Overall, from the analysis of the models is becomes evident that supervised methods perform the best when the analysis requires defining a quantitative measurement pertaining to the text and the predictions granted be the model need to be thoroughly validated using different dictionaries (Wang & Nordmark, 2015). As the current papers implies concentrating on working with the textual interpretation, it is best to concentrate on the unsupervised methods. In this light, unsupervised methods are the most useful for characterizing the texts and searching for the meaning behind the allocated results in the context of a particular environment and, hence, this is what will serve best for the analysis in the current paper.

With the mentioned disadvantages of keywords and scaling and document clustering methods in mind, the LDA approach was chosen to conduct further analysis of the data in the current research paper as it allows to manually define the topics (in comparison to STM, where the summary of the corpus is defined automatically) (Hillard, Purpura & Wilkerson, 2008).

As, the topic of the current paper involves investigating video game development industry and promotion success factors in relation to the real quality of the game, the paper explores two types of developers - companies financially supported by the publisher and independent game developers. The research focuses on the quality of the final product and on the response of the game among the consumers.

3.3 Descriptive statistics for the dataset

With the rapid development of digital distribution technologies, main promotional signals sent by the game developing companies are perceived through the lens of social media and digital platforms. Thus, the main source for data collection is the largest digital distribution platform of PC games called Steam that is used by gamers worldwide to purchase and play games and that contains an enormous database of video games and players' reviews (Steam, 2019; Wang & Nordmark, 2015).

Video games can be considered as `experience goods' which means that for this type of product a consumer cannot define the quality before that product has actually been consumed (Boatwright, Basuroy, & Kamakura, 2007). Hence, the reviews critics leave about the game allow to estimate their consumption experience whether it was a pleasurable one or disagreeable. Moreover, according to Ivory (2006), the content of the games is rightfully reflected in the reviews and this reflection is valid.

Reviews fairly often include criticism aspect which is used by the developers of video games as a source of information of the inconsistencies in the development process or to track the appealing and unsuccessful features of the game when developing new games. Such criticism may include the links to the creator's previous works, comparison to other products or the estimation of the evolution of the developer as an artist (as video games sometimes are referred to as a form of art (Schmierbach, 2009)). Good criticism identifies the main problems and even offers solutions to make the performance better, its ultimate goal implies improving the subject in question (Wyatt & Badger, 1990).

Game reviewers convey their own judgement and perception of the game and they can be objective to the product or biased depending on the type of the game and the platform where their review is published. The largest base of the online reviews is Metacritic.com which aggregates the reviews from different sources and provides scoring evaluation system. It stores not only player's reviews, but also entries from professional game critics. The good criticism, however, requires motivation, most often, the monetary one (Gifford, 2013). Hence, in the current research it is best to concentrate on the games that were reviewed by the players who most certainly played the game in order to avoid bias or gathering the reviews of people who were paid to write good or bad reviews in pursuit of the promotional goals (as bad publicity is also a mean of attracting attention). The only platform that restricts any user, besides the ones who purchased the game, to evaluate the product is digital distribution platform Steam.com. Only a player who had some experience playing the game can leave the review. Thus, in this work the reviews will be gathered from this platform as they are readily available and are targeted primarily to a general readership. The reviews there, however, include only players' opinion and no criticism from the professionals from the industry. That is where Metacritic.com can be useful for the research as it provides insights from the professional game critics. The platform provides a weighted average score form the critical publications about the game, appointing different weights to different sources. For the algorithm to give a game a score, there must be at least four reviews provided. The metascores range from 0-100 and are highlighted in different colors for easier estimation. The following figure presents the scores scale.

Figure 2. Metascore break down for the Metacritic.com*

The data collected consists of 650 video games produced by notable game development corporations and indie game developers they released between 2016 and 2018. Such a time span was taken because the trend of declining game quality had been noted by different reviewers only in the recent years (Wang & Nordmark, 2015). Overall the population consists of 3250 reviews. Apart from the reviews, the data contains other characteristics of the game, such as the original game descriptions (as they influence players' decision to purchase the game and include the signals sent to players by the companies' marketing campaigns), the overall number of reviews of the game, the percentage of the positive reviews and the metascore. The description of each variable can be observed in Appendix 3.

Appendix 4 provides general overview of the data, where it is possible to see the distribution of games developed by country and year for both types of developers. It is important to note that USA and Japan occupy the leading positions for both types of production. The main difference in countries distribution is that independent developers seem to be more diverse in geographical sense that results into the fact that this type of developers has spread around the world, whereas corporations are mostly situated in America and Europe. Talking about the year the games where produced, it seems that for both types of developers game development was quite stable over the period of analysis.

As the data contains numerical data on the number of reviews, percent of positive reviews, metascores, it is crucial to provide general statistics on the variables in question. The following tables represent the main statistical characteristics on the data.

Table 2 Corporations' games summary statistics

From the numbers on the games produced by corporations it is evident that both critics and common users on average mark the games with no such a high score of satisfaction - only around 60 percent of people on average leave positive review for such games. The highest number of reviews was left for a game called PlayerUnknown's Battlegrounds and it has only 49% of positive reviews left by players, although metascore is 86 which means that critics left generally favorable reviews. The highest percent of positive reviews pertains to the game The Witcher 3: Wild Hunt - Blood and Wine which is the extension to the original The Witcher series and which also has been rewarded the highest score from the critics. The lowest rating by the players reviews got the game called Left Alive.

The following table shows summary statistics for the games produced by indie developers.

Table 3 Independent developers` summary statistics

It is evident that players leave reviews for the indie games less frequently that for the games of big developers, probably because they are bought less than well-known titles. The numbers of positive reviews do not vary so much as the same data for corporations, and people generally leave more positive reviews for indie games. In fact, there are several titles with ultimate 100% of positive reviews, such as Corinne Cross's Dead & Breakfast, Lucah: Born of a Dream or Graze Counter, it is also important to note here that games that got 100% of positive reviews do not have a lot of reviews - fluctuating between 30 and 60, so the reason behind their high scores could be simple insufficiency of reviews. Nevertheless, even though they were critiqued by not so many players, the metascores from the professionals of the industry confirm that these games in fact of good quality. For instance, Doki Doki Literature Club which has high player reviews also scored 78 by the critics.

From the data presented in the table it can be observed that the numbers for both types of developers are quite different. The visual representation of the differences in the number of reviews, percentage of positive reviews and metascores can be spotted on the boxplots in Appendix 5. The graphs show that the difference between the types of developers exists, hence, it is important to check whether this difference is statistically significant. T-test allows to check this significance between two groups. However, before applying the test, it is crucial to check whether the data is normally distributed. Figure 3 shows that the data on most of the categories is normally distributed, although the distribution for the percentage of positive reviews for indie developers seems to be a little bit skewed, which, as was discussed earlier, can be because of the fact that most of the indie games do not get reviewed enough, but this can be corrected by looking at the graph of metascores which provide opinions from the specialists in the industry.

There are two possible tests to estimate the normality of distribution: Kolmogorov-Smirnov (K-S) test and Shapiro-Wilks (S-W) test. As the analysis is applied through the RStudio, it is better to use the Shapiro-Wilks test as the sample used for analysis is not so large in size and S-W provides more powerful estimation on smaller samples.

Figure 3. Normality of distribution check

What is more, in case of the current analysis, S-W test will yield better estimation as the test grounds on the condition that the correlations between the data in the sample and corresponding scores of normality are checked, whereas K-S test does not provide such estimations (Shapiro, Wilk & Chen, 1968). The output suggests that in all of the cases the p-value is greater than 0.05, which means that the null-hypothesis that the distribution is normal holds (Figure 4).

Figure 4. Shapiro-Wilk normality test results

From this point, seeing that the data is normally distributed, it is needed to check the hypothesis with Welch Two Sample t-test that is basically t-test but allows to compare data that has different variances that is extremely relevant judging by summary statistics on the data gathered. For all of the following tests the relevant hypothesis are: H0: there is no statistically significant difference between two categories; H1: the difference is statistically significant.

Figure 5. T-test for number of reviews

Figure 4 presents the results on the test of difference between the number of reviews for different types of developers. As p-value is less than 0,05 it leads to the fact that games developed by corporations are reviewed more often than the ones made by independent developers. Based on that it is possible to suggest that indie games are bought more rarely than games developed by corporations and, thus, get less reviews.

The next test checks the difference between the percentage of positive reviews for different types of developers.

Figure 6. T-test for the % of positive reviews

The difference if again significant, because p-value is less than 0,05, hence the assumption that games produced by independent developers bring more satisfaction to players than games produced by corporations (with the mead difference of 33.34154). However, players are not the only ones that consider independent games better: it is also relevant for professional critics (where mean difference between groups equals to 15.87384):

Figure 7. T-test for metascores

Both tests allow to judge the quality of the games produced by independent developers and corporations. As the means in both cases of rated marks are higher for indie games, it results in to the fact that these games possess better quality (based on the overall assessment by players and critics) than huge companies' games.

3.4 Text analysis with LDA

As current research exploits LDA approach, it is important to point out that it relates to topic modelling. Topic modelling treats the textual data - in this case players' reviews and game description - as the mixture of word-units that occur in the sample with a certain probability. Hence, a single word is the unit of analysis. In the core of the method stands Bayesian hierarchical mixture model that classifies the topics based on the co-occurrence patterns of words. It is crucial to note that this model does not regard grammar or word order which opens more opportunities for in-depth statistical analysis of the patterns of words that occur. The model first forms quantified text matrix where for each review the occurrence of words is distinguished. Based on that matrix the authors manually determine the number of topics that occur overall in reviews and based on that the algorithm then compose a new matrix where each review is assigned with the probability of each topic. As the result, such matrix gives a quantitative representation of probability to find each topic in each review.

Finally, the model represents hierarchy inside of the variables: on the first layer each review has a probability distribution of topics and on the second - each topic is a mixture of occurring words. From that point, it is possible to manually define the topics by the combination of words that with a higher probability (or rank) that can be met inside. After that, the review topics are interpreted and compared with the topics found in marketing campaigns of financially aided developers and independent game developers. This allow to contrast the responses of players and inspect the quality of the games.

LDA is implemented by using package `topic-models' in R software with Gibbs sampling which is more commonly used for small data sets. During algorithm implementation each topic gets a special label and a definition that best represents its content. The label is determined in such a robust way that would serve as a good representation of the data. The authors independently choose lists of words that need to be included into the topic and then calculate Cohen's kappa that helps to make the decisions of word inclusion into the topic relevant and fair. Each topic is defined and supported by the examples from the original reviews.

As the result, the aforementioned methods allow to identify which digital promotion success factors are the most influential and efficient in terms of reaching the target audiences in the game development industry that are utilized by different types of developers. Another important issue that is possible to consider with employed methodology is the estimation of the game quality that grants an opportunity to assess the scale on which the developers are able to conceal the real quality with the promotion factors they exploit.

The process of reviews and description selection can be observed in the Appendix 6. The important thing here to mention is that the first review in the table is the most helpful one and the following ones come in the order of the decrease of said helpfulness. This crucial for the research as in that way it will be possible to distinguish which characteristics are more important for the players.

3.5 Game descriptions analysis

The main appeal for the customers comes from the producer, hence, it is crucial to analyze how developers attract players' attention. Work with textual data starts with stemming and cleaning the text from any uninteresting for the research words. The wordcloud below allows to see which words are most frequently occur in the descriptions (presented on the figure 7 are the words that occur at least 55 times in the entire dataset). It is important to note that after stemming the words only have their base as all the suffixes were removed. Nevertheless, it does not hinder the process of analysis, and eliminates the repetitions inside the corpus, which is helpful for further topic allocation.

Figure 8. Wordcloud for game descriptions (1 - corporations, 2 - indie)

When creating a document-term matrix, the program allocated 325 documents and 6998 terms overall for corporations and 6840 terms for independent developers. By applying LDA it was found out that the optimal number of topics that illustrate the descriptions of games is 6 for corporations and 4 for independent developers (the larger number causes overlaps in data and makes it harder for interpretation). The following graphs depict the top words that most frequently occur in the topics. Appendix 7 presents a slightly different set of words that were elicited when comparing topics between each other. It is evident on the Figures 8 that there are repetitions of most common words in each topic such as mode, battle, feature and others. It is possible, however, to evaluate the differences in the probabilities to meet each word in the topic which is calculated with the following formula: , where represents the probability of the word to be presented inside of the topic, and the indexes represent the number of the topic. The output when applying the analysis look like that:

Figure 9. Word-distribution per topic (example)

Figure 10. Word distribution in the topics (corporations)

By comparing all 6 topics between each other (and for indie games 4 topic), the authors came up with sets of words that are more representative and common in the topics based on the difference in probabilities (Appendix 7).

Figure 11. Word distribution in the topics (independent developers)

The output of the LDA generated probabilities of each document to be assigned to each topic. The table below presents an example of probabilities of topic distributions for the first 20 documents, where the highest probability is marked in red (each document in this case is a separate game).

Figure 12. Probability of topic assignment to documents

As the result, it is possible to count how many times each topic can be met in the corpus, making it possible to understand which topics are used more frequently.

Figure 13. Topic distribution throughout descriptions (corporations)

The distribution of topics in the corpus will be examined later in results section, where each topic will be defined and, thus, such graph will have more sense; the presented graph in Figure 11 is an example of the output that can be feasible for further interpretation.

3.6 Reviews analysis

Each game has 5 reviews in the data table. The first review is the most helpful one as estimated by the users, and the next ones are in descending order of helpfulness. The methodology of analysis stays the same as for the descriptions of the games, thus, some of the elements of analysis will be slighted, presenting just the viable for further analysis outputs. When creating a document-term matrix for each review in order of helpfulness for both type of developers, the numbers of terms occurring where the following:

Table 4 Number of terms in the reviews

Judging by the number of terms in each review, the most helpful reviews in most of the cases tend be longer and use more terms then less helpful ones.

After applying LDA with optimal 4 topics and calculating the difference in probabilities of word occurrence in the topics, the algorithm yielded the following words to be the most common for each review (Appendix 8 provides graphical representation of most common words and word-topic distribution for corporations, and Appendix 9 - for independent developers).

Table 5 Word occurrence in topics of reviews

A more elaborate analysis of each topic as well as giving definitions and finding connections to the promotion factors will be discussed further in the results section.

3.7 Assumptions check for regression analysis

Current research implies identifying the effect each promotion factor has on shaping the decision of the customer to buy a game and the perception of the quality of the game. Because of the fact that independent variables (topics of descriptions and reviews) are categorical, the best type of the analysis to test the influence would be multinomial logistic regression. This type of analysis is used in cases when there is a need to predict the placement of a categorical value in (or category membership probability on) a dependent variable based on several independent variables. The technique implies the same approach as binary logistic regression with the exception of an opportunity to test the dependencies on more than two independent variables. It uses the maximum of the likelihood evaluation to estimate the probability of occurrence of a categorical membership.

Even though multinomial regressions analysis does not necessitate checking as many assumptions as simple linear regressions, in order to apply this type of regression analysis, several conjectures must be checked. First, independent variables should be evaluated on the matter of multicollinearity. It means that there should not be perfect correlation between independent variables. Unfortunately, when dealing with multinomial logistic regression, R software package does not provide a specific function for this type of regression to extract variance inflation factor (VIF), so it is feasible to extract values from the model and then manually calculate VIF using the formula:

As the result, the following estimations were calculated:

Table 6 Multicollinearity detection

From the table it is evident that none of the variables possess VIF higher than 10, which signifies that there is no multicollinearity between the independent variables, hence, the regression model can be created.

One of the reasons why multinomial logistic regression is used to create a model for categorical dependent variable is the fact that it does not require the data to be strictly normally distributed and does not need testing for linearity or homoscedasticity. However, the most important assumption for this type of regression that needs to be tested is the assumption of Independent of Irrelevant Alternatives (IIA), estimating the independence of choice inside the dependent variable which means that choosing the membership in one of the categories will not impact that same choice in the other categories. The independence of choice is tested through the Hausman-McFadden test. The simplest description of the test would be the following: the test compares two multinomial logistic regression models, where two different choices inside a category are estimated on the nature of their probability to be chosen if all other parameters are equal (More on the Hausman-McFadden test can be found in Vijverberg, 2011). The output of the test is presented below.

Figure 14. The results of the Hausman-McFadden test for both datasets

(fdt1 data - corporations, fdt2 - independent developers)

The test implies two hypotheses, where the null hypothesis states that the IIA holds and alternative one where IIA is rejected. The result is evident from p-value, which is greater than 0.05, meaning that the null hypothesis stands and the choice between the alternative inside a dependent variable is actually independent. Quite low Chi-squared in this case also can be interpreted as non-perfect separation between different predictors, which is also important for this type of regression analysis as perfect separations between them would have yielded too unrealistic coefficients that exaggerate the real relationships between variables.

When all assumptions are checked it is possible to build a regression model which output created the following estimates for the dependency between description topics and number of reviews (that in our case can be estimated proportional to the number of purchases):

Figure 15. Multinomial logistic regression coefficients for description topics

From the coefficients it is evident that none of them cause drastic change in the dependent variable when others are held constant, however, it is crucial to note that change in several topics is more significant than in the others (i.e. with the “content” and “reach” topics). The model is relevant for interpretation as the McFadden which in the multinomial logistics regression is treated in the same way as the in simple linear regression. The fit of the model is also significant with a small p-value meaning that the model predicts better estimates compared to the null model on the 0.001 significance level.

Based on the coefficients in the model, the equation would be the following: . The coefficients show the degree of change in the number of reviews for the change in each variable, holding others constant.

The same modeling approach was conducted for the review topics in comparison to metascores and per cent of positive reviews, the coefficients are presented in the picture below.

Figure 16. Multinomial logistic regression coefficients for review topics

In the case of reviews all of the topics are significant. For example, the change in the second and fourth topics can impact the percentage of positive reviews and the change in all of the topics can influence the metascores. In case of the model for the reviews the McFadden and low p-value that means that the model has quite good fit and estimates the variables well on the 0.001 level of significance. With the same logic applied the coefficients in the model would be the following:

. Each of the factors identified for topics of descriptions influence as well as for the reviews will be explained further in the Results section of this paper.

Coefficients for the same type of regression were also calculated for indie games data (Appendix 10).

4. Results & Discussion

RQ1: What are the promotion factors that impact customers' decision to purchase a video game which remains a digital product?

As a result, for the corpus of text of game descriptions that pertains to corporations there have been found six different topics that are used more frequently. Concerning the text documents of reviews for each type of the developers, the algorithm allocated four topics mentioned by players. For each of the sets of words that the allocated topics include, the researchers strived to find the best factor attained in the literature that would fully describe what the topic is about.

For the mass-marketed game developers, it turned out that the game descriptions tend to focus more on the narrative, the setting of the game or the characters available to play. These developers relate to the atmosphere that the game is supposed to create for the players. The words such as “character”, “hero”, “customization” or “interact” refer to carrying a more personal approach to the player, appealing to the world the game has to offer and what a user can experience while playing. This topic includes the general characteristics of what helps to build the imaginary world that the players anticipate and how far can the players go in customizing this world for their personal needs. Therefore, the topic is named after the factor “Personalization”.

The other detected in the game descriptions topic is related to the first one but is more attached to the storyline. Adams (2014) mentioned an importance of understanding the preferences of the player and the need to create the environment that is the most suitable for a certain category of players. For example, for female players there is a special imaginary world with a sensitive story line and an opportunity to collect things during the game. For male players, their world needs to be more logical and systematical that provoke critical thinking and analysis. In the sets of words for this topic the terms with the highest probability were precise and depicted the main surroundings of the game. Words like “gameplay”, “city” and “location” are crucial in understanding the content of the game and the things that are useful in creating the foundation of the imaginary setting. This topic is assigned to the “Content” factor stressing the importance on how the message is delivered to the target audience and how the game can be perceived by the players, provoking sensual type of reactions, rather than experiential ones like in the first topic.

The third topic includes the words that are used to describe the game with definite terms such as “adventure”, “horror” or “action”. These words are just like the filters that a person sets when deciding on the film to watch or a song to listen, and often display the genre of the game. The factor “Credibility” defines this topic best as it eliminates the vagueness and sets proper tags on each of the game. This feature is vital in digital industry as the intangibility of the product must be defined by accurate terms in order to avoid the confusion of the customer and provide the most comprehensive information possible to ensure the best customer experience. When exploiting this factor companies establish a baseline that outlines the expectations of the player form the game.

Topic 4 is best described by the “Reach” factor that emphasizes the importance of mentioning the features that best reach the target audience. This can be words such as “fast-paced”, “co-op” or “mature” that help to connect to the audience by offering some insight of what they may get from playing the game. The words are helpful in attracting most of the target audience as they remain defying in depicting the interests of the players. By using such words in the descriptions, developers determine the players who will enjoy playing this game, they define the community of other players who potentially could have the same interests as the customer. This feeling of belonging to the community with the same interests may often be the deciding factor for making a purchase. It also allows to establish contact between the expectations of the players and what the game has to offer.

The next topic is characterized as the “Incentive” factor due to the words included in the topic that relate to the features that come as a bonus in the game. Words “upgrade”, “rank” and “reward” are traced in game descriptions as they possess the meaning of supplement designed to enhance the value of the game for the player. The incentives have always been a significant driver in purchasing decisions of customers whether it is discounts or warranty. This topic allows to create a sense of achievement for the players and justifies the reasons to buy the game. In the digital industry the incentive is even more appreciated because of the vulnerability and intangibility of the offered product or service as players may feel that they get some recognition from what they spend their time on.

“System”, “online” and “expansion” are the most popular words that create the last topic that is referred to the “Complementarity” factor. This information is commonly put in the description in an attempt to add to the environment that the game possesses more value than the user already knows. Complementarity stimulates a customer to think beyond the limit and encourage them to expect more from a game. In many ways this factor creates the vision of the game in the players mind by establishing the features that the player additionally gets form purchasing the game, showing how the player can adapt current personal preferences and skills to the in-game environment.

Table 7 Topic definitions for corporations (descriptions)

Comparing the factors established in the theoretical foundation to the specifics of the video game market allowed to distinguish related to the market conditions definitions for each of the factors. The Table 7 presents the relevant for each topic definitions with the examples most common words and of the typical descriptions that occur in them.

By attributing the allocated definitions to the topic in the dataset, it is possible to get more representative results. The following figure shows which topics are more frequently used by corporations. It is evident that corporations more commonly employ Reach factor in the promotion of their games by appealing to customers' possible personal interests.

Figure 17. The distribution of topics for corporations description corpus

One of the research questions implied identifying the most influential topic that makes players choose games. The degree of influence can be recognized by using the measure of change in R-Squared in the regression model when adding more predictors to the estimation. The higher the percentage of change is, the more influential is the predictor. The following table shows the estimates that illustrate the change in R-Squared for the multinomial logistic regression that was built to test the relationships between the topics used in the description and the number of reviews (that in the case of the current research were treated proportional to the number of purchases). quality value strategy marketing

Table 8 Increase in R-Squared (%) for the regression model (corporations, descriptions)

The highest increase occurs with the Reach and Content topics meaning that they have the most significant influence on the number of reviews and, hence, on the number of purchases. It is important to note that in the regression model itself the topics 4 (Reach) and 6 (Content) had the lowest p-value and were also considered most influential, which gives additional prove of their importance in the model. When corporations appeal to players through suggestions of the possible interesting one-of-a-kind feature that connect to their interest, it results in the increase of the reviews (purchases) of the game. However, it does not mean that players necessarily enjoy the game they purchased. It was found out that the higher the number of the reviews for a specific influential topic is, the lower is the percentage of positive reviews. This can be observed on the graph below. It is clear that the factor (or topic) with the highest influence (“Reach”) simultaneously experiences an inverse proportionality when looking at the change in the number of reviews and the percentage of positive ones. At the same time with the decrease of influence of the topics comes a more independent distribution of values inside the factor.

As for the independent developers the situation changes in terms of the word inclusion into the factor definitions. The first topic concentrates around the atmosphere of the game, the music and art style, even though it still pertains to the “Reach” factor as it establishes a connection with the player.

Figure 18. A change in the review number and per cent of positive reviews between factors for corporations

The second detected topic remains close to the “Content” by an attempt to create an imaginary world that players are seeking to obtain. Words “world”, “story” and “level” serve to establish the imaginary world that players are looking for. Third topic is most associated with the “Credibility” factor that consists of words such as “puzzle”, “mystery” and “explore”. As mentioned before, these words help to defy the terms of the game and eliminate the ambiguity of the description. Factor “Incentive” is represented by the fourth topic tracked among the game descriptions of the independent indie developers with words such as “campaign”, “mode” and “network”. The Table 9 presents more illustrative definitions and examples that best describe the topics.

Table 9 Topic definitions for indie games (descriptions)

From the table it becomes clear that independent developers exploit the same factors when trying to appeal to the customers, albeit fewer in number. The factors also hold different words inside and convey a slightly different meaning. For example, the factor “Reach” exploits the appeal to the style of the game, its visual representation, not like the targeted approach on consumers' interests that corporations use. The frequency of topic usage also differs which is shown on the graph below.

Figure 19. Topic distribution for independent developers

The difference in the topic distribution between two types of developers is quite apparent. The factor least exploited by corporations is the most frequent in descriptions of indie games (“Credibility”), and two most utilized by corporations' factors are not in favor with independent developers (“Content” and “Reach”). However, before talking about the effects of such difference, it is crucial to measure the importance of each factor for the percentage of positive reviews and metascore values. This is again evident from the change in when the variables are added to the model.

The first dissimilarity that is important to note is that each factor has higher influence on the variables than in the corporations' dataset, as there are fewer topic identified used by independent developers. The second crucial fact is the existence of similarity between the importance of factors “Content” and “Reach” that possess significant influence in both datasets.

Table 10 Increase in R-Squared (%) for the regression model (indie games, descriptions)

However, the nature of such influence is quite hard to determine due to the skewness of the data on indie games towards more favorable reviews and generally low numbers of reviews as the customers tend to review indie games less than games of corporations. Thus, it is only safe to state that the “Content” factor is the most important in the model and allows to describe the main part of variance when independent variables are concerned. Nevertheless, even though it is significantly more influential on the number of purchases, it is not exploited frequently enough by the independent developers to attract more players.

When moving on to the reviews' analysis for both types of developers, it is first vital to establish which review is the most important in the model. The initial assumption before data collection implied that the first most helpful review on the website will hold the majority of the necessary information for the analysis. However, without checking this assumption statistically, there is no way to eliminate completely the reviews that had lower helpfulness estimation by the users. The following graphs, nevertheless, confirmed the assumption that the first most helpful review has the highest influence in the model. The following Figure contains graphical illustration of the mean decrease in accuracy (MSE) which shows how the elimination of one variable will worsen the model by increasing MSE. And the first review seems to be the most important one as its elimination will drastically influence the accuracy in the model.

The other types of reviews (apart fromt the first one) have played an important role in LDA model creation and defining the topics occuring in the reviews. For the analysis, however, it is better to take only the first most helpful review for the games as it is the most important in the model and in the majority of the cases contains the most relevant information.

From this point, the main difference between the content of the descriptions and reviews must be established. On the one hand, descriptions contain the information the companies want users to see (marketing signals), on the other hand, reviews are user resposes to the in-game experience, hence, they are more focused on the usability principles that were identified in the theoretical background. Players more frequently talk about what they liked or did not, about their emotions and their general feel of the game. From their responses it is possible to make some conclusions on the relationship between companies appeal tactics and users reaction.

Figure .20 Testing the importance of the type of the review (on the left - corprations dataset, on the right - independent developers)

The authors of this paper strive to allocate the precise factors from each review, thus, by contemplating the literature, it was decided on the best definition of the factors that appear in the reviews (Table 11). The fist topic refers to “Customization” that concentrates on adjusting the general seetings of the game such as changing the difficulty, screen resolution or combination of keyboard shortcuts. Customization, as it has been already discussed, allows a player to reap the best from the game by modifying it to the personal needs and interests. The second topic can be best described as “Mechanics” due to the fact that this factor unites the features that help players to navigate in the game such as tutorials, instructions and manuals. Next topic refers to the responses the player gets through the in-game navigation. This is explained by the responses that each move of the player receives through either the support of developers or the output that the game provides. Hence, this topic is best described as “Responsiveness”. Last topic “Gameplay” received its name by the fact that the in-game coordination has to be justified while each move is controled. Words that best descibe the factor remain “mission”, “health” and “map”.