Reputation offsets judgments based on social biases among Airbnb users

Bruno Abrahaoa,1, Paolo Parigib, Alok Guptac, and Karen S. Cooka,1

aDepartment of Sociology, Stanford University, Stanford, CA 94305; bInstitute for Research in the Social , Stanford University, Stanford, CA 94305; and cData Division, Airbnb, San Francisco, CA 94103

Contributed by Karen S. Cook, July 20, 2017 (sent for review March 15, 2016; reviewed by David Lazer and Chris Snijders) To provide social exchange on a global level, sharing-economy Social biases are a result of a number of mechanisms that leverage interpersonal trust between their members are difficult to measure. In this work, we make social biases on a scale unimaginable even a few years ago. A challenge to amenable to investigation by focusing on a form of social bias this mission is the presence of social biases among a large hetero- that naturally maps into a quantifiable interpretation and that geneous and independent population of users, a factor that hin- we expect to be at work in these environments. At the same time, ders the growth of these services. We investigate whether and this source of bias is well understood in the social sciences so that to what extent a sharing-economy platform can design artificially we can rely on previous literature, instead of opening up a new engineered features, such as reputation systems, to override peo- dimension of complexity. To this end, we focus on homophily (2– ple’s natural tendency to base judgments of trustworthiness on 6), the higher likelihood that people trust others who are similar social biases. We focus on the common tendency to trust oth- to themselves. ers who are similar (i.e., homophily) as a source of bias. We test McPherson (4) proposed a theory of how homophily struc- this argument through an online experiment with 8,906 users of tures modern societies using a construct of social space defined in Airbnb, a leading hospitality in the . Blau’s theory of preferences (6). Each individual occupies a posi- The experiment is based on an interpersonal investment game, tion in the social space whose coordinates are a function of his in which we vary the characteristics of recipients to study trust or her sociodemographic characteristics. The more features two through the interplay between homophily and reputation. Our individuals share in common, the more likely they are to form findings show that reputation systems can significantly increase relationships based on mutual trust. the trust between dissimilar users and that aversion has an To operationalize homophily in a structured way, we use inverse relationship with trust given high reputation. We also Blau’s construct of social space to induce and measure the effect present evidence that our experimental findings are confirmed by of homophily in an experimental setting whose volunteers are analyses of 1 million actual hospitality interactions among users active members of the sharing economy. (At the time of writing, of Airbnb. the online experiment is accepting participants for demonstra- tion purposes at stanfordexchange.org.) online trust | reputation systems | sharing economy | social biases | risk Building on this baseline, the heart of our experiment is the measurement of the extent to which another source of informa- tion that can be artificially engineered could potentially alter the new wave of companies, emerging under the banner of Athe sharing economy (1), is profoundly altering the way we interact and exchange with one another. These Internet-based Significance services are driving a major change in our cultural and technolog- ical landscapes and have achieved astounding success, enabling We investigate the extent to which artificial features engi- users to share their own personal resources, such as their vehi- neered by sharing-economy platforms, such as reputation sys- cles, real estate properties, time, or skills. A growing number of tems, can be used to override people’s tendency to base individuals trust the sharing economy with a variety of services to judgments of trustworthiness on social biases, such as to satisfy their needs, to generate income, or, more simply, to meet trust others who are similar (i.e., homophily). To this end, new people. Examples of sharing-economy transactions include we engaged 8,906 users of Airbnb as volunteers in an online hiring a “tasker” from Task Rabbit to run errands, sharing a experiment. We demonstrate that homophily based on sev- “couch” with a perfect stranger through CouchSurfing, hiring a eral demographic characteristics is a relatively weak driver of “driver” on Uber, or staying in someone’s home while traveling trust. In fact, having high reputation is enough to counteract using Airbnb. homophily. Using Airbnb data, we present evidence that the Users in the sharing economy seek to connect with others effects we found experimentally are at work in the actual plat- engaged in activities on the same platform. Compared with form. Lastly, we found an inverse relationship between risk exchanges via traditional e-commerce companies, where trans- aversion and trust in those with positive reputations. actions are relatively anonymous, the sharing economy exposes

us to the more personal character of such interactions. This Author contributions: B.A., P.P., and K.S.C. designed research; B.A., P.P., and A.G. per- inevitably prompts attention to the users’ sociodemographic formed research; B.A. and P.P. contributed new reagents/analytic tools; B.A., P.P., A.G., characteristics as factors that drive selection. and K.S.C. analyzed data; and B.A., P.P., and K.S.C. wrote the paper. As a consequence, social biases figure as major hurdles to the Reviewers: D.L., Northeastern University; and C.S., Eindhoven University. growth of sharing-economy services, as they influence users’ per- Conflict of interest statement: A.G. is a data scientist at Airbnb who performed the exper- ceptions of trust and risk. To enable trust between strangers so iments that rely on the company’s private data. Paolo Parigi began working at Uber after that everyone can exchange with anyone, beyond cultural and the research design, experiment execution, data analysis, and writing of the study were social boundaries, these companies face daunting obstacles in completed. their attempts to minimize these biases. Freely available online through the PNAS open access option. In this study, we investigate whether and to what extent a Data deposition: The data necessary to replicate our experimental results are available sharing-economy platform can design technological features to through the Stanford Exchange Project at stanfordexchange.org/project. counteract natural behavioral tendencies that may lead to social 1To whom correspondence may be addressed. Email: [email protected] or biases. This question is of central importance in the social sci- [email protected]. ences more broadly, but also in the engineering of platforms that This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10. aim to enable trust. 1073/pnas.1604234114/-/DCSupplemental.

9848–9853 | PNAS | September 12, 2017 | vol. 114 | no. 37 www.pnas.org/cgi/doi/10.1073/pnas.1604234114 Downloaded by guest on September 25, 2021 Downloaded by guest on September 25, 2021 iiatadrtr odprino h rdt netdi them in invested par- credits the the of with portion prisoner’s cooperate good the a could of return receiver and spirit ticipant the the In (18), receiver. game that dilemma to given and three by by receivers the with rewards. interaction potential single-shot offering invest- a favorable in seek outcomes to participants ment for incentives results. gave the search then via receivers’ way We partners potential the the with users resembling displayed presents features, platform We reputation with credits. receivers and previous of demographics accumulate a the selected in to playing information randomly were incentives users their other other entered These to had round. profiles, belonging who five users as them Airbnb showed presented We we credits. of which number a with started widely the of 17). variation (16, a game designed investment we inter- studied end, measure this to To methods trust. standard personal pre- use economics to with To behavioral psychology on social comparability research transactions. and from for drew one-time allowing we literature, while in vious scenario the result this in usually mimic relationships However, economy time. sharing over exchanges multiple in (15). factors confounding that other decisions isolating on data while behavioral trust, collecting involve plat- of Airbnb’s aim to the external with experiment form, an in participate to unteers suc- in most listings world’s with the companies, of sharing-economy one cessful Airbnb, with collaboration in ratory unnec- and process capture selection to the avail- difficult complex. the making essarily among is without pair alternatives it every able between but preferences thought selections, their user’s the quantify their of by outcome users the process presents observe the platform we that the alternatives, data with when structured through Moreover, to photos displays. opposed through platform as indirectly signals, and other subjectively or heterogeneous, them isolate, highly infer to are users difficult factors and These are such quantify. etc., Features, or attractiveness, trust. categorize, with preference, confounded color are of as features that to alternatives users manip- exposes direct the platform to The amenable researchers. not by is ulation that challenge represents research platform major sharing-economy a a on directly reputation or to arrangements decisions. institutional their users preexisting inform while on standardized, services be directly of to rely pool diverse particu- the cannot and large as of too well is is as offer, case population, they user latter the The as (12–14). interest, lar economy sharing 11), (10, the crowdsourcing and (10), markets in platforms online interplay reputation. the and study by homophily This structured between space. trust social interpersonal in measures dissimilarity directly those of between degrees rep- trust which increase different to to with Moreover, capacity extent the opportunistic the (9). have of systems trust of quantification utation increased limited safeguards been in has result as there would serve which systems behavior, reputation that uncertainty. paralyzing distant generate and not unknown with does interacting alters the We the in that experience systems individual. and direct reputation of these quality that lack of the with because is of transactions it that indicator potential hypothesize an in is entailed person risk a with associated “rate” to and (7). review members to other users of allow system behavior to reputation the use the platforms on which focus 8), this we (7, counteract end, this and To homophily tendency. by natural structured trust of perception baa tal. et Abrahao h rdt atcpnsivse napol eemultiplied were profile a in invested participants credits The and of role the played participants experiment, the In engage and trust develop may sellers and buyers Traditionally, labo- online large-scale a designed we challenges, such to Due distance social of influence the of function a as trust Measuring in systems reputation studied extensively have Researchers evidence weak only presents research previous Nonetheless, rating aggregate the that is systems reputation of premise The 9 onre.W nae ,0 ibbuesa vol- as users Airbnb 8,906 engaged We countries. >190 ilo hospitality million >2 h nvraedsrbtoso h eorpisoe h partic- the over and demographics ipants, the of distributions participation). univariate entire the participate their to completed registered US and (6,714 as responded identified 8,906 who which users of Airbnb residents, 100,000 to invitations sent We Design as Experimental increases higher. interaction gets partners hospitality the select a of who have reputation users the to of whom diversity the with that others inter- evidence real-world show million guide 1 We to containing reputations. actions. data experiment high Airbnb the of of through analysis case gained our the insights dissimilar used in we trust to Lastly, relationship and trust inverse aversion strong of a risk show between extension we toward Third, the space. tendency social in in the users results counterbalance This that construc- homophily. expectations the possible of makes Second, system tion decisions. reputation the trust that make show we participants when extent nificant reputa- facet, and demographic each experimentally. varied features of we tion importance par- as how relative reputation, understand or the to homophily and was trust goal assign Our ticipants generated advance. actually in we profiles experiment, these the in participating also users the in placed they trust credits. much return how to for receiver receivers proxy each indirect in loses an investment as position participant of served vulnerable the amount a case the in investments, which themselves through (in put keep credits participants and these As defect credits). of or all credits) her or increase most participant the help (to rsrbdrls epae h eevr’polsa oildis- social at profiles receivers’ the tance placed We rules. prescribed in accumulated win- credits of of game. chances number investment The the the (USD). to dollars proportional US were 100 offered ning for we judgment, each good prizes, of 100 and exertion attention the capture for incentives To judgment provide their trust. of involving use decisions users best the engage making make to when to was attempt experiment would they online that while our so of safety, chal- design major the their a in Thus, as lenge cost. such minimizing eliminate to and factors, to pressed satisfaction imperative critical maximizing are is with users it associated as platform, decisions, the possible In best factors trust. the for make to except platform, external the are on that decisions make users way the attrition. participation reducing con- while Blau’s their experimentally operationalize for struct conveniently features to these us chose of allowing We pieces simplicity, state. four home and with status, us marital gender, providing age, demographics—namely, by their profile reflecting information short a built ticipant aig n h ubro eiw nArn.Tesa rating star The Airbnb. on star reviews of of number number the average and the ratings features—namely, reputation two ran- in screen order. participant’s from dom the on distances category simultaneously profiles at five different placed the were a receivers in The d participant. features the demographic of those the of Lastly, all category. different changed feature a selected to randomly additional par- one having the by of pant at that profile from The differed ticipant. category feature’s turn, selected In domly marital state). same US the same gender, the same and the status, group, partici- age the same of the attributes (e.g., demographic pant the of individu- all matched of two receiver definition distance which mathematical Accordingly, the on distance. to features equivalent Hamming is of This (6). number differ the als as space social 0 = epeeteiec,fis,ta oohl sa lyt sig- a to play at is homophily that first, evidence, present We Airbnb other as participants to profiles the presented we While egnrtdec fteptnilrciesacrigto according receivers potential the of each generated We possible, as much as reflect, experiment the that required We nadto odmgahc,tegnrtdpolsincluded profiles generated the demographics, to addition In , d 1, n w fte eepae at placed were them of two and 2, rmtepriiat endi h otx fBlau’s of context the in defined participant, the from alsS2 Tables PNAS | – etme 2 2017 12, September S4 d 2 = nlz efslcinba.Ec par- Each bias. self-selection analyze ssrcl ate rmtepartici- the from farther strictly is d 4 = en httepol had profile the that meant d | 1 = o.114 vol. d en htoeran- one that meant 0 = d al S1 Table 4 = en htthe that meant | eshowed We . o 37 no. presents | 9849

SOCIAL SCIENCES is a postinteraction subjective of an alter. It consists dence of the connection between risk attitudes and trust has been of the assignment of zero to five stars, where the number of stars weak (17). Moreover, research that has addressed this question is proportional to the degree of positiveness. The ratings a mem- has been limited to laboratory experiments or small datasets. ber receives are averaged over all of their raters, rounded to the Given the opportunity to study this question using a large half unit, and presented in the member’s profile on the platform. population, we introduced a risk-assessment question before the Similarly, an interaction grants the two parties the opportunity to investment game. We worded the question as: “A lottery ticket mutually provide free-form written reviews. Due to the difficulty costs 100 (USD) and people win with 50% chance. How much of manipulating textual contents of reviews experimentally, we should the prize be for you to choose to buy a ticket?” Players restricted our attention to the number of reviews a user received. could enter any numerical value, which corresponded to the min- We manipulated these two dimensions in a structured way to imum reward that would make the participant take the risk of buy- study their effects on trust. Among the five profiles participants ing a ticket. The prize value 200 (USD) had the expected value saw on the screen, four had reputation features with similar val- of net gain equal to zero (after paying off the ticket) and corre- ues, chosen independently at random for each participant’s ses- sponded to the minimum rational value. Thus, values >200 (USD) sion, which we refer to as the baseline reputation. These were measured risk aversion proportional to their magnitude. In Risk the profiles at social distances d = 0, 1, 2 and one of the pro- Assessment Question, we summarize the distribution of answers files at d = 4. The other generated profile at distance d = 4 had (Table S1) and argue that our measure captures risk behaviors in one of the reputation features randomly selected to be switched accordance with previous research (Table S5) (19, 20). to either a better or a worse value than baseline (see Game Design Details for how we manipulated the numerical values of Multilevel–Multivariate Analysis reputation). For convenience, we refer to the profile that has a We had five measurements (investments) on each observational different reputation feature than the baseline as being at dist- unit (participant). As a result, the five investments were cor- ance d = 5. related, which we accounted for by nesting investments within We randomly assigned users to two possible worlds. In world subjects in a multilevel model. We fitted the model using a 1, the profile at d = 5 not only had the largest distance from the multivariate regression with 10 independent variables, one for participant, but also a weaker reputation than all other profiles each investment in the combination (d, w) of profile distance (the baseline reputation). In this case, reputation did not com- d : {0, 1, 2, 4, 5} and world w : {1, 2}. The investments a partic- pete with the tendency toward homophily. In world 2, the pro- ipant made had different sources of mutual correlation. For file at d = 5 had a better reputation than the baseline reputation. instance, the sum of the investments had to be at most 100 credits. This induced a tension between placing trust in the most distant We accounted for these by computing the model fit with an uncon- profile with a better reputation or in the other profiles closer to strained covariance structure that learned from the data the cor- the participant in social space. Fig. S1 shows a partial view of the relations and independent variances across measurements (21). screen users see in the experiment, and Fig. S2 shows a diagram As a first-order approximation, we fitted the empty model (i.e., that exemplifies the structure of a user’s session. without explanatory variables) with 10 intercepts. The five inter- We gave participants a single “wallet” with 100 credits, which cepts for each world corresponded to the average distribution they could keep or invest in receivers in whatever way they chose. of investments among the five profiles across all participants Therefore, participants could gain or lose credits through their (complete pooling). Fig. 1 shows a plot of the mean estimates, investments. Because this was a one-time game, it was easy to together with the mean number of credits saved, for worlds 1 show that the Nash equilibrium was not to invest any amount, and 2. Table S6, model 1 shows the numerical estimates from the since the dominant strategy for receivers was not to return any model fit. amount. (Nevertheless, we observed such rational behavior only We were mainly interested in the additive effect of the number in rare instances.) of different coordinates between two individuals’ feature vectors, It is argued that risk is a component of trust in general, and or their Hamming distance. However, any real-world sociode- some definitions of trust include risk (8). Even though previous mographic feature inevitably produces heterogeneous effects on research has attempted to relate trust and risk, the empirical evi- trust (e.g., gender may affect investments more than marital

A World 1 B World 2

20 20

Reputation Reputation

Baseline Baseline

Lower Higher 10 10 Average invested Average Average invested Average

0 0

Savings 0 1 2 4 5 Savings 0 1 2 4 5 Distance from the participant Distance from the participant

Fig. 1. Empty model estimates of average investment in profile at distance d and average savings. (A) In world 1, the second profile at distance d = 4 (here identified as d = 5) has a worse reputation than baseline. (B) In world 2, the profile at distance d = 5 has a better reputation than the baseline.

9850 | www.pnas.org/cgi/doi/10.1073/pnas.1604234114 Abrahao et al. Downloaded by guest on September 25, 2021 Downloaded by guest on September 25, 2021 baa tal. et Abrahao values 2. Fig. variable dependent each for specifically covariates the for effects. control trust- to these aimed for perceived interactions of cross-level preferences the indicators prefer- as Thus, as worthiness. such in “older” biases, or rooted “married,” necessarily effects “female,” not in are resulted which have ences, could distance social variables. 2 between and interactions 1 cross-level level the the characteristics, (iii) and subject’s) “S”; (or with annotated participant’s the to the corresponding to ( variables corresponding “P”; 1 with annotated level in characteristics, (i) structured profile’s as: sets model three into multilevel categorized a be can They covariates. explanatory including by model into variables. empty effects the these take extended all to we explain Thus, account, investments. not the may in itself variance by the of distance Hamming and does), status aibe eecnee n tnadzdt aealcoeffi- all make to investments, average reduce standardized values Negative and comparable). cients centered were distri- point variables the starting values. these of a changed estimate as variables explanatory profiles this S6, the the used (Table how over studied we errors and investment Thus, estimation mean 3). to of and bution up 2, model, 1, empty models the in those (22). model choice McFadden’s on based data simultaneously estimated profiles In we five world). the cases, (by in these investments the In on effects region. joint and For status, session. marital participant’s at gender, the profiles in always sepa- all not symmetries worlds) was example, the the this of to interactions, each due cross-level (in of possible profiles case variable five the explanatory the In each of rately. of each effects in the trust show on to us allowed This

(P) age*((P)−(S)age) Variable h utvraemdletmtdteefcsascae with associated effects the estimated model multivariate The increase to used we features demographic the that expected We S7 Table i.2peet h fet fteecvrae (integer-valued covariates these of effects the presents 2 Fig. with consistent are model full the of intercepts the that Note (S) risk in(1001,10000) (S) risk (S) married ,(P)single (S) married 7 hc orsodt h mletaeaeivsmn ifrnebtentopolswt aeierptto,mnstosadr errors. standard two minus reputation, baseline with profiles two between difference investment average smallest the to correspond which ±1.37, (S) risk in(500,1000) (S) risk (S) single, (P)single (S) female, (P)male (S) risk in(10001+) (S) risk h fet ftecvrae soitdwt h atcpn S n rfie P ntemliait utlvlmdl h ahdlnshv the have lines dashed The model. multilevel multivariate the in (P) profiles and (S) participant the with associated covariates the of effects The (S) male, (P)male (P) age−(S) (P) log(reviews) (P) rating =5 (P) rating =4 (S) single hw ito h nusw sdt omthese form to used we inputs the of list a shows (S) male (P) age al S8 Table epeeta lentv nlsso the of analysis alternative an present we d 0 = 505 0 −5 xcl ace h participant’s the matched exactly World 1 ee variables 2 level ii) (d , w ). Coefficient tto,mnstosadr ros hti,acefiin that coefficient a is, rep- That baseline errors. with standard difference profiles two smallest two minus the between utation, to investment way, average conservative in most the in spond, (26.1, value value (F F smaller explained much a variance with to respect with 5668.8, one dominant the Work. at Is Homophily that model. empty reputation the and from derived homophily previously regarding we significantly to conclusions enough the strong produced not alter effects were model these multivariate that the aver- in the intercepts) their in (model and changes investments the group next, age argue demographic we by as differ- However, are behavior interactions. there investment that in implies which ences distance, social by explained with profiles homophily. the by in driven profiles. investments system still the closest was reputation of reputation the the magnitude baseline of the of influence 2, all world strong in in the despite invested that average Note the the in than investment higher average The at the homophily. profile (i.e., beyond distance trust builds social the to than at respect average profile with at on alternative investment profile less equivalent demographic the received the Furthermore, reputation on they worse average. investment was the on profile lower the received, the participants, away the from farther dimensions the is, That 1A Fig. a Results as investments average on to significantly observed not of was we function did effects goal features main main the the of Our alter heterogeneity them. the that increase show values positive whereas h ahdlnsi i.2hv h values the have 2 Fig. in lines dashed The that beyond variance exhibited variables explanatory The P hw hthmpiydmntdivsmn decisions. investment dominated homophily that shows d d 505 0 −5 < d 5 = 4 = .Ti a olwdb h ubro reviews of number the by followed was This 0.001). nteepymliait model. multivariate empty the in PNAS ossigtebs euain a significantly was reputation, best the possessing , .Qiesrknl,Fg 1B Fig. strikingly, Quite ). World 2 World | h oait pol itne a yfar by was distance” “profile covariate The etme 2 2017 12, September P < 0.001). | o.114 vol. hw htreputation that shows ±1.37 | Effect on o 37 no. d n corre- and 5 = Joint Effect d =5 d =4 d =2 d =1 d =0 | with 9851

SOCIAL SCIENCES exceeds these boundaries potentially produces an effect that (0.52) credits. Symmetric to world 1, the change in the number could alter the conclusions we derived from the empty model. of reviews of the baseline reputation did not affect the average A first glance at Fig. 2 reveals that most of the coefficients are investment in these profiles. contained within these boundaries. Comparing the effects of number of reviews between the two Fig. 2 shows that participant’s gender “(S) male” in both worlds, we see that high reputation resulted in larger investment worlds had small positive effects on all profiles. Marital status increases in cases in which holding the best reputation was an “(S) single” had effects that were not significantly different from exception among the alternatives (world 2). Surprisingly, these zero. For age, the older the profiles [“(P) age”], the more credits exceptions were the profiles that were the farthest away from the they received. One SD (14 y) above the mean (39.7) had positive participants in the social space. effects for all of the profiles with coefficients ranging from 0.93 The coefficient of the joint effect estimated for variable “(P) (0.44) to 2.29 (0.81). The effects associated with region had small rating = 4” represented an increase of 1.74 (0.83) and 1.13 (0.37) values that varied together across different profiles (omitted in credits for worlds 1 and 2, respectively, in mean investment (ref- Fig. 2 for clarity). erence “no rating available”). The corresponding increases for As these effects changed the investments roughly uniformly profiles with five-star ratings, “(P) rating = 5,” was 2.21 (0.82) across the profiles, these effects did not cause significant changes and 0.99 (0.36) for worlds 1 and 2, respectively. This shows that in the differences between the investment means. varying between 4 and 5 stars did not cause a significant differ- We note that the preceding effects did not change homophily ence in average investment, as participants may have considered trends due to the inclusion of interaction effects between partic- them equally high. ipants’ characteristics and those of the profile in the model. In In the full model of world 1, the difference between the Fig. 2, these variables are labeled with both S and P, such as “(S) mean investments comparing the profile with baseline reputation female, (P) male” for gender. Recall that we included these inter- receiving the smallest average investment (d = 4) and the profile actions to capture preferences that are not necessarily biases. at d = 5 with the lowest reputation was 5.73 (1.89). In world 2, the Indeed, in both worlds, male profiles received on average up to difference between the profile with baseline reputation receiv- 3.38 (0.50) fewer credits than females, while not married profiles ing the largest mean investment (d = 0) and the profile at d = 5 received on average up to 2.50 (0.40) fewer credits than mar- with the best reputation was 15.46 (1.68). In Fig. 2, we see that ried profiles. Age difference exhibited a nonlinear relationship. none of the effects were large enough to cancel out the shifts As the profiles got older than the participant, homophily came produced by reputation and alter our conclusion with respect to into play, and the positive effect of the profile’s age decreased trust increases (world 2) or reductions (world 1). This shows evi- significantly, as indicated by the interaction of profile’s age with dence that the is a strong signal that shifts trust the age difference between the profile and the subject. beyond homophily, thereby overriding the effects of assessments Without controlling for these preferences (no interaction of social distance. effects), the model exhibited effects associated with demographic features that canceled out the homophily effects produced by Risk. Fig. 2 shows the effects of answers to the risk-assessment social distance in the case of males or singles. For illustration, question on the investments in each profile. We grouped in Table S6, we included the effects of gender and marital status responses by ranges, where the higher the range, the more risk- for the models that included the interactions (model 3) and that averse we classified a participant to be. These are the covariates with interactions removed (model 2). with prefix “(S) risk in range,” where the reference level is the As the group effects of investment behavior were not large range [200, 400], the low end of rational values. enough to alter the trends produced by profile distance, we show In world 1, we saw little or no effect associated with risk atti- evidence that homophily figures as a major driving force, struc- tudes on the investments in any of the profiles, except for small turing decisions of whom to trust with investments. negative effects on the investments in those with baseline repu- tation and weak similarity with the participant (d = 2 and d = 4). Trust via Reputation. We first focus on the effects of reviews in The effects ranged from a reduction of −1.74 (0.61) to −2.85 Fig. 2. In world 1, an increase in the log-transformed number (0.48) in average investments, with slightly stronger effects pro- of reviews, “(P) reviews (log)” resulted in a statistically indistin- portional to the level of risk aversion. guishable increase in mean investment in profiles with baseline The most striking results were related to world 2. In this case, reputation, between 2.03 (0.47) and 3.20 (0.36) credits. Although risk attitudes did not correlate with the average investments in the profile at d = 5 in this world always had fewer reviews than any of the profiles, except in the profile at d = 5 (with significance baseline, the variation in its number of reviews did not affect P < 0.001). These effects were among the strongest we found the average investment it received. In contrast, in world 2, an (Fig. 2, Right, the bottom three items). The decrease in mean increase in the number of reviews increased the mean invest- investment ranged between 3.91 (0.75) and 8.16 (0.71) and was ment in the profile at d = 5, with the best reputation by 5.42 inversely proportional to the degree of risk aversion. This shows

Fig. 3. Real-world data from Airbnb show that an increased reputation of the host in the form of rating (graph) and number of reviews (x axis) results in greater diversity of guests who selected them (y axis).

9852 | www.pnas.org/cgi/doi/10.1073/pnas.1604234114 Abrahao et al. Downloaded by guest on September 25, 2021 Downloaded by guest on September 25, 2021 3 tt ,ArhoB okK(06 oe maac n aigsystems. rating and imbalance Power (2016) K Cook B, Abrahao B, State 13. exchange. online in status to power From (2012) K Cook B, Abrahao B, State 12. monitor- partial Combinatorial (2014) W Chen J, Lui R, Kleinberg B, eBay. Abrahao T, from evidence Lin and Theory, 11. : in Reputation (2006) J Wooders D, Houser 10. baa tal. et Abrahao atices nsca itnefrhsswt ag ubrof number by large explained a signifi- with a simply hosts see for not distance not social is did in Fig. interactions—we effect increase in of cant graphs this number (across that increased ratings an Note and 3) dis- changed. Fig. social 3) reviews in average of in graph request. number the the each held for booking as (within clearer trend experiment of became a which the time increase, saw to the in tance We and at platform. derived reviews actual host we of the the intuition number of the the rating Strikingly, star on the conditioned on Across varied, y. hosts 20 and and 10, distance not.) 5, social did trends 3, of the values namely but belong- absolute changed, group, the as experiments, age people different same two the the considered age to different within we using and ing were analysis which equal, the they repeated within and if (We thresholds available equal other. each ages were of two y guests 10 considered and We hosts otherwise. for 1 values if 0 the distinct and from users status, locations.) between distance: marital occur collect social usually not interactions of does of hospitality (Airbnb dimensions purposes gender. For two and study. considered age the we as period analysis, guests by same this stay the and to over requests searching million place 1 through taking studied We partners request. a select making They process. real-world, selection a intuition. same guide the to extract and experiment analysis the data the use large-scale Airbnb’s to through from sought gained we observe interactions, insights to historical possible of not database is internal this Although pair them. every for the of a preference option of every degree or to the and access host considered had participant a we partner’s data, select experimental their the they which In guest. when in differ dimensions characteristics of demographic number the user’s to the reduce attention may reputation on reliance that suggested iment Analysis. Data Real-World reputations. did negative aversion in distrust risk pro- with Interestingly, correlate information to system. seem positive reputation not the the trusted by they vided the less risk-averse more the strongly The a gen- participant, aversion. had in risk trust with reputation, trust high correlation reduced of negative case with the correlated to not restricted eral; was aversion risk that .Sidr ,Wei 20)Oln rgamn markets. programming Online (2009) J Weesie C, research. empirical Snijders of review 9. model-based A : risk in Trust (2010) T Earle 8. systems. Reputation (2000) E Friedman R, structure. Zeckhauser social K, of Kuwabara social P, theory Resnick in macrosociological Homophily A 7. feather: (1977) a P Blau of Birds 6. (2001) JM Cook L, and Smith-lovin substantive affiliation. M, of McPherson A ecology An 5. process: (1983) M social McPherson a 4. as Friendship (1954) R, Merton P, rent Lazarsfeld and errands, 3. my (1955) G do Simmel and car, 2. my drive can you Baby, 2015) 29, (January J Stein 1. i.3soshwteaeaesca itnebtenguests between distance social average the how shows 3 Fig. as distance coded we features, demographic these of each For social the in participants active the are guests Airbnb, On fte1t nentoa AICneec nWbadSca Media, Social CA). Alto, and Palo Web Press, on (AAAI Conference many AAAI International 10th the of Science Web on Conference Int’l York). ACM New Machinery, 4th the of ings Beijing). Learning Machine applications. on its Conference and tional feedback linear with game ing Strat Manag Econ J 166–186. pp York), New World Online the in ships Anal Risk ACM networks. 18–66. pp York), New Nostrand, (Van C Page T, Abel analysis. methodological 2017. 9, August Accessed vol-185-no-4-u-s/. stuff. my 43:45–48. vial thttp://time.com/magazine/us/3687285/february-9th-2015- at Available Time. 30:541–574. nuRvSociol Rev Annu oflc n h e fGopAffiliation Group of Web the and Conflict 15:353–369. h usl aeFudto eiso rs,(usl Sage, (Russell Trust, on Series Foundation Sage Russell The , reo n oto nMdr Society Modern in Control and Freedom 27:415–444. h nutosw andfo h exper- the from gained we intuitions The Poednso ahn erigResearch, Learning Machine of (Proceedings mScoRev Socio Am rceig fte3s Interna- 31st the of Proceedings AscainfrComputing for (Association TeFe rs,NwYork). New Press, Free (The Tut omn Relation- Forming eTrust: mJSociol J Am 48:519–532. d egrM, Berger eds , oon,Ger- Cologne, 83:26–54. Proceedings Commun Proceed- 7 atn ,Prg 21)Rs vrinadeggmn ntesaigeconomy. sharing the in engagement and aversion Risk (2015) P Parigi J, Santana 27. where Airbnb, on reputation online at look first A (2015) J Byers D, Proserpio G, Zervas 26. 3 dla ,Lc 20)Dgtldsrmnto:Tecs fairbnb.com. of case (1992) The A trust. Fausto-Sterling and discrimination: Race 25. (2010) S Digital Smith 24. (2005) M Luca B, Edelman behavior. 23. choice qualitative of analysis logit Conditional (1974) D McFadden (2000) 22. DM Mea- Bates periods. scale: evaluation J, and taking risk-attitude Pinheiro risk on 21. domain-specific experiment An A (1997) J Potters (2002) U, Gneezy E 20. Betz A, Blais of E, anatomy An Weber risk: 19. from (1992) trust W Poundstone Distinguishing (2010) 18. J Winter D, Schunk history. D, social Houser and reciprocity, 17. social Trust, (1995) and K McCabe trust, J, Dickhaut cooperation, J, Berg of 16. studies Experimental (2003) R rela- on Cooper of K, impact Cook The world: 15. the Disenchanting (2014) B State P, Parigi 14. ainGat1518(oBA,PP,adK.S.C.). and P.P., B.A., (to 1257138 Grant dation ACKNOWLEDGMENTS. and design research our on analysis. information data more and self-selection, of website. analysis experiment’s study, our the of page in entry participate the to on consent displayed we with terms us 2015). whose provide 11, to August invitees on required approved We 34470, (protocol Board Univer- Review Stanford Internal by sity’s approved and reviewed were methods experimental Our Methods and Materials significant a 24). play (23, features trust exam- racial determining For in that effects. other role suggests greater even literature or of the pictures observation ple, the through etc.— to ethnicity, lead indirectly religion, could class, inferred race, nationality, platform, be as the signals—such can by that explicitly associated demo- displayed but other effects study. not of inclusion sizable our characteristics the graphic in that very emphasize participated we found lim- homophily, that with we are population although results specific experimental Moreover, the our to platform, ited Airbnb actual the dissimilarity of to degree space. willing high social a were the exhibited in study who those our to trusting in trust by extend participants others, system, similar trust reputation to the individuals evidence of gathered tendency we the Although for platforms. social gap in the grown present organically bridge trust the to and operate trust generated reviews—may institutionally of between of rat- number system star the sites—the and reputation sharing-economy ings the of extension that by evidence and Airbnb, shows work arrange- Our institutional on preexisting ments. on predicated directly are rely cannot economy but sharing trust, the in operating Companies Discussion better. hosts got selected host the their of and to reputation guests the at tends as individuals between reputation for distances tolerance high social higher as farther saw is, we distance, That social interactions. shrink online real patterns our to in of generalized conclusions found our artifact main that our an that shows simply but This laboratory, not graph). were first 3, findings (Fig. experimental ratings low and reviews See in work at are effects these that evidence present we While Games 2017. 9, average. August above Accessed is stay every Men 14-054. SSRN 105–142. pp York), New (Academic, P Zarembka ed Econometrics, in York). New behaviors. Econ risk and perceptions 290. risk suring game. investment the Behav 209–244. pp York), New Sage, (Russell J Walker E, Ostrom eds exchange. tionships. uprigInformation Supporting HradBsns col otn avr uiesSho okn Paper Working School Harvard Boston) School, Business (Harvard . Bsc e York). New (Basic, 112:631–645. 10:122–142. 6:560–573. oilInform Social rs n eirct:ItricpiayLsosfrEprmna Research, Experimental for Lessons Interdisciplinary Reciprocity: and Trust PNAS rsnrsDilemma Prisoner’s cnBhvOrgan Behav Econ J 1:166–182. hswr a upre yNtoa cec Foun- Science National by supported was work This | yh fGne:Booia hoisAotWmnand Women About Theories Biological Gender: of Myths etme 2 2017 12, September nuRvSociol Rev Annu SSRN ie-fet oesi n S-PLUS and S in Models Mixed-Effects o ealdifraino u ape an sample, our on information detailed for vial thttps://ssrn.com/abstract=2554500. at Available . Duldy e York). New (Doubleday, 74:72–81. 36:453–75. ea ei Making Decis Behav J | o.114 vol. | o 37 no. ae Econ Games (Springer, Frontiers | 15:263– 9853 J Q

SOCIAL SCIENCES