Prediction 1: Social influence

Evolutionary models have formally examined the adaptiveness of social learning, defined as copying the knowledge or behaviour of other individuals, relative to asocial learning, defined as personally evaluating behaviours or knowledge with no influence from others [15]. One prediction of these models is that social learning should be used when asocial learning is particularly costly or ambiguous [16,17]. This prediction has received empirical support in both cultural evolution [18,19] and social psychology [20] experiments. Note that while this is a generally adaptive strategy, it can in certain cases lead to maladaptive ‘informational cascades’ where individuals copy inappropriate information from others without directly evaluating its effectiveness [21].

Given that it can be difficult to determine from its content alone whether a submission should be up- or down-voted, and also given that social information concerning others people’s voting decisions is freely available, we predict that Redditors should report using previous voting decisions as a guide to their voting decisions. A recent large-scale randomised experiment on a news aggregation website similar to Reddit [22] supports this prediction, finding that artificially up-voting a comment significantly increased the likelihood that actual users would subsequently up-vote that comment. Another study used replicate cultural markets to show that song preferences are susceptible to social influence, given that different songs became popular in different markets [23,24]. Our survey provides a test of whether Redditors are aware of the effect of social influence or not.

Prediction 2: Informational content

Evolutionary models predict that people should preferentially copy and transmit useful or relevant information [6,16,25,26]. Empirical studies have identified various criteria for ‘usefulness’ or ‘relevance’ such as that the information concerns social interactions [27,28], contains supernatural or non-intuitive concepts [29,30], or elicits emotional reactions of disgust [31,32]. For our purposes, we might predict that Redditors up-vote content that has characteristics such as wide anticipated appeal or uniqueness. One recent study conducted textual analysis of user reviews on a retail website [33], finding that the rated helpfulness of a review is predicted by the review’s length, detail and understandability (e.g. use of short rather than long words). Our study represents a test of whether Redditors are aware of these effects. We might also expect Redditors to up-vote content that they personally agree with, and down-vote content that they personally disagree with, given evidence that people typically transform information to fit pre-existing beliefs [14,34].

Prediction 3: Model-based bias

The fitness benefits of social learning can be further improved through selective learning from evidently skilful, successful or prestigious individuals, as they may be especially likely to possess useful information [12,35]. This has again been supported by experimental studies where participants preferentially copy successful individuals in various tasks [18,19,36,37] as well as copy individuals who others have looked at or shown deference in some way [38,39]. Here, we might predict that Redditors will report preferentially up-voting content that is submitted by individuals who they like and/or respect, and down-voting content that is submitted by individuals who they dislike and/or do not respect.

Prediction 4: Group norm enforcement

Norms are defined as “learned behavioral standards shared and enforced by a community” [40] p.218. Social psychologists have long demonstrated the powerful role that norms play in directing human behaviour and judgement [41,42], while recent developmental psychology studies show that from a very early age children internalise and follow norms for even arbitrary behaviours [43]. Chudek and Henrich [40] argue that culture-gene coevolution has resulted in this powerful norm-psychology given that there were likely fitness benefits to both coordinating group behaviour, and adopting majority behaviours that represent the accumulated wisdom of previous generations.

Voting on Reddit takes place within an online community that can by itself instil particular shared norms in its users. Redditors are encouraged to adhere to an informal set of values known as the “reddiquette”, in addition to various moderated rules that guide user behaviour inside particular subreddits (subreddits are lists of submitted comments and links on specific topics, such as ‘worldnews’, ‘politics’ or ‘movies’). Given the possibility that people have evolved to enforce the behavioural standards of their community [40], it seems reasonable to expect that Redditors may internalise and try to enforce their community values by up-voting or down-voting content on the basis of its compliance with Reddit norms. We therefore predict that compliance with Reddit norms will be an important influence on Redditors’ voting decisions.

Prediction 5: Enforcement of prosocial norms

Although in theory any behaviour can be stabilised as a social norm [44], there has been much interest in prosocial (or altruistic) norms which entail a cost to the individual and a benefit to the group [40]. Prosocial norms have been detected using economic games in which people willingly punish others who under-contribute to public goods [45]. This altruistic punishment has been observed across all human societies that have been studied [46,47], and especially in the Western societies from which the majority of our respondents come. One explanation for the widespread existence of prosocial norms is cultural group selection [6,16,40], wherein throughout human history groups with stable prosocial norms out-competed and replaced groups without such prosocial norms (although other theories based on purely individual benefit have also been proposed: [48]).

The occurrence of anti-social behaviour on the internet is well known (e.g. ‘trolling’). As noted above, Reddit is an established and successful online community, so we might expect Redditors to possess norms designed to prevent anti-social posts from disrupting that community. Our final prediction is therefore that Redditors will up-vote content that exhibits prosocial sentiments (e.g. praise or helpfulness), and down-vote content that exhibits anti-social sentiments (e.g. personal abuse).

Overview

To summarise, our aim in this study is to examine whether people report their online voting to be motivated by (i) the votes of others via social influence, (ii) the informational content of posts, (iii) the characteristics of the poster, via model-based bias, (iv) the enforcement of group norms, and (v) the enforcement of prosociality. All of these predictions derive from prior models and lab experiments that aim to identify the adaptive design features of human information use. Some predictions have already been tested in an online voting context (e.g. [22,33]), although not using self-report surveys. While self-report surveys have many weaknesses, and people may well not be aware of the actual reasons behind their voting decisions [49], it is instructive to know the explicit motivations of online voters, and whether these match with the results of online experiments [22,23] and corpus analyses [33].

We take a two-stage approach to our survey. To avoid guiding our respondents towards the predictions above, we first take advantage of a subreddit called “Theory of Reddit” where frequent voters post their motivations for up- and down-voting content. From these we assembled a list of 29 commonly-stated reasons, which contained the predicted reasons but also a range of others. We then surveyed 489 Redditors to see which of these 29 were most important, how they clustered together, and whether the characteristics of Redditors (e.g. their age, gender, or time on the site) predicted their evaluation of each.

Ethics statement

The study was approved by the Durham University Department of Anthropology Research Ethics Committee. All participants viewed an informed consent web page and agreed to it by ticking an electronic box before proceeding with the study.

Participants

Respondents were recruited through notices posted in subreddits that are concerned with how Reddit communities work (www.reddit.com/r/TheoryOfReddit), anthropology (www.reddit.com/r/Anthropology) and surveys (www.reddit.com/r/SampleSize). These subreddits were chosen because their users were expected to be interested in being part of the study. Participants were invited to follow a link to an online survey, which was hosted between 12^th to 20^th November 2013 using SurveyGizmo (www.surveygizmo.com) survey software. The sample consisted of 489 Redditors (236 females, 248 males, 5 gendered as “other”). The participants had a mean age of 26 years (s.d. = 7.78) with an age range of 18 to 64. The majority of participants were located in the United States (68.1%), Canada (8.0%), United Kingdom (5.9%) or Australia (3.5%), with the remaining 14.5% coming from 33 countries none of which individually exceeded 1.6% (8 participants).

Materials

We distributed a survey to each participant that contained qualitative and quantitative questions in order to gain an insight into Redditors' motivations for voting, as well as brief demographic questions (see S1 File for full survey text). The first qualitative part of the survey asked participants to vote on a sample of 15 Reddit comments and then to write brief explanations for these decisions. The quantitative part of the survey contained generic questions about what influences the respondents’ usual voting behaviour on Reddit. This paper focuses only on the quantitative results from the second part of the survey.

During the first phase, the search term “upvote downvote” was used inside the subreddit “Theory of Reddit” to look for posts where Redditors previously discussed their motivations for voting. Search results were sorted by relevance and four posts whose titles referred to motivations for voting were opened. One of the researchers (MP) then read and analysed these posts and their response comments, which included a total of 180 comments made by Reddit users. A dramaturgical qualitative coding framework described in [50] was used to consider the users’ objectives, conflicts, tactics, attitudes, emotions and subtexts when analysing their written motivations for voting. This preliminary investigation identified 29 recurring themes upon which the quantitative questions in the survey were based. These characteristics encompassed traits that were necessary to address the predictions (e.g. content sounding intelligent, agreement or disagreement, aggression or consideration for other people etc.), as well as other unanticipated qualities. Participants were asked to rate the importance of these 29 characteristics as influences on upvotes and downvotes, placing their answers on a 5 point Likert scale ranging from “not important” to “very important”.

In addition to questions about voting motivations, respondents were asked to state their age, gender and Reddit membership duration category (1–6 months, 7–12 months, 1–2 years, 2–5 years, more than 5 years). They were also asked to state their frequency of Reddit visits, votes on posts, votes on comments, submissions of posts and submissions of comments on a four-point scale (daily, weekly, monthly, few times a year or less). Two variables were then created to represent the mean frequency of post and comment votes, and the mean frequency of post and comment submissions. The respondents’ location country was provided by SurveyGizmo. Respondents were asked to state the extent to which their overall voting decisions are influenced by emotional reactions to content, objective evaluations of the content’s quality, online reputation of the poster, and the length of time the poster had been a Redditor, where answers were given on a 5 point Likert scale ranging from “not at all” to “a lot”. Finally, there were two yes/no questions that asked respondents if they take more notice of highly up-voted content and content that is accompanied by Gold badges, which can act as indicators of commendation or popularity on Reddit.

Analyses

All analyses were conducted using R version 2.8.0 [51]. First, a correlation matrix was calculated for the 29 variables concerned with different content characteristics as influences on voting. Principal Components Analysis (PCA) was used to explore whether these responses were organised in ways that reflected a smaller number of broad underlying forces, using the ‘principal’ function in R package psych [52]. The number of extracted factors was based on Eigenvalues greater than 1, and orthogonal (varimax) rotation was applied to better identify each item with a single factor.

Factor scores were created for each individual by calculating the mean of the raw scores corresponding to all items loading on a factor, as recommended for exploratory studies such as this one [53]. Consequently, factor scores retained the scale metric of the original Likert items, allowing for easier interpretation in subsequent analyses. These factor scores were then used as dependent variables in a series of ordinal logistic regression models to see if the importance scores of different voting influences varied depending on the voters' personal traits. Ordinal logistic regression was used because the dependent variables were 5-point Likert responses and were therefore not normally distributed. We used the clm and clmm functions in the R package ordinal [54] (clm gave identical results to the more popular polr function for ordinal logistic regression, but also allowed us to model random effects via the clmm function). Thirteen predictors were included in the models. Continuous predictors were the frequency of the respondent’s Reddit (i) visits, (ii) votes and (iii) contributions, (iv) their age, and the extent to which respondents report their voting being influenced by (v) their emotional reaction to the post, (vi) their objective evaluation of the post, (vii) the poster’s reputation, and (viii) the membership duration of the poster. Categorical predictors were the respondents' (ix) location country, (x) gender, (xi) Reddit membership duration, and whether they take more notice of (xii) up-voted content and (xiii) gold-badged content. Country was entered as a random effect given that shared location may generate non-independence in responses. For gender, ‘male’ served as the baseline against which ‘female’ and ‘other’ were compared. For Reddit membership duration, ‘1–6 months’ served as the baseline against which ‘6–12 months’, ‘1–2 years’, ‘2–5 years’ and ‘more than 5 years’ were compared. To avoid inflated Type I error rates associated with stepwise regression [55], we ran an initial full model with all thirteen predictors and then removed predictors with p>0.05. Model comparison was then used to check that all remaining predictors significantly improved model fit; where they did not they were removed. The remaining best-fit models are presented here. The full data file is available as S1 Dataset.

Discussion

The aim of this study was to explore the self-reported motivations of frequent content voters on the popular aggregator website Reddit. We were particularly interested in whether these motivations might reflect features of human cognition and learning that have evolved in the offline world to process social information in an adaptive manner, as predicted by the formal models and experiments of cultural evolutionary theory.

An initial qualitative review of subreddit discussions in which Redditors themselves proposed reasons for up- and down-voting content yielded 29 potential motivations. These were then used in a quantitative survey of 489 active Redditors to assess their importance in a much larger sample, and identify individual differences between Redditors in their ratings of each. Principal Components Analysis revealed eight clusters of reasons. In descending order of perceived importance, these were Wish Others To See, Intelligence / Uniqueness, Reddit Norms, Prosociality, Unoriginality, Empathy / Humour, Disagreement Of Opinion, Social Influence and Attitude Towards User. The factor Unshared Experiences / Bad Memories was dropped due to low importance.

Some of these factors clearly map onto the predicted motivations for information sharing identified in the Introduction based on evolutionary principles. The factor Social Influence was predicted (Prediction 1) based both on theoretical analyses of the costs and benefits of social information use [15–17] and previous experiments showing social influence in online voting decisions [22]. However, Social Influence was here reported to be relatively unimportant compared to other factors (see Fig 1), with a mean importance score of just 1.66 (sd = 0.76) on a scale of 1–5. Previous randomized experiments [22] have shown that artificially introducing positive up-votes increased the chance of subsequent positive ratings by 32%, a not insubstantial effect. This mismatch may be because Redditors, like people in general [49], are unaware of the power of social influence over their behaviour. This combination of strong effect yet lack of awareness in those affected raises ethical questions over the power that social media websites, advertisers and other organisations may have over our online behaviour. We also found that younger respondents attributed a greater importance to Social Influence, a finding that was further supported by results showing that younger respondents were significantly more likely to report taking notice of highly up-voted content and Gold badges. It may be that younger people are simply more aware of the effect of social influence. Alternatively, younger people may genuinely be more susceptible to social influence, a possibility that deserves further study. As would be expected, and confirming the validity of our measures, respondents who attributed a greater importance to Social Influence also reported taking more notice of up-voted content and the reputation of the poster.

The following of group norms (Prediction 4) was reflected in the factor Reddit Norms, which concerned whether the submission obeyed the rules of, and is relevant to, the particular subreddit. Obviously human cognition has not biologically evolved to follow the rules of subreddits specifically, but evidence suggests that we have evolved to readily absorb and follow whatever local social norms we encounter [40–43]. Our findings reinforce this notion. Reddit Norms was rated of high importance (mean = 3.71, sd = 1.11; see Fig 1). Interestingly, norm compliance was rated as more important by younger respondents, longer-serving Redditors, and more frequent voters and posters. The effect of age is similar to that for Social Influence, with younger respondents more likely to acquire group norms via social learning. The other predictors may reflect an experience-dependent socialisation period, with more experienced Redditors having become more familiar with the norms of the Reddit community and thus more likely to have observed them in operation and being enforced. Reddit Norms was also rated as more important by female respondents, an unanticipated sex difference that deserves further investigation.

As predicted, norms specifically related to prosociality (Prediction 5) were rated as important by our respondents in the form of the Prosociality factor. Respondents were willing to up-vote (reward) submissions that reflect prosocial attitudes, and down-vote (punish) submissions that reflect abuse or aggression. Interestingly, the biggest individual difference variable that predicted Prosociality was whether the respondent used emotional responses as a guide to voting. This fits with recent studies showing that people are more cooperative when they use intuitive emotional processing rather than lengthy reflection and deliberation [45,56], and suggests that deep-seated evolved motivations for prosociality may shape online interactions.

Our second prediction (Prediction 2) concerned informational content, reflecting previous research suggesting that human cognition has evolved to bias information towards particularly salient representations, or ‘cultural attractors’, via ‘content biases’ [6,16,25,26,30]. In our study this appears to be reflected in more than one factor: Wish Others To See, Intelligence / Uniqueness, Unoriginality, Empathy / Humour, and Disagreement Of Opinion can all be seen as relating to the content of the submission. This suggests that there is no single criterion for informational content. The first factor, Wish Others To See (i.e. the up-voting of content that the respondent wants others to see), received the highest overall importance score (mean = 4.27, sd = 1.00), but is rather understandable given that this is the purpose of voting. This factor is not very theoretically informative given that it does not specify why the respondent wants content to be viewed by others, and how they select such content. Intelligence / Uniqueness and Unoriginality are more specific, suggesting that respondents use the perceived quality of a post as a guide to voting, e.g. whether it is intelligently coherent, or whether it is unique or original such that it will be novel to other Redditors. This reflects previous content analyses of up-voted reviews [33]. An important individual difference variable for both of these factors was whether the respondent uses objective criteria to guide their voting, which makes sense given that intelligence and originality can be objectively assessed.

Empathy / Humour also relates to the informational content of submissions, with respondents reporting being likely to up-vote content that refers to shared experiences, that elicits sympathy or support, that contains humour, or that the respondent agrees with (Table 1). Previous studies have shown that people are likely to remember and share emotionally salient information [31,32,57]. Individuals who report voting based on their emotional reaction to content give Empathy / Humour especially high importance. Female respondents also rated this factor as more important than male respondents, in line with previous behavioural and neurobiological findings that women show superior empathy than men [58] and women more than men find the production of humour attractive [59].

As part of Prediction 2 we expected that respondents should up-vote content that fits with their pre-existing attitudes and beliefs, and down-vote content they disagree with. In the study, however, agreement and disagreement did not cluster together: Disagreement Of Opinion factored on its own and was rated as relatively unimportant (mean = 2.69, sd = 1.23), while Agreement was grouped with Empathy / Humour. This may be a result of our self-report methodology, with respondents unwilling to admit to partisan voting based on their prior beliefs. However, it may be a genuine result driven by the community values that redditors are encouraged to adhere to, where down-voting based on personal opinion is particularly frowned upon.

Finally, our Prediction 3 that respondents would vote based on whether they like or respect the poster of the submission was not supported, since the factor Attitude Towards User was rated lowest in importance (mean = 1.60, sd = 0.83; see Fig 1). This is interesting given experimental evidence that people preferentially learn from high status, liked or knowledgeable individuals [35,38,39]. It may be that this is another influence that is outside explicit awareness. Alternatively, this may be a domain where content-based learning biases outweigh model-based learning biases. This is consistent with the finding that model traits such as the poster’s online reputation and Reddit membership duration were thought to be minimally influential by participants, whereas content-related traits were an important and recurring theme throughout the results.

It is important to acknowledge the limitations of our study. First, the reliance on self-report measures may not be representative of the true motivations behind voting behaviour, as we note above in relation to social influence. Second, the recruitment of respondents through targeted subreddits means that the sample may not be representative of the general user base of Reddit, or internet users in general. Third, the requirement to fill out a lengthy survey may have biased towards the recruitment of individuals with especially prosocial tendencies, which may have biased the results concerning prosociality in particular. For these reasons our findings should be treated with scepticism until they can be replicated with broader samples and more focused methods to address the emergent themes with greater precision.

In conclusion, this exploratory empirical study demonstrates that Redditors’ patterns of content voting can offer unique avenues of research into people’s interactions with information and its providers, and suggests a novel arena for testing evolutionary theories of social information use. Whereas expressions of fondness, agreement, aversions or punishments can be subjective in traditional settings, online votes provide clearly delineated forms of action that are well suited to empirical investigation.