Appendix I   
Modeling Technical Appendix

 

This appendix details the regression analysis used to examine the effects of censorship on video game revenues that is presented in chapter 4. The following provides a thorough description of the regressions that were estimated, the data that were used, the results that were produced, and some potential limitations of this approach.

The video game analysis used a conventional linear regression of per-user video game revenues on a series of variables reflecting censorship and other determinants of video game revenues. The data used for the analysis, which are detailed later in this appendix, represent a panel consisting of country-level information on video game revenues and country characteristics for the period of 2017 to 2019 for 141 countries. Throughout, the analysis distinguishes between and reports separate estimates for video games that were sold digitally and those that were sold physically.

The video game revenues regression took the following form:

$$revenue{s}_{it}= {\beta }_{1}incom{e}_{it}+{\beta }_{2}populatio{n}_{it}+ {\beta }_{3}interne{t}_{it}+ {\beta }_{4}censorshi{p}_{it}+Y_t+ {ϵ}_{it}$$

Per-user revenues for country i in year t (revenues_it) are modeled as a function of per capita national income (income_it), population (population_it), and internet penetration rates (internet_it). Censorship is included as an indicator variable (censorship_it) that equals zero if country i had high levels of censorship in year t and 1 if it had low levels. A series of year fixed effects (Y_t) are included to capture trends in video game revenues over time. Finally, an error term (ε_it) is included to capture otherwise unexplained variation in revenues.

The model was estimated via ordinary least squares, which is a standard regression technique. The estimation results include heteroskedasticity robust standard errors, which help to provide more accurate measures of standard errors if heteroskedasticity is a present. Heteroskedasticity refers to cases in which the variance of the error term is not constant over the range of the sample.

The variables included on the right-hand side of the regression explain per-user revenues for video games. Censorship, as highlighted in the report, was expected to reduce per-user revenues by limiting market access and constraining the content of games. Income was included because higher levels of income likely allow for greater spending on video games at higher prices, thereby raising revenues. Population captures possible effects that market size (in terms of potential users) might have on video game revenues. Unlike with censorship and income where there were clear anticipated impacts on revenues, the likely relationships between revenues and population are less obvious, though the inclusion of this term improves goodness of fit in the results below.[1] Finally, internet penetration captures the impacts that internet connectivity may have on game revenues. In recent years, video games have increasingly relied on the internet for both distribution and as part of their game play. For these reasons, we might expect it to have an increasing effect on revenues$—$especially for digital games. Additionally, internet penetration is likely a good indicator of technology adoption more broadly, which could similarly impact video game sales.

The data for the analysis were derived from several sources. Information on video game revenues and users was sourced from two Statista data series covering worldwide digitally and physically sold video games.[2] Throughout the analysis, digital and physical sales were treated separately in order to avoid potential double-counting with respect to users. Within the data, it is not possible to determine the extent to which the physical and digital users overlap, which precluded them from being combined into a single user-base. The Statista data provided information on both total revenues and user penetration rates for 148 countries beginning in 2017. Information on income, population, and internet penetration was sourced from the World Bank’s World Development Indicators database.[3] The three series drawn from this database were “Adjusted net national income per capita (current US$),” “Population, total,” and “Individuals using the Internet (% of population).” Per-user revenues were calculated from these two sources. First, the video game user penetration data were combined with the total population data to calculate the total number of users in each country and year. Second, total revenues were divided by the total number of users to generate per-user revenues.

Information on censorship was derived from Freedom House’s Freedom in the World report and its accompanying database.[4] The Freedom House database numerically rates countries on a range of different aspects of political rights and civil liberties. For the analysis, a rating category titled “Are there free and independent media?” was used to identify the presence of video game censorship. Although not specific to video games, the basis of this media rating is reflective of many of the types of censorship activities that affect video games. For example, the rating reflects responses to questions like the following:[5]

·         Are the media directly or indirectly censored?

·         Are works of literature, art, music, or other forms of cultural expression censored or banned for political purposes?

·         Does the government attempt to influence media content and access through means including politically motivated awarding or suspension of broadcast frequencies and newspaper registrations, unfair control and influence over printing facilities and distribution networks, blackouts of internet or mobile service, selective distribution of advertising, onerous operating requirements, prohibitive tariffs, and bribery?

The Freedom House media freedom rating is categorical and ranges from 0 to 4, where 0 denotes the least freedom and 4 denotes the most freedom. For the regression analysis, censorship ratings were used to define a single indicator for high or low censorship. Ratings of 0, 1, or 2 were considered “high censorship,” while ratings of 3 or 4 were considered “low censorship.” The decision on how to divide the ratings between these two categories was based on the scores for the key markets discussed as having notable censorship throughout this report. In particular, of the six key markets (China, India, Indonesia, Russia, Turkey, and Vietnam), all but Indonesia exhibited ratings of 2 or lower. Based on this, 2 was selected as the cutoff for high censorship. It should be noted that despite the discussions elsewhere in this report that identify Indonesia as a key censoring market, it was rated a 3 by Freedom House in all years, implying that it was considered “low censorship” in this specific analysis. Across all countries and years in the sample, about 58 percent of countries exhibited high levels of censorship and 42 percent exhibited low.

The decision to simplify the media ratings into a high/low censorship indicator was based on the fact that the ratings are categorical and cannot be treated like a continuous measure of censorship. While there is a strict ordering to the ratings, their magnitudes may not accurately capture the differences between each level in terms of their impact on video game revenues. For example, there is no clear reason to believe that the effect of increasing from a 0 to a 1 rating is necessarily equivalent to that of increasing from a 3 to a 4. Similarly, it is not clear if the difference between 1 and 3 should be exactly twice that of 1 and 2. For this reason, defining and using an indicator variable based on the ratings is more appropriate than using the ratings themselves.

The final sample used for estimation covered 141 countries for the years 2017, 2018, and 2019. The earliest year was determined by the Statista video game data. The latest year was determined by the national income data, which was not available past 2019 at the time of writing. The country coverage was primarily based on the Statista video game data although missing data in the other series prevented several of the Statista countries from being included in at least some years.

The regression estimates largely fit the expectations discussed above and are, in most cases, statistically significant at conventional levels (table I.1). For both digital and physical games, low censorship is associated with statistically significantly higher per-user revenues. Net national income per capita is also associated with higher revenues. A $1,000 increase in income per capita is associated with an approximately $1.26 increase in digital video game revenues per person and a $1.46 increase in revenues for physical games. For internet penetration, a 1 percentage point increase in the percent of the population using the internet is associated with an approximately $0.30 increase in digital game revenues but a $0.16 decrease in physical game revenues, suggesting there may be a substitution between the two mediums as people become internet users. The estimates for population show a significant relationship with digital game revenues such that a 1 million person increase in the population is associated with a $0.02 increase in per-user revenues. By comparison, there appears to be a negative relationship between population and physical revenues, but the estimate is not significant at conventional levels. Finally, because there was no constant included in either regression, the estimated values for the year fixed effects can be thought of as the regression intercepts for each year. For digital goods, the values increase in each of the three years, implying that average revenues have grown each year over the sample time period. For physical goods, 2017 exhibits the highest value, implying the largest average per-user revenues occurred at the beginning of the sample. Finally, the R-squared values for the digital and physical games were 0.680 and 0.785, respectively. These values suggest that the models fit the data well and explain a substantial portion of their variation.

Table I.1 Regression results for digital and physical video game revenues per user

P-values <0.001 indicate very small values; * indicates a value between 0.00 and -0.01.

Predictor variables	Digital video games coefficient	Digital video games standard error	Digital video games p-value	Physical video games coefficient	Physical video games standard error	Physical video games p-value
Income per capita (1,000 $)	1.26	1.10	<0.001	1.46	0.07	<0.001
Internet penetration (%)	0.30	0.04	<0.001	-0.16	0.02	<0.001
Population (millions)	0.02	0.01	<0.001	*	<0.01	0.105
Low censorship	3.88	1.98	0.050	3.02	1.26	0.017
2017 fixed effect	-4.84	2.17	0.026	3.12	0.85	<0.001
2018 fixed effect	-4.66	2.43	0.055	2.58	1.22	0.035
2019 fixed effect	-4.55	2.52	0.071	2.74	1.31	0.036

Source: USITC calculations.

Note: Standard errors are constructed as heteroskedasticity robust standard errors. P-values indicate the probability that the true population coefficient value is actually zero. Digital video games refer to fee-based video games distributed over the internet. Physical video games refer to console and PC games distributed over solid storage media, such as discs.

As with any empirical analysis of this type, there are some considerations and potential sensitivities that should be noted. First, there may be additional factors that significantly influence video game revenues but were not included in the analysis. The omission of these variables could bias the existing estimates if they were correlated with any of the variables that were included. For the purposes of this work, the most problematic type of omitted variable would be something that is correlated with the high/low censorship designation, as it could inadvertently affect the censorship estimates. For example, one possible consideration could be the role of other types of institutional quality other than censorship, which were not included but could impact revenues and be closely tied to censorship. However, the inclusion of additional measures of institutional quality could pose a second type of issue in the form of excessive correlation. While the Freedom House data do provide many series that reflect other potentially influential measures of institutional quality, none were included in this analysis because of concerns that they would be too closely correlated with censorship. Ultimately, the chosen specification attempted to balance these issues.

A second consideration is the use of an ordinary least squares regression. This approach inherently introduces certain assumptions about the data and the relationship between the dependent variable and each of the explanatory variables. In particular, the approach assumes a linear relationship between per-user video game revenues and each other term. While this approach is effective at identifying linear relationships between these variables, it may not fully capture any nonlinear relationships that may exist between them. However, given that there is no good information or guidance from the literature suggesting what complex relationships might exist between revenues and these variables, there was no obvious choice for how nonlinear relationships ought to be modeled. Thus, the linear assumption was the best assumption available.

The third consideration is that potential sensitivities exist within the chosen model specification. There were potentially multiple different ways that the selected variables could be included in the model. For example, the designation of high versus low censorship could have been based on a different cutoff value. Similarly, the income, population, and internet variables could have been included as logged values, as is often done in regressions. In each case, changes to the way these variables were included could have impacts on the estimate values. While preparing the analysis, multiple potential specifications were considered. The final specification presented here reflects that which produced the best model fit of the data from among the set of alternatives considered.

Finally, much of the post-estimation analysis conducted using the regression estimates$—$including the computation of total revenue effects and the impacts on U.S. video game revenues$—$was conducted based on some assumptions about the key video game markets. Many pieces of information on specific conditions in each of the key markets, in general and for U.S. firms, were largely unavailable. In light of this, these additional analyses were conducted by applying information about global video game sales and average effects of censorship to the specific markets. In the case of the total revenue calculations, the global average censorship effects of $3.88 and $3.02 per user may not perfectly reflect the individual impacts in each of the key markets. Depending on how the censorship occurs in each market, the local effects could be higher or lower than the global average. Further, the local impacts could depend on the types of games being sold in each market. For example, internet shutdowns likely impact online multiplayer games more severely than games without any online components. Similarly, China’s console ban likely had large impacts on many console-based games but more limited impacts on many digital mobile games. In the case of the U.S. revenue impacts, the calculations were based on the third-quarter 2021 U.S. share of the global revenues generated by the top 25 video game companies. This value may not perfectly reflect the shares that all U.S. companies had in each of the individual markets in 2019, thereby potentially over- or under-estimating the impacts of censorship on U.S. video game producers.

 

Bibliography

Freedom House. Freedom in the World, 2022. https://freedomhouse.org/report/freedom-world.

Freedom House. “Freedom in the World Research Methodology.” Accessed March 15, 2022. https://freedomhouse.org/reports/freedom-world/freedom-world-research-methodology.

Statista database. “Digital Video Games$—$Worldwide.” Accessed March 10, 2022. https://www.statista.com/outlook/amo/media/games/digital-video-games/worldwide.

Statista database. “Physically Sold Video Games$—$Worldwide.” Accessed March 15, 2022. https://www.statista.com/outlook/amo/media/games/physically-sold-video-games/worldwide?currency=USD.

World Bank. World Development Indicators. Accessed March 10, 2022. https://databank.worldbank.org/source/world-development-indicators.