Adverse Selection in Online Auctions: A Study of eBay Motors

City University of New York (CUNY) City University of New York (CUNY)

CUNY Academic Works CUNY Academic Works

Theses and Dissertations Hunter College

Spring 5-3-2024

Adverse Selection in Online Auctions: A Study of eBay Motors Adverse Selection in Online Auctions: A Study of eBay Motors

Ilanith Nizard

CUNY Hunter College

How does access to this work bene9t you? Let us know!

More information about this work at: https://academicworks.cuny.edu/hc_sas_etds/1177

Discover additional works at: https://academicworks.cuny.edu

This work is made publicly available by the City University of New York (CUNY).

Contact: AcademicWorks@cuny.edu

Adverse Selection in Online Auctions: A Study of eBay Motors

Ilanith Nizard

Submitted in partial fulfillment

of the requirements for the degree of

Master of Arts in Economics, Hunter College

The City University of New York

2024

May 3, 2024 Ingmar Nyman

Date Thesis Sponsor

May 3, 2024 Partha Deb

Date Second Reader

ii Adverse Selection in Online Auctions

Acknowledgments

This paper would’ve stayed only but a scramble of ideas if it wasn’t for the constant

guidance, support, and detailed attention given by so many.

Maman et Papa, merci pour tout ce que vous faites pour nous, sacrifiant tant

pour nous offrir toutes ces opportunités. Et merci de me demander tout les jours ’Tu

fais quoi comme études déjà ?’, ’T’écris une thèse sur la joie de vivre ?’ La réponse

est non, mais merci d’être fiers de moi , même si vous ne savez pas exactement quel

est mon domaine d’études (juste pour mémoire c’est de l’éco et pas du droit).

Shan, mon autre (et unique) neurone, merci pour toutes les performances

privée. J’apprécie toutes tes relectures, même si la plupart de cette thèse n’a au-

cun sens pour toi. You’re Irreplaceable.

Nava, merci de m’avoir fait membre honoraire de ton comité de candidature

universitaire, et Sam merci de t’assurer que je n’ai jamais un moment de paix ou de

calme pour travailler.

Tracy, ma pseudo grande sœur, merci de me garder saine d’esprit tout au long

de ce processus, never in my Wildest Dreams thought we (…I…) would finish but

personne d’autre avec qui j’aurais préférer le faire.

I am immensely grateful to Professor Ingmar Nyman for his steadfast support

and guidance throughout this thesis journey (and even before that). Your belief in

my potential and dedication to refining my initial ideas into tangible research have

been invaluable. Your mentorship has been instrumental in shaping the trajectory of

my academic journey, and I am forever grateful for all the advice you have provided.

To Professor Partha Deb, your expertise and keen insights have lent depth and

precision to my research. Your consistent provision of feedback, meticulous guid-

ance, and commitment to academic excellence has not only honed my arguments

but also instilled in me a drive for continuous improvement. Your dedication to

my success, extending beyond this thesis, has been evident. Thank you for always

pushing me to do my best.

I am deeply appreciative for Professor Kenneth McLaughlin and his helpful

suggestions through each stage; they have been instrumental in my growth as a

writer and as an economist, greatly shaping both my skills and the development of

this paper.

Lastly, I would like to thank my professors in the economics department at

Hunter College for instilling in me a passion for learning and providing the founda-

tional knowledge that underpins this work.

Abstract

This study provides evidence of adverse selection by examining the disparity in

the value of a photo between dealer sellers and non-dealer sellers in online used-

car auctions. It examines the impact of photos on three outcomes in auctions of

automobiles: bid prices, the number of bidders, and the probability of a trade. To

do so, it uses non-linear regression models and data from 80,000 completed used-

car auctions on eBay Motors from March to October 2006. The results show that

an additional photo will increase all three of the outcome variables and more so for

non-dealers than for dealers. The marginal effects of an additional photo on price,

number of bidders and the probability of sale are all significantly and substantially

greater for non-dealers than dealers.

iii

Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

Adverse Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

eBay Auctions Market . . . . . . . . . . . . . . . . . . . . . . . . 4

Empirical Studies on Adverse Selection . . . . . . . . . . . . . . . 5

3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

4 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

Auction Price . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

Market Thickness . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

Trade Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

Robustness: Heterogeneity . . . . . . . . . . . . . . . . . . . . . . 13

Robustness: Specification Checks . . . . . . . . . . . . . . . . . . 14

5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

Auction Price . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

Market Thickness . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

Trade Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

Robustness: Heterogeneity . . . . . . . . . . . . . . . . . . . . . . 20

Robustness: Specification Checks . . . . . . . . . . . . . . . . . . 22

6 Summary and Conclusion . . . . . . . . . . . . . . . . . . . . . . . 22

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

Figures

1 Marginal Effects on Price by Seller Type . . . . . . . . . . . . . . . 18

2 Marginal Effects on Bidders by Seller Type . . . . . . . . . . . . . 19

3 Marginal Effects on Trade Probability by Seller Type . . . . . . . . 20

Tables

1 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2 Regressions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3 Marginal Effects of Number of Photos on Auction Outcomes . . . . 18

4 Heterogeneous Effects . . . . . . . . . . . . . . . . . . . . . . . . 21

5 Specification Checks . . . . . . . . . . . . . . . . . . . . . . . . . 23

1 Introduction 1

1. Introduction

eBay Motors is the largest marketplace for used cars in the United States, offering

a wide range of brands and vintages. Due to pre-existing information asymmetries,

this is an environment prone to adverse selection. Sellers possess extensive infor-

mation about the cars they list, having firsthand knowledge about their vehicle’s

condition, maintenance history, and potential issues. Conversely, buyers often can-

not physically inspect the car before purchase. They therefore rely on information

provided by the sellers, obtaining mandatory details such as make, model, and year,

but often lacking access to a complete maintenance history or insights into subtle

issues.

This information gap may put buyers at a distinct disadvantage, compelling

them to make decisions based on incomplete or biased data. The information asym-

metry disrupts the trade of high-quality cars. Their limited knowledge prevents

buyers’ from recognizing and paying for quality. As a consequence, sellers of high-

quality cars exit the market since they cannot be compensated for the quality that

they offer.

Sellers on eBay Motors provide information to counter adverse selection. Sell-

ers share details through photos and text descriptions, which narrows the informa-

tion gap between themselves and potential buyers. Reputable sellers are particularly

effective in mitigating adverse selection because their established track record for

providing accurate information and maintaining brand reputation serves as a mar-

ket signal that collectively reduce the information asymmetry between buyers and

sellers, thereby mitigating the risks associated with adverse selection.

In his work, Lewis (2011) investigates the complexities of adverse selection

within online auctions, acknowledging the exacerbation of information asymme-

tries in such environments. He suggests that sellers’ ability to partially convey their

car’s quality through photographs may contribute to the sustained volume of trade

and transactions, despite adverse selection concerns. While Lewis hints that photos

may have a stronger impact on price for non-dealers, possibly due to the absence of

reputation as an alternative quality signal, my analysis seeks to empirically validate

the efficacy of seller-provided information in mitigating adverse selection. By ex-

tending Lewis’s analysis to examine auction outcomes, namely price, the number of

bidders, and the likelihood of a trade, I aim to complement his findings and provide

a deeper understanding of the role of information in online auctions.

This empirical analysis investigates whether there is evidence of adverse se-

lection in the eBay motors auctions’ market, with a distinction between dealer-

sellers and non-dealer sellers lacking an established reputation. The primary ob-

jective is to assess how the disparity in seller types influences the impact of infor-

mation provision. Ultimately, the observed differences in outcomes between dealers

and private sellers serve as evidence of adverse selection.

2 Adverse Selection in Online Auctions

To address this question, I use Lewis’s (2011) data on completed eBay Mo-

tors auctions for selected car models between March and October 2006. Non-linear

regressions control for the trade environment, including auction model, the week

when it was held, and the vintage of the car. I study the effect of a measure of in-

formation, photos, on three auction outcomes: price, the number of bidders, and the

likelihood of a trade. By analyzing the differential effects of information provision

across seller types, I aim to elucidate whether there is evidence of adverse selection

in this market.

The results shed light on the presence of adverse selection in online used-car

auctions by using the disparity in the value of providing photos between dealer and

non-dealer sellers. The significant difference in how much photos influence auc-

tion outcomes like pricing, number of bidders, and likelihood of trade across the

two seller groups captures the information asymmetry driven by adverse selection.

For private sellers, lacking established reputations, photos have a substantial posi-

tive impact, indicating buyers face more uncertainty about the true condition of the

product from these sellers. In contrast, photos have a relatively muted effect for

dealers with more reputation, as buyers likely already have reasonable quality ex-

pectations. This marked divergence in the value of additional information between

seller types provides evidence supporting the existence of adverse selection.

2. Literature Review

This literature review examines the concept of adverse selection in economic trans-

actions, focusing on its theoretical foundations and potential remedies. Akerlof’s

(1970) seminal work lays the groundwork for understanding information asymme-

try and adverse selection, with a specific illustration in the used car-market. Subse-

quent studies by Spence (1973), Oi (1971), Rothschild and Stiglitz (1976), Wilson

(1977), and Heal (1976) shed light on strategies like signaling, screening, and rep-

utation as potential remedies. The section then incorporates a focused examination

of the background and mechanisms of eBay auctions. Lastly, this literature review

looks into empirical studies by Houser and Wooders (2006), Peterson and Schnei-

der (2014), and Lewis (2011), that provide real-world insights into the impact of

reputation and adverse selection in the market for used cars.

Adverse Selection

Akerlof (1970) lays the theoretical groundwork for understanding information asym-

metry and its impact on market efficiency. The central concept in Akerlof’s model

is adverse selection, a phenomenon arising from the unequal distribution of infor-

mation between buyers and sellers regarding the quality of goods.

Akerlof illustrates adverse selection in the context of the used-car market. In

this market, sellers possess more information about their vehicles’ condition than

potential buyers, creating an information gap. As a consequence, buyers, despite

2 Literature Review 3

their eagerness, are unable to compensate high-quality sellers adequately, leading

to a lack of trade for high-quality items. This leaves a preponderance of low-quality

lemons in circulation.

The mechanics of Akerlof’s model rest on the rational response of market par-

ticipants to uncertainty. Sellers of high-quality goods, discouraged by the inability

to command fair prices, withdraw from the market, leading to a degradation of over-

all market quality.

Despite the challenges posed by adverse selection, potential remedies exist

to enhance trade efficiency. Three key mechanisms—signaling, screening, and

reputation—stand out as effective tools in addressing the challenges posed by in-

formation asymmetry in economic transactions.

Spence (1973) examines the first mechanism, signaling, which involves the

intentional communication of information by the informed party. Faced with the

challenge of being unable to trade at a fair price, high-quality sellers have an incen-

tive to communicate their information to buyers. Since low-quality sellers have an

incentive to mimic high-quality sellers, the signal must not only incur a cost but the

cost must correlate with the missing information. This higher cost for low-quality

sellers to send the signal is crucial in preventing them from effectively mimicking

high-quality sellers. Spence (1973) refers to signaling as the deliberate acts of in-

dividuals taken to communicate their intrinsic capabilities to other, less informed

parties.

The second strategy, screening, involves extraction of information by the un-

informed parties. Oi (1971), Rothschild and Stiglitz (1976), and Wilson (1977) offer

key insights into this mechanism.

Oi (1971) explores the intricacies of a two-part tariff, a pricing strategy that

combines a variable fee per unit of service and a fixed lump sum payment. The

variable fee per unit introduces a form of price discrimination, allowing the firm

to differentiate prices based on the quantity of service demanded, thereby extract-

ing information about consumers’ preferences and their varying willingness to pay.

The two-part tariff functions as a screening mechanism, enabling the monopolist

to discern and extract surplus from consumers willing to pay more for the service.

By separating the pricing into a fixed and variable component, the two-part tariff

effectively sorts consumers based on their demand characteristics, allowing the mo-

nopolist to capture a larger portion of the consumer surplus.

Rothschild and Stiglitz’s (1976) paper addresses the challenges arising from

information asymmetry between insurers and policyholders, elucidating how indi-

vidual risk characteristics impact insurance markets. Rothschild and Stiglitz illus-

trate how insurers strategically extract information about policyholders’ risk profiles

by offering multiple contract options.

Similarly, Wilson (1977) employs a game-theoretic approach to capture the

4 Adverse Selection in Online Auctions

strategic interactions between insurers and policyholders. His model focuses on

the concept of separating equilibria, wherein different risk groups opt for distinct

insurance contracts.

The final mitigating factor of adverse selection is the concept of reputation,

which imposes a cost on sellers who disappoint their customers. Sellers who prior-

itize reputation alleviate concerns related to adverse selection, as buyers trust that

these sellers are less likely to offer sub-optimal goods or engage in opportunistic

behavior.

Geoffrey Heal (1976) highlights the business mythology perspective, suggest-

ing that established businesses with a reputation to protect may prioritize maintain-

ing standards, thereby mitigating the risk of bad products displacing good ones in

the market. This concept forms the basis of this study’s identification strategy. Sell-

ers try to diminish information asymmetries by providing additional information to

potential buyers. Reputable dealers may already possess mechanisms to address ad-

verse selection, while non-dealers may rely more heavily on information provision

to foster buyer confidence and facilitate trade.

By examining how the provision of additional information influences auction

outcomes, this research aims to empirically test whether such information alleviates

adverse selection problems. Specifically, it investigates whether this extra infor-

mation increases the likelihood of trade for non-established sellers, where adverse

selection is more likely to hamper trade.

eBay Auctions Market

Hasker and Sickles (2010) provide an in-depth analysis of eBay’s marketplace evo-

lution, focusing on seller strategies, trust-building mechanisms, and buyer behav-

iors. They discuss eBay’s transition from an auction platform to offering fixed-price

listings and bargaining options, highlighting the diverse range of goods available.

The study examines eBay auction mechanics, emphasizing the proxy bidding sys-

tem and the importance of detailed item descriptions. It investigates how asymmet-

ric information influences buyer behavior and explores the role of seller feedback

in building trust. Additionally, the study examines alternative trust-building mech-

anisms like professional accreditation and certification. Methodologies include re-

gression analysis, field experiments, and statistical modeling to reveal market in-

sights and buyer-seller interactions.

Lucking-Riley et al. (2007) explore price dispersion in online auctions, fo-

cusing on eBay. They analyze auction data to assess the impact of seller reputa-

tion, auction duration, and listing format on final prices. Their findings indicate

that higher seller reputations correlate with higher prices due to increased trust,

longer auction durations lead to higher final prices due to more competitive bid-

ding, and auction-style listings result in lower final prices compared to fixed-price

listings. Additionally, they note that while a higher starting price increases bid price,

2 Literature Review 5

it decreases auction success, highlighting a trade-off. Furthermore, bidder behavior,

such as aggressiveness and timing, significantly influences auction outcomes.

Ockenfels and Roth’s (2006) paper offers a detailed analysis of bidding be-

havior in online auctions, focusing on late bidding and incremental bidding, also

known as sniping. Using empirical data from eBay auctions, the study uncovers key

insights into bidder strategies. The authors find that late bidding, more pronounced

on eBay, particularly in categories like antiques, is influenced by auction rules, with

eBay’s hard close encouraging sniping behavior. Moreover, the study explores fac-

tors driving incremental bidding, revealing a negative correlation between bidder

experience, measured by feedback number, and bid frequency per bidder.

Li et al. (2009) investigate the impact of auction features on bidding behavior

in online auctions, highlighting their roles in signaling seller credibility and product

quality. Using a modified Park and Bradlow model from marketing research, they

analyze key bidding decisions: whether to bid, who bids, when to bid, and how much

to bid. Quality indicators, like multiple picture postings and money-back guaran-

tees, increase bidder participation and early bidding, but result in bid shading and

lower final bid prices. Additionally, credibility indicators, such as seller feedback

ratings and third-party payment options, lead to higher participation probabilities,

shorter inter-bidding times, and reduced final bid prices, as they attract experienced

bidders. The study emphasizes the need for sellers to balance credibility and quality

indicators to establish trust in online auctions. However, it notes limitations in its

focus on affiliated-value auctions and suggests exploring private-value auctions and

longitudinal bidding patterns for further insights.

Empirical Studies on Adverse Selection

Peterson and Schneider’s (2014) empirical strategy involves comparing repair rates

between traded and non-traded cars, with a focus on distinguishing adverse selection

effects from sorting effects. Sorting refers to the voluntary trading of goods based

on observable characteristics, where buyers and sellers self-select into transactions

based on their preferences or needs. Their study estimates repair rates for individ-

ual car parts using a probit model, considering various car and owner characteristics

such as mileage and driver income. Their results reveal evidence of adverse selec-

tion for specific parts like transmissions and engines, while sorting effects manifest

for other parts such as vehicle bodies and air conditioning. Their paper also em-

ploys quantification exercises to estimate the aggregate impact of adverse selection

on turnover and new-car sales. They find that a monolithic view of quality may

oversimplify market realities, underscoring the need for a more detailed examina-

tion along narrower dimensions of quality.

Houser and Wooder’s study (2006) investigates the role of reputation in online

auctions using data from eBay’s Pentium III 500 processor auctions in 1999. The

study employs a two-step generalized least squares (GLS) procedure to analyze the

impact of seller reputation on auction prices. Key variables include seller reputa-

6 Adverse Selection in Online Auctions

tion metrics such as positive, neutral, and negative feedback, auction characteristics

like length and exclusion of low-reputation bidders, and product attributes such as

condition and packaging. The results indicate that seller reputation significantly af-

fects auction prices, with a 10% increase in positive comments leading to a 0.17%

increase in the final bid price while a 10% increase in neutral or negative comments

leads to a 0.24% decrease in the final bid price.

Lewis (2011) addresses the paradox of sustained trade volume in the face

of adverse selection due to information asymmetry, using data from eBay Mo-

tors. Lewis argues that sellers counteract adverse selection by communicating their

goods’ quality through detailed disclosures on auction web pages. Using hedonic

regressions and data from over 80,000 car auctions on eBay Motors, he studies the

role of seller disclosures on price. Results indicate that photos and text descriptions

significantly influence prices, substantiating the empirical importance of online dis-

closures. Lewis also emphasizes the role of disclosure costs, considering factors

like technology and time, in shaping the success of online markets for used-cars.

Extending Lewis’s (2011) work, this research contributes to existing litera-

ture by empirically examining the role of supplemental information—specifically,

photos—in mitigating adverse selection in eBay Motors auctions. Unlike existing

studies, which mostly look at how disclosures of information affect prices, this anal-

ysis focuses on three auction outcomes—price, the number of bidders, and the like-

lihood of a trade. This study also categorizes the data by seller type, distinguishing

between experienced and inexperienced sellers, to proxy for seller reputation. The

research explores how sellers lacking an established reputation strategically commu-

nicate their vehicle’s quality to shape buyer perceptions and reduce uncertainty. The

interpretation of differing effects of photos on these outcomes across seller segments

provides empirical evidence of adverse selection mitigation beyond traditional price

considerations.

Previous literature on adverse selection in online markets, exemplified by

Hasker and Sickles (2010), has presented conflicting findings regarding the impact

of seller feedback and reputation on auction outcomes. While some studies suggest

limited explanatory power of seller feedback in determining auction prices, others

indicate evidence of adverse selection, particularly regarding early feedback and

seller reputation. However, this study takes a different approach by investigating

how seller-provided information addresses adverse selection issues. Specifically, it

distinguishes between dealer and non-dealer sellers of used cars on eBay Motors,

using dealer status based on listing volume as a proxy for reputation rather than

seller feedback.

3. Data

This paper uses Lewis’s (2011) data on asymmetric information and disclosures in

the market for used cars. The data come from completed eBay Motors auctions for

3 Data 7

used cars. Lewis collects the data by downloading auction web pages for selected

car models over the eight months from March to October 2006, and uses individ-

ual auctions as the primary unit of observation. A matching algorithm extracts the

relevant variables from each auction’s web page HTML code (Lewis 2011, p. 1538).

I drop observations with non-standard or missing data and those pertaining

to new or certified preowned cars. To narrow the analysis to a subset of vehicles

that hold particular interest to potential buyers, I drop observations of cars manu-

factured before 1965. This selective exclusion follows a robustness check, which

indicates that the removal of those observations has no significant impact on the

estimated outcomes. Consequently, the sample primarily comprises vehicles that

closely match the preferences and characteristics of contemporary car buyers, en-

hancing the relevance of the findings to the modern-day car market.

The web pages contain standardized information that the seller must provide

when listing the vehicle. For each auction, I observe car characteristics that the

seller lists: model, model year, mileage, and the number of accessories (denoted as

options) (e.g., car radio, etc). The data also indicate whether the vehicle is under

manufacturer warranty, has been recently inspected, relisted, or featured.

I include the seller’s eBay feedback

, as well as the percentage of negative

feedback on a given seller’s page. The focus of this paper is voluntary disclosures of

information and the primary measure is the number of photos posted on the auction

web page. I also have binary variables for whether key text phrases—rust, scratch,

and dent—are used in the item description, and modifiers for how they are used.

For example, a negation is a phrase like “rust-free,” or “never seen any rust.” As is

clear from the summary statistics in Table 1, the data exhibit substantial variation in

information content.

On average, the cars are from 1992 and well traveled (about 90 thousand miles

on the odometer). There are 17 photos on a typical auction web page. Sellers are

usually experienced with average feedback scores of 147. A typical auction attracts

about five bidders, and the auction price averages at just about $11,000. However,

only about 28% of the cars sell because sellers can set reserve prices.

In columns two and three, I distinguish between dealers and private sellers, as

a proxy for seller reputation, where a dealer is defined as any seller who lists more

than one vehicle on eBay. Dealers and private sellers behave differently. Dealers

tend to list newer cars, about four years newer, that are more likely to be under

warranty. Dealers typically upload approximately seven more photos than private

sellers, a behavior that could be attributed to their advanced advertising skills. With

their experience and resources, dealers are adept at showcasing their vehicles ef-

fectively, potentially influencing their decision to include more photos to attract

Buyers and sellers assign numerical evaluations from their respective perspectives. After com-

pleting a transaction, both parties can rate each other positively (+1), neutrally (0), or negatively

(-1).

8 Adverse Selection in Online Auctions

Table 1. Summary Statistics

Full Sample Dealers Private Sellers

Auction Price 10, 947.511 12, 224.588 8, 746.741

(12829.117) (14285.355) (9, 428.260)

Number of Bidders 5.292 5.129 5.572

(4.243) (4.238) (4.237)

Trade Probability 0.285 0.238 0.365

(0.451) (0.426) (0.481)

Miles (in thousands) 89.810 83.201 101.181

(89.018) (83.628) (96.547)

Vintage 1992 1993 1989

(12.596) (12.417) (12.516)

Photos 17.032 19.441 12.635

(10.873) (11.628) (7.570)

Number of Options 5.335 5.307 5.382

(4.920) (4.949) (4.870)

Seller Score 147 167 113

(558.357) (565.407) (544.437)

% Negative Feedback 1.610 1.700 1.441

(6.010) (5.768) (6.432)

Warranty 0.191 0.233 0.118

Inspection 0.240 0.250 0.223

Rust 0.188 0.165 0.227

Rust Negation 0.094 0.088 0.104

Scratch 0.167 0.188 0.132

Scratch Negation 0.056 0.062 0.047

Dent 0.122 0.118 0.130

Dent Negation 0.050 0.050 0.049

Relist 0.177 0.249 0.053

Featured 0.071 0.085 0.047

Buy-it-now 0.274 0.325 0.187

N 103,562 65,498 38,064

Note: Means and standard deviations (in parentheses). Standard deviation for categorical

variables are not reported. Column 2 is estimated on a sub-sample of dealers (list multiple

vehicles in the sample period). Column 3 is estimated on a sub-sample of private sellers

(list only a single vehicle in the sample period).

4 Methods 9

buyers’ attention and trust.

4. Methods

The goal of this paper is to assess how supplemental information from sellers, namely

photos, influences adverse selection in eBay Motors auctions. This section outlines

the analysis of repeated cross-sectional data on auctions using statistical methods.

I employ Generalized Linear Models (GLMs) to examine the three specified

auction outcomes: price, number of bidders, and trade likelihood. I use a log-link

with a gamma family specification to model auction prices, a Poisson specification

for the number of bidders, and a logit-link with a binomial family specification for

trade likelihood.

GLM models, like OLS, maintain robustness against violations

of distributional assumptions.

Auction Price

To estimate the effect of information revelations on price, I use a GLM model with

a log-link and gamma family specification. The advantage of a GLM model is its

ability to adapt to a wide range of data and distributions. Auction price data is often

right-skewed, deviating significantly from the normal distribution assumption that

underpins traditional regression techniques. A consequence of such skewness is

that predicted total expenditures from a linear regression model can be negative.

Even if negative predictions are not a concern, the excessive skewness can result in

unacceptably significant sample-to-sample variation in OLS estimates (McCullagh

and Nelder 1989, p. 287). In light of these challenges, GLMs provide an elegant

solution. By modeling the variance as a function of the mean, GLMs effectively

captures the distribution of auction prices.

The gamma regression takes on the following form:

E(p

) = exp(γ

+ µs

+ γ

+ τ

· s

+ τ

· s + x

β + α

+ θ

+ δ

)

(1)

where E(p

) is the expected value of the price in auction i at time t, γ

counts

the number of photos included, µs

is a binary variable for private sellers, and x

is a vector of car characteristics listed by the seller. The regression includes an

interaction term between photos and private sellers to estimate the marginal effect of

a photo and test for significant differences between coefficients across seller types.

In addition to the main effects of photos and seller status, the regression model

Estimating a GLM log link with a Poisson family is identical to estimating a Poisson maximum

likelihood model, and estimating a logit link with a binomial family is equivalent to estimating a

logistic regression model.

10 Adverse Selection in Online Auctions

also includes a squared term for photos, τ

, and an interaction term between the

squared photos and dealer status.

The squared term for photos introduces a quadratic effect, implying an in-

verted U-shaped relationship between the number of photos and prices. This sug-

gests that while adding photos initially boosts prices, there comes a point of dimin-

ishing returns, beyond which the marginal benefit of additional photos diminishes,

and prices may even start to decrease.

I include car model, θ

and model year effects, α

in the model to account

for unobservable factors that influence auction prices within specific car models and

over time. Model effects control for unique car model characteristics or desirability

factors that affect the number of bids these cars receive. Time effects, which in-

clude week of listing effects (δ

), allow for the capture and control of time-specific

variations and seasonal demand fluctuations that could influence auction bids.

The regression is clustered on the seller name to address the presence of simi-

larities among observations with a common seller. Clustering based on seller names

aids in identifying these correlations and adapting the standard errors of the regres-

sion estimates to account for this clustering structure.

The regression coefficients are not easily interpretable given the complexity

of the specification (quadratic and interactions within a nonlinear specification).

However, indicators like coefficients’ sign, statistical significance, and curvature of

squared variables offer some information.

I report Average Marginal Effects (AME), a statistical measure that provides

insights into how changes in one variable affect another, averaged across the dataset

while disregarding specific values of the predictor variable. The AME is calcu-

lated by first computing the marginal effect, which is the partial derivative of the

regression function with respect to the predictor variable, for each observation in

the sample. This marginal effect measures the instantaneous rate of change in the

response variable as the predictor variable increases by one unit. The marginal ef-

fect can vary across observations because the regression function is non-linear. To

obtain the AME, the marginal effects are calculated for each observation, and then

averaged across all observations. This provides an overall estimate of the average

change in the response variable associated with a one-unit change in the predictor

variable, accommodating potential variations in the existing values of the predictor

variable featured in the sample.

Furthermore, I conduct a joint statistical significance test between dealers and

private-sellers to assess whether the overall difference across seller types is signifi-

cant on average.

I also plot the Average Marginal Effects (AMEs) for each seller type across a

range of photo quantities. These quantities are determined by percentiles, covering

from the 10th to the 90th percentile of the photo distribution in the data. This ap-

4 Methods 11

proach facilitates a visual representation of the nonlinear relationship, particularly

showcasing the diminishing marginal returns of photos, and how this relationship

varies across different seller types. The illustration also helps in analyzing the sig-

nificance of these differences among seller types by evaluating whether the confi-

dence intervals overlap at each photo quantity. When the confidence intervals do

not overlap, it is likely that the differences between seller types for that specific

photo quantity is statistically significant.

Estimates showing a significant increase in price associated with an increase

in the number of photos serve as evidence that photos constitute quality information.

This suggests that the inclusion of photos positively influences the perceived quality

of cars in the market. The observed increase in prices indicates that buyers attribute

higher value to cars with more visual information. Additionally, increase in prices

due to additional photos suggest the trade of higher-quality cars.

A significant difference in the effect of photos across seller types would un-

derscore the multifaceted role of seller-provided information in online auctions.

Dealers, backed by their reputations, may rely less on photos to convey quality,

as buyers may not require as much supplementary information due to the credibility

their established reputations provide. Therefore, for dealer, the marginal effect of

adding each additional photo should be relatively small. In contrast, private sellers,

who lack such reputational capital, likely benefit more substantially from provid-

ing additional photos, which could serve the dual purpose of conveying quality and

alleviating buyer apprehensions about the car’s condition.

Market Thickness

To analyze market thickness, I replace the dependent variable in equation (1) with

the number of bidders in the auction and estimate a Poisson regression of the fol-

lowing form:

E(b

) = exp(γ

+ µs

+ γ

+ τ

· s

+ τ

· s + x

β + α

+ θ

+ δ

)

(2)

where E(b

) is the expected value of the number of bidders in auction i at time

t, γ

counts the number of photos included, µs

is a binary variable for private

sellers, and x

β is a vector of car characteristics listed by the seller. The regres-

sion includes an interaction term between photos and private sellers to estimate the

marginal effect of a photo and test for significant differences between coefficients

across seller types. In addition to the main effects of photos and seller status, the

regression model also includes a squared term for photos, τ

, and an interaction

term between the squared photos and dealer status.

I base the selection of the number of bidders as a measure for market thickness

on Kagel and Levin’s (1986) study, as they propose that in auctions with common

12 Adverse Selection in Online Auctions

values, like those observed on eBay, the number of bidders reflects both competition

and available information, thus acting as an indicator for market thickness.

The choice of a Poisson regression is suitable for modeling count data, specif-

ically non-negative integer values like the number of bidders. Poisson models ac-

count for the presence of excess zeros and address over-dispersion in count data

analysis (McCullagh and Nelder 1989, p. 194). Additionally, it’s generally assumed

that bidders join auctions without prior knowledge of the goods’ values, following a

stochastic process for their entry decision. This random entry process aligns with the

assumptions underlying Poisson models (Hasker and Sickles 2010, p. 20), making

the Poisson regression a suitable choice for analyzing auction behavior and bidder

participation.

Excess zeros often occur in auction data when certain auctions fail to attract

any bidders, either due to the nature of the items being auctioned or the quality of

the auction listings. Additionally, over-dispersion is a common challenge in ana-

lyzing count data, especially in auction settings, where bidder behavior can vary

significantly. Multiple unobserved factors (e.g., changes in income or shifts in con-

sumer sentiment) may influence participation and contribute to the variability in the

number of bidders. By accounting for excess zeros and addressing over-dispersion,

the analysis gains robustness and enhances the efficiency of coefficient estimates

(McCullagh and Nelder 1989, p. 198-199).

Similar to price, the regression estimates are not easily interpretable and AMEs

report the estimates in the original units of the response variable, number of bidders.

In this setting, information plays a crucial role. Sellers who do not disclose

information deter potential bidders from participating in the auction. The disclosure

of information introduces a two-pronged issue. On the one hand, revealing negative

aspects may discourage some bidders due to the perceived poor condition of the

car. On the other hand, honest disclosures are expected to attract more bidders,

as they can better assess the car’s condition and make informed decisions about

market participation. The seller must trade off the positive and negative effects of

disclosure, a challenge that is particularly pronounced for private sellers, as adverse

selection concerns may deter potential buyers.

Trade Likelihood

To further measure whether adverse selection exists in this market, I estimate a GLM

model with a logit-link and binomial family specification of the following form:

E(t

) = Λ(γ

+ µs

+ γ

+ τ

· s

+ τ

· s + x

β + α

+ θ

+ δ

)

(3)

where E(t

) is the expected value of a trade in auction i at time t, γ

counts the

number of photos included, µs

is a binary variable for private sellers, and x

4 Methods 13

is a vector of car characteristics listed by the seller. The regression includes an

interaction term between photos and private sellers to estimate the marginal effect of

a photo and test for significant differences between coefficients across seller types.

In addition to the main effects of photos and seller status, the regression model

also includes a squared term for photos, τ

, and an interaction term between the

squared photos and dealer status.

A logit model allows for the quantification of the likelihood of a sale under

varying conditions, which encompass the presence of photos and other car charac-

teristics.

Regression coefficients can determine the significance and the direction of the

effects but do not provide a direct estimation of their magnitude. I report the AMEs

to estimate the average change in the predicted probability of a trade for a one-unit

change in the independent variable.

Estimates showing a The differential impact of photos on trade likelihood be-

tween private sellers and reputable dealers highlights how information provision

counters the adverse selection issues that distinctly affect lesser-known private sell-

ers in the absence of such quality signals. Providing more photos has a much larger

positive effect on the probability of completing a transaction for private sellers ver-

sus dealers, demonstrating how this circumvents the adverse selection barrier private

sellers uniquely face due to their lack of reputation.

Robustness: Heterogeneity

Following my analysis, I test for sources of heterogeneity, which in the context of

used car auctions, can reveal how different characteristics of a car influence auction

outcomes. Specifically, I look at four dimensions: the presence of a warranty, the

age of the car (whether it is newer), recent inspection history, and whether the car

is a more recent model (manufactured from 1990 onwards). I analyze the AME

of photos on the four aforementioned sub-samples for all three auction outcomes,

categorized by seller type. Comparing these AMEs with those from the original

regressions helps detect any variations in photo effects attributed to the presence of

heterogeneity.

The presence of warranties in used car auctions serves as a crucial signal of

car quality, particularly for high-quality car traders who may otherwise exit the

market due to inadequate compensation for their vehicle’s quality. In this context,

warranties function as a signaling mechanism to convey the higher quality of cer-

tain cars. This signal is effective because it meets two essential requirements: it

is observable to uninformed parties, in this case buyers, and it carries a cost that

is prohibitively high for owners of low-quality cars. By offering warranties, sellers

effectively communicate their confidence in the vehicle’s condition, reassuring buy-

ers and attracting bids. Importantly, the cost associated with providing a warranty

serves as a credible signal of quality, as it is financially burdensome for sellers of

14 Adverse Selection in Online Auctions

low-quality cars to bear. Consequently, the use of warranties can result in a sepa-

rating equilibrium in the market, where buyers can discern between high and low-

quality cars, facilitating the trade of high-quality vehicles with greater transparency

and confidence.

The next sub-sample examines whether the car is newer, where the term refers

to cars with fewer than 4000 miles traveled on the odometer and an age of five

years or less. This cutoff selection follows search of industry standards and common

practices and is reflective of the typical understanding of what constitutes a new car.

Newer cars typically exhibit a narrower range of uncertainty regarding their quality.

Buyers tend to feel more confident in assessing the average quality of newer cars

due to their relatively pristine condition and limited wear and tear.

Analyzing whether the car has been recently inspected serves as another po-

tential source of heterogeneity in auction outcomes. A recent inspection acts as a

signal of the car’s quality, potentially alleviating buyer concerns about the need for

immediate post-purchase inspections. In many jurisdictions, inspections are manda-

tory for vehicle registration, ensuring that inspected cars meet safety and quality

standards. Therefore, a recent inspection implies that the car is in better condition

and less likely to have underlying issues, providing buyers with greater confidence

in their purchase.

Lastly, considering cars manufactured from 1990 onwards as a source of het-

erogeneity serves to make the sample more homogeneous focusing on vehicles gen-

erally used for transportation purposes rather than being sought after as collector’s

items.

For each specified sub-sample, estimates suggesting a reduction in the impact

of photos on auction outcomes, indicate a diminished value of photos as a means to

communicate information. This decrease in photo disclosures’ significance is ex-

pected to apply consistently across both dealer and private seller types. The dimin-

ished value of photos in these contexts reflects factors such as the higher perceived

quality of newer cars, the assurance provided by warranties and recent inspections,

and the focus on transportation-oriented vehicles in post-1990 manufacturing. Con-

sequently, the convergence in the effects of photos between different seller types

indicates a standardized and transparent marketplace where the influence of addi-

tional photo disclosures is less pronounced, and other factors may take precedence

in buyers’ decision-making processes.

Robustness: Specication Checks

I then conduct a series of specification-checks to validate the robustness of the es-

timated outcomes in my analysis. Specifically, I examine five elements: the inclu-

sion of observations from 1950 onwards, the incorporation of text descriptors and

modifiers, the impact of relisted cars, the significance of featured listings, and the

presence of a ”buy it now” option. For each specification check, I report the overall

5 Results 15

regression AME, comparing it to the AME obtained from the original regressions.

Estimates showing similar AME across different specifications serve as successful

specification checks, indicating that variations in model specifications or sample

compositions do not significantly affect the estimated outcomes.

Following Lewis’s (2011) approach, I include observations from the 1950-

1964 period to ensure that the exclusion of those vehicles in my analysis does not

significantly affect my estimated outcomes.

I then examine the impact of text descriptors and modifiers on auction out-

comes. This check assesses whether the textual information provided in the listings

independently influences auction outcomes, apart from visual cues.

Relisting refers to the re-posting of an item for sale after its initial listing ex-

pires without a successful sale. Incorporating this variable as a specification check

enables examination of whether relisted vehicles demonstrate different auction out-

comes compared to their initial listings.

Featured listings receive special placement and visibility on the platform, po-

tentially attracting more attention from buyers. Considering whether a vehicle list-

ing is featured allows evaluation of whether enhanced visibility affects auction out-

comes. This specification check ensures that our analysis accounts for the potential

impact of promotional features on buyer engagement and final sale prices.

The Buy-It-Now option enables sellers to set a fixed price at which buyers

can purchase the item immediately, bypassing the auction process. Incorporating

this variable in the analysis enables assessment of whether the availability of this

option influences auction outcomes.

5. Results

Tables 2 and 3 present the estimates and marginal effects, respectively, for the three

model specifications. The results consistently reveal a positive effect of disclosures

of information, in the form of photos, on the success of an auction. An additional

photo is associated with a higher auction price, increased bidder participation, and a

greater likelihood of trade for both seller types, yet the effects are markedly stronger

for private sellers. This difference underscores how providing information is es-

pecially critical for sellers without an established reputation to overcome adverse

selection issues that reputable dealers do not face to the same degree.

Auction Price

The first two columns of Table 2 present the results of the regression model that

estimates the influence of photos on auction prices. While the coefficients do not

offer straightforward interpretation because of the complexity of the specification

(quadratic and interactions within a nonlinear specification), indicators like coeffi-

cients’ sign and statistical significance offer some information. Table 3 presents the

16 Adverse Selection in Online Auctions

Table 2. Regressions

Price Price

∗

Bidders Bidders

∗

Trade Trade

∗

Car Characteristics

Log Miles

−

1557

−

1552

0123

0121

1320

1323

(0.0000) (0.0000) (0.0005) (0.0005) (0.0000) (0.0000)

Options 0.0122 0.0123 0.0083 0.0073 0.0083 0.0025

(0.0000) (0.0000) (0.0000) (0.0000) (0.0259) (0.4864)

Photos 0.0156 0.0232 0.0165 0.0201 0.0157 0.0283

(0.0000) (0.0000) (0.0000) (0.0000) (0.0013) (0.0000)

Photos Squared

−

0002

−

0003

−

0003

−

0003

−

0004

−

0003

(0.0000) (0.0000) (0.0001) (0.0001) (0.0016) (0.0231)

Dealer 0.2008 −0.0365 −0.3602

(0.0000) (0.3607) (0.0000)

Dealer * Photos −0.0128 −0.0042 −0.0122

(0.0000) (0.2997) (0.1378)

Dealer * Photos Squared

0001

0000

−

0000

(0.0001) (0.7618) (0.9862)

Seller Characteristics

Seller Score −0.0076 −0.0091 0.0026 0.0048 0.0719 0.0866

(0.0006) (0.0000) (0.5242) (0.2324) (0.0000) (0.0000)

% Negative Feedback

−

0039

−

0042

−

0029

−

0022

0070

0109

(0.0000) (0.0000) (0.0753) (0.1606) (0.0434) (0.0011)

Model Year Effects Yes Yes Yes Yes Yes Yes

Listing Week Effects Yes Yes Yes Yes Yes Yes

Car Model Effects Yes Yes Yes Yes Yes Yes

N 69,370 69,370 80,346 80,346 80,346 80,346

Notes: All three models are GLM models. The Gamma model regresses the auction price on the characteristics of the car and the

seller. The Poisson model regresses the number of bidders on those same characteristics. The Logit model estimates the likelihood

of a trade. Columns with a star include an interaction term to test for significant differences across seller type. The sample size is

higher in the Poisson and Logit models due to the inclusion of auctions that did not result in a sale but contain relevant information

about the auction. p-values, shown in parentheses, are calculated using standard errors that are robust to seller-level clustering.

5 Results 17

average marginal effects derived from the regressions to further interpret the impact

of information on auction outcomes.

The positive coefficient on photos suggests that additional photos generally

result in higher prices. The negative coefficient on the squared term for photos

indicates a quadratic effect, implying an inverted U-shaped relationship between

the number of photos and auction prices. This suggests that while initially, adding

photos increases auction prices, there reaches a point of diminishing return, beyond

which the marginal benefit of additional photos diminishes and may even become

negative. Because the regression model is nonlinear, these quadratic effects are best

visualized using estimates of the marginal effects. These are described below.

The positive and significant coefficient associated with the dealer variable in

column 2 indicates that, on average, dealers secure higher prices compared to non-

dealers, reflecting potential buyer perceptions of dealer reliability or trustworthi-

ness. While the regression model includes interaction terms between dealer status,

photos, and photos squared, the interpretation of these coefficients is better under-

stood by examining the marginal effects.

The covariates include car features such as mileage and options, as well as

seller characteristics like feedback score and percentage of negative feedback. Most

coefficients align with expectations and attain statistical significance, evaluating the

model’s overall significance through hypothesis tests at the 5% level. Per Lewis’s

(2011) analysis, auction year, listing week, and car model effects are all statistically

significant at the 5% level.

An additional option, such as a radio, increases the auction price. Conversely,

for each percentage increase in logged miles traveled, the price decreases. The nega-

tive and significant coefficient for the percentage of negative feedback suggests that

as the proportion of negative feedback increases, auction prices tend to decrease.

This indicates that potential buyers may perceive sellers with a higher incidence of

negative feedback as less reputable or trustworthy, leading to lower auction prices.

While a higher seller score could be interpreted positively as a sign of seller rep-

utation, it may also signify a high volume of sales, potentially leading sellers to

prioritize volume over profit margins. This dual interpretation results in conflicting

signals, with the coefficient estimation being negative in this case. This suggests

that sellers prioritizing sales volume and potentially accepting lower profit margins

may contribute to decreasing auction prices.

The first line in Table 3 displays the corresponding average marginal effects

derived from the regressions in Table 2.

The average marginal effect of photos on price is $96. All else equal, an addi-

tional photo increases auction price for dealers by approximately $70 and for non-

dealers by around $139. The difference across seller types is statistically significant

(p-value = 0.0000).

18 Adverse Selection in Online Auctions

Table 3. Marginal Eects of Number of Photos on Auction Outcomes

Overall Dealers Private Sellers

Price 96.1810 70.0894 138.6443

(0.0000) (0.0000) (0.0000)

Number of bidders 0.0398 0.0333 0.0729

(0.0000) (0.0000) (0.0000)

Probability of sale 0.0007 0.0007 0.0043

(0.1204) (0.1747) (0.0000)

Notes: p-values, shown in parentheses, are calculated using standard errors that are robust to seller-

level clustering.

A graph of the marginal effect of photos on price reveals distinct relation-

ships for the two types of sellers, as depicted in Figure 1. This visualization allows

observing the diminishing marginal returns to additional photos, with the marginal

effect consistently lower for dealers relative to non-dealers. This suggests that for

any given increase in the number of photos, dealers tend to experience a smaller

increase in auction prices compared to non-dealers. The graph exhibits an inverted

concave shape, indicating that while adding more photos increases prices for both

seller types, the marginal impact of each additional photo decreases.

Figure 1. Marginal Eects on Price by Seller Type

100

150

200

Change in Price

5 10 15 20 25 30

Number of Photos

Non-dealer

Dealer

Market Thickness

Poisson regressions estimate the impact of information metrics on the number of

bidders, in columns 3 and 4 of Table 2. The second line of Table 3 displays the

corresponding average marginal effects derived from the regressions.

5 Results 19

Figure 2. Marginal Eects on Bidders by Seller Type

.02

.04

.06

.08

Change in Number of Bidders

5 10 15 20 25 30

Number of Photos

Non-dealer

Dealer

On average, an additional photo increases the number of bidders by .04%.

More specifically, an additional photo increases the number of bidders by .03% for

dealers and .07% for private sellers. The difference across sellers is significantly

different from 0 (p-value = 0.0055).

Seller characteristics, specifically the seller score and the percentage of nega-

tive feedback, show no statistical significance and exert no influence on the number

of bidders in an auction.

Similar to the marginal effect on auction price, a graph can depict the marginal

effect of photos on the number of bidders, with distinct relationships observed for

the two types of sellers, as shown in Figure 2. The inverted concave shape indicates

diminishing marginal returns of additional photos in attracting more bidders. This

marginal effect is consistently lower for dealers compared to non-dealers, suggesting

that for any given increase in the number of photos, dealers tend to experience a

smaller increase in the number of bidders relative to non-dealers.

Trade Likelihood

A logit regression model estimates the likelihood of a car sale based on informa-

tion metrics in columns 5 and 6 of Table 2. The third line of Table 3 shows the

corresponding marginal effects.

An additional photo, on average, increases the likelihood of a trade. The av-

erage marginal effect of photos on trade likelihood is not significant. However, the

effect is significant for private sellers. In that case, including a photo increases trade

probability by .004%. The difference across seller type is statistically significant (p-

value = 0.0001).

Figure 3 shows the marginal effect of photos on the trade probability, with

20 Adverse Selection in Online Auctions

Figure 3. Marginal Eects on Trade Probability by Seller Type

-.002

.002

.004

.006

Change in the Probability of Sale

5 10 15 20 25 30

Number of Photos

Non-dealer

Dealer

distinct relationships observed for the two types of sellers. The inverted concave

shape indicates diminishing marginal returns of additional photos in increasing the

trade probability. This marginal effect of each additional photo on trade probabil-

ity is consistently lower for dealers compared to non-dealers, suggesting that for

any given increase in the number of photos, dealers tend to experience a smaller

increase in trade probability relative to non-dealers. Additionally, for most ranges

of the number of photos, the confidence intervals do not overlap, providing a rea-

sonable indication that the estimated marginal effects are statistically different from

one another.

Robustness: Heterogeneity

Table 4 presents the average marginal effects by seller type for all three auction out-

comes, calculated from regressions estimated on sub-samples identified to induce

heterogeneity. The AME provide insights into how these heterogeneous character-

istics affect auction outcomes differently for dealers and private sellers, contributing

to an understanding of the factors driving variability in the auction process. These

estimates are to be compared with the original AME presented in Table 3.

The effect of photos on auction outcomes for cars under manufacturer war-

ranty tends to be smaller compared to the overall sample estimates. This aligns

with theoretical expectations, considering the warranty’s role as both a contractual

enforcement mechanism and a signaling tool to mitigate adverse selection. Further-

more, the difference in the effect of photos between dealer and non-dealer sellers

appears to be smaller in this sub-sample. This convergence suggests that the pres-

ence of a warranty may reduce the informational advantage typically enjoyed by

dealers.

For cars that are newer, recently inspected, and a more recent model, the re-

5 Results 21

Table 4. Heterogeneous Eects

Warranty Newer Car Inspection Recent Models

Dealers Private Sellers Dealers Private Sellers Dealers Private Sellers Dealers Private Sellers

Price 12.5854 70.9828 31.9119 66.4103 43.9731 108.2008 38.3296 108.6842

(0.2213) (0.0000) (0.3806) (0.2667) (0.0000) (0.0000) (0.0000) (0.0000)

Number of bidders 0.0227 0.0588 0.00721 0.0589 0.0500 0.0850 0.0246 0.0717

(0.0010) (0.0000) (0.5507) (0.0099) (0.0000) (0.0000) (0.0000) (0.0000)

Probability of sale 0.0004 0.0027 -0.0009 0.0006 0.0013 0.0049 0.0003 0.0043

(0.2555) (0.0035) (0.2885) (0.8218) (0.0386) (0.0000) (0.5181) (0.0000)

Notes: Columns 3 and 4 are estimated on a sub-sample of newer cars, where newer is defined as cars with fewer than 4000 miles traveled on the odometer and

an age of five years or less. Columns 7 and 8 are estimated on a sub-sample of recent models referring to cars manufactured from 1990 onwards. p-values,

shown in parentheses, are calculated using standard errors that are robust to seller-level clustering.

22 Adverse Selection in Online Auctions

sults align with theoretical prediction. The quality assurances associated with these

features should reduce the impact of photos on auction outcomes compared to the

overall sample, a pattern that largely holds in the data.

Moreover, the difference in the effect of photos between dealer and non-dealer

sellers tends to be smaller in these subgroups. This convergence is attributed to the

reduced informational advantage that dealers typically hold in auctions involving

such vehicles. Additionally, buyers in these contexts may prioritize attributes be-

yond just the number of photos when evaluating the value of newer cars or those with

recent inspection histories. Consequently, the effect of photos on auction outcomes

may show less variability between dealer and private sellers in these particular sce-

narios.

Robustness: Specication Checks

Table 5 presents the overall AME sample estimates for the specification check sub-

samples. These subsets consist of classic cars (vehicles from 1950 onwards), those

with text descriptors and modifiers, as well as those with certain listing features like

relisted, featured, or buy-it-now options. These estimates are compared with the

overall estimates presented in Column 1 of Table 3. This comparison allows for an

assessment of whether variations in model specifications and sample compositions

significantly impact the estimated outcomes, thereby validating the reliability of the

analysis.

For classic cars, the average marginal effects stay relatively similar to the ones

estimated in the regression. These findings indicate robustness, suggesting that ex-

cluding these observations from the overall analysis does not significantly impact

the estimated outcomes. The results for the other specifications are expected to be

similar, as these subsets are all a small fraction of the sample, indicating that their

exclusion does not heavily influence the overall analysis.

6. Summary and Conclusion

Analyzing eBay Motors data, this paper assesses the presence of adverse selection

by examining how the impact of providing observable information through photos

differs between dealer and private sellers. It uses the disparity in photo value across

seller types as an indicator of the prevalence of adverse selection. To capture how

information asymmetries affect market behavior, the study analyzes photos’ influ-

ence on auction price, number of bidders, and trade probability.

The results consistently reveal a positive effect of disclosures of information,

in the form of photos, on the success of an auction. Including an additional photo

has a significantly greater impact on private sellers compared to dealers, with the

effect being twice as large for both auction price and bidder participation. Moreover,

for private sellers, the influence of additional photos on the probability of trade is

6 Summary and Conclusion 23

Table 5. Specication Checks

Classic Cars Text Descriptors Re-list Featured Buy-it-now

Price 100.9010 95.8542 103.3647 94.7370 114.7016

(0.0000) (0.0000) (0.0000) (0.0000) (0.0000)

Number of bidders 0.0399 0.0383 0.0414 0.0385 0.0428

(0.0000) (0.0000) (0.0000) (0.0000) (0.0000)

Probability of sale 0.0006 0.0005 0.0013 0.0007 0.0009

(0.1265) (0.2606) (0.0022) (0.0859) (0.0399)

Notes: p-values, shown in parentheses, are calculated using standard errors that are robust to seller-level clustering.

24 Adverse Selection in Online Auctions

nearly six times greater. This disparity in the impact of supplemental information

suggests evidence of adverse selection in the market, highlighting the role photos

as a constituent of quality information. Photos serve as a direct transfer of quality

information and help mitigate adverse selection, particularly for sellers without an

established reputation, where their importance should be greater.

The paper has several important limitations. Firstly, potential measurement

error exists in the categorization of sellers as dealers or private sellers. The distinc-

tion relies on the frequency of appearances in the sample: dealers are those selling

more than one car, and private sellers are those selling only one. This categoriza-

tion lacks a definitive criterion, and different definitions of dealer and private seller

categories could introduce variations.

Additionally, an interpretation limitation emerges due to the absence of infor-

mation about the quality of the photos. While the quantity of photos is observable,

there is no information about the photo quality. Higher-quality photos can act as

a direct transfer of information from sellers, shaping buyer perceptions of the cars

condition.

Lastly, the data’s focus on eBay Motors auctions may limit the applicability

of the results to other contexts or periods.

This empirical evidence is consistent with the theory of adverse selection.

The information’s impact on trade likelihood is more pronounced for private sell-

ers, while it shows less influence on sellers with an established reputation. This

distinction underscores the importance of information for those facing greater ad-

verse selection concerns. Conversely, for sellers with a reputation, where adverse

selection is less prevalent, the impact is less discernible.

References

Akerlof, George A. 1970. “The Market for Lemons: Quality Uncertainty and the

Market Mechanism.” Quarterly Journal of Economics 84 (3): 488–500.

David Lucking-Reiley, Naghi Prasad, Doug Bryanz and Daniel Reevesz. 2007.

“Pennies from eBay: The Determinants of Price in Online Auctions.” Journal of

Industrial Economics 55 (2): 223–233.

Hasker, Kevin and Robin Sickles. 2010. “eBay in the Economic Literature: Analysis

of an Auction Marketplace.” Review of Industrial Oganization 37 (1): 3–42.

Heal, Geoffrey. 1976. “Do Bad Products Drive Out Good?” Quarterly Journal of

Economics 90 (3): 499–502.

Houser, Daniel and John Wooders. 2006. “Reputation in Auctions: Theory, and

Evidence from eBay.” Journal of Economics and Management Strategy 15 (2):

353–369.

References 25

Kagel, John H. and Dan Levin. 1986. “The Winner’s Curse and Public Information

in Common Value Auctions.” American Economic Review 76 (5): 894–920.

Lewis, Gregory. 2011. “Asymmetric Information, Adverse Selection and Online

Disclosure: The Case of eBay Motors.” American Economic Review 101: 1535–

1546.

McCullagh, Peter and John A. Nelder. 1989. Generalized Linear Models. 2. London:

Chapman and Hall.

Ockenfels, Axel and Alvin Roth. 2006. “Late and Multiple Bidding in Second Price

Internet Auctions: Theory and Evidence Concerning Different Rules for Ending

an Auction.” Games and Economic Behavior 55 (2): 297–320.

Oi, Walter. 1971. “A Disneyland Dilemma: Two-Part Tariffs for a Mickey Mouse

Monopoly.” Quarterly Journal of Economics 85 (1): 355–374.

Peterson, Jonathan and Henry Schneider. 2014. “Adverse Selection in the Used-Car

Market: Evidence from Purchase and Repair Patterns in the Consumer Expendi-

ture Survey.” RAND Journal of Economics 45: 140–154.

Rothschild, Michael and Joseph Stiglitz. 1976. “Equilibrium in Competitive Insur-

ance Markets: An Essay on the Economics of Imperfect Information.” Quarterly

Journal of Economics 90 (4): 629–649.

Shibo Li, Kannan Srinivasan and Baohong Sun. 2009. “Internet Auction Features

as Quality Signals.” Journal of Marketing 73 (1): 75–92.

Spence, Michael. 1973. “Job Market Signaling.” Quarterly Journal of Economics

87 (3): 355–374.

Wilson, Charles. 1977. “A Model of Insurance Markets with Incomplete Informa-

tion.” Journal of Economic Theory 16 (2): 167–207.