Are attractive people more likely
to get what they want?
An empirical analysis on beauty
discrimination
Student Name: Nikolaos Fournarakis
Student ID: 10884025
Supervisor: Dr. Jeroen Van de Ven
Master Thesis Organization Economics (ECTS: 15)
MSc Business Economics
Track: Organization Economics
Calendar Year: 2014-2015
A m s t e r d a m S c h o o l o f E c o n o m i c s / A m s t e r d a m B u s i n e s s S c h o o l U n i v e r s i t y o f A m s t e r d a m
1. Introduction
Whether attractive people are subject to favorable behavior relative to less attractive ones is a topic that researchers from various disciplines have examined. The importance of this discrimination is hidden in its very nature, which seems to be “insidious” (Belot, Bhaskar, & van de Ven, 2008). In other words, attractive people seem to be treated differently for reasons that vary from consumption value (Belot et al., 2008), which is the utility that one gets while interacting with attractive people on the long run, to perceptions of ability and persuasiveness.
In the contemporary job market, social media and social norms drive more and more candidates to provide photographs of themselves to potential employers, either by including them in the Curriculum Vitae’s or by using LinkedIn. Other job-seeking platforms, such as Monsterboard.nl and Magent.me among others, nudge their members to upload pictures of themselves, by telling them that they have increased chances of getting noticed by employers. There seems to be a new trend in the job seeking procedure, which will render physical appearance a more salient element of the seeker’s profile.
This paper is inspired by the relevant economic literature but examines beauty discrimination in a different setting: in Facebook. More specifically, the paper tests whether more attractive FB members are more likely to receive a response when they request for help by fellow members compared to less attractive ones. Although not directly related to the job market, the purpose of the paper is to test whether beauty discrimination is a systematic bias, present in different contexts. If consumption value is an “insidious element” that affects how employers choose future employees, where more objective criteria should be drivers of choice, then beauty discrimination should be less insidious in a more relaxed setting such as Facebook.
Previous studies provide evidence of beauty discrimination, in settings where monetary outcomes are involved, such as monetary prize in TV shows (Belot et al., 2008) or the choice of employees, which is a form of investment on behalf of the employer (Doorley & Sierminska, 2012; Mobius & Rosenblat, 2006). Players and employers seem to prefer attractive people, regardless if they actually are the ones that will help them reach the optimal outcome. Thus, helping behavior in Facebook is used as a proxy of a general favorable behavior towards attractive people.
In other words, this paper aims at expanding the role of beauty discrimination in one more setting where the counterparts’ attractiveness is common knowledge. Helping behavior requires effort on behalf of the responders without considerable returns and beauty discrimination is introduced as an alternative motive to altruism. If attractive people are more likely to receive help, get employed and receive higher wages, then being attractive can be considered as an important asset in one’s life.
Data is acquired by two student dorms in the area of Amsterdam, whose inhabitants participate in Facebook pages that give them the opportunity to interact with each other for various reasons. One of these reasons is requesting help in the form of small favors or information, and only these posts are used as a source of data. Facebook is an ideal environment to acquire such data, since it offers a great number of subjects and interactions among them. Moreover, all interactions in Facebook are already coded with unique IDs that can help researchers organize and process a substantial amount of data. The results show that attractiveness is not the main driver of behavior in the setting examined. Although there seems to be a weak positive correlation between attractiveness and the chances of receiving help, the variable that explains the response behavior is the nature of the posts. That is, whether they are favor- or information-oriented. Posters who request information are way more likely to receive help, regardless of their appearance. Providing information requires less effort and involvement on behalf of the responders, suggesting that the students in the sample act altruistically when it is more convenient for them.
The paper is structured in the following way. Section 2 cites multidisciplinary findings from related literature in economics and psychology in order to introduce the reader to the notion of beauty discrimination, possible sources and its impact as measured empirically. Section 3 describes the setting and the data acquired, the research question, the methodology and the model that is used. The results are presented in section 4, and the limitations and suggestions for further research are in Section 5.
2. Related Literature
Beauty discrimination is the term describing the idea that not all people look the same and, as a result, not all people receive the same treatment, either in terms of behavior or in more sensitive ones, such as wages. Hamermesh & Biddle (1994), in their seminal
paper, found a correlation between attractiveness and labor market earnings across a variety of occupations. Attractive individuals earn 5 percent more than those with average looks, and less attractive individuals earn 5 to 10 percent less. This is the so called “beauty premium”. Long before this study, Heilman & Saruwatari (1979) found that attractiveness consistently proved to be an advantage for men but was an advantage for women only when seeking a non-managerial position. What is interesting is that beauty discrimination seems to be something stable; Doorley & Sierminska (2012) provide evidence, using data from the German labor market, which indicate that the male beauty premium is present throughout the wage distribution, while the female beauty premium is concentrated at the bottom of the wage distribution.
Trying to disentangle the beauty premium researchers have identified and examined different channels that could justify the link between beauty and better treatment. One of such was the channel between beauty and increased ability, but evidence show contradictory results. For instance, Pfann et al. (2000) find that among Dutch advertising companies, those with better looking executives have higher revenues, while Landry et al. (2006) find that attractive female solicitors are more productive fundraisers. On the other hand, when it comes to activities for which productivity seems irrelevant to beauty, evidence suggest that good-looking people do not perform better than less good-looking ones (Mobius & Rosenblat, 2006). More specifically, Belot et al. (2008) showed that, in the setting of a TV show where the lead player of each round decides which one of the remaining players to eliminate, these lead players seem to prefer more attractive players, even if their scores are lower than other contestants; only 27 percent of the least attractive players make it to the final round, against 49 percent of the most attractive ones.
Researchers, thus, focused on alternative channels that may justify the beauty premium. These channels are linked to how beautiful people are perceived by others and include the “beauty-is-good” stereotype according to which people perceive beauty to be correlated with intelligence, social skills and health (Eagly, 1991; Feingold, 1992; Mobius & Rosenblat, 2006). In addition, Becker (1958) introduced the term “taste-based discrimination”, according to which employers and customers derive utility from interacting with physically attractive employees, and thus employers choose to pay them higher wages. Rosenblat (2008) provides evidence for the so called “negotiation channel theory” which suggests that physically attractive workers receive higher wages because they negotiate more effectively. This theory renders beauty discrimination a rather
systematic bias, since physically attractive workers tend to be perceived as more persuasive, even when the message they deliver is very similar to the ones delivered by less attractive people Workers. In the contemporary labor market, where interactive media (such as pre-recorded videos required for applicants) or professional social networks (such as Facebook or LinkedIn) are becoming more and more popular recruiting tools among employers, beauty discrimination can evolve into an even more important issue. For instance, Belot et al. (2008) suggest that beauty discrimination is different than race or gender discrimination because attractiveness is more subjective and, thus, public awareness about its prevalence is limited.
In other words, it is not quite clear whether attractive people actually perform better in professional settings or are perceived better by potential employers, resulting in increased wages or opportunities. Attractive people may be more confident (Mobius & Rosenblat, 2006), which might also be reflected through personality traits or tendency for grooming that evidence suggest that could also contribute to the beauty premium in addition to beauty measures (Robins et al., 2011). Another interesting aspect of beauty discrimination is the “kernel of truth” hypothesis, which implies that the physical attractiveness stereotype can become a self-fulfilling prophecy: for example, teachers expect better looking kids to outperform their peers in school and devote more attention to children who are perceived to have greater potential (Hatfield & Sprecher, 1986). Education is a trait that has been proven to reduce the effect of beauty discrimination when examined together (Doorley & Sierminska, 2012), thus beautiful people might have developed certain traits that could further boost their career opportunities. After all, the beauty premium in their research provides evidence that not only the beauty premium is higher for those with college education, but there is also a penalty for women with low education, concluding that “beautiful people also have other characteristics which increase their wage”.
The purpose of this paper is to test whether beauty discrimination can be considered a systematic behavioral bias by examining human behavior in an alternative setting: helping behavior in Facebook groups. Although this setting is not appropriate to test for the “beauty premium” in the labor market, it offers the unique opportunity to check whether attractive people are treated differently in the context of communities.
Moreover, this paper contributes to relevant literature by introducing Facebook as a possible source to acquire data in order to examine human behavior. It offers multiple advantages, such as the fact that users consider it a relaxed environment where they voluntarily provide personal information, traits of their personality and most importantly, researchers can have access to communication channels, where the identity of the users, the messages they exchange and the environment in which the interaction takes place, are already coded and easy to import to data analysis software. Each user, post, comment and like are given unique IDs, which makes the identification of the users and of their activities quite easy. Nevertheless, Facebook is a dynamic environment, and for the data to provide systemic patterns of behavior, sometimes personal information about the users must be obtained. This not only raises issues of privacy, but it also requires patience and creativity on behalf of the researcher in order to categorize the data in an efficient way. The limitations concerning this study will be presented in the Discussionsection.
3. Data - Methodology The setting
The setting of the study is two Facebook pages of students’ dorms in Amsterdam, which are used by their members for multiple reasons. The main reason is for members to request for help, which could vary from tools needed to using someone else’s printer, and for information about nearby markets and pharmacies among others. It is a very direct way of communication among the inhabitants of the dorms and it is widely used among dormitories in the Netherlands.
Each member uploads a status with their request on the page and waits for other members to respond. Responders can choose to reply by choosing either to comment under the status, where their answer is shown in public, or to send a personal message, which is private. Personal messages are the main limitation in this study, since there is not access to this data. Thus, some of the requests that received no reply in the setting of the page might have been resolved in a more private level.
The allocation of the students in the apartments is randomly determined. More specifically, the students, after having been admitted in an academic program by one of the partner universities, and after having been granted an apartment in a dorm, they
themselves choose the location of the dorm and the number of the room1. Note that no information is provided about the precise location of the room in the building, so students cannot apply any criteria when selecting an apartment.
Partner universities reserve rooms in these dorms for their students and guests. Inhabitants of the dorms are divided in three categories, which are “Master, Bachelor and Exchange students”, “PhD students” and “Guests”. Anyone who is not associated with the partner universities is not allowed to apply and subletting is prohibited by the contract signed, making the sample rather homogeneous. The duration of the stay can be either one (half year) or two (whole year) semesters.
The Facebook Pages
For each one of the dorms there is one Facebook page that facilitates communication among the inhabitants. As mentioned before, these pages are used for various reasons. For the purpose of the study, attention will only be given to requests; the statuses that members upload asking for a favor or information. Fellow members can voluntarily offer to help, either by just asking the question or by providing tools and other objects or by offering small services, such as a printing a fight ticket. The starting day of data collection is September 1st 2014. This date is chosen because by this time all inhabitants are supposed to have arrived due to the start of the courses in Dutch universities. The closing date is March 31st 2015. On January 2015 some inhabitants left due to the end of the first academic semester (especially exchange students) and were replaced by others. Nevertheless, their activity in the pages is recorded and thus can be included in the analysis.
Description of data
Before describing the statuses of interest, it would be useful to provide some basic statistics for the pages’ total activity, in terms of “user engagement” and “post engagement”, which are summarized in the Table 1. The table shows that both pages are similar in terms of engagement, which is rather low considering the ease of communication that these platforms offer to the members.
1 For more information about the process, visit the DUWO (www.duwo.nl/) and DeKey
(www.dekey.nl) websites. They are the housing providers of the dorms under examination.
7
Table 1
Statistics for overall page activity
User Engagement Post Engagement
Never
Posting Members Never
Commenting Members Posts commented
Dorm 1 47.12% 39.27% 64.98%
Dorm 2 42.79% 39.90% 60.49%
The category “Posts” includes several types (like event suggestions, videos, links etc.) but the dominant type of posts is “Status” which accounts for 90% of the posts in Dorm 1 and 85% in Dorm 2. I focus solely on statuses about requests, from which I exclude the ones that fall under one of the following categories:
• Statuses about matters that affect all inhabitants in the same way (such as taxes and bills, internet problems, laundry issues etc.)
• Statuses that referred only to Dutch speakers, either because it was a request for translation or because the status was written in Dutch
• Statuses that referred to selling or buying things, since this kind of transactions could be simply driven by demand and supply
• Requests that were privately resolved and is mentioned in a comment under the status. This is because the identity of the responder is not known.
• Requests by members that did not have a profile picture that could be used for ratings, because it would distort the variable of interest, which is in a scale from 1 to 7.
After the exclusion, 261 posts (114 in Dorm 1 and 147 in Dorm 2) were kept by a total of 139 posters (58 in Dorm 1 and 81 in Dorm 2). These statuses will be referred to as “requests” from now on. From the side of the commenters, in total 379 comments (169 in Dorm 1 and 210 in Dorm 2) were posted by 194 commenters (76 in Dorm 1 and 118 in Dorm 2). Summary statistics for the posters and the commenters are available in Table 2, which shows that the two pages are similar in terms of user engagement, providing us with a rather homogeneous sample for our analysis.
In both dorms, there seems to be a small group of members that accounts for a disproportionate part of the activity. More specifically, in both pages, the top 20 of the posters and the commenters accounts for more than 50% of the total posts and comments respectively, implying that a “community effect” might be present. In other words, it is
plausible that these active members might be treated in a different manner than the rest of the members due to their strong presence in the pages. For the purpose of the study, a binary variable Top Commenter is added to the model. The role of this variable is to indicate whether the poster is also one of the Top 20 commenters and to control for reciprocity or reputational effects; if a member is quite often willing to help, then he/she might have more chances of getting a response at their own requests.
Unfortunately, it was not possible to acquire information concerning the demographics of the dorms, since the house providers claimed that they do not possess this kind of information. This is a limitation that mainly affects the “rating of attractiveness” process, which is described in the next section.
Table 2
Summary Statistics of Posters’ and Commenters’ Activity
Mean Median Standard Deviation Max Min
Dorm 1 Dorm 2 Total Dorm 1 Dorm 2 Total Dorm 1 Dorm 2 Total Dorm 1 Dorm 2 Total Dorm 1 Dorm 2 Total Posters 1.93 1.88 2.04 1 1.5 1 1.32 1.15 1.44 7 6 7 1 1 1 Commenters 2.24 2.56 2.19 2 2 2 1.73 2.38 1.64 9 7 9 1 1 1
Notes: for posters summary statistics represent activity in terms of posts; for commenters summary statistics represents activity in terms of comments-responses
Rating Attractiveness
Our main independent variable of interest is the ratings of attractiveness for the posters. The ratings were acquired through an online survey that was sent to students in different countries, including Germany, Greece, Netherlands and Italy among others. No raters from the area of Amsterdam were used, in order to avoid familiarity with the faces, which would have caused a bias. Since demographics of the dorms were not available, it is not possible to claim that the sample of raters is representative to the dorms’ population in terms of gender and country of origin. Nevertheless, the sample of raters consists only of students (Bachelor, Master and PhD) and since only students are allowed to live in the dorms, the age of the raters resembles to the one of the inhabitants.
The total number of raters recruited was 125. A separate survey was conducted for each dorm, ensuring that the posters of each page would be rated by the same set of raters while varying the order with which the photographs appeared. Thus, each rater rated at least the total number of at least one dorm’s members (58 for dorm 1 and 81 for dorm 2), but some of them completed both surveys, with a maximum number of 139 ratings. The raters were presented with photographs of the posters and they were instructed to rate them according to how attractive they found the person in the picture in a scale from 1 to 7. They were also told to use the benchmark average attractiveness in the population at 4. The pictures used were the profile pictures of the posters at the time of the post (if one poster had posted two requests in different times with different pictures, I used the one that showed the person more clearly) and they are not always simple facial pictures. Facebook is a social medium and in many cases members use pictures that could be artistic, show landscapes from trips or other activities of the person. This could create another limitation to our research, since apart from the attractiveness of the person, the attractiveness of the photo should also be taken into consideration, since it might have affected the perception of attractiveness for the raters. Nevertheless, it seems a reliable measure for the purpose of this study, since these pictures are the only visual proxy of the posters’ attractiveness that the commenters have, assuming that they have not met in person before. Thus, attractiveness is used as a broad measure, which includes not only the physical attributes of the depicted person but also other attributes that might reflect personality traits or interests that the raters found attractive. Last but not least, as it is mentioned in the literature review, some pictures also include the body of the posters which signals different messages than the face although face ratings are generally found to be the best predictor (Currie & Little, 2009; Peters, Rhodes, & Simmons, 2007). Descriptive statistics for the attractiveness ratings are available on Table 3.
Overall, the consistency of ratings among the raters is rather high. Cronbach’s Alphas are 0.97 for dorm 1 and 0.94 for dorm 2. This suggest that although the sample of raters was heterogeneous, there seems to be a strong agreement on the attractiveness patterns across the raters.
Table 3
Descriptive Statistics For Attractiveness Ratings
Mean 4.17 Minimum 2.52
Median 4.12 Maximum 5.89
Standard Deviation 0.72 Sample Variance 0.52
Number of observations 139
Table 4
Average Ratings per subjects’ gender
Dorm 1 Dorm 2 Total
Male Female Male Female Male Female
Average
Rating 3.8 4.35 3.77 4.25 3.79 4.28
As can be seen in Table 4, male subjects are systematically given lower ratings in relation to females in both dorms. There are some possible explanations for this phenomenon, but they are hypothetical. For instance, women might be more likely to pay attention to their profile picture than men, in accordance to the general tendency of women to pay more attention to physical appearance than males do. Another explanation could be that raters might be more reluctant to give high ratings to same-sex subjects. For example, in Greece, there are still strong stereotypes about sexual orientation and, thus, Greek male raters are more likely to rate male subjects in a modest way, thus, inflating the equivalent ratings for women. On the other hand, it could always be the case that females in both dorms are generally perceived as more attractive than males.
Figure 1: The vertical axis measures the frequency with which each rating appears in the surveys. In the horizontal axis are the nominal rating values.
Figure 2: The vertical axis measures the averaged rating values. The bars Figure 3: The vertical axis measures the nominal rating values. indicate the average ratings per gender (per dorm and total) The bars indicate the average values for male and female raters (per
dorm and total)
3.4 3.6 3.8 4 4.2 4.4
Male Female Male Female Male Female Dorm 1 Dorm 2 Total
Average Ratings per subjects' gender
0 500 1000 1500 1 2 3 4 5 6 7 Fr eque nc y Rating Points
Histogram of Ratings
0.00 1.00 2.00 3.00 4.00 5.00Male Female Male Female Male Female Dorm 1 Dorm 2 Total
Average Ratings per raters' gender
The research question
So we arrive to our research question: are attractive people more likely to receive help or information by fellow members than less attractive people?
In other words, is behavior in the Facebook pages also driven by taste-based discrimination? Is the requestor’s attractiveness one of the criteria that responders take into consideration before they decide whether to help or not? The main difference between this study and the ones that examine beauty discrimination in other settings - for instance Belot et al. (2008) examine beauty discrimination in a TV show and Mobius & Rosenblat, (2006) do the same for job interviews – is that in our case the nature of the interaction is quite different. Asking for help cannot be compared to the context of cooperating in cases where mutual cooperation is required (such as prisoner’s dilemmas in game theory and economic experiments) or getting hired for a job. All these settings involve long-term interactions between the subjects, but this setting examines beauty premium without offering the possibility to translate the outcomes in economic terms, since the magnitude of helping someone is difficult to be measured in quantitative terms. The purpose of this paper is to expand the presence of beauty discrimination in a different setting. If more attractive members indeed receive more favorable behavior by co-members, then more evidence will be added to the literature beauty discrimination, rendering it a rather systematic bias of behavior. One would expect that since people discriminate in cases where monetary outcomes are at stake, they are likely to discriminate in favor of attractive people in settings with more personal motives.
The setting is ideal to examine these interactions, since all members live in the same building, thus, access to each other is rather easy. Moreover, both the favors and the information requested are quite simple (e.g. simple tools, irons and printers), which means that many inhabitants are expected to be able to help. Trying to disentangle the motives that drive the behavior of the responders, beauty discrimination seems to be a possible criterion that the responders take into consideration before they decide how to act.
The following hypotheses are formed:
Hypothesis1: More than average attractive members will receive comments by more
responders than less attractive members. (Tobit Regression)
Hypothesis 2: More than average attractive people are more likely (in terms of odds) to
receive a response to their request (Logistic Regression)
The variables
In order to construct the model that will allow us to test the hypotheses I created the following variables:
Dependent Variables:
• ComBin: this is a binary variable that reflects whether the posters received a comment for their request (used in the Conditional Logit model)
• ComNumber: this is a continuous variable that reflects the number of comments received for each request (used in the Tobit model)
Variables of Interest (personal attributes): Logit and Tobit Model
• Rating: the average rating of each poster I (the variable of interest)
Tobit Model 2
• Attractiveness Group (AG): As explained in the previous section, the ratings are also grouped in the “Below Average”, “Average” and “Above Average” groups, according to the distance of the rating to the mean (4.17±𝑆𝑆.𝐷𝐷.
2 ). Three dummy variables are created to
represent each group. The Average group will be used as the reference group.
Physical Traits Variables
• Female: binary variable indicating the gender of the poster
• Caucasian2: binary variable indicating whether the poster has white skin color. This
variable was added to test for different attitudes towards members with different skin colors in an effort to control for differences in perception of attractiveness.
Behavioral Trait Variable
• TopCom: this is also a binary variable, which gets the value 1 when the poster is also one of the Top commenters of each of the Facebook pages. The top commenters are the members with a total of at least 4 comments. This number is determined by adding one standard deviation to the total average number of comments and by rounding the number up; 3.83 rounded up to 4 (see Table 2). For each poster that has a total of 4 comments, the variable TopCom takes the value 1. It captures a behavioral aspect of the Posters in terms of activity in the page (in contrast to the previous variables of interest that correspond the physical attributes.)
Control Variables:
• Dorm: binary variable that indicates in which dorm the poster leaves (and thus in which page he/she is active)
• Object/Favor: binary variable indicating the nature of the request. It takes the value 1 when the request is about favors or a material object, and 0 when the request is about information,
• Peak: this variable is derived by the histogram of comments’ frequency in an hourly basis3. Each page had different times during the day that (on average) more comments
were posted. I decided to control for this effect, since requestors that posted during the peak hours of comment might have more odds of getting an answer, irrespective from their physical appearance.
To test Hypothesis 1, I run a Tobit regression, where ComNum (Number of Commenters per post) is used as the dependent variable. Tobit is a censored regression model and it is called like this because the dependent variable has been “censored” below a certain cut-2 There seems to be a discrepancy between the technical (anthropological) term and the way that the
term is used nowadays. More specifically, according to Dictionary.com, the scientific term includes “characteristic of one of the traditional racial divisions of humankind, marked by fair to dark skin, straight to tightly curled hair, and light to very dark eyes, and originally inhabiting Europe, parts of North Africa, western Asia, and India”. Nevertheless, the established use of the term reflects just white people and this is the way it is used in my analysis. (Source: Dictionary.com)
3 Histograms and other statistics for the activity in the Facebook Pages were acquired by Grytics.com.
16
off. In other words, the distribution of the number of commenters follows a discrete distribution at zero and a continuous distribution. The model supposes that there is a latent (or unobservable) variable𝑦𝑦𝑖𝑖∗. This variable linearly depends on the variable of interest via the beta coefficient which determines the relationship between the 𝑥𝑥𝑖𝑖 and the latent variable 𝑦𝑦𝑖𝑖∗ just as in a linear model (Mcdonald & Moffitt, 1980). The variable 𝑦𝑦𝑖𝑖 is defined to be equal to the latent variable 𝑦𝑦𝑖𝑖∗ whenever the latent variable is above zero and zero otherwise. The model could also be presented in the following way:
𝑦𝑦𝑖𝑖=
This censoring property of Tobit regression renders it an appropriate specification strategy for the analysis, since the dependent variable has 30% of its values clustered at zero. The OLS regression produces inconsistent coefficients, which is fixed by the methodology introduced by Tobin in 1958 (Amemiya, 1979). It is commonly used for analyzing data with large number of zeroes. Anastasopoulos, Shankar, Haddock, & Mannering (2012) use a multivariate a left censored Tobit regression with clustering at zero to estimate accident rates of specific injury-severity levels. This is because accidents of a specific injury severity may not be observed either because simply none have occurred or because data is not available, because accidents that cause low injury severity do not get reported if no material damage is also caused. One can find resemblance with the Facebook pages’ case, where some request either remain unanswered or receive responses through private messages; in both cases the dependent variable gets the value zero in both cases. In Tobit regression, the equivalent effect is assumed to be the same, clustered at zero-observations. With this regression I want to test if requests posted by attractive people activate more members to comment and receive increased response rate
𝑦𝑦𝑖𝑖∗ If 𝑦𝑦𝑖𝑖∗> 0
0 if 𝑦𝑦𝑖𝑖∗≤ 0
The Tobit model is the following:
𝑦𝑦𝑖𝑖 = 𝑎𝑎𝑖𝑖+ 𝛽𝛽1𝑅𝑅𝑎𝑎𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑖𝑖 + 𝛽𝛽2𝐹𝐹𝐹𝐹𝐹𝐹𝑎𝑎𝐹𝐹𝐹𝐹𝑖𝑖 + 𝛽𝛽3𝐶𝐶𝑎𝑎𝐶𝐶𝐶𝐶𝑎𝑎𝐶𝐶𝑅𝑅𝑎𝑎𝑅𝑅𝑖𝑖 + 𝛾𝛾1𝑇𝑇𝑇𝑇𝑇𝑇𝐶𝐶𝑇𝑇𝐹𝐹𝑖𝑖+
+𝛿𝛿1𝐷𝐷𝑇𝑇𝐷𝐷𝐹𝐹𝑖𝑖 + 𝛿𝛿3𝑃𝑃𝐹𝐹𝑎𝑎𝑃𝑃𝑖𝑖+ 𝐹𝐹𝑖𝑖
where 𝑦𝑦𝑖𝑖 is the number of different commenters under each post-request 𝑅𝑅 and 𝐹𝐹𝑖𝑖 is the error term.
In the second Tobit model, the attractiveness rating is replaced by the Attractiveness Group (AG) as the independent variable.
Our variable of interest in the Logit model is the attractiveness rating and Hypothesis 2 is tested in the following way; I estimate a logistic regression model for the odds of getting a response, using ComBin (whether there was a response or not) as the dependent variable. In logistic regressions, the coefficients of the variables are expressed in odd ratios. The coefficients estimated are the Maximum Likelihood Estimators and are the parameter values most likely to have produced the data (H. Stock & W. Watson, 2010). The logistic regression model takes the natural logarithm of the odds as a regression function of the independent variables. With one variable of interest X, this takes the form ln[𝑇𝑇𝑜𝑜𝑜𝑜(𝑌𝑌 = 1)] = 𝑎𝑎 + 𝑏𝑏1𝑋𝑋, where ln stands for the natural logarithm, Y is the outcome
and Y=1 when the event happens, 𝑎𝑎 is the intercept term, and 𝑏𝑏1 represents the regression coefficient, the change in the logarithm of the odds of the event with a 1-unit change in the predictor X. A useful way to think of the odds ratio is that 100 times the odds ratio minus 1, i.e., 100X(odds ratio -1), gives the percent change in the odds of the event corresponding to a 1-unit increase in X. If this value is negative, then the odds of the event decrease with increasing values of X; if positive, the odds increase. (LaValley, 2008). These ratios can also be interpreted in the following way. Let us assume that we want to estimate the odd ratios of success in a certain condition. If the odd ratio is 4, then
Physical Traits Behavioral Trait
Control Variables
one can argue that the odds of success are 4 to 1. On the other hand, if the odd ratio is 0.25, the odds of success is 1 to 4.
Logistic regression is suitable for situations where the dependent variable is not a continuous, but a categorical one, such as ComBin, and they are mainly used to ascertain the probability that an event will happen. Using a categorical outcome variable violates the assumption of linearity in linear regression. Logistic regression deals with this problem by using a logarithmic transformation on the outcome variable which allow to model a nonlinear association in a linear way.
As mentioned before, I also added a behavioral trait, which captures the effect of increased activity of the individual poster also as a commenter (thus, a member who is often willing to provide help or information to others).
4. Results Initial Analysis
We first test Hypothesis 1 by comparing the average number of commenters under each post per Attractiveness Groups (AG). We divide the attractiveness ratings in 3 major groups: Below Average, Average and Above Average (attractiveness). In order to create the groups, I added and subtracted from the mean of the ratings (4.17) half of the Standard Deviation (0.72
2 ). Every member with an average rating less than 3.81 is grouped
as below average, with a rating above 4.53 is grouped as above average and in the interval between as average attractive. The equivalent mean numbers of commenters (per post) per AG are 1.2, 1.4 and 1.2 respectively. After testing the differences between the means, no difference between the groups, in all three possible combinations, are significantly different from zero4. It seems that the Average group activates more participation on behalf of the responders, but still the difference is not significant enough to support that this effect it is related to posters’ attractiveness.
4 The p-values from the independent group t-test were much higher than the critical values. The
independent group t-test compares the means of two groups for the same variable, which the number of commenters per post. These values range from 0.48 to 0.89.
19
Figure 4: The vertical axis measures the average number of commenters per post uploaded. The bars on the horizontal axis indicate the average values for the different AGs (Below Average, Average and Above Average)
Figure 5: The vertical axis measures the percentage of posts that received at least one response. The bars on the horizontal axis indicate the values for female and male members.
When it comes to gender differences, males received on average at least one response in more posts than females did (35% compared to 25%). The average number of commenters per post is 1.4 for male and 1.2 for female posters. The difference between the means is not significant; the p-value is 0.3, thus the null hypothesis cannot be rejected. When it comes to the dorms, the equivalent means are 1.3 and 1.2. Again, the
1.1 1.2 1.2 1.3 1.3 1.4 1.4
Below Average Average Above Average
Histogram of average number of commenters (per post)
per AG
0% 10% 20% 30% 40% Female MalePercentage of posts that received at least one response
(per gender)
difference between the means is not significant, suggesting similar response rates in both of them5.
The Tobit Model
I start the analysis with the Tobit regression. Tobit regression coefficients are interpreted in a similar manner to OLS regression coefficients; however, the linear effect is on the uncensored latent variable, not the observed outcome. The regression confirms the initial analysis results; AGs below and above average receive on average less comments than the Average group (reference group omitted to avoid omitted variable bias), but the significance of the coefficients are rather low and not significant. When adding more variables, the “Below Average” variable’s coefficient further declines. One can notice a disproportionate negative “jump” after the physical traits variables are introduced (Caucasian and Female). The same happens with the inclusion of the control variables. The opposite happens to the “Above Average” variable’s coefficient after the inclusion of the control variables. One can notice that the coefficient increases, approaching zero which refers to the group “Average”. As the explanatory power of the model increases, the two groups (below and above average) follow different directions. The “Above Average” converges with the “Average” group and the discrepancy between them, in terms of the number of commenters, narrows. The behavioral variable “Top Commenter” does not influence the coefficient of the variables of interest, suggesting that it is weakly correlated both to attractiveness and response rates.
What is interesting is that although females account for around 70% of the posts, the coefficient of the Female variable is negative (although not significant). Recall, also, that females consistently received higher ratings by the independent raters, which makes the sign of the coefficient even more surprising. This case will be further analyzed at the Logit regression section.
After looking at the Chi Squared values it becomes more obvious which one is the main force that drives behavior by the responders; the value increases greatly after the control variables are added and the coefficient of the Object/Favor variable clearly shows that responders prefer information posts to favors. The Chi-Square tests whether at least one 5 Results derived from independent group t-test. An independent group’s t-test is used when one
wants to compare the means on a dependent variable (such as attractiveness ratings) for two independent groups (dorm 1 and 2 in this case).
21
of the predictors' regression coefficients is not equal to zero. The critical value of the chi square distribution with 8 degrees of freedom is 18.467 which is smaller than 29.024. This suggests that at least one of the coefficients is statistically different from zero, which in our case is the Object/Favor coefficient. Logic wise it makes perfect sense due to the effort required on behalf of the responder; information is easy to provide without any other involvement, which is not the case for favors or material objects.
Continuing the analysis, I run another Tobit regression, replacing the AG variables with the nominal averaged attractiveness scores. The results are quite similar. Again, the attractiveness ratings seem uncorrelated with the number of commenters involved in the posts and the model’s explanatory power is limited until the control variables are introduced. The Object/Favor variable coefficient is once more the only significant and it raises the chi-squared value disproportionally a lot compared to the rest of the variables. The rest of the variables seem to be uncorrelated with the response rates.
Table 5
Number of Commenters (per post) per Attractiveness Group
Attractiveness (1) Physical Traits (2) Behavioral Trait (3) Control Variables (4) Attractiveness Groups Below Average -0.153 -0.242 -0.238 -0.293 (-0.27) (-0.28) (-0.28) (-0.26) Above Average -0.214 -0.227 -0.234 -0.119 (-0.29) (-0.29) (-0.29) (-0.28) Female -0.264 -0.273 -0.223 (-0.27) (-0.27) (-0.26) Caucasian -0.247 -0.253 -0.25 (-0.35) (-0.35) (-0.33) Top Commenter 0.162 0.201 (-0.27) (-0.26) Object/Favor -1.223*** (-0.24)
Peak Comment Hours 0.217
(-0.24) Dorm 1 0.312 (-0.22) Constant 1.030*** 1.466*** 1.441*** 2.063*** (-0.19) (-0.4) (-0.4) (-0.43) Chi Squared 0.609 2.348 2.699 29.024 N 261 261 261 261
Notes: dependent variable: number of commenters per post per AG; coefficients of Tobit regression; * p<0.05, ** p<0.01, *** p<0.001
Table 6
Number of Commenters (per post) per Attractiveness Rating Attractiveness (1) Physical Traits (2) Behavioral Trait (3) Control Variables (4) Attractiveness Rating 0.041 0.103 0.103 0.202 (-0.16) (-0.17) (-0.17) (-0.16) Female -0.265 -0.276 -0.251 (-0.27) (-0.27) (-0.26) Caucasian -0.244 -0.251 -0.252 (-0.35) (-0.35) (-0.33) Top Commenter 0.162 0.21 (-0.27) (-0.26) Object/Favor -1.252*** (-0.24)
Peak Comment Hours 0.198
(-0.24) Dorm 1 0.298 (-0.22) Constant 0.748 0.892 0.867 1.143 (-0.67) (-0.72) (-0.72) (-0.69) chi2 0.066 1.772 2.121 29.384 N 261 261 261 261
Notes: dependent variable: number of commenters per post; coefficients of Tobit Regression; Standard errors parentheses; * p<0.05, ** p<0.01, *** p<0.001
The Logit model
The specification used for Hypothesis 2 is the logistic regression model. As can be seen in Table 7, in the first two regressions including the attractiveness rating and the physical traits, the correlation between getting a response and physical appearance seems to be positive but weak. Nevertheless, when the control variables are also taken into consideration in the model, both the size and the significance of the Attractiveness variable increase raising the odds at almost 1.5 times, but still the effect is not significant enough to reject the null hypothesis.
Also, when asking for favors or material objects, the odds of getting a response are again much lower than when requesting for information. This confirms the results of the Tobit regressions, according to which the main driver of the responders’ behavior is the nature of the request and not the physical or behavioral attributes of the poster. This is illustrated once more from the chi-squared values, which increase significantly after the inclusion of the control variables, and more specifically the Object/Favor one, are included.
Female members are once more less likely to receive a response, although they are more likely to upload a post in the first place, but this time the effect is significant. A possible explanation is that women generally ask for more favors than males do. 71% of the female members asked for a favor, which is 132 out of a total of 184 requests, while the equivalent percentage for men is 68%, with 53 posts out of 77. Since women ask more often for favors, and since favors are less likely to be offered, the coefficient of the Female variable could be a self-evident sequence. Nevertheless, the fact that after the variable “Object/Favor” is included, the coefficient of the variable “Female” remains relatively stable, indicates that they are not correlated.
Figure 8
From Figure 8, it can be derived that attractive females are more likely to receive a response by at least one fellow member relative to average and below average attractive females. This means, that the pattern of behavior tested in Hypothesis 2 is consistent for females; the more attractive, the more likely it is to receive at least one response, but still the odds are lower compared to male members. This is irrelevant to attractiveness and it
0% 5% 10% 15% 20% 25% 30% 35%
Below Average Average Above Average Females
Percentage of posts that received a response by at least on responder per AG (Females)
is a gender-specific effect, which suggests that males are just more likely to receive a response in our setting in general.
I decided to run an extra regression, introducing an interaction term for the beauty rating and the nature of the request. The results for the rest of the variables are quite similar and the interaction term has a positive (but not significant) coefficient. This implies that these two variables probably do act complementarily; one unit increase in attractiveness increases the odds of getting a response if the post is about favor, holding other factors constant. If the sample size was bigger, then effect might have been more significant.
Table 7
Odds of getting a response
Attractiveness (1) Physical Traits (2) Behavioral Attribute (3) Control Variables (4) Attractiveness Rating 1.077 1.223 1.223 1.499 (-0.37) (-0.96) (-0.95) (-1.7) Female 0.569 0.563 0.497* (-1.79) (-1.82) (-1.97) Caucasian 0.793 0.777 0.728 (-0.58) (-0.62) (-0.69) Top Commenter 1.207 1.52 (-0.57) (-1.12) Object/Favour 0.168*** (-5.56) Peak 0.866 (-0.41) Dorm 1 1.036 (-0.11) chi2 N 0.07 261 1.77 261 2.12 261 29.38 261
Notes: Logistic Regression; Exponentiated coefficients expressed in Odd Ratios; t statistics in parentheses; p<0.05, ** p<0.01, *** p<0.001
Results
All in all, the nature of the request is the main driver of behavior. As mentioned before, taking into consideration the amount of effort required on behalf of the responder, the outcome of the analysis is reasonable. Favor-oriented posts demand more engagement than the information posts, since the responder has to meet with the requestor, spend time to help him or take the risk of borrowing something to someone who might not be very intimate with.
Attractiveness seems to be weakly correlated with the response rates and the average looking members are the ones with more odds of “getting what they want” compared to below and above average attractive people. The interaction term between the nature of the post and the attractiveness rating is positive, which implies that these two variables might work in complementary way, but the results are not significant enough to support this theory.
Females are less likely to receive a response compared to males, but the effect is not significant. As mentioned above, this could be due to the sample size, which is not big enough to provide the desired internal validity of the results. Nevertheless, in an intra-group level analysis, more attractive women are more likely to receive a response, confirming Hypothesis 2 (at least for females) but with rather weak evidence.
5. Discussion Limitations
The main limitation of this paper is the fact that the personal messages cannot be controlled for. If responders choose to contact the posters in a private way is something that cannot be identified due to privacy reason, which renders feasible the scenario that more interaction has taken place in the Facebook pages setting. If the responder’s goal is to actually meet the requestor, in case he/she is attracted by him/her, then it is more likely that a personal message will be sent. Moreover, since the comments are visible by everyone in the group, if a very attractive person requests for something, some members might prefer PMs so that they to do not reveal publicly their intentions. The fact that females receive on average fewer responses may be a proxy for this hypothesis, since it would make sense if male responders approached attractive females on a personal level. But these are just hypothetical scenarios; according to the data collected, the nature of the
request is the main predictor of responders’ behavior and it is a rational outcome: the less effort required, the more likely it is for someone to respond.
A second important limitation is the sample size; the total number of 261 posts might not be enough to capture the correlation between response rates and attractiveness. If more observations were included the coefficients and odd ratios of the Ratings and the AG might have been different, since the explanatory power of the model increases substantially after the inclusion of the control variables. I could not acquire data from more student dorm due to privacy reasons; my request to gain access to relevant Facebook pages was rejected four times, because the administrators did not feel comfortable with the idea that other members’ data would be available to someone they do not know.
It would also be insightful if the activity of members without a profile picture was also included. Less attractive people are less likely to have a profile picture, causing an endogenous self-selection effect that biases the results. Not including a profile picture might signal to other members that the person is unattractive. If these members actually receive fewer responses compared to AGs included in the analysis, it would strengthen the conclusion that unattractive people are less likely to receive help. It would be insightful to compare the overall activity of members with and without profile pictures in the Facebook pages, both from the side of posters and commenters and see how they behave and if it is more likely for them to post or to respond. Excluding these members means that the pool of unattractive members might not be representative of the population, making any generalizations based on these results arbitrary.
All in all, the external validity of the outcomes is rather limited. Let us not neglect the fact that the sample is only two student dorms in Amsterdam that for several reasons might be different to each other. For instance, only one of the dorms has a common room where inhabitants can socialize. This offers the opportunity to students to meet more people and form stronger bonds among them, whereas in the other dorm inhabitants from different floors might have never met. Apart from this specific example, the way that people choose to socialize might differ from dorm to dorm, due to demographical reasons such as country of origin, age, gender ratio and cultural differences among others.
Further Research
Facebook and social media in general are complex and dynamic settings, where there is substantial diversity in human behavior and its driving forces and incentives. Although attractiveness probably is not one of the leading forces (at least in the dorms’ pages), it would be very useful and insightful to identify the strongest ones. For instance, it is possible that the behavioral trait examined here, which is the activity of members as commenters or how much these members are willing to help, might be altered if the top engaged members were visible to everyone. If there was a separate section somewhere in the page where these members were presented, or if a reputational status was introduced, the significance of the Top Commenter variable could be substantially larger. Introducing a reputation status or a bracket with the top engaged members could be a quite good opportunity for an A/B test, through which comparisons could be made for the activity in each group. If half members were introduced to these reputational mechanisms, unbeknownst to them, while the other half would continue using the page as it is, useful insight could be acquired.
Overall activity in Facebook could also be used as a proxy in our case; the Facebook members that in general use Facebook extensively may be more likely to comment or post anyway, because they use Facebook as a medium to socialize in the first place. They have more chances of seeing the posts earlier and also more likely to respond, if they are used to using Facebook also for online purchases or participate in other groups or pages, where similar interactions take place.
Last but not least, if the main limitation, i.e. the personal messages, could be controlled for, the results would be undeniably more robust. The fact that above average attractive members receive on average fewer responses than the average looking ones is suspicious and unfortunately no clear relevant answer can be given. Access to such information probably will never be available due to privacy reasons, but the connection between such pools of information concerning human behavior and universities should evolve closer and stronger in the future. In Facebook all information is coded and members have a unique ID. The only reason that someone might need access to the users’ profile would be if they examined physical attributes of them; otherwise anonymity can be maintained and it would be no use to the researcher to breach this moral contract.
References
Amemiya, T. (1979). The estimation of a simultaneous-equation Tobit model.
International Economic Review, 20(1), 169–181. http://doi.org/10.2307/2526423
Anastasopoulos, P. C., Shankar, V. N., Haddock, J. E., & Mannering, F. L. (2012). A multivariate tobit analysis of highway accident-injury-severity rates. Accident;
Analysis and Prevention, 45, 110–9. http://doi.org/10.1016/j.aap.2011.11.006
Ashmore, R. D., Makhijani, M. G., Longo, L. C., & Eagly, A. H. (1991). What is beautiful is good, but . . .: A meta-analytic review of research on the physical attractiveness stereotype. Psychological Bulletin, 110(1), 109–128.
http://doi.org/10.1037/0033-2909.110.1.109
Belot, M., Bhaskar, V., & van de Ven, J. (2008). Beauty and the sources of discrimination, 47(3), 851–872. http://doi.org/10.1353/jhr.2012.0020
Currie, T. E., & Little, A. C. (2009). The relative importance of the face and body in judgments of human physical attractiveness. Evolution and Human Behavior, 30(6), 409–416. http://doi.org/10.1016/j.evolhumbehav.2009.06.005
Doorley, K., & Sierminska, E. (2012). Myth or Fact ? The Beauty Premium across the
Wage Distribution. Iza, 129(6674), 29–34.
http://doi.org/10.1016/j.econlet.2015.01.033
Feingold, A. (1992). Good-looking people are not what we think. Psychological
Bulletin. http://doi.org/10.1037/0033-2909.111.2.304
H. Stock, J., & W. Watson, M. (2010). Introduction to Econometrics (3rd Edition) (Addison-Wesley Series in Economics). Addison-Wesley, 1, 840.
Hamermesh, D. S., & Biddle, J. E. (1994). Beauty and the Labor Market. The
American Economic Review, 84(5), 1174–1194. http://doi.org/10.2307/2117767
Heilman, M. E., & Saruwatari, L. R. (1979). When beauty is beastly: The effects of appearance and sex on evaluations of job applicants for managerial and nonmanagerial jobs. Organizational Behavior and Human Performance, 23(3), 360– 372. http://doi.org/10.1016/0030-5073(79)90003-5
Landry, C. E., Lange, A., List, J. A., Price, M. K., & Rupp, N. G. (2006). TOWARD AN UNDERSTANDING OF THE ECONOMICS OF CHARITY: EVIDENCE FROM A FIELD EXPERIMENT. Quarterly Journal of Economics, 121(2), 747–782. http://doi.org/10.1162/qjec.2006.121.2.747
LaValley, M. P. (2008). Logistic regression. Circulation, 117(18), 2395–9. http://doi.org/10.1161/CIRCULATIONAHA.106.682658
Mcdonald, J. F., & Moffitt, R. A. (1980). The Uses of Tobit Analysis. The Review of
Economics and Statistics, 62(2), 318–321. http://doi.org/10.2307/1924766
Mobius, M. M., & Rosenblat, T. S. (2006a). Why beauty matters. American Economic
Review, 96(1), 222–235.
Mobius, M. M., & Rosenblat, T. S. (2006b). Why beauty matters. American
Economic Review, 96(1), 222–235. http://doi.org/10.1257/000282806776157515
Peters, M., Rhodes, G., & Simmons, L. W. (2007). Contributions of the face and body to overall attractiveness. Animal Behaviour, 73(6), 937–942.
http://doi.org/10.1016/j.anbehav.2006.07.012
Pfann, G. a, Biddle, J. E., Hamermesh, D. S., & Bosman, C. M. (2000). Business success and businesses’ beauty capital. Economics Letters, 67(2), 201–207. http://doi.org/10.1016/S0165-1765(99)00255-4
Robins, P. K., Homer, J. F., & French, M. T. (2011). Beauty and the Labor Market: Accounting for the additional effects of personality and grooming. Labour, 25(2), 228–251. http://doi.org/10.1111/j.1467-9914.2010.00511.x
Rosenblat, T. S. (2008). The beauty premium: Physical attractiveness and gender in dictator games. Negotiation Journal, 24(4), 465–481. http://doi.org/10.1111/j.1571-9979.2008.00198.x
Schuessler, K., & Becker, G. S. (1958). The Economics of Discrimination. American
Sociological Review. http://doi.org/10.2307/2088646
