HYPOTHESES

Hypothesis I: Because of non-normally distributed results in driving performance (Eriksson and Stanton 2017), clusters found within drivers’ data (Yi et al. 2019b), and high inter-driver variability (Miyajima and Takeda 2016), we expected to find distinct customisation setting clusters among participants, as seen in Chapter 14.

Hypothesis 2: Based on the results of the simulator study which showed a clear distinction in drivers’ perception of the different visual displays (Chapter 14), we expected to see clusters for the visual display settings again.

Hypothesis 3: Assuming consistent human preferences and traits (Roberts and DelVecchio 2000), we expected to observe a high similarity between participants’ simulator and on-road customisation settings despite anticipating some adaptation due to the changed environment.

RESULTS

The results are structured by theme for overview reasons. Their implications for the hypotheses will be inferred in the discussion.

Overview Customisation Settings

In total, 20 participants (83%) deviated from the default settings and all of them were unique in their combination. The mean interface density was M=68% with a standard deviation of SD= 15% and a range of R=41%. To measure the dispersion of drivers’ preferences, the coefficient of variation was calculated (CV = 22%).

Binary Customisation Settings

An overview of the selected and unselected binary settings is shown in Figure 19.7. It shows that the majority of participants (79%) decided to keep the road view option turned off on the cluster as well as the central infotainment screen, indicating low' appreciation for it. This is supported by the fact that in the case of the cluster road view setting, 5 out of 10 participants w'ho experienced it in trial 2 decided to deselect it afterwards.

When comparing the different interfaces, HUD, cluster, and ID, we can observe that in total 28 HUD settings, 7 cluster settings, and 35 ID settings were turned off. This means that participants decided to keep the cluster the densest interface.

Ordinal Customisation Settings

The descriptive statistics of mean, standard deviation, and the number of participants who turned the setting up or down, along with the number of levels, are shown in

Selected binary customisation settings in final customisation

FIGURE 19.7 Selected binary customisation settings in final customisation.

TABLE 19.2

Descriptive Statistics for Ordinal Customisation Settings

Ordinal Setting

Mean

SD

Change in Number of Participants and (Levels)

Ambient

5.83

1.76

15(20)—19(0) f 0(0)

Audio

5.29

1.10

14(9) — 19(0)1 1 (2)

Haptic

2.04

0.69

15(5)—16(0)13(4)

Questions

3.83

1.37

1 1 (2) —8(0)1 15(30)

Table 19.2. The number of questions was the ordinal setting which decreased the most w'ith regard to participants as well as levels. This means participants considered the default number of five questions to be too high.

Cluster Analysis of Customisation Settings

Clustering Participants

The first hierarchical agglomerative cluster analysis grouped participants based on their customisation profile. The separate analyses of the two distance measures Euclidean and Gower resulted in similar clusters, shown as a tanglegram in Figure 19.8, indicating that the scaling of the ordinal settings mitigated the issue of mixed categorical and numerical variables for the Euclidean distance approach.

Due to its wide application in psychology (Clatw'orthy et al. 2005; Nosofsky 1985), the Euclidean distance was used in the further analysis to determine the number of clusters with NbClust. Because of the specific characteristics of our data, partly dichotomous and mixed types, we were able to apply 20 out of the 30 indices included in the NbClust package. Based on the majority rule, the best number of clusters is tw'o when limiting the maximum plausible number of clusters to six. However, the margin was very small with three indices indicating two clusters, and two indices indicating five clusters. The subsequent analysis with ‘dynamicTreeCut’ resulted in

Clustering participants

FIGURE 19.8 Clustering participants: Tanglegram Gower and Euclidean distance.

Clustering participants

FIGURE 19.9 Clustering participants: Dendrogram using Euclidean distance.

no observable clusters. Figure 19.9 shows the corresponding dendrogram with red rectangles indicating the two possible clusters.

While analysing the interface density and other possible explanations for the two clusters, we discovered that binary interface density can explain the clusters (see Table 19.3), more specifically the selected binary options for the ID. In the smaller cluster, nearly all binary options for the ID were turned off by participants.

Clustering Binary Interfaces

The second hierarchical agglomerative cluster analysis grouped binary interface settings based on their psychological similarity perceived by participants. Out of the 20 performance indices calculated using NbClust, seven supported two clusters (see Figure 19.10), five favoured four clusters (see Figure 19.11), and five advocated six

TABLE 19.3

Interface Densities of Identified Clusters

Interface

N

Mean Interface Density

Binary Settings (%)

Ordinal Settings (%)

Overall Settings (%)

Heavy

18

79

50

73

Light

6

50

55

51

Clustering binary interfaces

FIGURE 19.10 Clustering binary interfaces: Dendrogram with two clusters.

Clustering binary interfaces

FIGURE 19.11 Clustering binary interfaces: Dendrogram with four clusters.

clusters. The dynamicTreeCut package indicated no observable clusters. According to the majority rule, the optimal number of clusters is two.

Comparing Simulator and On-Road Study

Comparing the simulator to the on-road customisation settings, we can observe that average interface density was non-normally distributed (simulator: Z)(24)=0.178, /> = 0.049; on-road: D(24) = 0.251, /><0.001) and increased by 13.12%. However, a Wilcoxon signed-rank test revealed that the difference is insignificant (Z=-1.543, /> = 0.123). When analysing the average number of questions, we can also observe an increase from 3.5 in the simulator study to 3.83 in the on-road study. This difference was not significant though (Z=-0.772, /> = 0.44). A significant increase from simulator to on-road study was found for the ambient setting (Z= -2.895, /> = 0.004), indicating that participants preferred a stronger ambient lighting during the on-road study. All this shows that there are indications for participants’ desire for more salient interfaces for on-road driving.

To compare the extent to which participants were grouped similarly in the simulator and on-road study, a tanglegram, as shown in Figure 19.12, was created. We can observe that nearly all participants, with the exception of two, were not grouped in similar clusters. This indicates that the consistency of clusters was low over the course of the two studies and that intra-driver variability was high.

 
Source
< Prev   CONTENTS   Source   Next >