I finally broke down and spent the money to get the audio from this presentation, and thought I'd give a summary plus some additional info that I dug up after listening.: "Analyzing Virtual Worlds: Next Step in the Evolution of Social Science Research". American Association for the Advancement of Science (http://www.aaas.org/). 2009 AAAS Annual Meeting February 12-16, 2009, Chicago Illinois. The research team analyzed three years of data (over 60 terabytes) from the complete server logs and click-streams of SOE's MMORPG game, EverQuest 2, tracking every action performed in-game. EQ2 has more than 300,000 players who average 26 hours per week playing the game. Because of the intense level of involvement and multi-player environment of the game, the researchers were able to study human behavioral dynamics using the game as a proxy. Dmitri Williams -- University of Southern California Noshir Contractor -- Northwestern University Jaideep Srivastava -- University of Minnesota Scott Poole -- University of Illinois, Urbana-Champaign Sponsors: National Science Foundation, United States Army, Sonic Lab Northwestern University, Sony Online Entertainment, and others. The research team was quite varied. Aside from the primary investigators, specialists in other disciplines such as economists and anthropologists were also involved. Database analysis was handled at the National Center for Supercomputing Applications at the University of Illinois, Urbana-Champaign. Past studies have relied on surveys and self-selected samples. This project involved actual server data from EverQuest 2. It's estimated that in the Western world (ignoring Asia), there are about 40-50 million MMORPG player accounts, mostly fantasy RPGs. Some areas of interest for study are how (or if) the virtual world mirrors real world human interactions. For example, networking behavior, economic behavior, how groups work, how people talk, the way people get into conflicts, the way people learn, etc. Mapping Principle: when do behaviors in a virtual space act as a proxy for behaviors in real life interactions? In some cases, the results from virtual actions map closely to observed real world results, while in others the results do not correlate well. It's not scientifically possible to say with certainty how accurate the mapping is without further study. Prior work in this space was primarily ethnographic and survey-based research. This approach has problems due to self-selection and also difficulty in generalizing results. The current project allowed the researchers to work with a huge body of data. This was also correlated with in-game surveys of random players, allowing some interesting depth of insight. Who plays how much and why? It's not kids, not pasty white teenage boys in their mother's basements. This study found the average players were adults in their thirties, there are more players in their 30s than in their 20s. The age range gravitates to older audiences. EQ2 players tend to play a lot of hours per week. Older players tend to play more hours per week than younger players. Hours spent playing were less than average hours of TV watching per week, and gaming tends to displace consumption of other media such as TV. Players don't watch less news, but they do watch less entertainment television and go out to movies less, replacing their entertainment consumption with gaming. An unexpected finding was that players tended to be healthier than the general population, as measured by their Body Mass Index (BMI). They tended to be slightly more depressed, but less anxious. The healthiest participants in this sample were the older women. This led to interesting thinking about gender differences in gaming. Surveys revealed that while women were in the minority of players, they tended to be the most intense players, which is counter-stereotypical. The women were more likely to be bisexual, more likely to be disabled, and were a little bit older on average than the men. Researchers have fairly strong notions of gendered behavior in a society, which is to say that they acknowledge both nature and nurture, and realize that people's behaviors are shaped by gender roles into which they are socialized. Scientists realize and recognize that there are roles and expectations of types of behaviors that people feel they must stay within, and therefore they tend to expect women to behave certain ways and men to behave differently. The research team in this study made hypotheses about these behaviors and found that, in fact, men are indeed the majority of players in this space, which was not surprising, because frankly, these games are mostly made by men, for men. However, when the researchers looked at who the "hard-core" players are, as defined by the intensity of play and the attitude towards play, the hard-core players were the women. Women play more hours per week (four more hours per week), and are much less likely to quit a game. Women were also much higher on measures of satisfaction and interest in the game than male players. When researchers looked at the players playing at the very high end, the players playing the largest amounts, the women are much more highly represented than they are in the overall player population. This led to an interesting side-finding with some implications for social science research and this area in general, which is the difference between how much time a player self-reported that they played, versus how much they actually played as reflected in the server logs. When people were asked how much they play, they estimated between 24 and 26 hours, but they actually played 25 to 29 hours per week. This is interesting and notable for two reasons: (1) It suggests that the past 30 years of scientific estimates of game play in prior studies relying on self-reported estimates are incorrect, and may be off by as much as 10% if these findings are generalizable, and that marginal findings in this research are suspect. (2) The women clearly have a bigger gap in these two forms of data than men do, which suggests that women are either lying to themselves or lying to researchers moreso than the men are, which may confirm that these are gendered behaviors and gendered expectations. If video games are, in fact, "for boys", then you would expect women to underreport more, as there might be a social stigma or shame value attached. The observed behavior seems to fit this explanation. Men and Women Playing Together Another side finding is what happens when men and women or boys and girls play together. When men and women play together, the women are happy and the men are less happy. When they play apart, the men are happier and the women are less happy. Thus, what's good for the goose is not good for the gander. The women like to be playing in relationships, the men are less happy when doing so. Roleplaying Another interesting finding came when looking at roleplayers. These are "Massively Multiplayer Online Role Playing Games", which suggests that people are pretending to be something they are not, taking on the "role" of a character they are playing. In fact, the study found that relatively few people actually participate in roleplaying. A "moderate" level of roleplaying might be 15% of the players. When people are asked who is an "intense" roleplayer, that tended to be about 5% of the players, and these people are very, very different than the rest of the players. The roleplayers play for different reasons, and seem to have different outcomes. Why roleplayers roleplay, what do they get out of roleplaying, the kinds of identity issues, these are all rich avenues for future investigation. The roleplayers are psychologically much worse off than non-roleplayers. They tend to be more depressed, they tend to have been diagnosed with more ailments and illnesses, they are more likely to be physically disabled. Gameplay does not make them this way: interviews with these players tends to reveal use of gaming and roleplaying as coping mechanisms for dealing with pre-existing problems. It is hypothesized that roleplaying is an adaptation for people who are less accepted in real life spaces, and who go into gameworlds to be "who they truly are" or who they want to be. Investigating the demographics of roleplayers, intense roleplayers were more likely to be non-heterosexual, non-mainstream-religion, and minorities. The further out of the mainstream population they were, the more likely a player is to engage in roleplay. This suggests that the game space may be a little less judgemental and more accepting. Economics Economist Ted Castronova of Indiana ("Virtual Worlds: A First-Hand Account of Market and Society on the Cyberian Frontier"). Looking at the EQ2 data, Castronova investigated whether people behave "rationally" in the in-game economy, as they would be expected to do in real life. They investigated whether they could at least make predictions about basic economic aggregates, based on data culled out of the logs. The first step was calculating the GDP of one of these worlds, and also calculated price levels. They discovered that these were, over time, fairly steady, but could be a bit more volatile than normal non-economic-crisis real world trends. The Quantity Theory of Money (the theory that money supply has a direct, positive relationship with the price level) does apply in Norrath as it does in real life. Networks in Virtual Worlds Noshir Contractor investigated networks in virtual worlds. During the session, an article from Science magazine was passed out discussing how, with the advent of virtual worlds, computational social science can be performed on an unprecedented scale (David Lazer, Alex Pentland, Lada Adamic, Sinan Aral, Albert-Laszlo Barabasi, Devon Brewer, Nicholas Christakis, Noshir Contractor, James Fowler, Myron Gutmann, Tony Jebara, Gary King, Michael Macy, Deb Roy, and Marshall Van Alstyne. "Computational Social Science". Science 6 February 2009: 721-723.). EverQuest 2 is a multi-dimensional network, in which some of the nodes are people, some are documents, etc. There are different kinds of relationships, chatting, trading, buying, selling. It's of interest how these various nodes and structures are co-evolving and what kind of structures they lead to. To perform an analysis of networks in EQ2, the team investigated motivations. Why did people create links with someone else? Why send a chat message? Why would a player want to group with others for quests or other excursions? Why trade with someone? Why send someone in-game mail? Some of this is based on self-interest: I need something you have and I am going to economically maximize my own goal. Another could be based on social exchange: I want something from you, you want something from me, we have complimentary expertsie that brings us together. A third is that neither player wants anything from the other, but that together they have a better shot at beating or killing a certain monster, the collective action or mutual interest. A fourth is contagion, in which a new player wants to connect with those who are already really well connected, those who are rich in network links are going to get richer in such links. Fifth is balance, in which I want to play with someone not because they're particularly good at it, but because they're a friend, or a friend-of-a-friend, and I want to be able to keep balances between my friendships. Sixth is homophily, in which a player seeks others that are the same age as themselves, or the same gender, or other attributes that are common to me and that provides the networking motivation. Seventh is proximity, in which a player wants to play with others in their geographical area -- although you'd think that geographic proximity wouldn't matter in a virtual world, they discovered that it does have an impact on who plays together; or electronic proximity, in which players who have both been in the game a long time interact; temporal proximity or cohort effect in which players started playing at the same time and as a result have more connections with each other. From a network analysis point of view, each of these motivations has a unique structural signature. In a real world, messy network, researchers can use statistical tools to extract the extent to which each of these structural signatures is present in the network. For example, one could determine that a given network's exchanges are based upon 10% self interest, 15% social exchange, 20% by homophily, etc. This allows the researcher to decompose the network and understand the motivations that are driving the network. People playing EQ2 may be interacting both virtually and also within a shared physical space, and things that happen in the online world may spill over into being upset with the other in the real world also, or vice versa. It's been theorized that virtual worlds would expand our ability to interact with people widely, but we also see that it acts in some ways as an extension of the regular offline world and its interactions. Types of Relationships in EverQuest 2 Interpersonal interactions/Partnership interactions, in which two players come together to play together in a combat activity. What motivates two players to interact with one another? Why do two players use instant messaging within the system? Transactional interactions. What motived two players to trade with one another, which requires a "face to face" meeting in the game world? Why does one player send in-game mail, which may be just a message or may include an item? Network analysis was performed by looking at one week of activity on the Antonia Bayle server: Player age range from 5.6 to 71 years of age. IM was a very small network, trade was the largest. The overall structure of the network was quite similar to offline networks. Hypotheses: Individuals are not likely to engage in interactions randomly in the virtual world. The study confirmed this hypothesis. There individuals that are "hubs" in the network, individuals who are more likely to engage in interactions. The study determined that people who were "hubs" engaged much more in partnerships and chat, but less in trade interactions or in-game mail. Friend-of-a-friend, in which individuals who both interact with the same third party are more likely to interact with one another. The study determined this was the case in EQ2, especially for partnership and chat interactions. Those who have greater real world geographical proximity are more likely to engage with others from their real-world geographical area online as well. Counterintuitively, people who are close by (10km) are 5x more likely to interact in game than those far apart (100km or more). This may be due to the fact that these people have offline networks, such as being real-world friends, which induces them to play together more in-game. Those with greater real world temporal proximity, that is, those in the same time zones, will interact more. Players in the study were 1.5 times more likely to play with others in their own time zone. Researchers also examined homophily interactions, seeing whether people were more likely to play together based on gender, age, or in-game experience. There are not good age cues in-game to let people know the real life ages of other players, so age was not a factor in the formation of relationships. People of the same gender also were not more likely to engage in interaction. Females have a strong tendency to interact with male partners, but a negative female to female motivation, even when you control with the gender disparity overall within the game. Network structure data may be able to help SOE to identify illegal plat sales, and also to boost player retention. Large Dataset Data Mining Jaideep Srivastava is an engineer and computer scientist interested in data mining and machine learning. Large volumes of data provides challenges in analysis. From an algorithmic and computational basis, large dataset data mining is an interesting challenge. In trying to understand behavior using the EQ2 dataset, focus was placed on how items interest to the marketing world, such as customer churn, could be applied to virtual worlds. There is an opportunity to observe the player behavior down to every click, and from that there's a chance to understand behavior from that at a higher level with great accuracy. The new environment creates changes in behavior. For example, in the real world people do not often abandon a full shopping cart, but this is not at all uncommon at virtual retailers such as Amazon.com. Behavior online changes based on the technology and new modes of the new environment. Customer churn has been extensively studied in real world interactions by ISPs, phone networks, subscription-based services etc. because the cost of retaining an existing customer is far less than acquiring a new one. Churn analysis is focused on identifying customers who are likely to churn, and determining how to retain those customers. For SOE, it's of interest to determine who are "key" players, because a person who is an "influencer" or "hub" who leaves to go to another game may take all their friends with them. For example, a popular game such as World of Warcraft can suck population from smaller games such as Lineage and Lineage 2. The study looked at determining whether a given player was likely to churn or not likely to churn. Traditionally, marketing research into churn has looked only at each individual in isolation, because they don't have access to their networks to understand well the social influence of one person on another, although cell phone companies have started to be able to model such data. It's quite valuable to understand how interactions between players affect subscription and retention decisions. In examining the EQ2 dataset, researchers were able to look not only how much a player played in the previous week or quarter, but also who that person played with. The amount of time a person plays correlates with their engagement with the game. With this large a data set, multiple analysis passes through the data are not feasible. And most networks are not greatly scaleable.