6. OBJETIVOS Y MEDIDAS PARA LOS ELEMENTOS CLAVE U OBJETO DE
6.2. COMUNIDADES DE MUSGOS Y HELECHOS DE LAS REGATAS
7.4.2 Negative effects
It is necessary to verify that no participant suffered physically from the experi-ment. The initial analysis of the results considers the NEGATIVE group of questions, measuring negative effects induced by the system, such as nausea, eye strain, or headache.
We carried out this analysis with a bivariate boxplot, or bagplot [Rousseeuw et al., 1999]. The bagplot (see Figure7.7) is a bivariate version of the traditional boxplot. Its main components are a bag, which comprises 50% of the data points, and a fence, which separates inliers from outliers. The region outside the bag but inside the fence is called the loop. It is usual to present a mark at the median in a boxplot. The equivalent of this mark in the bagplot is represented by the white region in our plots. A bagplot of the NEGATIVE score versus the TEMPLE score (Figure 7.7), indicated that participant 23 was an outlier, reporting feeling much worse than the other participants. This participant was therefore discarded from the study. All others obtained a NEGATIVE score less than 2.17 (minimum possible value = 1), which can be considered as having experienced little or no negative effects during the experiment. Also, the presence score seems relatively independent of the experienced negative effects. Suffering from (little) negative effects does not seem to reduce the feeling of presence. Conversely, a low presence score is not necessarily associated to the experience of negative effects.
7.4.3 Impact of sound rendering condition on presence
The mean scores in each presence category of interest, obtained for each SOUND CONDITION, are given in Table7.3. Following an ANOVA analysis, all scores failed to achieve the 0.05 significance level. Hence, no significant effect was observed for the sound condition over all subjects.
7.4.4 A model for the perceived presence
The probability density function of each of SPATIAL, PERCEPTUAL REALISM, and TEMPLE scores (see Figure 7.8 for the probability density function of TEMPLE) suggest that there are in fact two groups with a normal distribution centered around different means. The data can thus be modeled as a special form of a Gaussian mixture model (GMM). The package Mclust [Fraley and Raftery,2003]
allows one to find coefficients of a Gaussian mixture from the data by selecting the optimal model according to the Bayesian information criterion (BIC) applied to
2 2.5 3 3.5 4 4.5 1
2 3 4
5 23
TEMPLE
NEGATIVE
STEREO HYBRID
WFS
Figure 7.7: Outliers in the NEGATIVE vs. TEMPLE relationship. The bivariate boxplot (or bagplot) generalizes the univariate boxplot.
The graph includes two colored regions, called the bag (yellow) and the loop (orange). The interior of the bag includes 50% of the data, and the loop includes all the points outside the bag but inside the fence, thereby outlining potential outliers [Rousseeuw et al., 1999].
Table 7.3: Presence questionnaire mean scores for each category by SOUND CONDITION (and standard deviations in parentheses).
Note: the degrees of freedom associated to the ANOVAs are vari-able.
STEREO HYBRID WFS F p-value
SPATIAL 2.90 2.83 2.47 0.50 0.900
(1.2) (1.12) (0.73)
PERCEPTUAL 2.71 2.71 2.66 0.01 0.991 REALISM (1.1) (0.87) (0.74)
TEMPLE 3.34 3.23 2.94 0.79 0.487
(0.79) (0.87) (0.54)
SWEDISH 5.15 4.97 4.57 0.89 0.773
(0.78) (1.34) (0.83)
7.4. RESULTS FROM POST-SESSION QUESTIONNAIRES
1 2 3 4 5 6
0 0.1 0.2 0.3 0.4
TEMPLE
Probabilitydensity[/]
Figure 7.8: The probability density function of TEMPLE scores.
an expectation-maximization (EM) algorithm initialized by hierarchical clustering for parameterized Gaussian mixture models.
The algorithm was run on the data defining the TEMPLE score and the resulting optimal model contains four Gaussian components. The probability that a given participant is not correctly classified using this model ranged from 0 to 5.8 × 10−3 (mean = 1.2 × 10−3). This demonstrates the good quality of the classification.
The four groups, referred to by the factor CLASS, are given by the algorithm in descending order of the number of participants they contain: 15, 10, 5, and 2. The mean presence scores for each CLASS category are given in Table 7.4.
Table 7.4: Presence questionnaire mean scores for each category by CLASS (and standard deviations in parentheses). Note: the degrees of freedom associated to the ANOVAs are variable.
1 2 3 4 F p-value
SPATIAL 2.25 3.81 1.71 3.64 19.49 < 10−5 (0.57) (0.78) (0.29) (0.51)
PERCEPTUAL 2.08 3.66 2.40 3.20 17.07 < 10−5 REALISM (0.46) (0.49) (0.91) (0.57)
TEMPLE 2.82 4.05 2.28 3.74 39.88 < 10−5 (0.39) (0.34) (0.22) (0.03)
SWEDISH 4.78 5.73 4.00 4.00 6.24 0.107 (0.97) (0.73) (0.53) (0.00)
Figure7.9shows an analysis of the TEMPLE score depending on the classification CLASS. Groups 1 and 3 tend to have a lower presence score than the groups 2 and 4.
An analysis of variance was carried out on the TEMPLE score with the fixed
1 2 3 4 2
3 4
MP
HP
LP
HP
CLASS
TEMPLE
Figure 7.9: Boxplots of the TEMPLE score vs. the GMM classifi-cation CLASS. The widths of the boxplots are proportional to the sample size: 15, 10, 5, and 2 from left to right. The labels LP, MP, and HP are introduced in the text.
factor CLASS (four levels). The factor showed a significant effect (F2.72,27.88= 39.88, p < 10−5). Subsequent post-hoc comparisons (Tukey’s HSD test), with an α level of 0.05, showed that groups 2 and 4 do not differ significantly (they form a homogeneous subset), while groups 1 and 3 are significantly different and both differ from the aforementioned set of groups 2 and 4. In the following sections, group 3 will be referred to as LP (low presence, 5 subjects), group 1 as MP (medium presence, 15 subjects), and the combination of groups 2 and 4 as HP (high presence, 12 subjects).
7.4.5 Further analysis in each group
An analysis of variance was carried out on the TEMPLE score with the fixed factor SOUND CONDITION (three levels) for each presence group defined in the previous section. The factor showed a significant effect on group HP (F1.95,8.99 = 6.85, p = 0.016). However, SOUND CONDITION failed to reach statistical significance for group MP (F2.00,11.90 = 0.11, p = 0.896) and for group LP (F1.00,3.00 = 0.69, p= 0.468).
Subsequent post-hoc comparisons (Tukey’s HSD test, α = 0.05) on the group HP showed that conditions STEREO (4 participants) and HYBRID (5 participants) do not differ significantly (they form a homogeneous subset), while condition WFS (3 participants) differs significantly from each of the conditions STEREO and HYBRID.
7.4. RESULTS FROM POST-SESSION QUESTIONNAIRES
Figure 7.10 shows an analysis of the TEMPLE score for each sound condi-tion in group HP. Presence scores are (statistically) lower in the WFS group than in the two other groups. A similar analysis was performed on the SPATIAL, PERCEPTUAL REALISM, and SWEDISH scores. These failed to achieve the 0.05 signif-icance level. This result, combined with the result on the TEMPLE score, indicates that the impact of sound reproduction is spread across different components of presence rather than confined to the components “Spatial presence” and “Percep-tual realism”.
STEREO HYBRID WFS
3.5 4 4.5
SOUND CONDITION
TEMPLE
Figure 7.10: Boxplots of TEMPLE score vs. SOUND CONDITION for participants in group HP. The widths of the boxplots are propor-tional to the sample size: 4, 5, and 3 from left to right.
7.4.6 Discussion
SOUND CONDITION as an independent variable fails to predict the obtained presence score for all participants. Rather, the participants are classified in three groups according to their presence score. The first group has a low presence score (LP), the second has a somewhat higher presence score but also a higher variability (MP), and the third has a high presence score (HP).
SOUND CONDITION has a statistically significant impact for the group HP. In this group, the HYBRID soundtrack is not statistically different from the original STEREO version, which means that the slight difference in content between the two soundtracks did not impact on the reported sense of presence.
When comparing the results for the HYBRID and the WFS soundtracks, one can see that there is a statistical difference in reported sense of presence which is to
the advantage of HYBRID. In this condition, sound objects were limited to the space between the virtual speakers, and since the participants were at the sweet-spot, objects between the two virtual sources were fairly well localized in azimuth.
Therefore, one could hypothesize that presence is lessened when the auditory ob-jects extend beyond the screen boundaries. Indeed, the virtual loudspeakers in the HYBRID condition were located near the screen borders, and Figure 7.3 shows the spread of the mean ITDs increasing with SOUND CONDITION, from STEREO to WFS.
Further studies with different source material would be required to substantiate this hypothesis.