded).
49