the e ects of student access to social networking …web.stanford.edu/~larmona/fb_proposal.pdfon...

22
The Effects of Student Access to Social Networking Sites on Labor Outcomes Luis Armona [email protected] June 13, 2017 Abstract This proposal analyzes the impact of access to social networking sites on future labor outcomes. Specifically, I study the effect of availability of facebook.com to university students during their postsecondary education on both employment and earnings 6-10 years after they first entered school. Using the differential timing of the rollout of the Facebook network to colleges during the site’s infancy, combined with panel regression methods, I can identify the causal impact of time exposed to Facebook while in school on later outcomes. I show reduced form regression results that identify substantial effects on student outcomes. I then introduce a 2-stage structural model of social networking in school followed by a labor market on the predetermined friendship graph that may be able to tease out more nuanced effects of the social network exposure. 1 Introduction Few Internet websites that exist today are as uniquitously known as Facebook.com. Facebook is a social networking site (SNS) users visit to share information and media with their online friends, As of November 2016, 68% of all adults in the United States are active Facebook members. 1 . It provides users with an extremely low-cost way to keep in touch with friends and acquaintances, even if they are thousands of miles away. This proposal examines the potential effects of student access to Facebook in postsec- ondary institutions on the labour outcomes of these individuals later on in life. It is well documented that “weak ties”, relationships with those who are informal acquaintances but not close friends, are key in providing individuals with employment opportunities [7]. Face- book as a platform offers an easy-to-use, widely adopted electronic social network for people across the world to remain in contact with others they have previously met. In theory, this social network should decrease the costs of maintaining weak ties dramatically, since users can communicate within the site rather than having to track down an address, physical or electronic, that may no longer be valid. This cost reduction should be particularly salient for those who adopted Facebook during its infancy, around 2004 when the website first began. 1 Source: http://www.pewinternet.org/2016/11/11/social-media-update-2016/ 1

Upload: others

Post on 10-Jul-2020

1 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: The E ects of Student Access to Social Networking …web.stanford.edu/~larmona/fb_proposal.pdfon later outcomes. I show reduced form regression results that identify substantial e

The Effects of Student Access to Social NetworkingSites on Labor Outcomes

Luis [email protected]

June 13, 2017

Abstract

This proposal analyzes the impact of access to social networking sites on future laboroutcomes. Specifically, I study the effect of availability of facebook.com to universitystudents during their postsecondary education on both employment and earnings 6-10years after they first entered school. Using the differential timing of the rollout of theFacebook network to colleges during the site’s infancy, combined with panel regressionmethods, I can identify the causal impact of time exposed to Facebook while in schoolon later outcomes. I show reduced form regression results that identify substantialeffects on student outcomes. I then introduce a 2-stage structural model of socialnetworking in school followed by a labor market on the predetermined friendship graphthat may be able to tease out more nuanced effects of the social network exposure.

1 Introduction

Few Internet websites that exist today are as uniquitously known as Facebook.com. Facebookis a social networking site (SNS) users visit to share information and media with their onlinefriends, As of November 2016, 68% of all adults in the United States are active Facebookmembers. 1. It provides users with an extremely low-cost way to keep in touch with friendsand acquaintances, even if they are thousands of miles away.

This proposal examines the potential effects of student access to Facebook in postsec-ondary institutions on the labour outcomes of these individuals later on in life. It is welldocumented that “weak ties”, relationships with those who are informal acquaintances butnot close friends, are key in providing individuals with employment opportunities [7]. Face-book as a platform offers an easy-to-use, widely adopted electronic social network for peopleacross the world to remain in contact with others they have previously met. In theory, thissocial network should decrease the costs of maintaining weak ties dramatically, since userscan communicate within the site rather than having to track down an address, physical orelectronic, that may no longer be valid. This cost reduction should be particularly salient forthose who adopted Facebook during its infancy, around 2004 when the website first began.

1Source: http://www.pewinternet.org/2016/11/11/social-media-update-2016/

1

Page 2: The E ects of Student Access to Social Networking …web.stanford.edu/~larmona/fb_proposal.pdfon later outcomes. I show reduced form regression results that identify substantial e

At that time, the few other SNSs that were available, such as Friendster and Myspace, ei-ther were subsequently shutdown or experienced a massive decrease in traffic / active users.Comparatively, Facebook today is the largest SNS in the world, with almost 2 billion usersworldwide. In this sense, those who adopted Facebook early would have gotten “lucky” andfriendships added on this early network would still be accessible today since in all likelihoodthose who adopted early are still active users. Differences in access to Facebook in this timeperiod would be akin to variation in access to any electronic social networking technology,when examining the impact on longer horizon outcomes, since no other social network main-tained relevance throughout the period of interest. For this reason, this proposal’s researchsubject concerns more than just what the impact of the website Facebook specifically is, butrather what the impact of access to social networking technology as a whole can do to labouroutcomes. In this proposal, I examine the effect of Facebook adoption by college students,who can use the SNS to document and maintain friendships made while in school, on laboroutcomes later in life. The central hypothesis of the proposal is that Facebook adoption willresult in easier maintenance of friendships, particularly weak ties, which could then resultin increased employment opportunities, thereby having a positive effect on employment andwages for students who adopted this technology, compared to those who either did not orwere unable this technology.

This proposal uses the differential timing of access to Facebook, accomplished via regis-tration tied to .edu email addresses, to identify the effect of Facebook access/adoption onlabour outcomes 6 to 10 years after students first enrolled in college. Due to conditional exo-geneity of this timing, I am able to identify causal effects of Facebook access from regressionmethods. However, I do not have data on Facebook adoption rates, so I can only look at theeffects of access, or the ability to sign up for Facebook. There is large evidence that diffusionrates of Facebook within my sample were substantial. Jacobs et al. [9] look at adoptionrates in the same sample of schools I examine 1-2 years after the introduction of Facebook,and find adoption rates in the neighborhood of 80-100%. This suggests that adoption wasnear-universal and the relevance of the Facebook network in campus life would have beenconsiderable. This provides evidence that Facebook access should be a good proxy for Face-book adoption, as it is unlikely some schools were given access but did not begin adoptionuntil a later date. In a subsequent section, I briefly consider a structural model that drawson the same Facebook data used by Jacobs et al. to endogenize decisions to form friendshipson the Facebook network, and how that would effect labour outcomes later in life. But fornow, I restrict my attention to the reduced-form / intent-to-treat effect of access to Facebookon labour outcomes, with the understanding that near-universal adoption of Facebook onceit was introduced means the access measure should be a good proxy for actual adoption.

There is a broad literature dissecting both the structure of the Facebook social networkand the relevance of social networks on real-world outcomes. In a series of papers [14] &[15], Traud and coauthors examine the structure of the Facebook social networks of theschools in our sample. They document high clustering on observables such as college dormand major, and short diameters in these networks, consistent with the general attributes ofsocial networks described in Jackson [8], suggesting these networks can be reflections of theactual social networks describing relationships between students on campus. They also findsignificant heterogeneity in network structure between universities. Ugander et al. [16] per-form a large-scale study of the entire Facebook social graph in 2011, and confirm that these

2

Page 3: The E ects of Student Access to Social Networking …web.stanford.edu/~larmona/fb_proposal.pdfon later outcomes. I show reduced form regression results that identify substantial e

attributes of local density / low average distance persists in the structure of the network onceit was a widely adopted technology used by the general U.S. population. In regard to labouroutcomes, Calvo [1] is one of the first papers to examine the implications for labor outcomeswhen they are based on social network structures. In a simple toy model of homogeneousagents and employment offers diffusing to neighbors , he finds that employment rates areincreasing in one’s degree, but decreasing in the degree of one’s neighbors, suggesting bothcomplementary and substitability effects of the network on employment prospects. In asymmetric setting, aggregate employment rates on the networks display an inverse U shape,so there is an “optimal” degree to maximize employment prospects of agents in the network.Calvo & Jackson [3] add an inequality component to this employment model. Calvo & Jack-son [2], enrich the model with wages and prove in their model positive correlation of wagesbetween agents connected on the network, and that their model can generate significantlong-run inequality of wages between similar agents if they are given different sets of initialconditions. Empirically, Lachever [10] finds that post-war employment outcomes are posi-tively correlated among those who served together in the same military groups. Enikolopovet al. [5] uses an identification strategy similar to mine that shows access to an SNS in Russiaincreased participation on collective action. Mayer & Puller [12] estimate a structural modelof friendship formation using Facebook data to demonstrate the inefficacy of university in-tervention to limit segregation in student social networks . Finally, Jacobs et al. [9] looks atthe Facebook network structure of same sample of schools considered in this proposal, to tryand establish global trends in network formation across these school-specific social networks.They conclude the vast heterogeneity in the networks suggests that the Facebook networksat these schools are not being drawn from a single network assembly process, but also con-clude the school-level Facebook networks obey some the classical relationships between size,distance, and degree distribution that are documented in social networks.

2 Data

This research proposal primarily draws on 3 datasets. Throughout, my sample consists ofthe incoming freshmen cohorts from 1999-2004 at the 100 first universities to which Facebookwas made available. These 100 universities are the only universities for which I have data onthe introduction timing of Facebook networks. The 1999-2004 sample was chosen becauseit encompasses all years for which there would be any variation in the treatment (time inuniversity with Facebook access), since before 1999 no individuals had access to Facebookduring their entire college tenure in our sample, and all freshmen cohorts after fall 2004had complete access at these universities. In addition, 1999 is the first year we have all thenecessary data to estimate our model.

The first dataset is employment and earnings data by university from the U.S. Depart-ment of Education’s College Scorecard Data (CSD), available for public usage at https://collegescorecard.ed.gov/data/.This dataset consists of cohort-level data on the earnings distribution (conditional on em-ployment) of students who took any federal aid (either direct grants or through subsidizedloans) via their tax returns 6 to 10 years after the year they first enrolled in college. Theseare the students who participated in so called “Title IV” federal programs while at the uni-versities they attended. Moments provided on the earnings distribution include percentiles

3

Page 4: The E ects of Student Access to Social Networking …web.stanford.edu/~larmona/fb_proposal.pdfon later outcomes. I show reduced form regression results that identify substantial e

(10th,25th,50th,75th,90th), means, and standard deviations (all of these are from the dis-tribution of earnings conditional on employment). In addition, they provide informationon the overall employment rate of the cohort for those who are currently not enrolled in apost-secondary education institution 2. Unfortunately, due to confidentiality concerns, theCSD pools together cohorts so that the data are reported on a bi-annual basis for the co-horts who entered the university in academic years t and t + 1. That is, each observationin the CSD earnings data are the moments for the earnings distribution of both those whoentered the school in year t and t + 1. This makes our identification of isolating the effecton a specific cohort from Facebook exposure significantly more difficult. I attempt varioustechniques later in the proposal to recover the original effects on single cohorts.

The next dataset used is a supplementary dataset directly downloaded from the Depart-ment of Education’s IPEDS3 website which provided supplementary information on Title IVstudents in the universities in our sample. Specifically, I use the number of students in eachannual cohort from first-time, full-time students that recieved any federal aid to approximatethe share of students from each cohort in the bi-annual earnings data. The only reason thesestudent counts might differ from the true share of students in each cohort in the earningsdata is:

1. Some students’ information was unable to be located at the time of tax filings. Itseems unlikely that there would be noticeable, systematic reasons the weights wouldbe biased in relation to this missing data problem.

2. Some students’ in the earnings data were enrolled part time at the university. Whilethis may be concerning for affecting our inference on the earnings distribution, sincepart-time students are likely to have different earnings, the 100 schools in our samplehave a relatively high full-time enrollment rate, as they are mostly universities andpart-time enrollment is much more of a phenomenom at state schools and communitycolleges. In our sample period of cohorts entering in the fall of 1999 through 2006, theaverage part time enrollment rate is only 5.4%, according to IPEDS.

Otherwise, uage of this data seems largely appropriate for inferring the relative distributionof each cohort in the pooled cohort data.

The final dataset used is the list compiled by Jacobs et al. [9] which lists the date ofintroduction of Facebook into each of the first 100 universities Facebook rolled out to. Inits inception, Facebook originally was hailed as an “exclusive” network that only allowedindividuals with .edu email addresses at invited universities to register with the network,and it was primarily designed to interact with one’s fellow students at their specific university.The Jacobs data contains the date students with .edu addresses corresponding to each ofthe first 100 universities Facebook was opened too were allowed to register on the website. Italso contains an uncertainty range in days for values that were imputed based on knowledgeof the rollout to be within a certain range of dates. The data is shown in Table ??.

2These individuals who are currently enrolled in a post-secondary education institution at the time offiling taxes 6-10 years later are technically not employed, but this is not necessarily a result of poor labormarket outcomes, so the CSD excludes them from employment rate calculations

3downloaded from here https://nces.ed.gov/ipeds/home/usethedata

4

Page 5: The E ects of Student Access to Social Networking …web.stanford.edu/~larmona/fb_proposal.pdfon later outcomes. I show reduced form regression results that identify substantial e

3 Reduced-Form Model

My model is that a cohort-level statistic S of the labor outcome random variable Y for students at university i for cohort t, h years after cohort t enrolled in the university is impacted bythe amount of years the cohort was exposed to the Facebook social network (denoted by τ),after controlling for fixed effects at the university, national cohort,and horizon level:

Si,t,h = αi + δt + γh + βτi,t + εi,t,h (1)

For example, Ys,i,t,h might be the wage earnings of a student from cohort t of universityi, h = 6 years after they first enrolled, and S might be the mean: Si,t,h = S(Ys,i,t,h) =1

Ni,t

∑s Ys,i,t,h. I define the exposure variable τ to be the number of years an incoming fresh-

men cohort would have access to Facebook while they were studying for a 4-year bachelor’sdegree. My metric for access is the date of introduction of Facebook relative to the academicyear (AY) correpsonding to an incoming freshmen class. The AY is the standard used forannual statistics in U.S. Department of Education data, beginning on July 1st and endingon June 30th. The access variable is constructed assuming the freshmen cohort enters atthe beginning of the AY t, July 1st. Since the estimation strategy is only going to capturerelatives differences between the exposure variable τ across cohorts, it is not particularlyimportant when I choose the start date to be, even if it is likely freshmen cohorts do notarrive on campus until August or September. I also assume that each student graduateson-time (4 years) and so leaves campus exactly 4 years after they entered4. Because I donot have good data on the timing of degree completion for the students in my sample, Imust assume this to be consistent across universities. Explicitly, I define access length toFacebook τ as a function of Facebook introduction date at university i, di, as follows:

τi,t =

0 if 0 > di−StartDatet

365d−StartDatet

365if 0 ≤ d−StartDatet

365≤ 4

4 if 4 < di−StartDatet365

From now on, I suppress the h index for ease of notation/brevity 5, but this dimensionof the outcomes investigated should not be forgotten. Recall I do not observe the relevantstatistics at the cohort level, but rather the statistics of the pooled cohort random variableS(Ys,i,t,t+1) where Ys,i,t,t+1 = ωYs,i,t + (1 − ω) ∗ Ys,i,t+1 and ω is the share of students incohort t from university i of the pooled cohort for incoming classes t and t + 1. Dependingon the statistic, given our original equation 1, we can transform this equation to a suitableone we can test on our data. For example, if the statistic of interest is the mean, then

4One exception is for Northeastern University for which I assume students stay on campus for 5 years. It iswell documented that (see for example here) bachelor degrees offered at this school were almost exclusively 5-year programs during the sample period. Results do not change if I exclude Northeastern from my estimationsample.

5though one can just think of the same Y indexed by a different h as a completely different randomvariable that will therefore have different estimated fixed effects/coefficients

5

Page 6: The E ects of Student Access to Social Networking …web.stanford.edu/~larmona/fb_proposal.pdfon later outcomes. I show reduced form regression results that identify substantial e

S(Ys,i,t,t+1) = ωS(Ys,i,t) + (1− ω) ∗ S(Ys,i,t+1) and our testable equation would be

Si,t,t+1 = S(Ys,i,t,t+1) = αi + ωδt + (1− ω)δt+1 + ωβτi,t + (1− ω)βτi,t+1 + ωεi,t + (1− ω)εi,t+1

≡ αi + δt,t+1 + βτi,t,t+1 + εi,t,t+1

(2)

For other statistics, such as the variance and quantiles, I approximate the relationship ofthe pooled cohort random variable to be linear in the weighted sum of the individual randomvariables (as was exact in the case of S being the mean):

S(Ys,i,t,t+1) ≈ ωS(Ys,i,t) + (1− ω)S(Ys,i,t+1)

So I can then plug in this statistic on the left-hand side of equation 2 to estimate theeffect of Facebook exposure on other earnings statistics in a given cohort using the pooleddata. I use the log of these statistics in order to test hypotheses that Facebook exposureleads to a percentage changes in the statistics of interest, rather than a level increase. Thisspecification is the one I mainly focus on, which effectively specifies the cohort earningsstatistic as a multiplicative model in relation to Facebook Exposure, i.e.:

log(Si,t,t+1) ≈ αi + δt,t+1 + βτi,t,t+1 + εi,t,t+1 (3)

Interpretation of the coefficient β from this equation is as follows. Suppose the statistic ofinterest is the mean earnings. The thought exercise is that, for a given (Title IV) studententering a cohort at a university, if they treat their earnings post-college as an i.i.d. draw fromthe distribution of earnings of Title IV students from their university, in a counterfactualenvironment where they were exposed to Facebook access for 1 additional year while incollege, they would expect a β × 100% increase on average in the level of their earnings(if β > 0). This specification seems more reasonable than one where an additional yearof Facebook exposure leads to a constant increase in the level of an outcome like meanearnings, especiallly since we have schools in our sample as diverse as Harvard Universityand UC Riverside, whose graduates earn such different amounts at baseline (for the incomingfreshmen in 2001, graduates in our sample from these universities earned $103,000 and$37,000, respectively, in 2007, on average) that a constant dollar effect from access to asocial network seems far-fetched.

3.1 Identification

As discussed in the previous section, Facebook introduced its social network to the worldby iteratively rolling out to preselected universities whose students needed to sign up with a.edu email address for the relevant school. The first 100 schools, those used in the sample,and their rollout dates are shown in Tables 1 and 2. One can clearly see that this is nota randomly chosen sample of schools from the universe of postsecondary institutions in theUnited States. Specifically, the initial recipients of the Facebook network were exclusivelycomposed of the United States’ most elite universities. To assume that these universitieswere randomly chosen to have a greater exposure to Facebook would be ill-advised andwrought with endogeneity/ selection bias.

6

Page 7: The E ects of Student Access to Social Networking …web.stanford.edu/~larmona/fb_proposal.pdfon later outcomes. I show reduced form regression results that identify substantial e

A far more believable assumption is that, conditional on the university exposed to thetreatment (Facebook access), any particular cohort within that university is randomly chosento participate in the initial rollout of the Facebook social network. That is, it is fairlyplausible that Facebook did not coordinate to rollout their network to particular cohorts atthese elite universities. For example, it seems unlikely that Facebook decided to withold theirrollout to UT Austin until the 2004 academic year because they internally determined thatthe Class of 2004 that had just graduated that June would not take up their social network.Instead, UT Austin was likely rolled out to as a whole later than other universities becauseit was determined that UT Austin students were less likely to sign up for Facebook thanhistorically elite universities such as Stanford and Yale, or Facebook’s founders, who werestudents at Harvard, had less ties to UT Austin so it was a less natural place to expand to.The independence of the timing of which cohort is exposed to Facebook at a given universityis my key identifying assumption, and allows me to estimate the causal effect of Facebookexposure on cohort outcomes, using university-level fixed effects in my regression analysis tocapture unobservables in the earnings of students who graduate from the same university.In the causal inference potential outcomes framework established by Rosenbaum & Rubin[13], my assumption amounts to an unconfoundedness assumption on earnings and timingof Facebook access, conditional on the cohort’s university of origin:

Ys(t)⊥τi,t|αi, ∀i, t

where Ys(t) is the potential outcome of student s that is exposed to Facebook access for tyears, τi,t is the measure of Facebook access for a particular cohort t at university i, and αi

captures any unobservable effects on outcomes stemming from university i. I supplementthis with cohort-level fixed effects δt,t+1 in my estimation to ensure my regression analysis ofdifferential returns for cohorts with greater access to Facebook is not capturing differencesin the national labor market at the time these cohorts graduated and entered the workforce.Because I am also limiting the sample to cohorts that experienced some variation in total timeexposed while in school, we can understand any estimated effects as those corresponding tochanging the intensive margin of exposure to Facebook. With only 100 schools and a rangeof 8 months in the data for which Facebook was accessible at these schools, a regressionrelating the extensive margin (any access to Facebook while in school) seems underpoweredand futile, particularly with pooled cohort fixed effects.

Figure 1 graphically demonstrates the unconfoundedness distinction I am making. InPanel (a), I show the relationship within each cohort between observed average earnings atthese universities and exposure time to Facebook, with a superimposed LOESS [4] smootheron each set of cohort data. One can see that the conjecture of elite universities recievingFacebook first is clearly reflected in the data; running a simple OLS regression on the ob-served data would yield a coefficient on the order of $60,000 / year for an additional year ofexposure to Facebook. This clearly just reflects that schools that recieved Facebook earlywere schools whose graduates that already had a propensity to earn more than other schoolsin the sample. On the other hand, Panel (b) shows the relationship, estimated again with aLOESS smoother, between average earnings levels and years exposed once we have partialedout university and cohort fixed effects from the outcome and treatment variable. It is agraphic depiction of the relationship between the outcome and treatment as in Equation2. OLS on these residuals would yield a coefficient equivalent to that from estimating the

7

Page 8: The E ects of Student Access to Social Networking …web.stanford.edu/~larmona/fb_proposal.pdfon later outcomes. I show reduced form regression results that identify substantial e

joint multivariate regression expressed in Equation 2 by the Frisch-Waugh Theorem [6]. Inthis sense, the smoother can be understood as a non-parametric expression of the causalrelationship between treatment and mean earnings. The graph tells us that time exposedmatters more for those cohorts with relatively early access to Facebook while in school, asis evident from the flattening of the smoother for those with an above median orthogonal-ized treatment variable. Nonetheless, the relationship is still positive, but significantly lessextreme in its predictions for the effect of Facebook access compared to the observationalrelationship.

4 Reduced-Form Results

I now discuss the empirical regression results from running Equation 3 on my data of pooledcohort earnings variables and Facebook introduction to campuses. For all of these tables, Icluster standard errors at the university level, so that arbitrary correlation between cohortsat the same university is allowed [11]. Table 3 displays the results for the effects of Facebookaccess on the mean and variance of the earnings distribution of students 6 years after theyfirst entered college. I show the results for these moments both on the earnings distributionconditional on employment and unconditionally (but still excluding those who are seekinghigher education at the time of earnings data collection). As discussed before, all regressionson moments of the earnings distribution (mean,variance,percentiles) are done on the naturallog of the earnings statistic, so that the coefficient(s) may be interpreted as percent increasesin these left hand side variables. The estimates document a significant causal effect fromFacebook exposure: for the earnings distribution of cohorts, average earnings increasing by20% with an additional year of exposure, and the variance increases by 60%. These effectssuggest that, overall, the effect of Facebook is positive on the earnings of individuals froma given cohort, but it also appears to suggest that inequality within cohorts dramaticallyincreases as well, as measured by the earnings variance. Whether we condition on employ-ment or not does not seem to matter much, since employment rates are quite high (92% onaverage in our sample).

I next look at whether the effects are heterogeneous at different locations of the earningsdistribution, conditional on employment 6. Specifically, I look at 5 quantiles of interestand the effect of Facebook exposure in Table 4 to see if this positive effect on earnings isuniform across the distribution (so if there is a “location shift” but no difference in shape), orwhether some earnings quantiles change more than others from additional Facebook access.The regression results imply that the effect is global in the sense that all of the earningsquantiles have a statistically significant, positive coefficients, but there appear to be someregions that increase more than others. Specifically, it appears that highest earners fromthese universities, as measured by the 90th percentile of earnings, benefit the most fromadditional access to Facebook. The coefficient for the 90th percentile regression suggests anincreas of 30% from an additional year of Facebook exposure. On the other hand, the median,which can be here thought of as a more robust measure of centrality of the distribution thanthe mean, increases by a more modest/reasonable 13%. These two combine suggest that the

6I only have percentile data on the earnings data conditional on employment. As Table 3 demonstrated,we would likely expect similar estimations if I had the unconditional percentiles

8

Page 9: The E ects of Student Access to Social Networking …web.stanford.edu/~larmona/fb_proposal.pdfon later outcomes. I show reduced form regression results that identify substantial e

result in the previous table is at least partly driven by outliers in the earnings distribution,but the positive result is robust to using the median of the earnings distributions instead ofthe mean.

In a similar vein, Table 5 documents the relationship between Facebook access and morerobust measures of earnings dispersion. Specifically, I look at the interquartile range(IQR,75th - 25th percentile) and the interdecile range (IDR, 90th - 10th) as robust alternativesto the variance to measure disparity of earnings within university cohorts. We again see theeffects on disperson are more modest for the robust alternatives than initial estimates. Inparticular the effect on the IDR is a 30% increase from an additional year of Facebook, and aneven more modest 17% for the IQR. These results again suggest that the tails of the earningsdistribution may at least partly driving the results from before, since the results from thesealternatives are less pronounced, and in fact the effect on the “central mass of students”appears to be lower, as illustrated by the choice of more robust measures of dispersion (i.e.trim more of the tails of the distribution).

Table 5 also shows estimates of effects on two proportion variables of interest: the em-ployment rate (% working of those not enrolled in higher education) and the fraction ofstudents earning more than $25,000 a year in income. Note that both of these variables arenot logged since they are already rates and so we can directly estimate the coefficients inpercentage terms. Both show no effect of additional Facebook exposure. While we mighthave expected employment rate in particular to be positively effected by access to Facebook,given that wages are across the board positively impacted and the theoretical literaturepoints to correlation between the two [2], the suggestion here is that social ties maintainedvia Facebook might not be strong enough to push students out of a period of unemployment,but are strong enough to, say, get individuals already employed a marginally better outsideoffer from friendships. Or at least Facebook is acting as a way to diffuse information aboutjob opportunities that these employed friends may not have known about otherwise.

I now turn to heterogeneity in the effect of Facebook access on demographics subgroupswithin the pool of Title IV students for which the earnings data is available. Specifically, thedata available is for mean earnings 6 and 10 years after entry, and cuts across 3 strata: fam-ily income of student (low,middle, and high income; corresponds to $0 to $30,000, $30,000to $75,000, and $75,000+, respectively), financial dependency status upon entry to college,and gender7. In the medium-term (6 years after entry, or 2 years after leaving college) theeffects are fairly uniform across income groups and from 20-26%, no more than a standarderror away from each other. However, this uniformity dissipates for long-term (6 years aftercollege/10 years after entry) average earnings; now the effect is significantly stronger forstudents from low income families. This finding suggests that, initially when leaving college,the effect of Facebook is evenly spread across the cohorts, but over time, weak ties may bemore and more dependent on SNSs such as Facebook. The effect may be most dramaticfor those from low socioeconomic status (SES) at that point. A narrative that is consis-tent with this finding is that those with a low SES background previously had difficultymaintaining long-term relationships with all of their classmates and became segmented into

7Note that these groups are not comprehensive in the sense that for some students this data is unavailable.One can construct a missing demographic group but this would consist of very few students per cohort so itwas excluded from the table.

9

Page 10: The E ects of Student Access to Social Networking …web.stanford.edu/~larmona/fb_proposal.pdfon later outcomes. I show reduced form regression results that identify substantial e

different “cliques”, since these students often sort into different career sectors, such as thenon-profit sector, than other students in their cohort. But with access to friendships col-lected on Facebook during their time in college, they are able to maintain relationships withtheir cohort-at-large and thereby become informed of employment opportunities “median”students might have received, resulting in a relatively larger boost in their earnings. Alongother demographics, there is mild evidence of heterogeneity by gender (males benefit more)and substantial heterogeneity in response of average earnings by dependency status (thosewho are independent when entering college benefit far more from Facebook), which is con-sistent with independents using college as a means to obtain an earnings premium, whereasdependents may be more concerned with using college to enter a career they find interesting.

My final specification documents the persistence of additional access to Facebook onearnings 6 to 10 years after entry. For this part, I limit my analysis to the AY 1999-2000 andAY 2001-2002 pooled cohorts, since they are the only cohorts we have data for all of thesehorizons on, and I want to be consistent across the years after entry to be sure differentialeffects by horizon are not generated by differences in the composition of the cohorts eachyear after entry. For this table, I reshape the data so that each observation is the earningsof a pooled cohort from a given university at horizon h years after entry. I allow here foruniversity×horizon fixed effects, so that, say Harvard University students may on averageearn a different amount 6 versus 7 years after entry, and this may be different than thedifference between Stanford University students 6 versus 7 years after entry. In addition, Iallow for horizon×pooled cohort fixed effects, to account for both differences in the nationallabor market at the time of measurement and differential effects of experience on wages (ascaptured by years in the labor market). The full regression specification is:

log(Si,t,t+1,h) ≈ µi,h + κt,t+1,h + βτi,t,t+1 + εi,t,t+1,h

Table 7 documents the estimated effects. First of all, we see that excluding the 2003-2004 AYpooled cohort does little to the effect for 6 years after entry documented in Table 3, whichis a sort of robustness check that the findings are not heavily sample-dependent. Second, wesee some evidence of dissipation of the effect of Facebook access for longer horizon earnings,from 23% at the 6-year horizon to 12-15% later. So while the effect is clearly persistent,its magnitude decreases, which suggests a decaying effect of Facebook access over years onthe labour market. The earnings variance effect also decays, more dramatically than theeffect on mean earnings: it is insignificant from 7-9 years after entry and then significantagain at the 10-year horizon, but nearly halved in magnitude relative to the 6-year case.Facebook access has a universally zero effect on the employment rate, and there is evidenceof significant positive effects on the later horizon fraction of students earning greater than$25,000 a year, that is of comparable magnitude across horizons despite varying significance.

5 Structural Model

I now briefly propose a structural model of SNS friendship formation, followed by labormarket interactions that may be able to uncover more nuanced effects of network structureon earnings distributions . This friendship formation model is largely based on the networkformation model introduced in Mayer & Puller [12], which itself is partly based on the model

10

Page 11: The E ects of Student Access to Social Networking …web.stanford.edu/~larmona/fb_proposal.pdfon later outcomes. I show reduced form regression results that identify substantial e

introduced in Jackson & Rogers [2]. The friendship graph is initialized at time t = 0 (periodshere are months) to have zero connections and is the size of the undergraduate populationat the university. In the first period, everyone meets other individuals randomly so that mi

meetings are formed in the first period. This can be enriched to have the meeting likelihoodbased on sharing similar attributes, such as cohort. After meeting with other agents/ nodeson the graph, meetings are turned into friendships / connections on the graph depending

on the friendship function f(Xi,Xj, ~β), where meetings turn into friendships if fi,j ≥ 0, i.e.marginal utility from the friendship is positive:

f(Xi,Xj, ~β) = ui,j + β0 +∑k

βk × (Xi,k == Xj,k)

Where k is an index for an observable characteristic of the meeting nodes, so that thefriendship function depends on nodes sharing characteristics, and ui,j is an i.i.d. idiosyncraticrandom noise term. After these initial sets of friendship formations, in each period, studentsmeet other students via their neighbors, but also at random. They meet enough friends offriends (nodes separated by 2 degrees) so that mf meetings through neighbors are formedeach period (the neighbors of neighbors are chosen randomly), and mr meetings are formedcompletely randomly across all the nodes in the graph they are not yet friends with, so thatmr +mf meetings are formed in periods t ≥ 1 for each node. Meetings are again tranformedinto friendships depending on the friendship function f . the process iterates identically inlater periods. In periods where t mod 12 = 0, corresponding the end of an academic year,the nodes that are in the most senior cohort graduate and leave campus, so they remainon the friendship graph but cease forming connections in subsequent periods 8. In the nextperiod, when t mod 12 = 1, an incoming freshmen cohort enters the graph. They beginas zero-degree nodes, whose attributes are randomly sampled (with replacement) from theset of all nodes on the graph. The size of the cohort can be determined from IPEDS dataon undergraduate cohort sizes. In the first period, they again form mi meetings uniformlyat random, and thereafter behave as other nodes. This process repeats for an indefiniteamount of periods, with cohorts being replaced each year by a new freshmen class, but thetotal graph node count expands indefinitely since older cohorts are able to maintain theirconnections due to the SNS technology.

Once nodes graduate, they enter the labour market. Initially, upon graduation, theyreceive employment with probability pe, and conditional on employment, draw a wage froman initial wage distribution F (w). F (w) can either be discrete (by creating a grid of w’s) orcontinuous (by assuming a gamma or log-normal distribution, and estimating the underlyingparameters). Unemployment corresponds to a wage of 0. The agents now only care abouttheir connections to nodes who are also already in the labour market, and the graph isstatic so that no new meetings are taken by nodes post-graduation (since they are now “off-campus”). The model then follows the labour market network model in Calvo and Jackson[2]. Each period, nodes on the labour market recieve a job offer from the wage distributionw ∼ F (w) with probability po directly; They take the offer if w > wi, their current wage.Otherwise, they pass the information on to one of their neighbors in the labour market

8I could alter the model so that these nodes simply have a smaller amount of meetings each period, orno longer form random meetings with others but still do network-based meetings

11

Page 12: The E ects of Student Access to Social Networking …web.stanford.edu/~larmona/fb_proposal.pdfon later outcomes. I show reduced form regression results that identify substantial e

randomly, with probability pp, conditional on the new wage being greater than the neighbor’swage, w > wj, j ∈ Ni(g). If no neighbors are unemployed or have a lower wage, the offer isnot shared on the social network9. After the set of offers are generated and distributed ineach period, employed agents become unemployed (fired) randomly with uniform probabilitypb. we let this economy iterate over many periods like this, which itself is growing due tonew cohorts graduating from school.

I can estimate the parameters mi,mr,mf , ~β within a network using data from Traud etal. [15], a snapshort of the Facebook network graphs correspond to the schools in my sampleas of September 2005. This data contains friendships (edges) and node attributes such asmajor, dorm, cohort, etc. I can match the model to moments in the data, such as proportionof Facebook friends who share the same gender, moments of the degree distribution, etc., by asimulated method-of-moments approach. Each simulated network would have their simulatedmoments evaluated relative to the observed moments at different periods depending on whenFacebook became accessible on their campus relative to September 2005, the snapshot ofthe observed network graph. After estimating this school friendship formation model, Ican simulate the friendship formation stage and then use earnings data from the CSD tochoose parameters F (w), pb, pe, pp to match moments of the simulated labour market earningsdistribution with the earnings distribution at a particular horizon 2-6 years after exitingcollege. Of course, this will now require us to assume the earnings data of Title IV studentsis a good approximation of the earnings distribution of the student population at large.

Conditional on estimating these parameters, I could then calculate relationships of in-terest such as how wages depend on the degree of an individual, or how much inequalitybetween as well as within schools this model might be able to generate, from differences inschool network structures. Computationally, this estimation procedure is likely to be costly.However, it is my belief that this structural estimation could verify some of the reducedform findings by uncovering the dependence of an earnings distribution on the underlyingstructure a given school’s SNS graph.

6 Conclusion

This proposal demonstrates a statistically significant and economically large in magnitudeeffect of access to a social networking site while in college on future earnings of students.We find that the tail of the earnings distribution plays a role in how large of an effectwe estimate, but the findings are preserved when examining alternative measurements morerobust to outliers. Effects on the earnings distribution display some interesting heterogeneityby income, and are persistent as far out as 6 years on the labour market. How seriously shouldwe take these results? Recall that the universe of students for which we have earnings datais those who receive federal financial aid. Clearly, these students are of lower socioeconomicstatus than the average student who does not take federal aid. The heterogeneity by incomewithin Title IV students suggests that the effects on the larger population would be dampenedrelative to the results reported here. As discussed before, low SES students probably have themost to gain from SNS technology, since they can keep in touch with the cohorts from their

9It might be interesting to let friends pass offers to their friends if none of the original recipient’s neighborswould experience an increase in their wages from it

12

Page 13: The E ects of Student Access to Social Networking …web.stanford.edu/~larmona/fb_proposal.pdfon later outcomes. I show reduced form regression results that identify substantial e

university and maintain information about job opportunities they may not have otherwisehad had they returned to their communities of origin where job opportunities are likely lesspromising. Using pooled rather than the single cohorts may impact the precision of ourestimates. Additionally, many of these results at the 6 year horizon do not hold up if weinclude all the available data, up to the 2006-2007 pooled cohort. However, in my model,most of these cohorts would be assigned the same amount of exposure to Facebook as thoseentering in AY 2004, but this is likely inappropriate since in fact students entering in 2007 areentering a setting where 1) Facebook is no longer the only SNS that is going to be relevant in10 years 2) The social networks have had 3+ years to develop when they enter college, whichis going to reflect a different “treatment” than those who entered at the time it was firstrolled out. The effect on these cohorts should not be the same as those who entered whenthere was no existing network in place. Regardless, I document important causal effects onearnings for at least the initial cohorts who were given access to Facebook. It is my hopethat the structural model briefly discussed above can uncover some of the more complexeffects of electronic social network formation on future labour outcomes.

References

[1] Antoni Calvo-Armengol. Job contact networks. Journal of economic Theory,115(1):191–206, 2004.

[2] Antoni Calvo-Armengol and Matthew O Jackson. Networks in labor markets: Wageand employment dynamics and inequality. Journal of economic theory, 132(1):27–46,2007.

[3] Antoni Calv-Armengol and Matthew O. Jackson. The effects of social networks onemployment and inequality. American Economic Review, 94(3):426–454, June 2004.

[4] William S Cleveland, Eric Grosse, and William M Shyu. Local regression models.Statistical models in S, 2:309–376, 1992.

[5] Ruben Enikolopov, Alexey Makarin, and Maria Petrova. Social media and protestparticipation: Evidence from russia. 2016.

[6] Ragnar Frisch and Frederick V Waugh. Partial time regressions as compared withindividual trends. Econometrica: Journal of the Econometric Society, pages 387–401,1933.

[7] Mark S Granovetter. The strength of weak ties. American journal of sociology,78(6):1360–1380, 1973.

[8] Matthew O Jackson. Social and economic networks. Princeton university press, 2010.

[9] Abigail Z Jacobs, Samuel F Way, Johan Ugander, and Aaron Clauset. Assemblingthefacebook: Using heterogeneity to understand online social network assembly. InProceedings of the ACM Web Science Conference, page 18. ACM, 2015.

13

Page 14: The E ects of Student Access to Social Networking …web.stanford.edu/~larmona/fb_proposal.pdfon later outcomes. I show reduced form regression results that identify substantial e

[10] Ron Laschever. The doughboys network: Social interactions and the employment ofworld war i veterans. 2013.

[11] Kung-Yee Liang and Scott L Zeger. Longitudinal data analysis using generalized linearmodels. Biometrika, pages 13–22, 1986.

[12] Adalbert Mayer and Steven L Puller. The old boy (and girl) network: Social networkformation on university campuses. Journal of public economics, 92(1):329–347, 2008.

[13] Paul R Rosenbaum and Donald B Rubin. The central role of the propensity score inobservational studies for causal effects. Biometrika, pages 41–55, 1983.

[14] Amanda L Traud, Eric D Kelsic, Peter J Mucha, and Mason A Porter. Comparingcommunity structure to characteristics in online collegiate social networks. SIAM review,53(3):526–543, 2011.

[15] Amanda L Traud, Peter J Mucha, and Mason A Porter. Social structure of facebooknetworks. Physica A: Statistical Mechanics and its Applications, 391(16):4165–4180,2012.

[16] Johan Ugander, Brian Karrer, Lars Backstrom, and Cameron Marlow. The anatomy ofthe facebook social graph. arXiv preprint arXiv:1111.4503, 2011.

14

Page 15: The E ects of Student Access to Social Networking …web.stanford.edu/~larmona/fb_proposal.pdfon later outcomes. I show reduced form regression results that identify substantial e

● ●

●●

●● ●

●●

●●

●●

● ●

●●

●●

●●

●●

● ● ●

●●

●●

●●

●●

●●

●●

● ●

●● ●

●●

●●

0.00 0.05 0.10 0.15 0.20 1.4 1.6 1.8 3.3 3.4 3.5 3.6 3.7

40

60

80

100

Time Exposed to Facebook (Years)

Mea

n E

arni

ngs

6 Ye

ars

Pos

t−en

try

(in $

1000

's)

Pooled Freshmen Cohorts●

AY 1999−2000

AY 2001−2002

AY 2003−2004

(a) Observed Relation

● ●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●

●●

● ●

●●

●●

●●

●●

●●

●●

●●●

● ●

●●

−10

−5

0

5

10

−0.1 0.0 0.1 0.2

Orthogonalized Time Exposed to Facebook (Years)

Ort

hogo

naliz

ed M

ean

Ear

ning

s 6

Year

s P

ost−

entr

y (in

$10

00's

)

Pooled Freshmen Cohorts●

AY 1999−2000

AY 2001−2002

AY 2003−2004

(b) Orthogonalized Relation

Figure 1: Relationship between Earnings Levels and Facebook Access While in College

15

Page 16: The E ects of Student Access to Social Networking …web.stanford.edu/~larmona/fb_proposal.pdfon later outcomes. I show reduced form regression results that identify substantial e

Table 1: Dates of Access to Facebook for 100 Universities in Sample from Jacobs et al.(2015), 1-50

University Date Ranking UncertaintyHarvard University 2/4/2004 1 0Columbia University in the City of New York 2/25/2004 2 0Stanford University 2/26/2004 3 0Yale University 2/29/2004 4 0Cornell University 3/7/2004 5 0Dartmouth College 3/7/2004 6 0University of Pennsylvania 3/14/2004 7 0Massachusetts Institute of Technology 3/14/2004 8 0New York University 3/21/2004 9 0Boston University 3/21/2004 10 0Brown University 4/4/2004 11 0Princeton University 4/4/2004 12 0University of California-Berkeley 4/4/2004 13 0Duke University 4/11/2004 14 0Georgetown University 4/11/2004 15 0University of Virginia-Main Campus 4/11/2004 16 0Boston College 4/19/2004 17 0Tufts University 4/19/2004 18 0Northeastern University 4/19/2004 19 0University of Illinois at Urbana-Champaign 4/19/2004 20 0University of Florida 4/25/2004 21 6Wellesley College 4/25/2004 22 6University of Michigan-Ann Arbor 4/25/2004 23 6Michigan State University 4/25/2004 24 6Northwestern University 4/25/2004 25 0University of California-Los Angeles 4/27/2004 26 3Emory University 4/30/2004 27 3University of North Carolina at Chapel Hill 4/30/2004 28 3Tulane University of Louisiana 4/30/2004 29 3University of Chicago 4/30/2004 30 0Rice University 4/30/2004 31 0Washington University in St Louis 5/2/2004 32 0University of California-Davis 5/20/2004 33 5University of California-San Diego 5/20/2004 34 5University of Southern California 6/23/2004 35 0California Institute of Technology 6/25/2004 36 0University of California-Santa Barbara 6/25/2004 37 0University of Rochester 8/4/2004 38 9Bucknell University 8/4/2004 39 9Williams College 8/8/2004 40 4Amherst College 8/8/2004 41 4Swarthmore College 8/8/2004 42 4Wesleyan University 8/8/2004 43 4Oberlin College 8/8/2004 44 4Middlebury College 8/8/2004 45 4Hamilton College 8/8/2004 46 4Bowdoin College 8/8/2004 47 4Vanderbilt University 8/21/2004 48 1Carnegie Mellon University 8/21/2004 49 1University of Georgia 8/21/2004 50 1

16

Page 17: The E ects of Student Access to Social Networking …web.stanford.edu/~larmona/fb_proposal.pdfon later outcomes. I show reduced form regression results that identify substantial e

Table 2: Dates of Access to Facebook for 100 Universities in Sample from Jacobs et al.(2015), 51-100

University Date Ranking UncertaintyUniversity of South Florida-Main Campus 8/21/2004 51 1University of Central Florida 8/21/2004 52 1Florida State University 8/21/2004 53 1George Washington University 8/21/2004 54 0Johns Hopkins University 8/21/2004 55 0Syracuse University 8/22/2004 56 1University of Notre Dame 8/22/2004 57 1University of Maryland-College Park 8/22/2004 58 1University of Maine 9/7/2004 59 3Smith College 9/7/2004 60 3University of California-Irvine 9/7/2004 61 3Villanova University 9/7/2004 62 3Virginia Polytechnic Institute and State University 9/7/2004 63 3University of California-Riverside 9/7/2004 64 3California Polytechnic State University-San Luis Obispo 9/7/2004 65 3University of Mississippi 9/7/2004 66 3Michigan Technological University 9/7/2004 67 3University of California-Santa Cruz 9/7/2004 68 3Indiana University-Bloomington 9/7/2004 69 3University of Vermont 9/7/2004 70 3Auburn University 9/7/2004 71 3University of San Francisco 9/7/2004 72 3Wake Forest University 9/7/2004 73 3Santa Clara University 9/7/2004 74 3American University 9/7/2004 75 3Haverford College 9/7/2004 76 3College of William and Mary 9/7/2004 77 3Miami University-Oxford 9/7/2004 78 3James Madison University 9/7/2004 79 3The University of Texas at Austin 9/7/2004 80 3Simmons College 9/7/2004 81 3SUNY at Binghamton 9/7/2004 82 3Temple University 9/7/2004 83 3Texas A & M University-College Station 9/7/2004 84 3Vassar College 9/7/2004 85 3Pepperdine University 9/7/2004 86 3University of Wisconsin-Madison 9/7/2004 87 3Colgate University 9/7/2004 88 3Rutgers University-New Brunswick 9/7/2004 89 3Howard University 9/7/2004 90 3University of Connecticut 9/7/2004 91 3University of Massachusetts-Amherst 9/7/2004 92 3Baylor University 9/7/2004 93 3Pennsylvania State University-Main Campus 9/7/2004 94 3The University of Tennessee-Knoxville 9/7/2004 95 3Lehigh University 9/7/2004 96 3University of Oklahoma-Norman Campus 9/7/2004 97 3Reed College 9/7/2004 98 3Brandeis University 9/7/2004 99 3Trinity College 9/24/2004 100 0

17

Page 18: The E ects of Student Access to Social Networking …web.stanford.edu/~larmona/fb_proposal.pdfon later outcomes. I show reduced form regression results that identify substantial e

Tab

le3:

Eff

ect

ofF

aceb

ook

Acc

ess

On

Ear

nin

gsM

omen

tsof

Coh

orts

6Y

ears

Aft

erE

ntr

yto

Col

lege

(1)

(2)

(3)

(4)

Mea

nE

arnin

gs|W

orkin

gM

ean

Ear

nin

gsE

arnin

gsV

aria

nce|W

orkin

gE

arnin

gsV

aria

nce

Acc

ess

toF

aceb

ook

(Yea

rs)

0.20

8∗∗∗

0.21

5∗∗∗

0.64

0∗∗

0.61

8∗∗

(0.0

47)

(0.0

50)

(0.2

70)

(0.2

51)

Obse

rvat

ions

300

300

300

300

R-S

quar

ed0.

370.

350.

130.

15O

utc

ome

Mea

n10

.84

10.7

721

.19

21.2

3O

utc

ome

SD

0.23

0.23

0.71

0.68

Univ

ersi

tyF

ixed

Eff

ects

Yes

Yes

Yes

Yes

Yea

rF

ixed

Eff

ects

Yes

Yes

Yes

Yes

OL

Ses

tim

ates

rep

orte

d.

Sta

ndar

der

rors

clust

ered

atth

euniv

ersi

tyle

vel

inpar

enth

eses

.Sig

nifi

cant

at*p<

0.10

,**

p<

0.05

,**

*p<

0.01

.D

epen

den

tva

riab

les

are

the

condit

ional

/unco

ndit

ional

logg

edm

ean

/va

rian

ceof

earn

ings

ofp

ool

edco

hor

ts6

year

saf

ter

they

firs

ten

tere

dco

lleg

e.C

ohor

tsco

nsi

stof

studen

tsw

ho

reci

eved

feder

alfinan

cial

aid

while

enro

lled

,an

dar

enot

enro

lled

atan

yp

osts

econ

dar

yin

stit

uti

onduri

ng

the

mea

sure

men

tof

earn

ings

.

18

Page 19: The E ects of Student Access to Social Networking …web.stanford.edu/~larmona/fb_proposal.pdfon later outcomes. I show reduced form regression results that identify substantial e

Tab

le4:

Eff

ect

ofF

aceb

ook

Acc

ess

On

Per

centi

les

ofE

arnin

gsD

istr

ibuti

on6

Yea

rsA

fter

Entr

yto

Col

lege

(1)

(2)

(3)

(4)

(5)

10th

Pct

ile|W

orkin

g25

thP

ctil

e|W

orkin

g50

thP

ctil

e|W

orkin

g75

thP

ctil

e|W

orkin

g90

thP

ctil

e|W

orkin

g

Acc

ess

toF

aceb

ook

(Yea

rs)

0.1

66∗

0.14

9∗∗

∗0.

131∗∗

∗0.

168∗

∗∗0.

298∗∗

(0.0

94)

(0.0

54)

(0.0

35)

(0.0

50)

(0.0

68)

Ob

serv

ati

on

s29

829

830

029

829

8R

-Squ

are

d0.

370.

290.

240.

360.

37O

utc

om

eM

ean

9.47

10.2

510

.72

11.0

711

.39

Ou

tcom

eS

D0.

290.

220.

190.

210.

27U

niv

ersi

tyF

ixed

Eff

ects

Yes

Yes

Yes

Yes

Yes

Yea

rF

ixed

Eff

ects

Yes

Yes

Yes

Yes

Yes

OL

Ses

tim

ates

rep

ort

ed.

Sta

nd

ard

erro

rscl

ust

ered

atth

eu

niv

ersi

tyle

vel

inp

aren

thes

es.

Sig

nifi

cant

at*p

<0.

10,

**p<

0.05

,**

*p<

0.01

.D

epen

den

tva

riab

les

are

,th

elo

gged

per

centi

les

ofth

eea

rnin

gsd

istr

ibu

tion

con

dit

ion

alon

wor

kin

g,fo

rp

ool

edco

hor

ts6

year

saf

ter

they

firs

ten

tere

dco

lleg

e.C

ohort

sco

nsi

stof

stu

den

tsw

ho

reci

eved

fed

eral

fin

anci

alai

dw

hil

een

roll

ed,

and

are

not

enro

lled

atan

yp

osts

econ

dar

yin

stit

uti

on

du

rin

gth

em

easu

rem

ent

ofea

rnin

gs.

19

Page 20: The E ects of Student Access to Social Networking …web.stanford.edu/~larmona/fb_proposal.pdfon later outcomes. I show reduced form regression results that identify substantial e

Tab

le5:

Eff

ect

ofF

aceb

ook

Acc

ess

On

Addit

ional

Ear

nin

gsSta

tist

ics

6Y

ears

Aft

erE

ntr

yto

Col

lege

(1)

(2)

(3)

(4)

Ear

nin

gsIQ

R|W

orkin

gE

arnin

gsID

R|W

orkin

gE

mplo

ym

ent

Rat

e%

Ear

nin

g≥

$25,

000

Acc

ess

toF

aceb

ook

(Yea

rs)

0.17

2∗∗∗

0.31

8∗∗∗

0.00

50.

021

(0.0

64)

(0.0

73)

(0.0

09)

(0.0

16)

Obse

rvat

ions

298

298

300

300

R-S

quar

ed0.

520.

440.

430.

36O

utc

ome

Mea

n10

.48

11.2

30.

930.

78O

utc

ome

SD

0.25

0.30

0.02

0.06

Univ

ersi

tyF

ixed

Eff

ects

Yes

Yes

Yes

Yes

Yea

rF

ixed

Eff

ects

Yes

Yes

Yes

Yes

OL

Ses

tim

ates

rep

orte

d.

Sta

ndar

der

rors

clust

ered

atth

euniv

ersi

tyle

vel

inpar

enth

eses

.Sig

nifi

cant

at*p<

0.10

,**

p<

0.05

,**

*p<

0.01

.D

epen

den

tva

riab

les

are,

resp

ecti

vely

,th

eem

plo

ym

ent

rate

,fr

acti

onea

rnin

g$2

5,00

0p

erye

ar,

logg

edin

terq

uar

tile

range

(75t

h-2

5th

per

centi

le)

ofea

rnin

gsco

ndit

ional

onem

plo

ym

ent,

and

the

logg

edin

terd

ecile

range

(90t

h-1

0th

per

centi

le)

ofea

rnin

gsco

ndit

ional

onem

plo

ym

ent,

for

pool

edco

hor

ts6

year

saf

ter

they

firs

ten

tere

dco

lleg

e.C

ohor

tsco

nsi

stof

studen

tsw

ho

reci

eved

feder

alfinan

cial

aid

while

enro

lled

,an

dar

enot

enro

lled

atan

yp

osts

econ

dar

yin

stit

uti

onduri

ng

the

mea

sure

men

tof

earn

ings

.

20

Page 21: The E ects of Student Access to Social Networking …web.stanford.edu/~larmona/fb_proposal.pdfon later outcomes. I show reduced form regression results that identify substantial e

Tab

le6:

Eff

ect

ofF

aceb

ook

Acc

ess

onT

itle

IVStu

den

tD

emog

raphic

Subgr

oups

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

All

Low

-Inco

me

Mid

dle

-Inco

me

Hig

h-I

nco

me

Dep

enden

tIn

dep

enden

tF

emal

eM

ale

Acc

ess

toF

aceb

ook

(Yea

rs)

0.20

8∗∗∗

0.25

7∗∗∗

0.26

9∗∗∗

0.22

8∗∗∗

0.18

3∗∗∗

0.29

9∗∗∗

0.22

4∗∗∗

0.21

1∗∗∗

(0.0

47)

(0.0

78)

(0.0

70)

(0.0

49)

(0.0

38)

(0.0

82)

(0.0

45)

(0.0

65)

Obse

rvat

ions

300

270

275

275

237

234

288

288

R-S

quar

ed0.

370.

240.

260.

320.

540.

340.

280.

33O

utc

ome

Mea

n10

.84

10.8

610

.83

10.8

510

.81

10.9

910

.76

10.9

3O

utc

ome

SD

0.23

0.26

0.23

0.22

0.20

0.31

0.21

0.25

Univ

ersi

tyF

ixed

Eff

ects

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yea

rF

ixed

Eff

ects

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

(a)

Ear

nin

gs6

year

sP

ost-

Entr

y

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

All

Low

-Inco

me

Mid

dle

-Inco

me

Hig

h-I

nco

me

Dep

enden

tIn

dep

enden

tF

emal

eM

ale

Acc

ess

toF

aceb

ook

(Yea

rs)

0.15

7∗∗∗

0.37

9∗∗∗

0.14

3∗∗∗

0.11

8∗∗∗

0.06

9∗∗

0.40

9∗∗∗

0.16

1∗∗∗

0.20

0∗∗∗

(0.0

28)

(0.0

64)

(0.0

53)

(0.0

30)

(0.0

28)

(0.0

68)

(0.0

34)

(0.0

40)

Obse

rvat

ions

300

278

279

285

239

235

288

288

R-S

quar

ed0.

220.

240.

110.

140.

080.

350.

130.

27O

utc

ome

Mea

n11

.14

11.1

211

.10

11.1

611

.12

11.1

911

.02

11.2

7O

utc

ome

SD

0.24

0.26

0.23

0.24

0.23

0.32

0.21

0.26

Univ

ersi

tyF

ixed

Eff

ects

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yea

rF

ixed

Eff

ects

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

(b)

Ear

nin

gs10

year

sP

ost-

Entr

y

OL

Ses

tim

ates

rep

orte

d.

Sta

ndar

der

rors

clust

ered

atth

euniv

ersi

tyle

vel

inpar

enth

eses

.Sig

nifi

cant

at*p<

0.10

,**

p<

0.05

,**

*p<

0.01

.D

epen

den

tva

riab

les

are

the

logg

edm

ean

earn

ings

amon

gva

riou

sdem

ogra

phic

subgr

oups

ofp

ool

edco

hor

tsof

studen

tsw

ho

reci

eved

feder

alfinan

cial

aid

while

enro

lled

6ye

ars

afte

rth

eyfirs

ten

tere

dco

lleg

e,co

ndit

ional

onem

plo

ym

ent.

21

Page 22: The E ects of Student Access to Social Networking …web.stanford.edu/~larmona/fb_proposal.pdfon later outcomes. I show reduced form regression results that identify substantial e

Tab

le7:

Per

sist

ence

ofE

ffec

tof

Fac

ebook

Acc

ess

On

Ear

nin

gsfo

rth

e19

99-2

003

Coh

orts

(1)

(2)

(3)

(4)

Mea

nE

arnin

gsE

arnin

gsV

aria

nce

Em

plo

ym

ent

Rat

e%

Ear

nin

g≥

$25,

000

Yea

rsP

ost-

entr

y=

Yea

rsof

FB

Acc

ess

0.22

9∗∗∗

0.61

1∗∗

0.00

50.

020

(0.0

47)

(0.2

62)

(0.0

09)

(0.0

15)

Yea

rsP

ost-

entr

y=

Yea

rsof

FB

Acc

ess

0.11

7∗∗∗

0.20

6-0

.009

0.03

9∗∗∗

(0.0

35)

(0.2

04)

(0.0

08)

(0.0

10)

Yea

rsP

ost-

entr

y=

Yea

rsof

FB

Acc

ess

0.05

0-0

.049

-0.0

040.

024∗

(0.0

32)

(0.1

94)

(0.0

09)

(0.0

14)

Yea

rsP

ost-

entr

y=

Yea

rsof

FB

Acc

ess

0.12

1∗∗∗

0.30

4-0

.007

0.02

9∗

(0.0

33)

(0.2

00)

(0.0

10)

(0.0

15)

Yea

rsP

ost-

entr

y=

10×

Yea

rsof

FB

Acc

ess

0.15

1∗∗∗

0.36

1∗∗

0.00

10.

022∗

(0.0

32)

(0.1

65)

(0.0

09)

(0.0

11)

Obse

rvat

ions

1000

1000

1000

1000

R-S

quar

ed0.

360.

110.

340.

26O

utc

ome

Mea

n10

.92

21.6

60.

920.

82O

utc

ome

SD

0.24

0.75

0.02

0.05

Tes

tof

Joi

nt

Sig

nifi

cance

0.0

07.4

44.0

04U

niv

ersi

tyF

ixed

Eff

ects

Uni*

Hor

izon

Uni*

Hor

izon

Uni*

Hor

izon

Uni*

Hor

izon

Yea

rF

ixed

Eff

ects

Coh

ort*

Hor

izon

Coh

ort*

Hor

izon

Coh

ort*

Hor

izon

Coh

ort*

Hor

izon

OL

Ses

tim

ates

rep

orte

d.

Sta

ndar

der

rors

clust

ered

atth

euniv

ersi

tyle

vel

inpar

enth

eses

(so

ther

ear

e10

0cl

ust

ers)

.Sig

nifi

cant

at*p<

0.10

,**

p<

0.05

,**

*p<

0.01

.D

epen

den

tva

riab

les

are

the

unco

ndit

ional

logg

edm

ean,

unco

ndit

ional

logg

edva

rian

ceof

earn

ings

ofp

ool

edco

hor

ts6

year

saf

ter

they

firs

ten

tere

dco

lleg

e,as

wel

las

the

emplo

ym

ent

rate

,an

dfr

acti

onea

rnin

g$2

5,00

0p

erye

ar.

Coh

orts

consi

stof

studen

tsw

ho

reci

eved

feder

alfinan

cial

aid

while

enro

lled

,an

dar

enot

enro

lled

atan

yp

osts

econ

dar

yin

stit

uti

onduri

ng

the

mea

sure

men

tof

earn

ings

.

22