The economic effects of Facebook

Social media permeates many aspects of our lives, including how we connect with others, where we get our news and how we spend our time. Yet, we know little about the economic effects for users. In 2017, we ran a large field experiment with over 1765 individuals to document the value of Facebook to users and its causal effect on news, well-being and daily activities. Participants reveal how much they value one week of Facebook usage and are then randomly assigned to a validated Facebook restriction or normal use. One week of Facebook is worth $67. Those who are off Facebook for one week reduce news consumption, are less likely to recognize politically-skewed news stories, report being less depressed and engage in healthier activities. These results are strongest for men. Our results further suggest that, after the restriction, Facebook’s value increases, consistent with information loss or that using Facebook may be addictive.

1. It is estimated that the impact of Facebook through advertising is $77.6 billion in the U.S. (Deloitte 2015). Evidence on the value of Facebook is given in Brynjolfsson et al. (2018).

2. We note that the BDM mechanism used in our study, and in other studies using the BDM or other mechanisms, involve hassle costs and some complexity that may affect values. Our participants face a one in two chance of experiencing a Facebook restriction, and this may reduce bias in value estimates when using elicitation mechanisms coupled with implementation uncertainty.

3. Tromholt (2016) uses a one-week, self-enforced Facebook restriction and finds a positive effect on overall life satisfaction.

4. The study is registered in the AEA Registry (AEARCTR-0003952)

5. The email text and online survey questions are in the Appendix, Sections A.1 & A.2 in ESM

6. Our procedures made clear to participants that they would be paid the random offer upon study completion to mitigate any uncertainty bias (Horowitz 2006).

7. For budgetary reasons and expected participation rates, the random counter offers were drawn with the following probabilities: (5, 15.14%; 7, 15.14%; 9, 11.14%; 10, 11.14%; 12, 11.14%; 14, 11.14%; 16, 7.14%; 18, 6.14%; 20, 5.14%; 21, 5.14%; 24, 0.64%; 25, 0.64%; 28, 0.14%; 30, 0.14%). The expected offer is $11.58.

8. Those who are ineligible for subsequent phases are not contacted.

9. Participants complete the survey in Phase 2, prior to random assignment to the Facebook restriction, and in Phase 3. One might be concerned that changes in outcomes are due to experimenter demand effects. First, participants are not aware they will complete the same survey questions a second time. Second, we find effects for some, but not most, of the outcomes, thus alleviating concerns of such an effect. Finally, in a later study with a similar design, Allcott et al. (2019) find similar results to ours while explicitly testing for demand effects.

10. We chose Breitbart given that its internet traffic as of March 2017 surpassed other major skewed news sources and was similar in magnitude to that of mainstream news sources such as The Washington Post according to data from alexa.com

11. See the questionnaire in Appendix A.3 in ESM

12. The university randomly generates the last four digits of a student’s ID number.

13. We updated the news pool to reflect headlines from the previous week.

14. Participants are asked to write down their valuation and informed that their payment today is unaffected by their response. Eligible participants from this second BDM go through the same process as in Phase 2, return for a third and final survey in one week, and are paid their counteroffers from the second BDM. We do not include this third survey in our estimates.

15. Eligibility for Phase 2 means that the submitted value was less than a randomly-selected counter-offer of no more than $30. This is by the design of the elicitation mechanism—so all those with submitted values higher than $30 were ineligible. Descriptive statistics for these groups are in Appendix Table A.5.1 in ESM. In Sect. 4.1 of the paper, we test the robustness of the results to this design-induced selection.

16. When we compare participants who submitted values less than or equal to $30, so they could have been eligible to participate in Phase 2, there is no significant difference by age or gender between those who ended up being eligible or ineligible based on the counteroffer. See Appendix Table A.5.1 in ESM

17. Participants were aware of this procedure prior to submitting their value of Facebook in the Phase 1 online survey. Holding this session on main campus minimizes travel costs that might have affected valuations for Facebook.

18. Appendix Table A.5.2 in ESM shows the balance of covariates across the treatment and control groups.

19. Participants did not interact with the study account in any way.

20. Appendix Table A.5.3 in ESM

21. Appendix Table A.5.3 in ESM. While we cannot say what proportion of news participants get from Facebook, 81% report opening up Facebook every day or several times a day to check their news feed.

22. We use the rankings on www.allsides.com. If a participant lists a news outlet that is not reported on allsides.com, we treat their preferred news outlet as missing. The top five first choice sources are CNN (28.1%), FOX (12.6%), BBC (8.3%), NYT (4.7%), and ESPN (4.7%). Breitbart was not listed as a first choice, however, news from this source could appear on a Facebook news feed.

23. Appendix Figure A.5.1 in ESM shows the distribution of these emotions separately.

24. We calculate the confidence intervals using bootstrap with 1000 replications.

25. Our design also explored the willingness to pay (WTP)—willingness to accept (WTA) gap in the BDM mechanism [see Knetsch et al. (2001), Plott and Zeiler (2005), Horowitz (2006), and Brynjolfsson et al. (2018) for a discussion of this phenomenon]. Half of the participants were asked the value in terms of selling participation in the study (WTA), “How much money would you need to be given to stop using Facebook for a week?” and half were asked in terms of purchasing participation (WTP), “What is the value of your weekly time on Facebook?” We find no significant difference in the reported value of Facebook from either solicitation method or by covariates across groups, so we pool the data in our analysis.

26. The distribution is trimmed at $540 because of a few outliers in the data—the maximum value is $2,153. We use the nontrimmed, full sample in our analysis.

27. There are differences across studies. Brynjolfsson et al. (2018) use an online sample, one out of every 200 participants are randomized into the Facebook restriction, and respondents who do not use Facebook are not screened out for their weekly estimate. Corrigan et al. (2018) use a series of second-price auctions with different samples and compensation schemes. Allcott et al. (2019) also use a BDM mechanism but with an online sample.

28. Women are typically found to be more risk averse than men. The risk-adjusted values of Facebook that we use assume that men and women have the same level of risk aversion. Women would need to be 37% more risk-averse than men for the difference to be significant at the 10% level, 41% more risk-averse than men for the difference to be significant at the 5% level, and 48% more risk-averse than men for the difference to be significant at the 1% level.

29. We did not ask questions on income but asked the zip code of where the participant lived at age 15. Using income data from this zip code, we find no significant difference in mean income for younger participants compared to older.

30. Appendix Table A.5.4 in ESM presents the Pearson correlation coefficients between the value of Facebook and several measures that characterize Facebook users.

31. In-state tuition at Texas A&M is $11,200 per year, or $350 per week, implying that participants value Facebook as much as 19% of the weekly cost of studying at the university. According to the College Board, the average university student in the U.S. spends $225 per week ($10,800 per year) on room and board. Facebook is then worth 30% of these expenses.

32. All but three treated participants stayed off of Facebook for the entire week. The three who did log back into Facebook did so only once for less than an hour to communicate for a student organization via the organization’s Facebook account. All three participants contacted the research team prior to logging in to inform us why they were logging back on. These participants are included in our intent-to-treat analysis. Instrumental variable estimates are 5% larger and slightly less precise.

33. It would be important to test for effects at different quantiles, but we do not have enough power to estimate meaningful comparisons at the tails of the distribution. To test for distribution equality, let \(F_{(1)}\) be the distribution of outcome \(y_{it}\) for the treated group and \(F_{(0)}\) be the distribution of the control group. According to Abadie (2002), we define \(F_{(1)}\) first order stochastic dominates \(F_{(0)}\) if \(\int _0^x dF_{(1)}(y) \le \int _0^x dF_{(0)}(y) \ \forall x \ge 0\) and \(F_{(1)}\) second order stochastic dominates \(F_{(0)}\) if \(\int _0^x ( \int _0^z dF_{(1)}(y) ) dz \le \int _0^x ( \int _0^z dF_{(0)}(y) ) dz \ \forall x \ge 0\)

34. Doing this involves a trade off between a Type I error and the power of the test (Anderson 2008). We control for the false discovery rate to adjust our p values and achieve a balance between these two factors.

35. For reference, both the unadjusted and adjusted p values are reported in Table A.5.6. All of our results remain statistically significant at the 5% level or less, with the exception of the probability of answering “Don’t Know” for skewed news, the healthy activities index and the change in the value of Facebook. We also do a more robust adjustment controlling for the family-wise error rate. When we use the free step-down method described by Anderson (2008), only the effects on Facebook use, news access through social media, news consumption and the correct answer of skewed news are statistically significant at conventional levels.

36. We tested whether the Facebook restriction had different effects for true headlines and the false headlines we created (by changing a few words) in the news quiz. For both types of headlines, the point estimates are similar to the main results and statistically insignificant.

37. Gender differences do emerge. While both men and women are less likely to be aware of the veracity of skewed news when off of Facebook, the effect is much stronger for men than women. This suggest that men, more than women, are exposed to politically skewed news when on Facebook.

38. Our research design restricted usage of Facebook for those in the treatment group, but participants were not restricted in their usage of other social media platforms. We validate that those in the treatment group did reduce their use of Facebook—by 1.7 h per day. Given a baseline Facebook usage of 1.9 h per day, this illustrates that the treatment group did comply with the restriction. While the treatment group refrained from using Facebook, we find that they did not increase their usage of other social media (e.g. Instagram, Snapchat, Tumblr, Twitter). This is consistent with studies finding low cross-platform usage for social media and a significant cost to switch to alternatives for one week (Pew Research Center 2016). Only one-third of Facebook users are active on other social media platforms, yet about 90% of users of other platforms are active on Facebook (Pew Research Center 2016).

39. The questions are: (1) Overall, how satisfied are you with life as a whole? (2) Overall, to what extent do you feel that things you do in your life are worthwhile? (3) How happy are you? (4) How often do you worry? and (5) How often do you feel depressed? An alternative approach could have been to use the Day Reconstruction Method (Kahneman et al. 2004; Kahneman and Krueger 2006), however, to keep the survey short, we opted for the five OECD questions.

40. Our results on life satisfaction are smaller than Tromholt (2016) who finds a significant effect of 0.26 standard deviations. The study’s Danish sample is older (average age of 34 years) compared to our U.S. sample (average age of 20 years), and participants were contacted daily by the researcher team to follow their assigned treatment status.

41. Subjective well-being measures can be sensitive to temporary events (e.g. the weather, long lines at a coffee shop, meeting somebody) (Krueger and Schkade 2008), nonetheless, because our participants are randomly assigned to treatment, random shocks should be evenly distributed and our panel estimation allows us to directly control for events that affect both groups uniformly across time.

42. There is evidence that eating out is associated with excessive calorie intake (Urban et al. 2016), a less healthy diet (Wolfson and Bleich 2015), increased hypertension (Seow et al. 2015) and a higher exposure to phthalates (Varshavsky et al. 2018), which have been linked to asthma, breast cancer, type 2 diabetes and fertility issues. Diet is correlated with an individual’s mental health (O’Neil et al. 2014).

43. Participants were asked on a scale 1–5 to what extent they agreed with a particular statement, where 1: Strongly Agree, 5: Strongly Disagree. We adjust the coding so a higher value indicates a “healthier” response.

44. The adjusted p value is 0.125. Our sample size allows us to detect effects up to 0.182 percentage points at the 5% level with a power of 80%.

45. A key characteristic of an addictive good is that its consumption exhibits “adjacent complementarity” (Becker and Murphy 1988; Gruber and Köszegi 2001), which means that past consumption increases the marginal utility of present consumption.

46. Appendix Figure A.5.2 in ESM shows that while the level of depression in the treatment group has decreased relative to control group, there is no evidence that suggests that treated participants are internalizing this benefit by lowering their value for Facebook.

47. In our sample, about 43% of women post comments on Facebook at least once a week, compared to 21% of men. Also, 23% of women post pictures at least once a week compared to 8% of men.

Authors and Affiliations

1. Department of Economics, Universidad de las Americas, De los Granados E12-41 y Colimes esquina, Quito, Ecuador

   Roberto Mosquera

2. Department of Economics, Texas A&M University, 4228 TAMU, College Station, TX, 77843, USA

   Mofioluwasademi Odunowo, Trent McNamara, Xiongfei Guo & Ragan Petrie

Corresponding author

Correspondence to Ragan Petrie.

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Reprints and permissions