Reference Price Shifts and Customer Antagonism: Evidence from Reviews for Online Auctions
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
ETH Library Reference Price Shifts and Customer Antagonism: Evidence from Reviews for Online Auctions Journal Article Author(s): Gesche, Tobias Publication date: 2022 Permanent link: https://doi.org/10.3929/ethz-b-000519005 Rights / license: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International Originally published in: Journal of economics & management strategy, https://doi.org/10.1111/jems.12473 This page was generated automatically upon download from the ETH Zurich Research Collection. For more information, please consult the Terms of use.
Received: 20 December 2019 | Revised: 10 March 2021 | Accepted: 1 December 2021
DOI: 10.1111/jems.12473
ORIGINAL ARTICLE
Reference‐price shifts and customer antagonism: Evidence
from reviews for online auctions
Tobias Gesche
Center for Law & Economics, ETH
Zurich, Zurich, Switzerland Abstract
This paper investigates how bidders in an auction become antagonized over
Correspondence
Tobias Gesche, Center for Law &
their successful bid. Using data from a large‐scale sales campaign on eBay
Economics, ETH Zurich, Zurich 8092, shows that auction buyers use the platform's feedback system to punish the
Switzerland. seller when they discover that the same item is later offered for a lower fixed
Email: tgesche@ethz.ch
price. Specifically, it finds that (i) the probability of receiving unfavorable
Funding information feedback is four times larger for auction sales than when the same item is sold
Swiss National Science Foundation by the same seller for the fixed price and (ii) that this probability is increasing
(SNSF), Grant/Award Number:
in the auction price, even though reviewing bidders shape this price them-
#P2ZHP1_171900
selves. Exploiting a temporal variation in how salient the fixed‐price offer was
and using text analysis tools on buyer comments shows that these effects on
feedback are best explained by ex post reference‐price shifts. An additional
survey experiment with exogenous variation in reference prices provides fur-
ther evidence for this channel.
1 | INTRODUCTION
Pricing is crucial for sellers and policymakers alike for reasons that go beyond the resulting allocations and
transfers. Longstanding evidence shows that not only the price itself but also the circumstances under which it is
determined affect how buyers evaluate a transaction (Frey & Pommerehne, 1993; Kahneman et al., 1986; Xia
et al., 2004). This is particularly relevant for online markets as their flexible nature allows one to vary sales
conditions and prices across buyers, either as means of experimentation (Einav et al., 2015) or user‐based price
discrimination (Mikians et al., 2012; Shiller, 2014). At the same time, however, they rely crucially on user‐
written reviews which might be negatively affected by differential pricing (see Tadelis, 2016). Online auctions
can accommodate these elements simultaneously: They allow differential pricing but also have the potential to
prevent dissatisfied buyers who are acting as passive price‐takers but who are, through their bids, consciously
and willingly shaping the prices they pay (Chandran & Morwitz, 2005; Hinz et al., 2011). However, this paper's
findings presented show that when auctions coexist with fixed‐price offers, they cause customer antagonism
through features which are typical of online markets—reputation systems and buyers who underutilize com-
peting offers.
Using data from a large‐scale online sales campaign, I report on the determinants of postsale behavior by
successful bidders toward the seller. During the campaign, the seller (a railway company) first used auctions to sell
several thousand units of an item (a voucher for an open‐destination rail journey). Two days later, the same seller
This is an open access article under the terms of the Creative Commons Attribution‐NonCommercial‐NoDerivs License, which permits use and distribution in any
medium, provided the original work is properly cited, the use is non‐commercial and no modifications or adaptations are made.
© 2022 The Authors. Journal of Economics & Management Strategy published by Wiley Periodicals LLC.
J Econ Manage Strat. 2022;1–21. wileyonlinelibrary.com/journal/jems | 12 | GESCHE
sold the same item on the same platform (eBay) for a fixed price. I find that auction buyers are four times more likely
to use the market platform's feedback system to give the seller unfavorable, public feedback than fixed‐price buyers.
Also, the more auction buyers paid, the more likely they are to punish the seller through unfavorable feedback.
These findings are hard to reconcile with the fact that the buyers themselves played a crucial part in determining the
auction price. By bidding accordingly, they could have easily prevented to pay a price over which they become
antagonized. Considering this and the fact that the reviewed item and seller were the same across sales mechanisms,
the feedback difference between the auctions and fixed‐price sales is also puzzling. To address these findings, an
explanation based on ex post reference‐price updates is suggested. The main insight is that a downward shift in
reference prices, as caused by observing a lower fixed‐price offer, leads successful bidders to negatively assess the
auction.
In line with this explanation, I show that the effect of a higher auction price on adverse feedback is concentrated in
the period during which the item had not been shipped, thus when buyers had not got new any transaction‐specific
information. However, successful bidders could learn about the fixed‐price sale within this period so observing this
offer could shift successful bidders' reference price and their feedback downward. This contrasts with behavior right
after this initial period, when items started to arrive and buyers' actual experience from the transaction and using the
voucher could also start to affect feedback. In these later periods, the negative price effect on feedback drops to zero and
stays flat. I also show that this temporal variation of the price effect can be linked to the feedback differential between
the two sales mechanisms: In the initial period before the items arrived—when reference prices were salient and the
price effect pronounced—the rate of unfavorable feedback is about 40 ppt (percentage points) higher for auctions than
for fixed‐price sales. Afterward, when the price effect diminished and reference prices mattered less, this feedback
differential shrinks to less than a quarter of its initial size.
The findings are also in line with a sentiment analysis of the comments written by auction buyers in
their reviews. This sentiment analysis shows that comments written in the early period, before the voucher
arrived, are characterized by more negative keywords and sentiments as compared with later. A similar contrast
appears for the sentiment in comments referring to the sellers' pricing policy as opposed to pricing‐unrelated
comments.
Further evidence for the role of reference prices comes from a survey experiment. It exogenously varies the
reference price which is observed by customers. The results first replicate the findings from the field that observing a
lower price ex post leads lower feedback. Second, they show that this negative effect is larger in magnitude than the
increase in feedback observed following the observation of a higher reference price observed ex post. This finding is in
line with loss aversion and, therefore, reference‐dependent transaction evaluations.
2 | RELATED L ITERATURE
This paper's findings shed light on and resonate with other findings on pricing in online markets. First, the ob-
servation that buyers and, through their feedback, also sellers experience reference‐price shifts negatively provides
one explanation for why online prices do not change as often as one would expect (Gorodnichenko et al., 2018). For
auctions specifically, the present results relate to the findings by Einav et al. (2018): Using a comprehensive data set
from eBay, they show that the share of auctions of sales on the platform has decreased from about 65% in 2008 to just
over 15% five years later. They attribute their results to the search costs necessary to make a good deal in an auction,
relative to the convenience of fixed‐price offers. Similarly, Bhave and Budish (2017) show that auctions did not
persist as an online sales mechanism for event tickets, even though they cleared the market and prevented spec-
ulation much more effectively than fixed prices. In line with these observations, this paper shows how not only
buyers but also the seller suffers directly from using an auction when buyers undersearch and then punish the seller
for their (perceived) losses.
Such negative feedback is a form of customer antagonism, that is, buyers converting negative emotions into
concrete actions against sellers (see Yi & Baumgartner, 2004). In this regard, the field experiment by E. T.
Anderson and Simester (2010) is close to this paper. They show that after a mail‐order sent out catalogs with
randomized discounts, buyers who had bought the discounted items before, at a higher price, subsequently ceased
to order from that mail‐order.1 In line with this paper's findings, this effect is most concentrated among the buyers
who had previously paid the most, relative to the discount. This paper provides an explanation based on reference‐
price shifts which accommodate these results and those presented here. In addition, it adds to the empiricalGESCHE | 3
literature on customer antagonism along three main dimensions: First, it demonstrates the relevance of customer
antagonism in online retail. Second, this is the first study which demonstrates that price‐based customer antag-
onism can also arise in auctions, that is, in a sales mechanism in which buyers have a crucial and active role in the
price‐setting process. Third, it shows that customer antagonism does not only manifests through demand (e.g.,
antagonized buyers boycotting a seller) but also through feedback systems, on which sellers and sales platforms
crucially depend.2
The finding that observing an ex post reference‐price shift spills over to assessments of transactions also connects
this paper to a lab experiment by Bushong and Gagnon‐Bartsch (2019). They show that negative surprises in outcomes
lead subjects to evaluate a transaction and the associated real‐effort task negatively.3 The present work presents field
evidence for such a mechanism. In another related work, Backus et al. (2017) show that frustrated bidders leave eBay if
they fail to obtain an item for which they have been the leading bidder for a long time. However, the behavior observed
in these works is caused by a sudden downward shift in outcomes for a given reference point. The present work shows
evidence for similar behavior, but triggered by the reverse channel: downward shift in reference prices for a given
transaction outcome.
Finally, this works connects to previous research showing that buyers in online auctions often bid more than
what is necessary to obtain the same item via a fixed‐price sale (see Ariely & Simonson, 2003; Malmendier & Lee,
2011). This could be due to overbidding, that is, buyers who know their valuation but bid above it (as it has been
well established for second‐price auctions in the lab; for an overview see Kagel, 1995). Alternatively, the vast
space of online markets (Brynjolfsson et al., 2011) can lead buyers to undersearch so that lower‐priced offers are
not considered.4 In this regard, there is some discussion on whether such behavior ought to be called “over‐
bidding,” as this expression is typically associated only with the behavior described in the first account (see
Malmendier, 2016; Schneider, 2016). Both cases, however, show that bidders act under cognitive limitations,
either in their bidding process or in their perception and use of alternative, cheaper offers. Besides confirming
this notion, this paper also indicates a channel through which such limitations harm not just buyers, but also
sellers and the informativeness of online feedback systems.
3 | DESCRIPTION OF T HE ONLINE SALE
3.1 | Context
In early August 2008, a large German railway company, in cooperation with the German branch of the internet auction
and sales platform eBay (“ebay.de”), conducted a sales campaign for rail tickets. For each day from August 1 to 10, a
special offer became available for 24 h. Of particular relevance for this paper are the offers on August 1 and 3. On these
2 days, the offered item was the same: a voucher for a return trip, second class on all domestic trains (except night
trains) operated by the railway company, which was a de facto monopolist in this segment. After its sale and payment
via money transfer, the paper voucher was shipped by post. To use it, buyers had to fill in their departure and arrival
stations; its specific use was therefore up to the client. At the time of the sale, the railway operated a webshop from
which the regular price (i.e., the price without a voucher) for any given itineraries could be easily retrieved. Buyer
could therefore have different valuations for the voucher, depending on how intended to use it for (the most expensive
eligible itineraries valued at 230€).
While the vouchers sold on August 1 and 3 were the same, the sales mechanism differed between the 2 days. On
August 1, vouchers were sold via auctions. The auction format was eBay's standard incremental auction, which is a
slightly modified, open‐bid, second‐price auction, starting from a price of 1€.5 Buyers could therefore influence the final
price through their bid and cap it in case they won the auction. In contrast, vouchers sold on August 3 were offered for
a take‐it‐or‐leave‐it, fixed price of 66€. Except for this difference in the sale mechanism, all the procedures and
transactions characteristics, for example, payment options, the seller, the sale platform, shipping procedures, and
shipping costs were the same.6
eBay allows and encourages its members to mutually review their transactions. Part of such a review is that
buyers can rate sellers along several dimensions and give an overall feedback rating which is either “negative,”
“neutral,” or “positive” (−1/0/1). The seller's feedback score, which is the sum of hitherto received feedback
scores, is prominently displayed beside a seller's account name. Each seller also has a seller profile which is
publicly accessible. It displays for each review a seller has received the following information for a limited time:4 | GESCHE
the review's overall rating (positive, neutral, or negative), the time when the review was left, the offer to which
the review refers, the username of the buyer who left the review, and a short text comment written by the
reviewing buyer. For each review, the feedback score of the reviewing buyer's account is also displayed next to
this buyer's account name. Note that at the time of the sales campaign, feedback from sellers for buyers—in
contrast to feedback from buyers for sellers—could only be positive or not be given at all, a rule that had been
introduced to prevent reciprocal feedback‐giving (see Bolton et al., 2013).
3.2 | Data description
As a seller's review history on eBay.de was publicly accessible, the reviews and associated data for this sales campaign
could be collected. This paper looks, for reasons which will later become clear, at data covering multiples of 6 days after
the initial offer date. Specifically, I use data for auction reviews collected in the 36 days from August 1. For data
referring to the fixed‐price sale, I use reviews left in the 36 days starting from August 3. Table 1 shows the summary
statistics for this data.
The table shows that the price paid by suction customers was, on average, 13€ higher than the fixed price. It
also displays the reviewing buyer's own feedback score, abbreviated by “Buyer's Score.” This score increases
with each positive rating a buyer has received in a previous transaction.7 Note that eBay members typically give
and receive positive feedback in more than 98% of all transactions (see Bolton et al., 2013; Nosko & Tadelis,
2015). This high share of positive reviews among eBay members allows one to take a reviewing buyer's feedback
as an approximation of this buyer's eBay experience, even though it is actually a lower bound on the number of
prior transactions by the reviewing buyer. As these scores are strongly dispersed, they are put into four cate-
gories based on whether they surpass the thresholds of 10, 100, or 1000, respectively. Although to some degree
arbitrary, one can think of reviewing buyers in the first category to be relatively inexperienced eBay members,
while those in the second (and third) category are (very) experienced members, respectively. Those reviewing
buyers with a feedback score larger than 1000 (and, thus, at least 1001 prior transactions) were likely profes-
sional sellers themselves, who acted as buyers here.
Overall, the reviewing buyers are fairly experienced: For both sales mechanisms, just 11% had a feedback score of
less than 10 and only around 3% of the reviews were written by supposed professionals; the rest lies in between. Finally,
the table displays the feedback received by the seller from the buyers. The share of 96.6% positive reviews for the fixed‐
price sale is slightly lower but not too far away from eBay's 98% share of positive ratings. At a level of 86.4%, this share
is drastically smaller, if the voucher was sold in an auction.8
TABLE 1 Descriptive statistics by sales mechanism
Auction mean (s.d.) Fixed‐price mean (s.d.)
Price 79.09 (13.64) 66.00 (0.00)
Buyer's score (reviewing buyer's feedback score) 207.99 (504.10) 182.79 (628.18)
≤10 11.1 11.3
≤100 (and >10) 43.8 47.0
≤1000 (and >100) 41.8 39.3
>1000 3.3 2.4
Feedback given for the seller 0.79 (0.57) 0.95 (0.20)
+1: positive 86.4 96.6
±0: neutral 5.8 1.7
−1: negative 7.8 1.7
Observations 3575 15,175
Note: Descriptive statistics for the price paid, the reviewing buyer's own feedback score, and the feedback left by the buyer in a review, grouped by reviews
for auctions and fixed‐price sales. A total of 25 (0.7%) of the auction and 131 (0.9%) of the fixed‐price buyers hid their feedback scores; buyer's feedback
score‐statistics omit these observations (see footnote 7).GESCHE | 5
3.3 | First findings
Table 1 shows that auctions have a 10.2‐ppt higher chance of receiving nonpositive (negative or neutral) feedback than
fixed‐price sales. Note that giving positive feedback is the overwhelming norm on eBay and several studies conclude
that any other feedback, including neutral feedback, is considered to be a bad evaluation (see Bolton et al., 2013; Cabral
& Hortaçsu, 2010; Cabral & Li, 2015; Dellarocas & Wood, 2008; Nosko & Tadelis, 2015). Given this and the fact that
except for the sales format, auction and fixed‐price sales were identical, this difference in feedback between the sales
mechanisms is remarkable.
One reason could be the heterogeneity in the reviewing buyers' market experience. For example, socially motivated
behavior—to which giving unconditional feedback belongs—has been shown to be moderated by market participants'
experience and their own reputational stakes (see List, 2003, 2006). Another potential driver for the feedback differ-
ential is that some buyers bought multiple vouchers. This allowed them to issue multiple reviews, which may have
disproportionately affected reviews for one of the sales mechanisms. To examine these possibilities, I regressed an
indicator for nonpositive reviews on a dummy indicating whether a review refers to an auction while controlling for the
reviewing buyer's reputation and market experience via its feedback score. To deal with the case of multiple reviews by
the same buyers affecting feedback disproportionately, this analysis was also repeated using only the first review from
each distinct buyer. The full results are shown in Table B.1 in Supporting Information Appendix B and show a
consistent picture, irrespective of whether controls are added or whether multiple reviews from the same buyer
are used:
Finding 1. There is a feedback differential between the sales mechanisms: Relative to fixed‐price sales' rate of
3.4%, auctions are four times as likely (10 ppt higher; p < 0.001) to receive nonpositive feedback.
This finding could of course be due to remaining unobserved heterogeneity between and selection into the two sales
mechanisms. For example, high‐valuation buyers may have deliberately selected into the auction to preempt a later
fixed‐price market. However, buyers can cap the price which they have to eventually pay—a satisfying rent can always
be ensured. Put differently, any explanation which assumes that participation and bidding in the auction is the result of
deliberate selection and anticipating behavior also suggests that the eventual price reflects such deliberate choices and
should not trigger unfavorable feedback.
This is, however, not what one finds when looking at the effect of the price on reviews with data from the auction
reviews only. As before in the analysis of the feedback differential, this analysis was performed with additional controls
for the reviewing buyer's experience and by repeated only using the first reviews from each reviewing buyer. The main
independent variable is the auction price minus the fixed price of 66€. The estimated coefficient shows that the auction
price affected reviews negatively (full results in Table B.2 in Supporting Information Appendix B):
Finding 2. There is a negative price effect within auctions: For each 10€ paid more than the fixed‐price offer, the
chance to receive nonpositive feedback increases significantly, by about 2.2 ppt (p < 0.001).
While these findings are not compatible with deliberate, rational selection into and bidding within auctions, other
selection‐based mechanisms (such as high valuation customers being more critical) could still explain these results. In
the following, I present a theory and several pieces of evidence for how several ex post reference‐price shifts can
account for these findings.
4 | REFERENCE ‐P RI C E SH I F TS
The above findings raise several questions. Given that the seller and the sold item were identical across the two sales
mechanisms, why is feedback for auction sales so much worse than for fixed‐price offers? Even more puzzling, why do
higher auction prices lead to unfavorable reviews, when bidders could have easily avoided to pay an unsatisfying price?
This section first shows how ex post reference‐price shifts provide an answer. It then provides empirical evidence
which is in line with this channel.6 | GESCHE
4.1 | Sources and consequences of ex post reference‐price shifts
To see how reference‐price effects work in the present context, consider a successful bidder in the sale, that is, a buyer
whose bid exceeded the later‐offered fixed price but not its own valuation for the voucher. In itself, this does not
provide a reason to be unsatisfied because the buyer can still realize a positive surplus. In fact, reference dependence
can even be an additional source of utility here as it captures the extra joy of making a bargain when the reference price
is the higher regular price for the planned itinerary (i.e., the seller's valuation). However, there is numerous evidence
which shows that buyers also consider the prices paid by others as reference points (see Amir et al., 2008; Ariely &
Simonson, 2003; Shafir et al., 1997; Simonsohn & Loewenstein, 2006; Weaver & Frederick, 2012).
If a buyer evaluates a transaction based on nonidiosyncratic features such as prices offered to others, the above
reasoning fails. In this case, observing a fixed price lower than the auction price reverts the previously perceived extra‐
utility from making a bargain. Instead, the difference between the auction price and a successful bidder's reference
price becomes negative once a lower fixed‐price offer is observed. In consequence, what felt like a bargain before is now
evaluated as a loss. To the degree that this experience also affects reviews, successful bidders will then rate auctions
worse than otherwise identical fixed‐price sales, even though the economic value for the successful bidder has not
changed and is still positive. In addition, reference‐price shifts can also explain the negative price effect: The larger the
price a buyer paid in the auction, the larger the distance to the new, decreased reference price and therefore the
experienced loss. In consequence, bidders can become antagonized over the auction price, even if they have ex ante
been satisfied with it and explicitly agreed to pay this price (Supporting Information Appendix C features a formal
model capturing such behavior; it also shows that loss aversion can re‐enforce this mechanism but is not a prerequisite
for it).
In the present context, there are several reasons why such an ex post reference point shift could occur for successful
bidders in the auction. One is that even though the whole sequence of the sales campaign's offers was fixed beforehand,
it was relatively hidden. For example, it was listed as such only on the seller's eBay page but not on the pages displaying
the actual auctions. Another reason is that advertisements for the sales campaign were focused on the respective day's
special offer (and, in the case of web banners, lead directly to the corresponding sales pages on eBay, bypassing any
listing of upcoming offers). While these factors limited ex ante information about the fixed‐price offers, successful
bidders could learn about them ex post, after the auction had ended, through several channels:
First, successful bidders received a confirmation email which listed the obtained item and the final price, together
with further information regarding the payment and shipping procedures. Right below this essential information, the
confirmation email also contained a list of the seller's upcoming sales. For those who obtained the voucher in the
auction on August 1, this confirmation email therefore featured a salient advertisement for the fixed‐price sale fol-
lowing 2 days later. Second, successful bidders might have continued to be affected by the accompanying advertisement
campaign, as it continued to advertise the respective day's special offer, including the advertisement for the fixed‐price
sale on August 3. Third, several news pages started to report on the rather particular sales mechanisms after the auction
had ended. In consequence, successful bidders could learn about these other offers and take them into account as a
new reference price when evaluating the auctions they won.
4.2 | Temporal variations in reference‐prices' relative salience
To test the channel proposed above, a variable which indicates whether a successful bidder's reference price shifted
would be ideal. However, this requires knowledge over what such buyers perceive—which is generally hard to measure,
in particular so for observational field data. As an alternative, a variable which indicates whether reference prices were
relatively more salient than other factors and changes in this variable can serve the same purpose. To obtain such an
indicator, one can exploit the specific time structure of the sales campaign and when information regarding the fixed‐
price offer arrived, relative to other information.
In the first days after winning the auction, when successful bidders could learn about the fixed‐price offer, their
reference prices could change accordingly. In contrast, transaction‐specific information regarding their recent acqui-
sition (e.g., actual experiences during the associated train ride) remained constant for a while. This is because the
shipment of the paper voucher took time and was initiated only after the buyer's money transfer had arrived on the
seller's account. Before the voucher had arrived by post, buyers could therefore not get any new information regarding
quality, seller behavior, or other idiosyncratic transaction features, relative to the information they had on the day ofGESCHE | 7
the transaction. To determine this initial time period during which only reference prices but not the user experience
could change, the review comments for auctions were manually checked for statements indicating that a voucher had
arrived. The first such statement dates on August 7. Reassuringly, this coincides with the timing of the first such
comment in a review for the fixed‐price sale (indicating that the seller had the same after‐sale logistics for these sales of
the same item). This suggests that in the first 6 days from the day of the auction date (August 1), no voucher and
without it, no new transaction‐specific information reached successful bidders whereas information about the fixed‐
price sale could reach them.
Figure 1 provides further evidence for this conclusion. It displays the temporal pattern of when reviews for the
auction were left. In the first 6 days (labeled as days 0–5), relatively few reviews occur: less than 2% daily, in total 5.2%,
of the 3575 auction reviews. On August 7, the date with the first comments clearly indicating the arrival of the voucher,
the daily rate suddenly spikes to almost 15% and remains high in the following days. This is consistent with the notion
that after the initial 6 days, new information regarding the buyers' idiosyncratic transaction experience—in addition to
information regarding the fixed price—could affect feedback and triggered further reviews. As this information was
absent during the initial 6 days, reference‐price shifts were relatively more salient in this period.
To see how reference point shifts vary by time, all comments from reviews for auction sales were manually
reviewed and marked if they refer explicitly to the seller's pricing strategy (i.e., first selling via an auction and then via a
fixed price). To verify that this manual procedure captured the right comments, I computed term frequency‐inverse
document frequencies (TFIDFs) for the single terms/words in this corpus of pricing‐related comments. This procedure
is commonly used to find keywords in a corpus of text by computing the frequency of each word (or term, TF—term
frequency) and then down‐weighting it by the logarithm of the fraction of comments containing the word (IDF—
inverse document frequency) to adjust for the fact that some words appear often in general. Reassuringly, the top five
keywords according to TFIDF scores for pricing related includes references to the fixed‐price offer (“buy‐it‐now price”)
and expressions of dissatisfaction (“annoying”). In contrast, applying the same method to comments for auction
reviews that were not marked as pricing‐related yields different, more positive key terms (e.g., “great” or “repeat”)
appear. Table A1 in Appendix A contains the complete results for the TFIDF‐based key terms and further details of the
text analysis.
After having confirmed that comments marked as pricing‐related do indeed reflect a higher salience of the com-
peting offer, we now turn back to when they occur. This will be crucial in identifying the relative salience of reference
prices and its influence on reviews by means of hard, objective information from our data set: the timestamp of reviews.
Figure 2 illustrates this timing by displaying the share of reviews with pricing‐related comments among all auction
reviews left on a given day. On the day of the auction and the day thereafter (days 0 and 1), no such comments are
observed. Then, on the second day after the auction, when the fixed‐price sale took place (day 2), the share of daily
FIGURE 1 Timing of auction reviews. Notes: Histogram of when auction reviews were left; bins of six consecutive days separated by
vertical lines8 | GESCHE
F I G U R E 2 Timing of pricing‐related comments in auctions. Notes: Black triangles = Share of auction reviews on a given day in which
the comment refers to the seller's pricing strategy. Gray rectangles and horizontal lines = 95% confidence interval and point estimates of
coefficients indicating the probability that such a comment occurs within each bin of 6 days
reviews which contain such comments jumps to more than 36% and stays high for the next 3 days (days 3–5). A week
after the auction, on day 6, when the vouchers started to arrive by post, the share of such comments falls and stays
relatively low. This corresponds exactly to the proposed pattern of how the salience of reference prices, relative to other
information, behaves over time and how this affects buyers' perceptions and their corresponding feedback.
This visual impression can also be confirmed statistically by regressing a dummy indicating whether a comment is
pricing‐related on a constant and five dummies, allowing to identify each of the 6‐day‐bins covering the 36 days. The
conditional means obtained from these estimates are shown as horizontal, black lines in Figure 2; the gray rectangles
indicate the associated 95% confidence intervals around these conditional means. They show that the share of pricing‐
related comments, which is around 23% during the first 6‐day‐bin, is significantly lower, by 14–20 ppt, for each of the
following five 6‐day‐bins (p < 0.001, full regression results in Table B.3 in Supporting Information Appendix B).
4.3 | Temporal decomposition of the negative price effect
The above findings are consistent with the notion that for the first 6 days, the news of the fixed‐price offer and the
associated reference‐price shift was a stronger determinant of reviews than that in the following days. They are also
consistent with typical models of salience, such as Bordalo et al. (2013): In the first 6‐day‐bin, the salience of price is
relatively high as, upon learning about the fixed‐price offer, it is incorporated in a bidder's price norm, which then
decreases and makes a higher price paid in the auction stick out more. In contrast, the salience of quality or user
experience is low in this period as it corresponds to the reference level (i.e., the expectations buyers had at the time of
the transaction). Once the voucher arrives, new information in this dimension can start to enter buyers' perceptions
increasing the salience of quality or user experience so that prices become less salient.
It then follows that the negative effect of the auction price on feedback should—if caused by a shift in the reference
price—be more pronounced in these first 6 days when the relative salience of these features was high as compared
with later periods. To test this, the following Probit model was fitted with the auction sale data. It allows one to
measure the price effect separately for the initial 6 days and the sample's five remaining 6‐day‐bins:GESCHE | 9
Pr [ fi ≤ 0 x i] = Φ α + β1 ⋅ Δ10 Pricei + βt ⋅ Δ10 Pricei × SixDayBin#ti
6
+ γt ⋅ SixDayBin#ti + δs ⋅ 1[
t =2
10 s].
6 3 (1)
1 Buyer ′s Scorei >
t =2 s =1
In the above, the dependent variable is whether the review for a transaction entails nonpositive feedback (i.e., if
feedback is negative or neutral: fi ∈ {−1, 0} ). The function Φ captures the normal distribution's CDF for the error term,
x i summarizes the independent variables for transaction i . The main independent variable is Δ10 Pricei , the difference in
the auction price to the 66€ fixed price, divided by 10. Additional independent variables control for whether the
reviewing buyer's feedback score Buyer ′sScorei surpasses a power‐of‐10 threshold but not the next highest. If these
were the only variables, this regression setup would be the one which leads to Finding 2, the 2.2‐ppt‐increase in
nonpositive feedback for each 10€ paid about the fixed price (measured via the estimate for β1 ).
But the regression model also includes five dummies denoted by SixDayBin#ti and their interaction with Δ10 Pricei .
Since SixDayBin#ti equals one only if a review is left in the tth bin of six consecutive days, starting from the auction
date, this interaction allows one to decompose the price effect over the sample's duration. Because the first 6‐day‐bin is
the reference category, the coefficient β1 now measures the price slope in this period only. The coefficients on the terms
which interact the price with one of the 6‐day‐bin‐indicators (β2 –β6 ) therefore capture the differences between the first
period's price slope and those of later 6‐day‐bins. They are the main variables of interest for testing the following
prediction: “As shown above, reference prices are relatively more salient in the first 6‐day‐bin than those in the latter
bins. Accordingly, the price slope during the first 6 days, estimated by β̂1 , should—if triggered by a reference‐price
effect—have a larger value than the slope estimates βˆ1 + βˆt for the later 6‐day‐bins. Thus, the estimates for βt with
t ∈ {2, 3, 4, 5, 6} should be negative.”
Note that the above reasoning does also not predict a gradually decreasing price effect over time. It rather stipulates
a sharp decline in the price‐effect's magnitude after 6 days and no further change thereafter. Furthermore, note that
these predictions do not rely on any manually coded variable such as the dummy indicating pricing‐related comments.
While this variable was used to figure out when reference prices were particularly salient during the first 6 days, any
tests of the resulting hypotheses are based on “hard” data from eBay's database (i.e., the auction's final price and a
review's timestamp). Thus, any findings which emerge despite such imprecise measurement should only become more
pronounced if more precise tools were available.
With this in mind, one can look at Table 2. It shows the marginal effects from estimating model (1). The non-
interacted price coefficient shows that in the first 6‐day‐bin, each 10€ paid more correspond to an increase of around
12 ppt in the probability of nonpositive feedback. After this period, the price effect decreases significantly and sharply
for all following periods, as indicated by the significant and negative interaction terms. In fact, this drop is almost as
large as the estimate for the first 6‐day‐bin. In consequence, the price slopes of later 6‐day‐bins, given by the sum of the
coefficient estimates for Δ10 Price and Δ10 Price × SixDayBin#t (i.e., βˆ1 + βˆt ) are never statistically different from zero
(except for βˆ1 + βˆ6 in column 4, with p = 0.086). This pattern corresponds to the prediction of a negative price effect in
the initial period, when reference prices were most salient, and its subsequent drop.9
These results are hard to align with several alternative explanations. For example, they are incompatible with a
theory of unobserved selection in which high‐valuation auction customers are also more price‐sensitive and/or more
critical. While such selection could explain the simple negative price effect of Finding 2, the selection of such customers
into auctions should then affect their stronger reaction to higher prices paid throughout the whole sample, not just
during the initial 6 days. A similar channel in which higher‐bidding (and therefore, on average higher‐paying) cus-
tomers are particularly harsh is also not supported by these results: such a channel would also predict a negative price
effect that is independent of the review's timing or, at least, continuously fades out but does not sharply drop after the
initial 6‐day‐bin. Furthermore, these findings do not imply that reference point shifts disappear after the first 6‐day‐bin.
In contrast to this initial period, however, their negative effects can be countervailed by (positive) experiences gathered
after the good has arrived and could be used. Then, the average price effect on negative feedback can become smaller or
zero in these later periods.
To produce further evidence on how the exact timing of reviews is connected to how reviewing buyers feel, an
additional text analysis was conducted. It is similar to the one for the pricing‐related comments but based on the
comments' timing. Applying TFIDF to the corpus of auction comments left in the first 6‐day‐bin identifies negative top10 | GESCHE
TABLE 2 Price effects for auctions reviews decomposed over 6‐day‐bins
yi = 1: Negative or neutral feedback
Dependent variable (1) (2) (3) (4)
Δ10 Price 0.125*** 0.125*** 0.116*** 0.116***
(0.024) (0.024) (0.027) (0.027)
Δ10 Price × SixDayBin#2 −0.117*** −0.116*** −0.108*** −0.108***
(0.025) (0.025) (0.028) (0.027)
Δ10 Price × SixDayBin#3 −0.117*** −0.119*** −0.110*** −0.109***
(0.025) (0.026) (0.029) (0.029)
Δ10 Price × SixDayBin#4 −0.105*** −0.105*** −0.096*** −0.096***
(0.031) (0.031) (0.032) (0.035)
Δ10 Price × SixDayBin#5 −0.130*** −0.129*** −0.152*** −0.151***
(0.029) (0.029) (0.035) (0.035)
Δ10 Price × SixDayBin#6 −0.092*** −0.098*** −0.076*** −0.081**
(0.035) (0.035) (0.023) (0.036)
SixDayBin#2 −0.107** −0.105* −0.131** −0.126**
(0.054) (0.054) (0.058) (0.057)
SixDayBin#3 −0.118** −0.116** −0.129** −0.128**
(0.056) (0.056) (0.060) (0.059)
SixDayBin#4 −0.078 −0.075 −0.076 −0.072
(0.062) (0.061) (0.065) (0.065)
SixDayBin#5 −0.129** −0.128** −0.092 −0.090
(0.061) (0.061) (0.072) (0.071)
SixDayBin#6 −0.075 −0.069 −0.136** −0.125*
(0.073) (0.074) (0.066) (0.066)
Buyer's score > 10 −0.031 −0.031
(0.028) (0.024)
Buyer's score > 100 −0.047 −0.039
(0.028) (0.024)
Buyer's score > 1000 0.002 0.028
(0.051) (0.050)
First reviews only No No Yes Yes
Observations 3575 3550 2283 2265
Note: Average marginal effects of Probit estimates obtained from regressing a dummy for nonpositive feedback on the difference between the auction price and
the fixed price divided by 10, a dummy for the 6‐day‐bin since the auction data, its interaction with the price variable, and dummies for the reviewing buyers'
own feedback score. The first row reports the estimated price slope for the first 6‐day‐bin during which reference prices were particularly salient. The next five
rows present the difference of that slope to the respective price slopes estimated for the subsequent 6‐day‐bins (which are all not significantly different from
zero). Columns (1) and (2) use all reviews and report standard errors clustered on the buyer level. Columns (3) and (4) only use the first review each buyer
posted and report robust standard errors. Significance levels: ***p < 0.01, **p < 0.05, and *p < 0.1.
key terms, such as “unnecessary,” “bad,” or “revoke.” In contrast, auction comments left during the following 6‐day‐
bins feature more positive key terms such as “great” or the desire to repeat the sale (“to repeat” and “more often”).
These differences are also reflected in different sentiment scores. Such scores are computed for each comment using a
pretrained dictionary. They yield a score on the [−1, 1] ‐interval with lower values indicating more negative sentiments.
Auction comments from the first 6‐day‐bin have, depending on the specification, average scores between 0.160 andGESCHE | 11
0.188. This is significantly less than the corresponding sentiment scores for the later 6‐day‐bins, which are more
positive and have average values between 0.313 and 0.315 (t tests: p ≤ 0.001; more details in Appendix A and Table A2
therein). Thus, the negative price effect and its temporal concentration in the first 6‐day‐bin are also aligned with more
negative sentiments in early feedback comments and a significant improvement thereafter.
4.4 | Temporal decomposition of the feedback differential
The above results show how the negative effect of the auction price on reviews is concentrated in the initial 6‐day‐bin
after the auction, when reference‐price shifts were most salient. In the following, I show that an explanation based on
reference‐price shifts is also able to organize the feedback differential between auction and fixed‐price sales.
I do so by estimating a model similar to (1) but now using the combined data from reviews for the auctions and the
fixed‐price sale. The other difference is that instead of the Δ10 Pricei ‐variable, an Auctioni ‐dummy is included. It
indicates whether review i refers to an auction and therefore measures the feedback differential between the sales
mechanisms. This dummy is interacted with indicators for the 6‐day‐bins, which allows one to measure the feedback
differential separately for each bin:10
Pr [ fi ≤ 0x i] = Φ α + β1 ⋅ Auctioni + βt ⋅ Auctioni × SixDayBin#ti
6
+ γt ⋅ SixDayBin#ti +
t =2
1 Buyer ′sScorei > 10 s].
6 3 (2)
δs ⋅ 1[
t =2 s =1
Table 3 presents the estimates. They reflect the same temporal pattern observed in the preceding analysis of
the price effect. In the first 6‐day‐bin, when reference‐price shifts were particularly salient and the negative
price effect within auctions was pronounced, the feedback differential between the sales mechanisms is also
particularly strong. During this period, the rate of nonpositive feedback in auctions is about 41 ppt higher than
that for fixed‐price sales (where during the first 6‐day‐bin 3% of transactions receives unfavorable feedback).
This difference decreases significantly, by about 27–39 ppt (i.e., to a quarter or less of its initial size) during later
6‐day‐bins.
These estimates imply that in the later 6‐day‐bins, auctions are still significantly more likely—between 2 and 13 ppt—
to receive nonpositive feedback (sum of estimates βˆ1 + βˆt , significant at p ≤ 0.036 except for t = 5). Thus, these results do
not rule out that another channel besides reference‐price shifts (e.g., customers generally disliking auctions), could have
contributed to the worse reviews for auctions. However, while such a channel might drive the 2–13 ppt baseline
differences, it does not explain why this difference quadruples to 40 ppt in the first 6‐day‐bin. In contrast, reference prices
and the temporal variations in their relative salience can account for this shift. Furthermore, these differences in the later
6‐day‐bins could also be due to reference‐price shifts: even though the relative salience of reference prices and their
shifts is not as high in later periods as in the first 6 days, this does not mean that they cease to affect reviews in the later
periods – reference price shifts are just not as isolated and strong as in the first 6‐day‐bin.
Note that these findings are inconsistent with the notion of auction customers being generally harsher reviewers
than other customers. In this case, one would not expect to see the feedback differential being concentrated in the first
6‐day‐bin. Also note that the estimates for the noninteracted 6‐day‐bin dummies, which account for temporal varia-
tions in reviews for fixed‐price sales (the baseline category) do not indicate this temporal concentration. This provides
further evidence that the variations in auction reviews are not due to some general, unrelated‐to‐reference‐point‐shifts,
temporal variation in reviews.
4.5 | Timing of feedback differences and comments
Further evidence for the link between reference‐price shifts and feedback differentials comes from analyzing the
comments which buyers wrote in their feedback. Figure 3 plots the difference in the daily rates of nonpositive feedback
between auctions and fixed‐price sales over the daily rates of pricing‐related comments in auction reviews (i.e., the12 | GESCHE
TABLE 3 Feedback differences between auction and fixed‐price sales decomposed over 6‐day‐bins
yi = 1 : Negative or neutral feedback
Dependent variable (1) (2) (3) (4)
Auction 0.410*** 0.405*** 0.410*** 0.403***
(0.046) (0.046) (0.043) (0.043)
Auction × SixDayBin#2 −0.315*** −0.310*** −0.326*** −0.319***
(0.047) (0.047) (0.044) (0.044)
Auction × SixDayBin#3 −0.337*** −0.334*** −0.335*** −0.331***
(0.048) (0.048) (0.046) (0.046)
Auction × SixDayBin#4 −0.293*** −0.286*** −0.276*** −0.268 ***
(0.053) (0.053) (0.050) (0.050)
Auction × SixDayBin#5 −0.391*** −0.386*** −0.375*** −0.369***
(0.051) (0.051) (0.050) (0.050)
Auction × SixDayBin#6 −0.288*** −0.290*** −0.327*** −0.325***
(0.064) (0.064) (0.055) (0.055)
SixDayBin#2 −0.013*** −0.013*** −0.010** −0.010**
(0.004) (0.004) (0.004) (0.004)
SixDayBin#3 0.001 −0.001 0.000 0.000
(0.005) (0.005) (0.005) (0.005)
SixDayBin#4 0.011* 0.011* 0.013** 0.013**
(0.006) (0.006) (0.006) (0.006)
SixDayBin#5 0.025*** 0.025*** 0.026*** 0.025
(0.008) (0.008) (0.008) (0.008)
SixDayBin#6 0.025** 0.026*** 0.026*** 0.027***
(0.010) (0.010) (0.009) (0.009)
Buyer's score > 10 −0.017** −0.014**
(0.007) (0.006)
Buyer's score > 100 −0.014** −0.010*
(0.007) (0.006)
Buyer's score > 1000 −0.004 0.004
(0.014) (0.013)
First reviews only No No Yes Yes
Observations 18,750 18,594 15,857 15,721
Note: Average marginal effects of Probit estimates obtained from regressing a dummy for nonpositive feedback on the difference between the auction price and
the fixed price divided by 10, a dummy for the 6‐day‐bin since the auction data, its interaction with the price variable, and dummies for the reviewing buyers'
own feedback score. The first row reports the estimated difference in the rate of nonpositive feedback between auctions and fixed‐price sales for the first 6‐day‐
bin during which reference prices were particularly salient. The next five rows show how this initial auction‐effect differs from those estimated for each of the
respective subsequent 6‐day‐bins. Columns (1) and (2) use all reviews and report standard errors clustered on the buyer level. Columns (3) and (4) only use the
first review each buyer posted and report robust standard errors. Significance levels: ***p < 0.01, **p < 0.05, and *p < 0.1.
same rate of pricing‐related comments as depicted in Figure 2). The units of observation are the 34 days starting from
August 3 in which feedback for both, the auction and the fixed‐price sale, could be left.
If every pricing‐related comment resulted in a nonpositive feedback for the auction and vice versa, then these two
measures would covary perfectly. In reality, such a clear signal is, of course, absent. For example, it could be that for
every pricing‐related comment, there are multiple buyers whose reviews are also negatively affected by reference pricesGESCHE | 13
F I G U R E 3 Feedback differences between auction and fixed‐price sales over pricing‐related comments. Notes: Numbers = Data points,
the number indicates days since the auction sale. *Vertical axis = Difference in the daily rate of nonpositive feedback between auctions and
fixed‐price sales (for each of the samples 34 days with feedback data for both sales mechanisms. Horizontal axis = Daily rate of pricing‐
related feedback for auctions. Line = Result of an OLS regression of feedback differences on pricing‐related comments and a constant. (The
data points “10” and “17” are slightly shifted for illustrative purposes; their actual position would almost overlap with the data point
“11”.) OLS, ordinary least squares
but only some of them are explicit about it (reflecting the proxy‐nature of the comments). Despite these obstacles, the
data show that the two daily rates are indeed strongly correlated ( ρ = 0.891, p < 0.001).
These findings are also reflected in the results from regressing the differences in the daily rate of nonpositive
feedback between the two sales mechanisms on the daily rates of pricing‐related comments and a constant. The
resulting regression line, obtained by ordinary least squares (OLS) and depicted in Figure 3, has a significantly positive
slope: each ppt‐increase in the daily share of pricing‐related comments increases the difference between that day's
share of nonpositive feedback between auctions and fixed‐price sales by 1.284 ppt (on average). The associated R2 of
0.795 implies that about 80% of the feedback differential between auctions and fixed‐price sales can be explained by
comments related specifically to the seller's pricing decision.11 Given eBay's generally high level of feedback and the
pronounced negative consequences of negative feedback (see footnote 8, these results also show that even a small
number of antagonized auction customers can have a relatively large impact on reputation and, therefore, sellers'
businesses.
Note that Figure 2 displays the single data points as numbers, which correspond to the days having passed since the
auction sale. Consistent with the preceding findings, the observations with the largest share of pricing‐related com-
ments and, therefore, the strongest effect on the feedback differential (“2,” “3,” “4,” and “5”) are the ones in the first 6‐
day‐bin, when reference prices were most salient. Notably, the positive relationship between the daily rates of pricing‐
related comments and the feedback differentials persists if one excludes the observations from the initial 6‐day‐bin out
(slope 1.009 with p < 0.001, R2 = 0.452). This shows that even after the initial days, pricing can have a negative impact
on auction reviews, even though this is, unsurprisingly, less pronounced than when those days with a high salience of
the fixed‐price offer are included. Together, these findings suggest that the seller's pricing decision (i.e., the use of a
fixed‐price sale after the auction)—and not just the effect of a higher price per se—triggered the observed differences in
feedback between the two sales mechanisms.1214 | GESCHE
4.6 | Discussion
The preceding findings are in line with a behavioral mechanism in which winning an auction and then observing a
lower‐priced fixed‐price sale creates disutility through an ex post reference‐price shift. This disutility then causes
customer antagonism and materializes in negative feedback for the seller.
With regard to alternative explanations, note that these findings do not depend on whether multiple reviews per
buyer are used. This speaks against reselling motives by bidders and common‐value characteristics of the auction being
relevant here. In the absence of such motives and characteristics, the winner's curse (Kagel & Levin, 1986; Thaler,
1988) cannot be the cause of such antagonism. Furthermore, the temporal variations of within one sales mechanism
(the price effect in auctions, see Table 2) mirror the temporal variations between two sales mechanisms (the feedback
differential between the auction and the fixed‐price sale, see Table 3). This fact is inconsistent with some sort of
customer selection along an unobserved variable driving the results. If this were the case, the temporal pattern
organizing the differences between sales mechanisms—caused by such unobserved selection—should not be the same
pattern that organizes the price effect within one sales mechanism (the auction) because, at this stage, selection has
already taken place.
The temporal pattern of the review does also not indicate a gradually decreasing effect over time. It rather shows
that right after the first 6 days, both the feedback differential and the price effect sharply decrease and then stay roughly
constant at a low level over the sample's remaining days. This observation is inconsistent with an alternative ex-
planation based on the notion that negative emotions often “cool off” over time (see Bosman et al., 2001; Lee, 2013;
Oechssler et al., 2017). Such a channel would predict a gradually decreasing price slope and feedback differentials. This
is, however, not observed. Furthermore, the concentration of the price effect in the first 6‐day‐bin and its sharp decline
afterward is inconsistent with the notion that having paid a higher price does in itself trigger unfavorable reviews.
Apart from the fact that buyers could have easily prevented paying an unsatisfying price by bidding accordingly, there
is no reason why such a pure price effect should apply only in the first 6 days but not in later periods.
5 | A SURVEY E XPERIMENT
The above findings are consistent with ex post reference‐price shifts affecting buyer feedback in auctions. To directly
test this, I conducted an additional survey experiment. It provides exogenous variation in the reference prices, in-
cluding the case when the new price is higher than the one which customers had previously paid. This allows one to
causally test for an asymmetric, stronger effect of a reference‐price shift downward as opposed to an upward shift, as
one would expect from loss aversion (because the former leads to a perceived loss while the latter induces a perceived
gain, respectively).13
For the survey experiment, a market research firm recruited an online panel of 298 US‐based customers, re-
presentatives across ages and gender. The survey then asked these subjects to imagine a situation in which they had
acquired, via an online auction platform, a voucher for $150, entitling them to a two‐night stay at a hotel chain. The
scenario then described that after they received the voucher code by email, they also received an email asking them to
leave feedback for the seller (called Travelmaster). Subjects then got a description of the feedback stage, in which they
were supposed to leave feedback for the seller. They were informed that the average feedback score is used to inform
other customers how good and trustworthy a seller is and that it is displayed next to a seller's name (see Supporting
Information Appendix D for the full description of the scenario). Specifically, subjects were asked to provide a feedback
score between 0 and 100 as a response to the following question: “Which rating would you give Travelmaster for the
$150 auction transaction with the voucher?” This formulation is similar to eBay's question, which asks customers to
“Rate this purchase” (rather than the seller or its pricing policies). To avoid ceiling effects and engage subjects in
deliberation over an appropriate rating, the scenario also stated that the voucher arrived somewhat late.14
While the above was the same in all treatments, the experiment varied information on an advertisement on the
feedback site. This advertisement was described as being for a sale of the same voucher. The advertised sale was fixed‐
price sale and carried out by a different company (called Journey Expert). Experimental treatments then varied,
between subjects, the reference price via the other sale's fixed price. Specifically, the fixed price was either higher ($225,
treatment GAIN), lower ($75, treatment LOSS), or the same ($150, baseline) as the auction price, thereby leading to an
implied gain, loss, or no change in the retrospective evaluation of the auction.15 Given the preceding results, one would
expect that observing a lower price ex post leads a drop in feedback. Conversely, a higher reference price should lead toYou can also read
