29 June 2016 Group 3
Dirk Damsma Academic year 2015/2016
Bachelors Thesis (Supervisor: Junze Sun) Semester 3 period 1
Identifying network effects in the Steam video game market
Kristaps Karnītis 10630368
Abstract
This paper aims to research to what extent network externalities are present in the Steam market for video games. The first section evaluates numerous papers discussing both theoretical models of network externalities, as well as empirical studies attempting to identify and measure network effects in real world industries. The theory is then used to construct a methodology and a multiple regression model for evaluating whether or not network effects persist in the Steam markets. The results of an analysis of 227
multiplayer games suggests that there is indeed a significant relationship between the consumer base of a game and its price, namely, a 1 percent increase in the size of a games consumer base implies a minimum of a 0,09 percent increase in the games market price on Steam. While statistically significant, the estimated model is subject to numerous biases, including simultaneous causality and omitted variables, therefore further research is necessary in order to confirm the internal validity of the model and external validity of the findings.
2 | P a g e
Statement of Originality
This document is written by Student Kristaps Karnitis (10630368) who declares to take full responsibility for the contents of this document.
I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it.
The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
3 | P a g e
Table of contents
1. Introduction p. 4
2. Theoretical framework p. 6
2.1. Experiential goods with network externalities effects:
An empirical study of online rating systems p. 6
2.2. The City Network Paradigm: Measuring Urban
Network Externalities p. 7
2.3. Bertrand Competition in Markets with Network
Effects and Switching Costs p. 7
2.4. Lock-in vs. Critical Masses – Industry change
under network externalities p. 8
3. Methodology p. 9
4. Data analysis p. 12
5. Summary p. 15
6. Conclusions p. 17
4 | P a g e
1. Introduction
Group formation and network effects or externalities affect not only numerous industries, such as telecommunications, social networking sites and video games, but also other disciplines of science, such as politics, religious affiliation and linguistics. While numerous studies have both derived theoretical models and analyzed network externalities empirically, academic understanding of network externalities is limited. This is due not only to limited data availability and measurement, but also issues with endogeneity and simultaneous causality. Moreover, network effects are hard to quantify, as their effect on market outcomes varies due to other factors such as switching costs and rate of adaptation in the given market. The aim of this paper is to identify and quantify the network effects that persist in the video game industry and more specifically - Steam markets. It can be argued that there are no significant switching costs in this market, aside from the videogames price.
The online video game industry and widely available multiplayer games are growing at an increasing pace, not to mention they have created numerous large and profitable firms such as Blizzard Entertainment, Valve, Steam and Rockstar Games (Yang J. & E. Mai, 2010, p. 1). A quote from Electronic Arts annual report hints at the importance of network effects in this community of interconnected players, “…creating new online games and cultivating new customers is an important part of our growth strategy”. Yang & Mai also claim that massively multiplayer online games involve direct network externalities, as a larger installed base of users leads to more utility to customers through increased connections, available servers, improved service, community created game content, reviews, guides and many other factors.
Afuah (2013, p. 2) claims network effects can be considered present, when the marginal utility a consumer derives from a good is dependent on the size of the goods existing consumer base. Numerous articles have both constructed theoretical models to derive the market outcome dynamics and producer incentives resulting from network effects being present. Numerous researchers have also studied the relevance of network effects in practice. In this paper the focus will be put on empirical studies of network
5 | P a g e
effects, as well as the relevant theory for constructing a methodology that reduces measurement error and biases caused by simultaneous causality and omitted variables.
The specific market to be studied will be Steam – one of the largest online video game retailers in the world, accounting for roughly 75 percent of all PC games bought online. The goal of this paper is to answer the question: to what extent does the size of the existing active consumer base affect a multiplayer video games price in Steam markets? Note that the Steam video game market is marked by low switching costs and a high willingness of consumers to adapt to new, more preferable products, which arguably makes it a favorable market for an empirical study in this field.
In order to answer the research question, a literature study will be conducted, inquiring about previous empirical studies and relevant theoretical models. In addition, an empirical study, including a multiple linear regression will be conducted with data
describing 227 multiplayer games on the Steam market.
Answering the given research question would provide a tangible method for
quantifying the strength of network effects in the video game industry and for other goods, which are priced in a free market with low switching costs. Moreover, observing the market outcome and the strength of the network effect in a market would allow to evaluate the predictive power of existing theoretical models, considering these two market factors. As a result, the research would elaborate upon academic studies conducted not only in the industry of video games, but industries with considerable network externalities and low switching costs in general.
The first section of this paper will focus on identifying and describing the relevant literature dealing with the subject of network effects. The second section will use the theoretical insights in order to maintain and define the regression model used in the paper. In the third part, the multiple linear regression will be conducted and implied empirical results will be evaluated. Lastly, a short summary, research limitations and conclusions will be presented.
6 | P a g e
2. Theoretical Framework
While there are few concrete studies of network effects in the video game industry, there are numerous empirical studies in other industries, such as social networking sites. Moreover, there are numerous articles developing elaborate models to predict market outcomes, given the strength of network effects and other market factors, such as switching costs. In this section, the most relevant literature will be introduced.
First of all, an empirical study of online rating systems for video games will be described to gain an insight into previous research that attempted to measure and quantify network externalities in the video game industry. The second article considered will be another empirical study researching the network effects within collaborating cities. The third subsection will introduce a theoretical model developed by Suleymanova & Wey, which can be used to describe the market outcome in industries with network externalities and low switching costs. The last section will consider the theoretical solution of the apparent paradox of network effects – the market leader should lock-in and become the only monopolist in the market, however in practice innovators still manage to outcompete incumbent firms even when considerable network externalities are present. All these studies will be described with the goal to gain knowledge about network effects and their empirical measurement.
2.1. Experiential goods with network externalities effects: An empirical study of online rating systems
Yang and Mai (2010, p. 5) deduced that direct network externalities are present in the market of massively multiplayer online games. Their findings show that when a game has proved its popularity, consumers trust its quality despite the introduction of negative word of mouth. It is argued that for new adaptors a large user base can effectively function as a search attribute to convey the existing users’ evaluation of product quality and overall
7 | P a g e
consumption experience. Moreover, it is presented that customers prefer a large amount of user reviews with lower overall score more than fewer reviews with a better score,
signaling that users value the existing amount of connections and size of community over the perceived review score of the game. The findings in Yang and Mai’s paper are all deduced through carefully constructed regression models. Their study will be used as an example for creating the methodology and regression model in this paper.
2.2. The City Network Paradigm: Measuring Urban Network Externalities A recent research paper elaborates upon the missing empirical evidence on city networks and their economic advantages (Capello R., 2000, p. 17). So-called urban
network externalities allow cities to benefit through co-operative behavior. The regression model estimated in this paper shows a significant positive relationship between urban policy implementation and the degree of connectivity to other cities. This result confirms the validity of the city network theory, by empirically showing that the performance of cities benefits, when the city is in a network with more connections to other cities. This study gives an empirical approach to identifying and measuring network effects, by quantifying both the extent of network connections and resulting benefits. As such, it will be a basis for the methodology of this paper.
2.3. Bertrand Competition in Markets with Network Effects and Switching Costs
In this paper, the authors discuss current market observations that lead to conclude that markets with network effects often exhibit a critical mass effect, such that a firm which reaches a certain size first, becomes the market monopolist thereafter (Suleymanova I. & Wey C., 2011, p. 8). In Bertrand competition, this dynamic can be modeled by
network effects and switching costs, where the market has four possible states. The
extremes are a monopoly or a market sharing outcome, however at any point both could be defined as preserved or reversed.
8 | P a g e
It is derived from the model that the stable state of the market depends on the ratio of the switching costs relative to the network effects. If the ratio is high and it is relatively costly to switch, then a unique market sharing equilibrium exists. In the alternate case, multiple equilibriums prevail. In both cases, the market will follow a preserved pattern, when switching costs are large and a reversed pattern, when network effects are large. Following the reasoning, there exists a region where critical mass effect occurs, such that the initially dominant firm is guaranteed to become a monopolist.
The model also unveils a conflict between maximizing consumer surplus and social welfare. While positive network effects imply that coordination towards a single good maximizes the network effects and social welfare, consumer surplus is maximized when firms are required to compete and minimize prices.
The key takeaway for this research is that in industries with low switching, there are multiple possible market outcomes that can prevail. Moreover, market shares of firms are more likely to follow a reversed pattern, implying frequent changes in market leaders. Indeed, in practice market leaders on Steam gain and lose momentum rapidly, while only some reach a size large enough to sustain sizable communities for years. Moreover, one must recognize that this market is prone to the conflict of maximizing consumer surplus and social welfare. As is, the market is filled with competing video game developers, however according to the model social welfare would be maximized if one game had a monopoly and maximized its network effect (note the models theoretical assumption of uniform consumer preferences).
2.4. Lock-in vs. Critical Masses – Industry change under network externalities Witt U. (1997, p. 2) presents a paradox – industries with positive network
externalities favor the firms that have a high market share, regardless of the inherent value of the solution they offer. As such, this would imply technological stagnation and
increased resistance against new innovators. Examples of such industries are computer operating systems, color television coding standards, typewriter keyboards and video recording devices. In these industries, the firms offering the most adopted solution can
9 | P a g e
lock-in and outperform competition, even if the product they offer is inferior to smaller competitors. As a result of a bias towards industry leaders, economic policy makers favor standardization or subsidize new technologies which face network dis-economies. In the paper Witt proposes a model which explains how a locked-in incumbent can be
outcompeted by an innovator, given a critical mass of potential adopters.
This model is relevant in understanding how past Steam market leaders such as Counter-Strike: Source (2004) has been outcompeted by modern innovators such as Call of Duty. As mentioned before – a market with low switching costs is prone to reversing current patterns, not to mention the Steam markets are marked by consumers who are willing to adopt the products of new innovators, even if the network externalities favor the much larger senior developers. As such this allows to conjecture that the data collected from the Steam markets will be composed of both consumers that follow the gains of network effects and early adopters, who favor joining smaller game communities.
3. Methodology
The previous section discussed numerous articles relating to both empirical and theoretical studies of network effects. As such, Yang & Mai’s studies of user review scores of massively multiplayer online games and Capello’s study of network
externalities between collaborating cities uncovered the basis of the methodology in this paper. The estimated model will attempt to measure the effect that the amount of
available connections that active players have on the consumers perceived utility,
measured by the willingness to pay. The goal of the analysis is to identify whether or not a statistically significant relationship persists between the observable proxies of the target variables. Namely, willingness to pay is measured by the price in the Steam market and available connections are measured by active players in the past two weeks.
10 | P a g e
The null hypothesis is therefore
H0: There is a significant positive relationship between the amount of active players and the games price in the Steam video game market
Following the reasoning of Capello’s paper, measuring the network connections between individuals, be it players or cities, is a direct way to approximate the extent of the effect that the network externality has. In this case, the network exposure will be measured by the active player count in the recent period, which is an acceptable
approximate of the amount of connections and network externality benefits that any new player will have available. Moreover, according to Suleymanova and Wey’s model, a market with low switching costs ensures that consumers are free to distribute according to the strength of current network externalities, ensuring that the price mechanism represents the value that consumers perceive from each individual game. In order to avoid biases at this step, games with discounts or short term offers will be discarded and only one-off price offers with significant maturity will be considered.
In order to test the null hypothesis, data on 227 strictly multiplayer games was collected from Steam’s market database SteamSpy.com on May the 29th, 2016. The database does not offer an option for directly exporting all data, so each data point was manually collected and rechecked with the database on May 30th to ensure no significant collection errors or shocks within the market had taken place. The collected variables are the title of the game, the genre of the game, years since release, beta dummy, single player dummy, market price in US dollars and active player count in the last two weeks. All variables were imported into a Stata database and prepared for analysis.
The next step will be preparing a multiple linear regression model. This model will be constructed in order to establish whether or not significant statistical relationships are present between the price as the dependent variable and other collected variables as the independent variables. Note that the games title and genre are purely descriptive and will not be considered in the analysis. After the data collection, an additional analysis was conducted in order to identify significant outliers in terms of price and active player base.
11 | P a g e
The evaluation yielded no noteworthy entries, so the whole dataset of 227 multiplayer games will be used in the analysis.
The model will use active players in the last two weeks in order to proxy the active user base and the current games price in the free market to proxy the consumers’ willingness to pay and marginal utility. In addition, three control variables are also included in the model, namely:
Years since release, which measures the games progress through the lifecycle Beta dummy, which controls for the price changes of early access games, with
considerably lower quality
Single player dummy, which controls for the price changes of games, which have an option to also be played offline and without other players
Moreover, the model will attempt to account for the diminishing marginal returns nature of the network effects, by using the logarithms of both price and active player count, as opposed to absolute values. The years since release, beta dummy and single player dummy variables are introduced in the regression with no additional
transformations.
The resulting multiple regression model is therefore:
Ln(Pi) = β1 + β2*Ln(2wkActPlrs)i + β3*YrsSncRls i + β4*Beta i + β5*Sngl i + εi
Where
Ln(Pi) = the logarithm of game i’s price in USD,
Ln(2wkActPlrs) = the logarithm of active player count over the last two weeks, YrsSncRls = Years since release,
Beta = Dummy accounting for early access games,
12 | P a g e
The regression data will be loaded into Stata and a robust regression will be
conducted in order to account for heteroscedasticity, which cannot be ruled out given the existing information. The entire process of transforming price and user base variables into logarithms and running the regression is included in the .do file along with the original dataset.
Once the regression is finalized, the validity of the models fit will be evaluated via the F-statistic, which is provided in the Stata output display. In the current methodology, the viability of the null hypothesis will be estimated via the coefficient β2. It is assumed that if the coefficient β2 does not statistically significantly differ from zero, there is enough statistical evidence to reject the null hypothesis. If, however the β2 coefficient is
statistically significant, further analysis will be necessary in order to infer whether or not this is conclusive evidence for a causal relationship or merely suggestive of an underlying phenomenon.
The statistical evidence in favor of the null hypothesis will be estimated using the student t-ratio, which will be calculated with Stata. If the p-value implied from the
calculated t-statistic does not fall below 0.05, it is assumed that the coefficient β2 is indeed statistically insignificant and the null hypothesis will be rejected, resulting in a definitive answer to the research question. The following section will consider the data analysis given the methodology and theoretical framework discussed in previous sections.
4. Data analysis
The first part of this section will consider the overall statistics of the regression, while the second part will evaluate the significance of individual coefficients and conclude the statistical outcome relevant to the null hypothesis and research question. The goal is to evaluate the models statistical validity and implied empirical findings. The main focus will concern the variable β2, identifying and measuring the network externalities that are suggested to persist in the market.
13 | P a g e
Table 1. Overall regression results Number of obs. 227
F (4, 222) 52.40 Prob > F <0.0000 R-squared 0.4728
As can be observed in Table 1, Stata output suggests an F value of 52.4, which significantly exceeds the critical value, implying a p-value of less than 0,001. Note however, that the multiple linear regression model accounts for less than 50 percent of the deviation of game market prices, implying there are still numerous omitted variables beyond the explanatory power of this model.
Table 2. Regression output by individual variables
Table 2 presents the data from the multiple regression model of this paper. As can be observed, the regression implies that all five estimated coefficients have statistically significant relationships with the logarithm of the price of the game (the estimated p-values are all well below the critical value of 0.05).
While it is not certain to what extent the estimated trends can be applied to
approximate how the independent variables affect the price of a game, this paragraph will describe the intuitive results of the regression in terms of variable interaction. Namely, following the implication of a derivative, a 1 percent increase in active user base results in approximately a 0,125 percent increase in the video games price on average. This
14 | P a g e
implies that an effective doubling of the active user base would ceteris paribus result in a 12,5 percent increase in the price of a video game. On the other hand, each year that a game spends in a market after release, results in a 11,5 percent decrease in its initial price. An early access game in its beta period is expected to have a 61,2 percent lower price than a full-release alternative ceteris paribus. An option to play the game in single player meanwhile is expected to result in a 64,8 percent increase of the games original price. The constant of the model informs that a newly created, full-release, multiplayer-only game with no active community is expected to have a price of e^1,84=6,26$.
As such, this model shows multiple intriguing and intuitively valid dynamics present in the Steam market. The implication of the model for the null hypothesis is that the coefficient β2, approximated for the variable of LnActPlrs, is indeed significant, implying that there is not enough statistical evidence to infer that there is not a relationship
between the size of the active consumer base and the video games price. Moreover, the estimated coefficients lower 95 percent confidence interval mark lies at 0,0887, implying that a 1 percent increase in the active user base results in at least an 0,09 percent increase in the games price. This implies that an effective doubling of the games active player count is expected to result in a roughly approximated minimal 10 percent increase of the price that developers are able to charge in Steams free market.
The given result is highly intriguing, but requires proper in-depth analysis and further testing to maintain similar relations can be estimated in other data sets. The result
coincides with the theory discussed and the dynamics observed in the Steam market, as well as matches the strategic expectations of game developers, such as Electronic Arts who put online consumer base expansion as their number one priority for increasing sales volumes and revenues per individual game.
15 | P a g e
5. Summary
This section will aim to summarize the theoretical and empirical findings of this paper and deduce the implications that this has on the null hypothesis and research question discussed in this paper.
The goal of this paper was to gain insight into the network externalities of the Steam market and maintain to what extent the size of the active user base affects the price of a video game on Steams’ markets. The theoretical framework of this paper uncovered details of two empirical studies in the area of video games and network effects, which were used to establish methods for identifying and measuring network effects in real world industries. The other two discussed articles gave the theoretical background in order to form expectations of the market outcome in an industry with network
externalities and low switching costs, such as the one of Steams online retail platform. The resulting insight gave a solid basis for arguing why the Steam market is populated by numerous competing parties, as this matched the theoretical predictions of Yang and Mai’s model. The industry is also predictably marked by frequent change of market leaders and large amount of customers willing to adopt the products of new innovators, despite the existence of large, experienced developers with sizable communities and considerable network effects.
These insights were then used to form the core multiple linear regression model and supplement it with control variables that counter the biases caused by various factors not attributable to direct network externalities. The final regression model then
quintessentially linked the logarithm amount of active players with the logarithm of the games price on the Steam market, which allowed not only to identify and measure the network externality present, but also account for its nature of diminishing marginal returns to scale. Finally, the control variables – years since release, beta dummy, single player dummy – accounted for the games lifecycle and overall quality respectively. This model was then estimated on the data of 227 multiplayer games collected from the Steam market database SteamSpy.com.
16 | P a g e
Arguably, the estimated model allowed to identify and measure the network effect present in the Steam marketplace. Namely, it was deduced that a doubling of the games active user base allowed the developer to increase the games price by an approximate 10 percent, suggesting a significant increase of the consumers perceived utility of the game. This result then provides tangible statistics for inferring an answer to the proposed research question. The analysis conducted in this paper suggests that there is indeed a statistically significant positive relationship between the video games existing active consumer base and its price on the Steam market, confirming the intuition proposed in the central research question. Arguably this also suggests the existence of a substantial network externality in the Steam online retail platform
As such, this result uncovers significant characteristics of the Steam market place and industries with network effects in general. Moreover, it provides a basis for creating framework that can be used to evaluate the network externalities in other industries with free markets and low switching costs.
There are however significant limitations to the dataset, regression as well as theoretical basis considered in this research. First of all, while the dataset is sufficiently large for statistical analysis, it is limited to one time-period and one video game market, making claims of external validity less viable. Secondly, the internal validity of the methodology is limited as well, as one may raise significant concerns about the issues of simultaneous causality and omitted variable bias. For example, it is rational to argue that a direct comparison between the network externalities of different games in fact creates a significant bias. One might also argue that the observed statistical relationship between price and active user base is indeed the result of higher quality games generating more demand and thus resulting in a higher price on the market and attracting more players due to their inherent quality, which is not sufficiently accounted for in the model. Moreover, an argument may be raised that developers themselves actively increase the prices of their games, when they observe an increase of the consumer base, while the real increase of utility that each additional consumer creates is in fact insignificant. This claim would also coincide with Yang & Mai’s research, which uncovered that consumers trust and
17 | P a g e
evaluate more popular games with higher user count as more valuable, even if their actual ratings are somewhat lower.
In fact, issues with endogeneity are incredibly difficult to account for and numerous more potential biases might be present in the estimated model, leading to a false
conclusion about the implications of statistically confirming the null hypothesis. As such, the relationship between two variables that simultaneously affect each other, namely, price and consumer base are difficult to interpret and make concrete conclusions about.
6. Conclusions
This section will aim to make conclusions about the theoretical, as well as empirical findings of this paper, as well as conclude about the practical implications, limitations and further research topic suggested by the analysis in this paper.
The discussed academic research provided a theoretical framework for understanding the dynamics of the Steam online marketplace, as well as observing the practical efforts of previous research and establishing a methodology for identifying and measuring network effects. The resulting regression model then allowed to gain a superficial insight of the relation between the price of a video game and the size of its consumer base in Steam markets. It was found that a 100 percent increase of the active consumer base over the last two weeks leads to a minimum of a 10 percent increase of the games current price. This result therefore allows to statistically reject the hypothesis that there is no significant relationship between the two variables and allows to form an answer for the central research question.
Namely, the findings of this paper allow to infer that there is indeed a significant statistical relationship between the active user base and the price of a video game on the Steam markets. This leads to conclude that there is indeed a tangible, identifiable and measurable network externality present in the Steam marketplace. Moreover, the model uncovered the statistical significance of multiple other variables, such as the product life
18 | P a g e
cycle, beta releases and single player availability, which account for little below 50% of the deviation in the prices in the Steam markets. It is reasonable to conclude that video games prices are in fact affected by numerous factors, some which are and some which are not discussed in this paper.
The resulting findings uncover intriguing and intuitively coherent relationships between the variables in the Steam market, however severe limitations also persist. Chief among which is the fact that the multiple regression model considered in this paper is subject to significant biases due to simultaneous causality, endogeneity, as well as omitted variables. Moreover, the legitimate comparability of different games in the market is also difficult to confirm, resulting in a lack of internal validity. In addition, while technically statistically significant, the observations in this set of 227 multiplayer games from the Steam marketplace are hardly applicable to the entire population of Steam games or video games in general, especially across different timeframes.
Even so, the research conducted in this paper provides significant insight into the Steam market dynamics, as well as the markets for video games in general. Further research should include more control variables, such as user ratings and game
characteristics to account for a games inherent quality. Moreover, instrumental variables should be used in order to eliminate some of the bias caused by simultaneous causality present between a games price and user base. In addition, further research should consider other platforms and consoles aside from Steam and PC’s, as well as inquire about the presence of relationships across wider timeframes. These studies should also elaborate upon the existing methodology, in order to create a framework that is applicable not only to the Steam market place, but also to industries with considerable switching costs and other factors that affect the relationship between network externalities and free market prices.
19 | P a g e
7. Bibliography
Afuah, A. (2013)
Are network effect really all about size? The role of structure and conduct. Strategic
Management Journal, 6(34), pp. 257-273.
Capello, R. (2000)
The City Network Paradigm: Measuring Urban Network Externalities. Urban Studies,
11(37), pp. 1925-1945.
Suleymanova, I. and Wey, C. (2011)
Bertrand competition in markets with network effects and switching costs. The
B.E. Journal of Economic Policy & Analysis, 11(1), pp. 154-212.
Witt, U. (1997)
"Lock-in" vs. "critical masses" - industrial change under network externalities.
International Journal of Industrial Organization, 4(15), pp. 753-773.
Yang, J. and Mai, E. (2010)
Experiential goods with network externalities effects: An empirical study of online rating system. Journal of Business Research, 7(63), pp. 1050-1057.
