How does anonymous online peer communication affect prevention behavior? Evidence from a laboratory experiment

While the importance of social networks for health behaviors is well-recognized, relatively little is known regarding the accuracy of anonymous online communication and its impact on health behavior. In 2012, we conducted a laboratory experiment in Boston, Massachusetts with 679 individuals to understand how anonymous online communication affects individual prevention decisions. Participants had to opt for or against investing in prevention over three sessions, each consisting of 15 experimental rounds. In the third session only, participants could share their experiences with a group of 1–3 other anonymous participants after each round. Groups exchanged an average of 16 messages over the 15 rounds of the third session. 70% of messages contained information about the subject’s prevention decision and the resulting health outcome. Participants were more likely to communicate when they prevented than when they did not, with prevention failures resulting in the highest probability of sending a message. Nonetheless, receiving an additional message reporting prevention increased the odds a subject would prevent by 32 percent. We find that participants tend to adopt the prevention behavior reported by others, with less weight given to the reported outcomes of prevention, suggesting that social networks may influence behaviors through more than just information provision.


Introduction
A growing body of evidence suggests that social networks may play an important role in people's health behaviors such as smoking [1], alcohol consumption [2], the types of foods eaten [3], and amount of exercise performed [4]. While some of these correlations within networks could emerge because people tend to associate with those who share similar characteristics or because of common environmental influences [5][6][7], there is also experimental evidence that social networks may directly influence people's health behaviors [8][9][10].
There are two primary channels through which social networks are likely to affect behavior. First, direct information exchange within social networks may enable people to learn about the PLOS  advantages and disadvantages of engaging in a health behavior. Recent studies suggest that information received from peers related to the efficacy of health technologies can encourage uptake of effective health products such as vaccines [11], insecticide-treated bednets for malaria prevention [12], and artemisinin-based combination therapy for malaria treatment [13]. Second, people may imitate the behavior of their peers because they want to adhere to what they perceive to be the norm within their social networks [14]. Research has shown that providing information to people on how their peers are behaving can influence health behaviors such as alcohol and tobacco use [15,16], dietary patterns [17,18], and physical activity levels [8,19].
In recent years, anonymous online social networks such as health-focused support groups and discussion boards have become an increasingly important source of health information exchange [20,21]. The relatively low costs of participating in online social networks can increase interactions with others [22], and can also allow access to tailored health information [22][23][24]. However, the anonymous nature of many online social networks may also reduce the reliability of the information being shared [22,23,25], as is often the case with vaccines [26][27][28]. Informational biases may also result from selective reporting of experiences [29]. Previous reviews have found mixed and generally small effects of online social networks on health behaviors [30][31][32][33] and there is little quantitative evidence on the impact of anonymous online information-sharing on health behaviors and health outcomes [34,35].
We used a laboratory experiment to assess the quantity and quality of anonymous health information shared online, as well as the impact of the information shared on the uptake of an illness prevention technology.

Ethics statement
Ethical approval for this study was granted by the Harvard T.H. Chan School of Public Health Internal Review Board (Protocol# 22455-02). Written consent was obtained from all participants.

Overview of study
The experiment was conducted in 2012 at a university laboratory with 679 adult participants recruited via flyers and emails from the Greater Boston area. The experiment was designed to simulate a setting where individuals need to make decisions about whether to invest a small amount of money in an illness prevention technology which reduces, but does not eliminate, the probability of falling ill (for the purposes of the experiment, we specify this as an income loss to simulate the monetary costs often experienced due to sickness). Each subject participated in 3 sessions consisting of 15 rounds each. Participants were given the following instructions: "You will be playing this game for 15 periods. In each period you will earn an income of US$ 10 if you stay healthy and an income of zero if you fall sick. The probability of falling sick in each period is constant at p = 0.[x] (a [x] in 10 chance) over the game period. In each period you can invest in a preventive health technology for $1 which lowers the probability of falling sick." [x] was either 3, 5 or 7 depending on the participant (see the section "Randomized Parameters" below). If participants chose to invest in the prevention technology, their incomes when they remained healthy and when they fell sick were $9 and $-1 respectively (see S1 Fig for details on the experimental design).
Participants were recruited by email using the lab's email list (a list of people who had participated in previous studies or expressed interest in doing so) and through Craigslist using the standard procedures of the experimental lab. The experimental session was about one hour and thirty minutes long. All participants received a show-up fee of US$ 10. Participants were informed that they would receive additional payments corresponding to the outcomes of one randomly selected round experienced with each of three sessions. Across the three sessions, participants could thus earn an additional income between $-3 (prevented and got sick all three times) and $+30 (no prevention, never sick). The average payout per person, including the $10 everyone received for participating, was $28.

Peer messaging
After the first two sessions of independent decision-making, participants were informed in the third session that they would be able to communicate with other players in their group by sending and receiving messages. Each participant was randomly assigned to a group of 1-3 other anonymous participant(s). All group members were present and visible to each other in the computer lab as they participated in the study at the same time. However, they would not have known specifically who among the participants were a part of their messaging group and we generally observed very little communication between subjects before or after the session.
The laboratory was set up for a maximum of 20 participantswho would be randomly grouped into groups of four. The vast majority of groups consisted of 4 people (154/177 groups or 87%), which was the default for the study. Smaller groups only happened if the number of subjects that showed up for a given laboratory session was not divisible by 4. So, for example, if 19 people showed up, there were four groups of 4, and one of group size 3. Overall, there were only 17 (9.6%) groups with 3 people, and 6 (3.4%) groups with 2 people.
At the end of each of the 15 rounds, participants simultaneously sent messages and received messages from other members of their group, using textbox windows.

Randomized parameters
The experiment randomly varied a number of parameters including the probability the participant fell sick in the absence of prevention (baseline illness rate), the effectiveness of the prevention technology in reducing the probability of falling sick, and the information participants received about the prevention technology. S1 Table includes a summary of the different experimental treatment arms.
The baseline probability of falling sick without prevention was either 0.30, 0.50, or 0.70, was known to all participants, and was the same for all members within a messaging group. The effectiveness of the prevention technology was also randomly assigned with some participants receiving a more effective technology (absolute risk reduction of 20, rather than 12, percentage points). In addition, some participants were randomized to receive public health messaging that encouraged investing in prevention.

Socio-demographic information
At the end of the session, participants completed a survey which collected demographic information, and details about their usual health and prevention behavior. They were asked if they were currently taking multi-vitamins (yes/no), whether they had gotten a flu shot in the last 12 months (yes/no), whether they believed all children should be vaccinated (yes/no) how many times they had visited a dentist in the past year and how often they use sunscreen (always/ often/occasionally/never). The health prevention behavior was collected after the experiment in order to avoid priming the participants' prevention decisions in the experiment by asking about their usual prevention behavior [36].

Message coding
We coded all messages exchanged into categories that we expect would have an impact on others' behavior: messages that provided some information about participants' experiences with or without prevention and messages that expressed participants' attitudes towards prevention. We used the following categories for the messages: "Informational", "Encourages Prevention", "Discourages Prevention" and "Conversational." A message was coded as "informational" if it contained at least some information about the subject's prevention decision and/or illness outcome. When messages contained information, we also coded whether the subject communicated their success or failure with prevention/non-prevention in that particular round (for example if the subject said "I invested and did not get sick") (S2 Table provides further examples).
Messages were classified as "encouraging" or "discouraging prevention" if they mentioned the advantages of prevention or the disadvantages of prevention, either generally or in reference to this specific prevention technology. "Conversational" messages were those that did not fit into any of the other categories (S2 Table provides further examples). Some messages were coded as both containing information and either encouraging prevention or discouraging prevention. Although the coding of the messages was performed by a single author, two coauthors independently re-coded 10% (280 messages) each in order to check the reliability of the coding. Across all types of messages, there was 94% agreement in categorization between the two coders (for both sets of messages that were re-coded). In cases of disagreement, the final coding decision was jointly made by the author team.
Using the information contained in the messages on whether the subject prevented in that round and the illness outcome, we calculated, by group, the reported illness rates with and without prevention. We compared these reported rates to the expected illness rates, given the assigned probability of falling ill in the absence of prevention, and the risk reduction when using the prevention technology. Since the baseline probabilities of falling ill varied by subject, we re-scaled all reported rates so that the expected probability of falling ill in the absence of prevention was 0.5, and the expected probability of falling ill with prevention was 0.34 (the mean expected absolute risk reduction within all group was 16 percentage points)

Analytical approach
In a first step, we used multivariable logistic models to assess the determinants of posting a message online in the third session. The primary variable of interest in this first step were participants' experiences (utilization of prevention and outcome) as well as the messaging behavior of others. In the regression models we controlled for the age, gender, marital status, parental status, education, income, ethnicity, occupation, usual prevention behavior of the subject, the public health messaging the subject received, the effectiveness of their randomly assigned preventive technology, and their randomly assigned baseline illness rate. Usual prevention behavior included whether the subject took vitamins, had a flu shot in the past year, favored child vaccination, the number of dentist visits she/he had in the past year, and whether the subject regularly used sunscreen. Standard errors were adjusted for clustering at the group level.
In a second step, we used multivariable logistic models with session and participant fixed effects to assess the effect of online messages on prevention. We included all three sessions in this analysis and each observation consisted of a subject-round. Our outcome of interest was a binary variable for whether a participant prevented in a given round. Our main independent variable in this analysis was the number of messages the participant received in the previous round which was coded as zero for all rounds in the first two sessions. We included participant fixed effects in order to account for individual-level factors that could influence the propensity to send messages as well as their propensity to engage in prevention. These regression models also controlled for the effectiveness of the prevention technology which varied for participants across the three sessions and included session fixed effects to control for general learning. Standard errors were adjusted for clustering at the individual level.
All analyses were conducted using Stata/SE version 14.0 (StataCorp, College Station, TX) [37]. Table 1 provides detailed descriptive statistics on the study sample. Participants were, on average, 31 years old. 52% of participants were male, 54% were white, 14% black, and 6% Hispanic. Only 11% were married and 14% had children. 53% of individuals were currently enrolled as university students. Accordingly, participants were on average highly educated (52% had a college degree or higher) with relatively low income-only 17% of participants had an annual income above $40,000. Given the high proportion of college students in our sample, however, the distribution of income may not accurately represent the socio-economic background of our participants.

Characteristics of study participants
In terms of their general prevention behavior, individuals reported a mean of 1.3 dentist visits over the past year and 88% believed that all children should vaccinated. However, only 39% used sunscreen "often" or "always", 44% were taking multi-vitamins, and 47% had had a flu shot in the past year. A slightly higher proportion of our sample reported that they engage in these prevention behaviors than the population as a whole [38][39][40][41].

Baseline prevention behavior
In the first two sessions, individuals independently made their decision about whether to invest in prevention. In these two sessions the mean prevention rate was 70%. Only 2% of participants (14/679) never prevented in these two sessions, while 25% of participants (170/679) prevented in every round.

Messaging behavior
In the last session, participants were able to share information with peers in their group. The average group size was 3.8 and the vast majority of groups (87%) consisted of four participants. Approximately 60% of participants (82% of groups) sent at least one message over the 15 rounds of the third session, and individuals (groups) sent, on average, 4.2 (16) messages each. This resulted in individuals' receiving, on average, 12 messages over the 15 rounds. The distribution of the number of messages sent by groups is shown in S2 Fig. Table 2 shows the multivariable regression results for messaging behavior in the third session. The coefficients on both a baseline illness rate of 0.5 and 0.7 (relative to the reference group with a baseline rate of 0.3) are large and statistically significant, suggesting that the general messaging propensity strongly increases with the risk of falling sick. Columns 2 and 3 of Table 2 show the associations between participants' prevention experiences and messaging behavior. The reference group in both cases is participants who did not prevent and did not fall sick in the previous round. In the first round (Column 2), when participants had not yet received any messages from others, only those who prevented and fell sick had a marginally significant higher odds of sending a message compared to the reference group. In rounds 2-15 (Column 3), previous prevention without getting sick was associated with increased odds of sending a message (OR = 1.47, 95% CI [1.12 1.94]), and those who prevented and fell sick were even more likely to send a message. In terms of the direct response to other messages, each additional message received from others in the previous round was associated with 3.46 times the odds (95% CI [3.05 3.92]) of sending a message. Fig 1 displays the proportion of people who sent a message, across all 15 rounds of the third session, by whether they prevented, and whether they fell sick. Overall, people who prevented were more likely to send a message, as were people who fell sick. The highest messaging rates were among people who prevented but had adverse health outcomes. Fig 2 summarizes the content of the 2806 messages that were sent over the 15 rounds. Approximately 70% of messages contained some information about the current and/or previous rounds, 8% of messages directly encouraged others to invest in prevention, and nearly 5% discouraged prevention. 21% of messages were conversational and unrelated to the laboratory experiment or to prevention. In Fig 3, we use the information contained in the messages to plot the distribution, across groups, of participants' reported illness rates when they did and did not use the prevention technology. In the 55/177 groups (31%) where a subject reported on non-prevention outcomes at least once, the median reported re-scaled illness rate without prevention (0.5) is approximately the same as the expected illness probability without prevention (Fig 3, Panel A). However, in the 97/177 groups (55%) where participants reported on prevention outcomes at least once, the median reported re-scaled illness rate with prevention was 0.16 points higher than the expected illness probability, which means that prevention on average looks substantially less effective than it is (0.5 median illness risk with reported prevention compared to the expected probability of 0.34) (Fig 3B). In Column 3, we divide these informational messages further into those that reported successful (healthy) or unsuccessful (sick) outcomes with prevention and non-prevention ("risktaking"). Receiving a report of successful prevention in the previous round was associated with a 53% increase in the odds of prevention in the current round (95% CI [1. 25 1.86]). However, reports of failed prevention were also associated with a slightly increased, though statistically non-significant, odds of prevention (OR = 1.10, 95% CI [0.91 1.33]). Reports of successful risk-taking were associated with a 43% reduction in the odds of prevention (95% CI [0.42 0.79]) but there was no statistically significant association between reports of failed risk-taking reporting of health outcomes with prevention. Recent evidence suggests that the quality of health information available online is mixed [23,44,45]. Our evidence indicates that even if individuals share accurate data, selective reporting could result in the aggregate information being biased.

Messaging and prevention
Our third main finding is that despite the reporting bias in favor of failed prevention, online information-sharing is associated with higher prevention rates. These positive impacts appear to be both the result of higher messaging frequency among participants who prevented (as shown in Fig 1 and Table 2), and a tendency for individuals to imitate the behavior of others rather than respond only to reported effectiveness. That is, we find that participants were more likely to prevent when they received reports of others preventing, regardless of the outcome, and participants were less likely to prevent when others reported not preventing, regardless of the outcome. Participants were most likely to engage in prevention when other participants reported to have successfully prevented in the previous period, and least likely to prevent when others reported having successfully not prevented, which shows that information on outcomes  Peer communication and prevention also plays an important role in adoption of a new technology. However, even failed prevention reports result in marginally higher propensities to engage in prevention which suggests that imitation is an important driver of behavior in this setting. Overall, the results presented are consistent with recent evidence that people both imitate others' behavior and respond to the expected benefits of the action, though we find that the first channel dominates the second in our setting [11][12][13][14]14]. Our results are also consistent with a recent online experiment examining adoption of a new application [46]. The authors found that invitations from friends to use the application-which could be considered an endorsement of the product-had a larger effect on adoption than simple notifications that their peers were using the application [46].
Our finding that the actions of peers have more influence on health behaviors than information on effectiveness may offer insight into why public health messaging often has small impacts on people's behaviors [47,48]. Our experiment is most relevant for understanding individuals' prevention behaviors when the health outcome can be observed shortly after investment in the prevention technology and adverse outcomes are relatively frequent regardless of prevention. These include, for example, flu shots, flossing to prevent cavities or handwashing to reduce the risk of foodborne illnesses. In other cases, such as the use of sunscreen, the true benefits of using prevention may not be visible for decades. The imperfect connection between the prevention behavior and the outcome could help explain why people are more responsive to others' prevention behaviors rather than the impact it had on their health. Our results may be less applicable for preventive behaviors such as wearing sunscreen, where the time frames are much longer and adverse health outcome being targeted is uncommon.
There are some limitations to this study. First many of the participants in this study were students and relatively well-educated and their prevention behavior and response to messaging may not necessarily be generalizable to other populations. Relatedly, the groups in this study were relatively small whereas most online communication platforms reach many thousands of people. In addition, this experiment was conducted with a hypothetical prevention technology. People may behave differently when confronted with the choice of investing in an actual prevention technology. People may also behave differently when their behavior is not being observed by researchers. Lastly, the effect of communication on health behaviors may be different when individuals are able to select the peer group with whom they communicate, particularly if they share common traits [49].
However, laboratory experiments also offer several advantages in studying online behavior: first the decision-making environment is tightly controlled, allowing us to determine how variation in different parameters affect messaging behavior. Second, peer groups in the experiment were randomly assigned which limits the likelihood that peer effects are due to common unobserved factors among the groups [50]. Third, by recording the information shared among participants we can determine both the accuracy of the information shared as well as how it influences behavior. Laboratory experiments have been previously used to study a number of different aspects of social learning [51][52][53].

Conclusion
As people increasingly turn to online sources for health information, there is growing interest in finding ways to harness social networks for improving health behaviors while limiting the negative impacts of mis-information. Our results suggest that online communication platforms do not have to present unbiased information in order to have positive effects on health behaviors. In some contexts however, the tendency towards imitating others' health behaviors could have negative health impacts; for example social norms may drive the overuse of antibiotics [54]. This problem may be exacerbated if people select social networks that offer them a skewed sample of experiences [55,56]. One potential policy implication of our results is that encouraging people to report on social networks when they have invested in prevention behaviors could encourage others to do the same, though the effectiveness of this strategy would need to be studied further. More work is also needed to understand where people seek information on health and how the source of information influences their behavior.
Supporting information S1 Dataset. Data on prevention and messaging from laboratory experiment. Individual data on prevention from all three sessions of laboratory experiment and on messaging from third session. (DTA) S1  Table. Examples of types of messages sent. All messages were coded into one of these four categories. Some messages were counted as both "informational" and either "encourages prevention" or "discourages prevention." (DOCX) S1 Fig. Summary of experimental design. The probability p that a person fell ill when they did not prevent (p) was either 0.3, 0.5 or 0.7. The probability that a person fell when they prevented (p') was lower than the probability of falling ill when they didn't prevent (p'<p).