Shchukina_JMPpdf
Shchukina_JMPpdf
1 of 11900
Sofia Shchukina shared this file. Want to do more with it?
  1. Shchukina_JMP.pdf
    Love at First Photo?How Platform Design Concentrates Romantic Interest on Dating AppsSofia Shchukina∗October 31, 2026Click Here For Latest VersionAbstractRomantic interest on swipe-based dating apps (SBDAs) is highly concentrated. Thiscan frustrate both popular users who feel overwhelmed by a glut of attention, and lesspopular users who feel dejected by a lack of it. Is the concentration of interest ex-acerbated by the platform design of dating apps? I provide causal evidence on theeffects of two common design features – the emphasis on visual elements in profiles,and limits on the number of profiles users can express interest in – on the concentra-tion of interest (“likes”), and show that both, as implemented under the status quo,concentrate interest. To do so, I design, build, and run my own experimental datingapp, which closely resembles conventional SBDAs and allows participants to form realmatches. Across three field experiments, over 1,000 participants use the app to makeover 70,000 matching decisions. Under conventional SBDA profile design, I show thatvisual profile components (i.e., photos and height) dominate matching decisions, andthat the concentration of interest is driven mainly by users’ reactions to profile visuals.Simply reducing the salience of visuals on profiles does not significantly affect whichprofiles receive likes, but requiring users to screen profiles on other components spreadslikes more evenly (Gini coefficient decrease of 0.05 points) without reducing their num-ber or the resultant number of matches. Finally, I show that limiting the number ofprofiles that users can express interest in, which many SBDAs do as part of a freemiumsubscription model, amplifies concentration (Gini coefficient increase of 0.1 points).∗I am deeply grateful to my advisors Leonardo Bursztyn, Marianne Bertrand, and Alex Imas for theirinvaluable feedback and guidance on this project. I would also like to thank Elias Hadj Ammar, MonicaBarbosa, George Beknazar-Yuzbashev, Matthew Brown, Francesco Capozza, Michael Cuna, Joshua Dean,Benjamin Handel, Günter Hitsch, Ali Hortaçsu, Emir Kamenica, Jonathan Kolstad, Aaron Leonard, Haoxuan“Noah” Liu, Ieda Matavelli, Shivani Pandey, Ricardo Perez-Truglia, Devin Pope, Nicolò Rizzotti, Evan Rose,Heather Sarsons, and participants at the University of Chicago’s Normal Lab, Booth Behavioral EconomicsLab, and the Advances with Field Experiments conference for helpful discussions. I gratefully acknowledgesupport from the Becker Friedman Institute. All experiments were pre-registered on AsPredicted (#281413,#281412, #274046). This project was approved by the University of Chicago’s IRB under approval numberIRB25-1894. Claude was used for some coding and copyediting, I reviewed all analysis and text. All errorsare my own.
  2. Shchukina_JMP.pdf
    1 IntroductionRomantic interest on swipe-based dating apps (SBDAs) is highly concentrated, particu-larly among men. On Hinge, the third-largest dating app in the world with approximately30 million users, the top 10% of men receive 58% of all likes1sent to men, while the bottom50% receive 4.3% of all likes.2This inequality harms users at both ends of the distribu-tion: 64% of men who had recently used dating apps reported feeling insecure about howfew messages they receive, while 25% felt overwhelmed by how many they received (Vogelsand McClain,2023). While some degree of inequality of interest is naturally to be expecteddue to correlated preferences over partner attributes, a popular view voiced by dating appusers and commentators is that the platform design of dating apps further exacerbates thisinequality. At the core of this argument is a perceived tension between platform and userobjectives: if apps are profit-maximizing entities, they should optimize for user engagement,not matches. Indeed, dating app platforms control many design choices that plausibly affecthow users allocate interest: for instance, signaling interest is nearly costless, profile designheavily emphasizes photos, users’ choice sets are very large, and platforms’ recommendationalgorithms determine which profiles are seen, how frequently, and by whom. If platformdesign affects the allocation of interest, then alternative design choices might spread interestmore evenly, reducing both overwhelm at the upper end and disappointment at the lower end.Whether and how platform design shapes the distribution of attention is important be-cause dating apps have become a very popular way for people to meet their significant others.In the U.S., over 40% of heterosexual couples meet online (Rosenfeld et al.,2019). Datingapps are significant contributors to this trend: the most popular dating app in the world,Tinder, has facilitated over 100 billion connections (“matches”) since its launch in 2012 alone(Tinder,2026). Matching on dating apps is therefore the first step in determining who mil-lions of dating app users will potentially meet, date, and marry. However, credible causalevidence on how platform design affects the distribution of interest is limited. One importantreason for this gap is the challenges to identification: because many different design featuresvary together and plausibly affect user decisions at the same time, it is difficult to cleanlyidentify the causal effects of specific features using observational data alone.In this paper, I study this question in the field by designing, building, and running my own1A “like” on a dating app is an indication of interest in matching with a user.2https://web.archive.org/web/20190216162750/https://qz.com/1051462/these-statistics-show-why-its-so-hard-to-be-an-average-man-on-dating-apps/ ; https://web.archive.org/web/20170806142615/http://hingeirl.com/hinge-reports/whats-the-biggest-challenge-men-face-on-dating-apps-a-qa-with-aviv-goldgeier-junior-growth-engineer/2
  3. Shchukina_JMP.pdf
    experimental dating app, which closely mimics conventional SBDAs in terms of design andfunctionality. This allows me to manipulate platform design in a controlled setting and ob-serve matching decisions at a level of granularity that is typically unavailable to researchers,while also preserving the high stakes of facilitating real matches. To recruit participants forthe experiments, I partner with speed dating and events organizations, who publicize thestudy through their social media and email networks. This approach allows me to recruita pool of participants who are single, interested in dating, and live in the same geographicarea, such that they could feasibly meet in person after matching on the app if they wantedto.3Across all three experiments, over 1,000 participants use the experimental app to cre-ate profiles, browse other participants’ profiles, and form real matches, yielding over 70,000observed matching decisions in total. Eligibility criteria required participants to identify asheterosexual and be aged 25-39;4moreover, participants were disproportionately likely tohave used dating apps before (99%, 98%, 59% respectively for each experiment), have se-rious dating intentions (93%, 91%, 68%), and be politically liberal (78%, 66%, 52%). Myresults therefore most closely reflect the experiences of these user demographics, and may notfully generalize to other types of SBDA users, such as non-heterosexual users or casual daters.All experiments proceeded in four consecutive stages. First, participants created theirprofiles to be used on the app, during a period of about 2 weeks. Profile content was similarto conventional SBDAs, and consisted of 2-6 photos, biographical details (e.g., height, job,education), a free-text description of the user (“bio”), and some information about hobbies,interests, and personality traits (see Figure2for an example). Once the profile creation pe-riod concluded, I randomly assigned which participants would view which other participants’profiles in the app.5All participants both evaluated (i.e., chose whether to indicate interest inmatching with) other participants’ profiles, and also had their own profiles evaluated by otherparticipants; for clarity, I will therefore refer to participants as ‘viewers’ when discussing theirevaluations of others’ profiles, and ‘profiles’ otherwise. Once random assignment of profilesto viewers had been completed, participants received a personalized link to access the app.Each participant was asked to evaluate a fixed number of profiles in different experimentalconditions, and had a week to complete their evaluations. Similarly to conventional SBDAs,viewers evaluated profiles one at a time, and were asked to indicate whether they were in-terested in matching with each profile or not. I will refer to these matching decisions as“like/reject decisions”, with “like” denoting an indication of interest and “rejection” denot-3Participants were not required to interact with their matches as part of the study, but were able to do soif they wished.4For the Like Limit experiment, the age range was 18-24.5Men viewed women’s profiles in the app and vice versa.3
  4. Shchukina_JMP.pdf
    ing the lack thereof. Participants were matched if they both liked each others’ profiles (i.e.,both indicated an interest in matching with each other). All matches were revealed after theprofile evaluation period had concluded, via a separate personalized link that was emailed toparticipants. I provided matched participants with each others’ email addresses, such thatthey could contact each other if desired.To understand how people make like/reject decisions on SBDAs at baseline, each experi-ment includes aBaselinecondition which is designed to mimic the SBDA decision environ-ment as closely as possible. One aspect of this included emphasizing photos as part of theprofile design. On popular SBDAs such as Tinder and Bumble, photos are typically moresalient than other profile components: photos are the first component displayed on the pro-file, and take up the majority of the initial screen space (see Figure1). I replicate this designchoice in theBaselinecondition. Further, on conventional SBDAs, users typically have toscroll down the screen to access other profile components, such as biographical details andbios. To assess how frequently users do so in practice, I hide all profile content other thanfirst initial, photos, age, and height behind a “Click to See More About Me” button, andtrack when the button is clicked. This allows me to observe when like/reject decisions aremade using only the components available above the button. I assume that the first initialdoes not affect like/reject decisions as it carries no information, and control for the effectsof age by looking at a subset of like/reject decisions where profiles fit viewers’ self-reportedpreferred age range for a potential partner. This strategy allows me to plausibly identifylike/reject decisions that were made on the basis of photos and height (“visuals”) alone.I find that viewers frequently reject profiles based on visuals alone, but rarely like pro-files based on visuals alone. In all three experiments, over 30% ofBaselinedecisions arecomprised of rejections based on visuals alone, while likes based on visuals alone representfewer than 5% of all decisions. This suggests that viewers use visuals to screen out profiles,and only look at other components for profiles that pass this initial screen. Conditional onclicking the button and thereby accessing all profile information, decisions are significantlymore varied: like rates conditional on clicking are 28% for female viewers and 50% for maleviewers.6Consistent with previous studies of matching decisions on dating apps, I find thatwomen like a significantly smaller proportion of profiles relative to men, both overall andconditional on either clicking or not clicking the button to access other profile components(Tyson et al.,2016;Lehmann et al.,2023).6Results from the Two-Stage Experiment, which has the largest sample of participants (total viewers = 617).The corresponding proportions in the Text First experiment are 26% and 45% for male and female viewersrespectively. The corresponding proportions in the Like Limit experiment are 30% and 38%.4
  5. Shchukina_JMP.pdf
    I then use the experimental app to identify how different profile components, which I clas-sify as either ‘visuals’ (photos, height) or ‘text’ (all other components), affect the allocationand resultant distribution of interest. This enables me to better understand how viewers’reactions to specific profile components affect the concentration of interest, and thereforeidentify which levers may be effective in reducing it. To identify the separate contributionsof visuals and text, I design and implement a novel two-stage liking process. Participantsbegin by evaluating 20 partial-information profiles in Stage 1. In theTwo Stage, VisualsFirstcondition, partial-information profiles only include the profile owner’s first initial, age,and visuals. In theTwo Stage, Text Firstcondition, partial-information profiles only includefirst initial, age, and text. Participants do not make like/reject decisions in Stage 1; instead,they choose whether they want to advance each profile to Stage 2 or not (I will therefore referto Stage 1 decisions as “advancement”). Participants could advance up to 12 partial profilesin Stage 1.7Once participants have completed all Stage 1 advancement decisions, they pro-ceed to Stage 2, where they view full-information profiles (which comprise of first initial, age,visuals, and text) for all the profiles they selected to advance in Stage 1. Participants mustadvance profiles to Stage 2 in order to be able to like them – not advancing a profile in Stage1 is therefore effectively equivalent to rejection. All participants evaluate profiles, and havetheir own profiles evaluated by other participants, in both two-stage conditions, with the or-der of conditions cross-randomized between viewers. This two-stage design therefore allowsme to isolate viewers’ independent reactions to profiles’ visuals and text by looking at Stage1 advancement decisions. 641 participants create profiles that are used in this experiment,and 617 complete the profile evaluation activity. Due to a gender imbalance in my sample,8I am underpowered to detect profile-level effects for women’s profiles, and therefore focus allprofile-level analyses on men’s profiles.To identify the individual contributions of visuals and text to like/reject decisions at base-line, I use a mixed-effects logit model to jointly estimateBaselinelikes and Stage 1 advancesin each of the two-stage conditions. The model structure allows for separate profile-specificand viewer-specific effects in each condition, which I respectively refer to as profileappealandviewerleniency. Higher profile appeal raises every viewer’s evaluation of the profile, whilehigher viewer leniency raises the viewer’s evaluation of every profile. I document three keyfindings. First, I find that visuals-only and text-only profile appeals are only weakly correlated(ρ= 0.2). This suggests that profiles whose visuals viewers find most appealing are generally7This limit was implemented to incentivize participants to engage with the partial-information profiles, ratherthan simply advancing all 20 partial-information profiles.875% female5
  6. Shchukina_JMP.pdf
    not the same profiles as the ones whose text viewers find most appealing. Second, viewersagree on visuals-only appeal much more than they agree on text-only appeal: differencesbetween profiles account for twice as much of the latent variation in advancement decisionswhen viewers see only the visuals (32%) relative to only the text (16%). Correspondingly,visuals-only appeal is more widely dispersed between profiles, relative to text-only appeal(standard deviations of 1.47 and 0.84 respectively). Finally, I show thatBaselinedecisionsare driven primarily by visuals. A profile’sBaselineappeal is strongly correlated with itsvisuals-only appeal (ρ= 0.92), but only moderately with its text-only appeal (ρ= 0.47).Using a linear projection ofBaselineappeal on visuals-only and text-only appeal, I find thatthe standardized weight on visuals-only appeal (0.86) is almost three times as large as theweight on text-only appeal (0.30). Together, these findings suggest that the concentration oflikes inBaseline(Gini coefficient = 0.42) is driven mainly by visuals, both becauseBaselinedecisions rely mostly on visuals and because evaluations of visuals are themselves more con-centrated.The Two-Stage experiment also allows me to identify the causal effects of screening onvisuals versus on text in matching decisions, by comparing (unconditional) like/reject deci-sions in the two-stage conditions. I find that, when viewers screen on text first rather thanvisuals, the total number of likes and resultant matches is unchanged, but likes are somewhatless concentrated. Using a mixed-effects logit model to jointly estimate like/reject decisionsin all three conditions, I find that the Gini coefficient of likes is 0.47 inTwo Stage, VisualsFirstcompared 0.42 inTwo Stage, Text First. I also document that profile appeal is almostunchanged betweenTwo Stage, Visuals FirstandBaseline(ratio of standard deviations =0.995, rank correlation = 0.98), which is consistent with viewers already screening on visualsinBaseline.The Two-Stage experiment allows me to identify the contribution of visuals and text tolike/reject decisions at baseline, and show that two-stage text-first screening can deconcen-trate interest while keeping the total number of likes and matches unchanged. However,multi-stage screening is not used in practice on major SBDAs today, and it has severalfeatures that may be undesirable to SBDA platforms. For instance, it requires viewers toevaluate the same profiles multiple times (partial-information followed by full-information),which increases decision time and may frustrate users. I therefore use an additional fieldexperiment to test whether a more practical, lighter-touch intervention, which shares severalmechanisms with two-stage text-first screening, can achieve a similar deconcentration effect.6
  7. Shchukina_JMP.pdf
    In a separate experiment, I study the causal effect of simply reversing the order in whichprofile components are presented. To do so, I introduce aText Firstcondition: in this con-dition, each profile shows the first initial, age, and text at the top of the profile, and hidesvisuals behind the “Click To See More About Me” button below. The condition is other-wise identical toBaseline. Participants are asked to evaluate 30 profiles in each condition,and also have their own profiles evaluated in both conditions, with the order of conditionscross-randomized between viewers. Interventions of this kind, which make photos less salientrelative to other profile components, are common on SBDAs: for instance, profiles on Hingeintersperse photos with text. The text-first intervention shares two mechanisms with two-stage text-first screening. First, it ensures that viewers are exposed to every profile’s text,including the text of profiles they would have otherwise rejected based on their visuals alone.Second, it leads viewers to form a first impression from the profiles’ text rather than thevisuals, which could affect their overall impression of the profile through primacy effects inimpression formation (Asch,1946;Anderson,1965). Comparing results of these two exper-iments therefore separates the role of selection from that of exposure and first impressions.421 participants create profiles, and 400 complete the profile evaluation activity; I am againunderpowered to detect profile-level effects for women, and therefore focus on men’s profilesfor profile-level analyses.I find that presenting text first increases the likelihood that viewers attend to the text,but does not change how many profiles viewers like, or which profiles they like. Unlike vi-suals, text is rarely used to screen out profiles: inText First, viewers open the visuals inover 90% of decisions, so that almost all decisions are made after seeing both components.Profile ranking is nearly identical between conditions (observed correlation of like rates =0.94), and the compression of differences in profile appeal is small and insignificant (pointestimate of 6%). Viewers respond to seeing the text first in different ways: some like substan-tially more profiles, while others like substantially fewer, but these responses are unrelatedto viewers’ baseline leniency, or the order in which they complete conditions. Together,the slight compression of appeal and the greater dispersion of viewer leniency spread likesslightly more evenly inText First: the Gini coefficient decreases from 0.50 inBaselineto 0.47inText First, and the share of likes received by the top 20% of profiles decreases from 53%to 50%. About half of this decrease is attributable to the slight compression in differencesin appeal between profiles, and half to the greater dispersion in viewer leniency when textis shown first. However, the small decrease in inequality comes at a cost: because viewersread text more frequentlyText First, decisions in that condition take significantly longerrelative toBaseline(53% increase in decision time, p-value<0.01). Overall, my results7
  8. Shchukina_JMP.pdf
    suggest that exposure to text and shifting the anchor of the first impression to text are insuf-ficient to change which profiles viewers like, and hence the concentration of likes. They alsoshow that simply placing text first in a profile does not induce viewers to use text as a screen.In the final experiment, I turn to studying a different platform design feature which iscommon on SBDAs: limits on how many profiles users can like in a given period (“like lim-its”). Many SBDAs use like limits as part of a freemium subscription model: for instance,Tinder and Hinge both limit free users’ daily likes, but remove the limit for paying users.9Like limits introduce an opportunity cost, since sending a like to any given profile now takesaway the opportunity to send a like to a different profile down the line. This plausibly affectslike/reject decisions through several channels: for instance, viewers might take more timeor acquire more information about a profile before deciding to send a like, only like theirmost-preferred profiles, or strategically shift their likes towards profiles that they believethey would be more likely to match with (e.g., profiles they have more in common with, orless popular profiles if they are assumed to be less selective).To study how like limits affect like/reject decisions and the resultant concentration of likes,a different group of participants is asked to evaluate 25 profiles each in theBaselineconditionand aLike Limitcondition in a separate experiment. In theBaselinecondition, viewers cansend as many likes as they wish, while in theLike Limitcondition, they are only able tosend 5 likes maximum. 225 participants create profiles, and 216 complete profile evaluation; Iagain focus on men’s profiles for profile-level results. The like limit reduces the average prob-ability of a like by 4.6 percentage points, about a third relative toBaseline, and reduces likesproportionally more for profiles with lower baseline appeal. For example, a profile at the 10thpercentile ofBaselineappeal loses 68% of its likes, while a profile at the 90th percentile onlyloses 23%. As a result, the limit concentrates likes on the profiles with the highest appeal:the Gini coefficient increases from 0.42 inBaselineto 0.52 inLike Limit, and the share oflikes received by the top 20% of profiles increases from 46% to 55% respectively. About halfof this increase comes from the compression of differences in leniency between viewers: thelike limit constrains viewers who are least selective at baseline the strongest, and less selectiveviewers are disproportionately the ones who like profiles with lower appeal. The other halfcomes from a slight widening of profile appeal (9%, statistically insignificant). Since womenare less likely to pay for premium subscriptions on SBDAs relative to men (Vogels and Mc-Clain,2023), like limits may amplify the concentration of likes that men receive in particular.9Note that these limits do not restrict the number of profiles that a user is able to view. Users can thereforereject as many profiles as they want – only the quantity of likes is restricted.8
  9. Shchukina_JMP.pdf
    The remainder of the paper is organized as follows. Section1.1situates this paper withinthe literature. Section2explains the experimental design and logistics. Sections3,4, and5present the results of the Two-Stage Experiment, Text First Experiment, and Like LimitExperiment respectively. Section6concludes.1.1 Related LiteratureThis paper contributes to prior work on mate preferences in online dating. Researchershave studied this using observational data on dating platforms (Hitsch et al.,2010a,b;Skopeket al.,2011;Bruch et al.,2016;Bruch and Newman,2018,2019;Levy et al.,2019;Bojd andYoganarasimhan,2022;Jang et al.,2025), lab experiments (Ranzini et al.,2022;Witmeret al.,2025), and field experiments where researcher-generated profiles are placed on realdating apps (Egebark et al.,2021;Ong and Wang,2015;Tyson et al.,2016;Neyt et al.,2019,2020,2022). I develop and use a new methodological modality – an experimental dat-ing app – which allows me to preserve full experimental control while also maintaining thehigh stakes of facilitating real matches, thereby plausibly identifying causal effects of specificprofile components in the field. My within-subject design further allows me to identify theeffects of platform design on the demand for specific profiles, while holding profile contentfixed and controlling for profile visibility (i.e., how many viewers saw a given profile).This paper also contributes to literature studying how platform design and market con-ditions affect behavior and decision-making in online dating. Previous studies in this fieldhave looked at the effects of perceived competition (Fong,2024), costly preference signals(Lee and Niederle,2015), demand information disclosure (Huang et al.,2022), assortmentoptimization (Rios et al.,2023), choice set size (Pronk and Denissen,2020), and choice ca-pacity (Jung et al.,2022). Closest to this paper,Ma and Gajos(2022) run a lab experimentwith a similar component reordering manipulation to my Text First experiment, and findthat making photos less salient on dating app profiles reduces racial bias. Moreover,Van derZanden et al.(2022) use eye-tracking technology to study which profile components peoplepay attention to, and find that the location of photos on online dating profiles (i.e., left sideof the page versus right side of the page) affects how likely photos are to capture initialattention. They conclude that component placement and emphasis could be used to redi-rect attention to specific components. In this paper, I test whether redirecting attention inturn affects matching decisions in the field. More generally, I contribute to this literature bystudying how two aspects of platform design – the emphasis on visual elements in profiles,and limits on the number of profiles users can express interest in – affect matching decisions9
  10. Shchukina_JMP.pdf
    and the resultant concentration of interest. Both of these features are commonly used ondating apps, but there is limited causal evidence on their effects on matching decisions in thefield.Methodologically, this paper contributes to experimental work where researchers developtheir own platforms as a tool for studying behavior in a field setting. This approach hasbeen used to study a range of topics, including social influence in cultural markets (Salganiket al.,2006), social comparisons (Chen et al.,2010;Aridor et al.,2024), providing advice tojob seekers (Belot et al.,2019), and markets as aggregators of information (Forsythe et al.,1992). Closest to my setting,Frost et al.(2008) develop a virtual platform which allowsparticipants to take part in virtual dates. To the best of my knowledge, my paper is the firstto use an experimental dating app similar to a conventional SBDA.2 Experimental DesignIn this Section, I describe the general design of my experimental dating app and theprocess which I use to run all experiments.2.1 Designing an Experimental Dating AppThe goal was to design an experimental dating app that mimics conventional SBDAsas closely as possible in terms of design and functionality, but also allows the researcherto maintain experimental control over the choice environment. The experimental datingapp was implemented using Qualtrics, with customized JavaScript, HTML, and CSS codeto develop design and functionality. Profiles were designed to resemble conventional SBDAprofiles in terms of both the design and content (see Section2.3for details). The app couldbe accessed via desktop or mobile. All experiments proceeded in four consecutive stages:profile creation, randomization, profile evaluation, and matching; each stage is described indetail in the sections below.2.2 RecruitmentAll experiments were conducted between February and April 2026. The Two-Stage Exper-iment was conducted in a large city in California, the Text First Experiment was conductedin a large city in the Midwest, and the Like Limit Experiment was conducted in a collegetown in New York state. Descriptive statistics about each sample are provided in Section3.2for the Two-Stage Experiment, Section4.2for the Text First Experiment, and Section10
  11. Shchukina_JMP.pdf
    5.2for the Like Limit Experiment respectively. To recruit participants for the experiment, Iworked with local speed-dating and events organizations, who publicized the study throughtheir social media and email networks.Eligibility criteria for each experiment included being single and open to dating, identify-ing as heterosexual, and falling within a specific age range (25-39 for the Two-Stage and TextFirst experiments; 18-24 for the Like Limit Experiment). Additionally, eligibility requiredresiding in the large city or its suburbs for the Two-Stage and Text First experiments, andbeing affiliated with a local university for the Like Limit Experiment. Experimental poolswere completely separate (i.e., Two-Stage Experiment participants were only able to matchwith fellow Two-Stage Experiment participants, etc.).Participants were informed that they would be participating in a research study, and thatall profiles which they would evaluate were made by real study participants in their area,so that they had the opportunity to make real matches. All participants who completed allparts of the study also received a digital $20 gift card.2.3 Profile CreationIn the first step of the study, participants created their profiles for the experimental dat-ing app. Figure2shows an example profile design in theBaselinecondition. To create theirprofile, participants were asked to provide their information via a Qualtrics form. Partici-pants could complete the form at any point during the course of the profile creation period,which lasted 2 weeks (with the exception of the Like Limit Experiment, where it lasted 4weeks). Profiles included the following information:Basic Information: Participants were required to provide their first name, age, height,job, education, and type of relationship sought (e.g., casual or serious). Since participants inthe Like Limit Experiment were college students, they provided their major and class yearinstead of job and education. Participants could also optionally include their religion andpolitics.Photos: Participants uploaded 2-6 photos of themselves.Bio: On dating apps, bios are users’ descriptions of themselves. Bios were free-entry andthus could include any information that participants wanted. As an example, the bio of the11
  12. Shchukina_JMP.pdf
    example profile which participants saw read as follows: “Sports and live music aficionado.Always up for a food or hiking adventure.”‘My Vibe’ Sliders: Participants were provided with a choice of 8 sliders to indi-cate where they lay on various personality traits, and were asked to select four to com-plete and display on their profile. All sliders had ten-point scales. Slider options in-cluded ‘introvert-extrovert’, ‘planner-spontaneous’, ‘indoorsy-outdoorsy’, ‘cooking-orderingtakeout’, ‘homebody-always out’, ‘small gatherings-big parties’, ‘early bird-night owl’, and‘coffee date-adventure date’. For example, a value below 5 on the ‘introvert-extrovert’ scale in-dicated introversion, while a value above 5 indicated extroversion (see Figure2for examples).Hobbies and Interests: Participants were asked to provide between 3 and 7 of theirhobbies and interests.Prompts: On dating apps, prompts are questions or statements that users can answer,designed to showcase their personality and work as conversation starters. Participants chose3 out of a possible 6 prompts to answer and include on their profile. Prompt options included:“What I want to know about you...”, “One thing on my bucket list...”, “Something interestingabout me...”, “Our first adventure together could be...”, “Things that bring me joy...”, and“My favorite place in [city]...” (see Figure2for examples).These profile components were selected to make profile content similar to real datingapp profiles today. Once participants had entered all profile information, they were shownwhat their profile would look like on the app, and had the opportunity to go back and makechanges if they wished.2.4 RandomizationFor each experiment, once the profile creation period was complete, I randomized whichparticipants would view which other participants’ profiles in which experimental conditions.Within each experiment, each participant viewed and evaluated a fixed number of profiles(60 in the Two-Stage and Text First experiments, 50 in the Like Limit Experiment).I implemented an randomization algorithm that would assign participant profiles to otherparticipants to view such that i) each profile was viewed an approximately equal number oftimes between-condition within-gender, and ii) mutual views were within-condition only (i.e.,if A was assigned to view B’s profile in theBaselinecondition, B would also view A’s pro-12
We use cookies to provide, improve, protect and promote our services. Visit our Privacy Policy and Privacy Policy FAQs to learn more. You can manage your personal preferences, including your ‘Do not sell or share my personal data to third parties’ setting using the “Customize cookies” button below.