In ACE 2013 Netherlands, pp. 572-575. Copyright Springer 2013. This is author’s version. The definitive version DOI: 10.1007/978-3-319-03161-3_56.
This pilot study looks at how the formal features of character-driven games can be used to explain player-character engagement. Questionnaire data (N=206), formal game features (in 11 games), and ordinal regression were used in the analysis. The results show that interactive dialogue and cut-scenes showing the romances between the player-character and another character relates to higher character engagement scores, while romance modeling and friendship modeling relate to lower character engagement scores.
Keywords: ordinal regression, player-character, engagement, identification
This is a pilot study that aims to isolate formal game features that contribute toward player-character engagement. The focus of this study is on single-player character-driven games. Lankoski’s theory  is used as the basis of the study. In order to evaluate connection between games and engagement, the formal game features are classified and then evaluated by using ordinal regression and self-reported engagement scores.
Lankoski  argues that player-character engagement relates to goal-driven engagement and empathic engagement. Here I focus only on empathic engagement. Lankoski suggests that character engagement depends on how the character is presented; the kind of access that the player has to the characters actions, thoughts, and emotions; and how the player evaluates the character in terms of morals and aesthetics. 
The following formal features where selected (c.f., ): dialogue vs interactive dialogue, moral choices (no–yes), supporting different play styles (no–yes), cut-scenes with romantically content – romance in cut-scenes, appearance customization (no; possibility to change some aspects of the character, e.g., cloths, hair style); customizing sex and appearance, character development (no, scripted character development, player-guided character development), player character dialogue is voice acted (no–yes), romance/friendship modeling (some modeling, complex modeling as in Dragon Age: Origins) and moral choices (no–yes).
Most of these features relate to the kind of access the player has to the character except for appearance customization, play styles, and character development that relate more to the aesthetical and moral evaluation of the character (c.f., ). It is possible that this list does not include all the features that are relevant to player-character engagement and, hence, the analysis below reveals only the information on these features in relation to the feature set.
Player-character engagement is measured by using a single 5-point Likert scale question: I identified with my player-character (1: totally disagree, 5: totally agree). Here the concept of identification is used in the questionnaire as I assumed that the term is more familiar to players than that of engagement.
A non-proportional quota sampling method was used. The target was to set to 20 answers for each game. In addition the target for female respondents was set to 25%.
Respondents were gathered by advertising the study in Facebook, Twitter, Google+, and pelilauta.fi as well as in two forums dedicated to Assassin Creed: Brotherhood and Uncharted 2: Among Thieves whenever there was a need to gather more answers about those games. In addition, to get more answers from females, the study was advertised in a Finnish girl gamer forum.
The data was gathered by using two questionnaires. One-half of the games was in the first questionnaire and the other half in the second questionnaire.
I selected popular, rather new games for the study to ensure that the sufficient amount of answers were obtained. The games are listed in figure 1. The games were played and judged base on whether they have the aforementioned formal features.
The data were analyzed by using ordinal regression and mixed effect models. R  and cumulative link mixed models (clmm) from the ordinal package  were used. The Gauss-Hermite quadrature approximation with ten points was used in the analysis. No structure (except that 1<2<…<5) were assumed in the ordinal scale. Subjects were modeled as a random effect.
Stepwise model selection by using Aikake’s Information Criterion (AIC) was used. The model with the lowest AIC was selected, However, simpler model was preferred when the models were not statistically different (by using the likelihood ratio test).
The total number of subjects in the study is 206. However, there can be overlap, as the data was collected through two anonymous questionnaires. The mean age of the respondents of the questionnaires is 28.49 (min = 13.00, max = 51.00). 68.9 % of the respondents are male and 31.1 % female. 66% of the respondents are from Finland, 15 % from Sweden, 4 % from the US, and the rest from different countries. 21.4 % of the respondents had obtained high school eduction, 5.5 % vocational, 23.3 % college, 24.0 % bachelor, 18.5 % masters, 6.0 % doctoral, and 1.1 % other education. Figure 1 shows the various features of the games.
The predictor variables of the optimal model are interactive dialogue, romance modeling, friendship modeling, and romance in cut-scenes and subject is a random effect.
The optimal model is significantly better (p<.001) than the null model containing only a random effect subject. This means that the data is explained better with the optimal model than by assuming that the player preferences would explain the data.
The random effect subject has a variance of 2.219 and a standard deviation of 1.489. The strongest positive effect is by interactive dialogue (2.1274, ). This means that the games having of interactive dialogue are estimated to have higher player engagement with their PCs. Showing romantic engagement in the cut-scenes emph(romance in cut-scenes: 0.6932, ) relates to higher scores in the terms of identification. The effect is considerably smaller than the effect of interactive dialogue. The games having romance (yes: -1.2650, ) or friendship modeling (some: -1.4392, ) have lower player-character engagement than the games without those features. As an effect, limited romance modeling (romance some: 0.0843, ) does not differ from no romance modeling in a significant fashion (because the 95 % confidence intervals cross zero). The effect of limited friendship modeling (friendship some) could not be estimated because the model design is column rank deficient if the level is not dropped.
As nonprobability sampling was used, the results are not directly generalizable by using probability theory. However, the optimal model indicates that the different backgrounds has no significant role in the results. However, the study includes only 11 games. The implementation of formal features can have an impact on the results. Finally, if some formal feature that is relevant to player-character engagement is not included in the list of features used in the model selection, the relevance of the feature cannot be evaluated (e.g., the set of games does not contain pure 1st person games such as Half-Life). To conclude, I believe that the results are somewhat generalizable to the population outside the sample, but it is likely that the results are tied to the implementation of the formal features within the games in this study.
The results indicate that interactive dialogue and showing romantic episodes in the cut-scenes relate to a higher player-character engagement. Interactive dialogue in all the games in this study contains dialogue options that can be used to present different types of personalities (e.g., in Dragon Age 2 one can select from diplomatic/helpful, humorous/charming, and aggressive/direct lines). This allows the players to modulate the character towards their preferences. This can contribute towards positive evaluation of the character and a higher engagement (c.f., Lankoski). Surprisingly, romance and friendship modeling relate to lower player-character engagement scores. This might relate to the quality of the modeling.
In this study I am able to connect formal game features to the identification self-evaluation scores by using ordinal regression. However, using only one question in the questionnaire is being simplistic. Using more nuanced measurements for evaluating player-character engagement remains to be done in a future work.
- Lankoski, P.: Player character engagement in computer games. Games and Culture 6 (2011) 291–311
- R Development Core Team: R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria. (2012) ISBN 3-900051-07-0.
- Christensen, R.H.B.: Ordinal—regression models for ordinal data (2122) R package version 2012.01-19 http://www.cran.r-project.org/package=ordinal/.