3.1. Objectives and Hypotheses
Amid fragmented literature and the growing need for increased road safety and fewer road fatalities, the purpose of the present study is to test and validate an intentional model, meant to explain the extent to which consumers intend to use the eCall device, or, in other words, the intention to purchase/use the eCall device for aftermarket/old cars. The purpose of the research is justified by the contribution made to the specialized literature but also by the practical implications at managerial level. Decreasing the number of fatal road accidents is a strategic objective of both developed and developing countries. The manifestation of this behavior can be investigated through the study of the stated intention.
Modelling the consumers’ intention to continue using a technology, to buy the eCall IVS for personal use, etc., involves several central objectives as well as secondary objectives. Taking into account the purpose pursued, respectively obtaining relevant data for substantiating the analysis of benefits and costs generated by the after-market installation of the eCall IVS device at European level, in this study, we focused only on central objectives. Further analysis of consumer behavior at the level of secondary objectives will be the subject of future studies.
Purchasing an eCall device is a behavior that is not entirely volitional; it depends not only on motivational factors. There are situations in which some individuals cannot buy and use the device, because they do not know what it is used for, or they cannot afford to buy it or install it on the vehicle. In addition, not all consumers have the skills and knowledge needed to use it. These inhibitory factors are brought together under the concept of perceived behavioral control within the TPB—Theory of Planned Behavior [
41], which includes both the facilitating factors and the perceived personal effectiveness of the consumer. But the mere adoption of an innovation does not implicitly imply continuity [
18]. The consumer can at any time give up the purchase of the eCall device, resulting in the discontinuation of the decision to use. Thus, a second direction of study is observed in the modelling of usage behavior, namely, the studies focused on the post-adoption phenomenon.
In defining the study hypotheses, we will start from the idea that “the consumers’ intention to use the eCall device”, or “the consumers’ intention to purchase the eCall device”, is a manifestation of the continuity of the decision to use the device. Within this model, the decision of individuals to purchase the eCall device will be given by the stated behavioral intention. Intentional research allows simultaneous investigation of the influences of several explanatory variables on the declared behavioral intention. Behavioral intentions have often been treated as the conative component of attitude, which is closely linked to the affective component, which has led to a close relationship between the individual’s attitude and behavioral intention [
36]. As a general rule, the more favorable the attitude of the individual towards a certain behavior, the stronger the intention of the individual to perform the actual behavior [
36,
42,
43].
One of the most difficult tasks of the researcher is precisely to identify the fundamental beliefs of the individual, able to determine much of his attitude towards an object or towards a behavior [
36]. In order to choose those fundamental beliefs in determining the consumers’ attitude towards the adoption of the eCall device, we used a TAM that identifies two fundamental beliefs of individuals: perceived usefulness and ease of use [
9]. Perceived usefulness, a concept of TAM, represents the user’s conviction that the use of an information technology entails several benefits [
9]. However, consumers’ beliefs about the usefulness of the eCall device are not only a determining factor in adoption [
44,
45], but also a factor in the continuity of this behavior.
The second fundamental belief of individuals postulated by TAM as having a determining effect on the attitude of individuals towards a computer system is the ease of use. This concept represents the extent to which users of a computer system perceive its use as effortless [
9]. According to TAM, the individual’s conviction about the ease of use of the information system determines his attitude towards the use [
46]. In other words, the easier it is to use a device, the more consumers have a more favorable attitude towards this device. Even if they perceive the usefulness of new technology, some users may refuse to adopt it precisely because it is difficult to use, hence the existence of a direct and positive link between ease of use and perceived usefulness [
46]. The purpose of the study is to test and validate an intentional model meant to explain the extent to which consumers intend to purchase the eCall device.
A list of the established central objectives and hypotheses is presented in
Table 2.
The conceptual model of the study involves defining and determining the measurements of 4 latent variables: the consumers’ intention to buy the eCall device, the attitude towards the use of the device, the perceived usefulness, and the ease of use. The first step in determining the measurements of the latent variable “consumers’ intention to continue use/buy eCall” is to define it at both general and specific levels.
A singular behavior involves an action, directed at a target, executed in a certain context, at a certain point in time [
36,
47]. The principle of compatibility of Fishbein and Ajzen implies that the measures of intention and behavior include the same action, target, context, and time, which have been defined either at a specific or general level [
36].
Even if the conceptual model does not imply the existence of an endogenous variable “effective behavior”, this correspondence will be followed as in the case of a longitudinal study. As defined above, “buy” is the action (use of the device), directed at a target (in order to satisfy individual wishes and needs), executed in a certain context in which the individual has a choice between using the device or driving a vehicle without a device, an action that takes place at a certain point in time (when the need or desire arises). According to the recommendations of Fishbein and Statsson [
48], consumers’ intention to buy the eCall device is a single (dependent), one-dimensional reflective construct, also called first order, containing a set of three items (intention as expectation, intention as plan, expressed intention). This three-item plural expression of intention also allows a limitation of errors specific to declared preference surveys.
Again, the principle of compatibility of Fishbein and Ajzen implies that measures of attitude and behavior include the same action, same target, context, and time, which have been defined either at a specific or general level [
36]. Thus, at the specific level, the attitude of the consumers towards the idea of using the eCall device will be studied, in order to satisfy some wishes or needs, and at the general level, the attitude of the consumers will be studied towards the eCall device. When the two indicators comply with the principle of compatibility, they should correlate with each other [
47].
Fishbein and Ajzen differentiate between attitude toward an object and attitude toward behavior [
36]. According to the authors, the attitude towards an object represents the affective evaluation of the individual towards a specific object; the attitude towards the behavior represents the individual evaluation of a specific behavior that includes the object. Theories in the field of information systems have accepted Fishbein and Ajzen’s (1975) definition of attitude as an evaluative affect that represents the positive or negative feelings of the individual towards the achievement of the target behavior [
9,
22].
Based on these considerations, attitude is the measure of the evaluative affect that an individual associate with the idea of using the eCall device (definition adapted to the specifics of the study after [
46]). Based on these considerations, the attitude variable is constructed as a one-dimensional reflective construct, with a set of four items that do not differentiate between the instrumental and the experiential component of the attitude.
Perceived usefulness is a concept of TAM and represents the individual’s conviction that using a technology brings certain benefits [
9]. Aftermarket adoption of an eCall device implies the use of a new technology that individuals would not be willing to accept without its benefits. Perceived usefulness is closely linked to another concept in the literature: the relative advantage, a specific concept of Innovation Dissemination Theory, representing the degree to which an innovation is perceived as an amplification or improvement of the current supply [
18]. As a mode of construction, the variable “perceived usefulness” was constructed as a one-dimensional reflective construct, consisting of a set of four items.
Ease of use is a concept of the TAM and represents the extent to which users of an information system perceive a low degree of effort to use it [
9]. Adapted to the study of the usage behavior of the eCall device, we will define the perceived ease of use as the consumer’s beliefs about the low degree of effort required to use the device. The measures of perceived ease of use were adapted after Davis (1989), considering the measures already adapted for the study of the ease of use. As a mode of construction, the variable “ease of use” was constructed as a one-dimensional reflective construct, consisting of a set of three items.
Through the elaborated questions it was desired to obtain information regarding the socio-demographic profile of the driver and their driving habits, but also their opinions, beliefs, and attitudes related to the use of an eCall device. Several types of questions were used, depending on the nature of the question: closed, dichotomous, and multiple answer questions. In the case of questions with scaled answers formed based on the 7-category Likert measurement scale, the respondents chose the level of agreement regarding the statements.
The first question, also called the filter question, had the role of selecting those respondents who meet the qualifications to be part of the investigated sample, i.e., whether they hold a driving license. The rest of the questions were grouped by topic, in sections, arranged in a logical structure. The pretesting of the questionnaire was carried out among 256 respondents, with the characteristics of the investigated population, to identify possible confusions in the elaboration of the questions.
The participants in the pre-survey of the questionnaire were asked to take notes regarding the duration of completion of the entire questionnaire and the difficulty level of the questions, but also to provide feedback regarding any improvements or errors/failures of the web platform. Upon administering the questionnaire, respondents were also briefed on the aftermarket eCall in-vehicle system, its features, and its functionalities, as the survey included a presentation of the system and was also administered at automobile tradeshows, transportation conferences and other events where the project partners were involved in the presentation of the eCall IVS. The structure of the questionnaire is presented in the
Supplementary Materials.
3.2. Sampling
Given that the purpose of the research is to study the consumers’ intention to purchase the eCall device, the target population was delimited by certain fundamental characteristics: the possession of a driver’s license and a vehicle (or at least using a vehicle at the work place). Thus, only those consumers who own a driver’s license and drive a car represent the target population of the research. There are situations, as in the present case, when there are no sampling frames for the target population of the research. There is no database of all persons with driving license valid in 2019 and their contact information, hence our use of filter questions to correctly identify the elements of the target population.
Depending on the time criterion, the sampling method is a traditional one, which implies that the entire sample must be established before the data is collected. Moreover, the sampling method is a non-return method, which allows an element to be included only once in the research sample. Depending on the degree of involvement of the researcher in the selection of the sample, the sampling method is an improbable one, in which case the probability of selecting a unit from the researched population is not known before the data collection.
This sampling method was chosen due to the impossibility of using a sampling framework. The cases in which the sampling framework cannot be determined require the use of an improbable sampling method [
49]. One of the most commonly used non-probability sampling methods is random sampling, which involves finding respondents in places where they can be found, places that are within the researcher’s reach [
50]. By determining the sample size, it was desired to comply with the requirements regarding the size in order to make it possible to extrapolate the research results to the entire population surveyed. To this end, we started from the concept of “proportion” that describes the studied community to determine the sample size in relation to the investigated attributes:
- -
n = the sample size
- -
t = coefficient associated with the probability of guaranteeing the research results (confidence level)
- -
p = the percentage weight of the components of the sample that are characterized by a certain attribute
- -
q = the percentage weight of the components of the sample that are not characterized by the attribute p, being determined as a relation (1 − p)
- -
e = margin of error
Given that the sample will be used to investigate consumers’ intention to use the eCall device, a “driving license” attribute was chosen as the key attribute. According to the Eurostat report [
37], in the EU there is a 294,966,256 vehicles stock. Based on these statistics, one can calculate the weight of the persons who would be targeted in purchasing the eCall device among the drivers; the current population in the EU stands at roughly 513,481,690 [
37]. Given that the sample is used to investigate consumers’ intention to use the eCall based on the 112 IVS, we used a “driving license” key attribute as a filter. Based on these official data, the value of attribute “
p” will be
p = 0.575, and that of attribute q will be q = 0.425. It was decided to work under a confidence interval of 0.05, corresponding to a probability of guaranteeing 99% results and a margin of error of +/−5%. Thus, the sample size will be:
Based on the above considerations, the research sample must include no less than 651 observation units to respect the principle of representativeness for the investigated population. Due to the fact that we managed to gather 689 valid surveys, the final sample size remained at 689.
It is also worth mentioning that at a standard
p value of 0.5, the sample value would be
n = 666 respondents; therefore, our final sample size remains valid. The differences and similarities between the chosen sample, consisting of 689 respondents and the reference sample of the Eurostat report [
37], were analyzed from the perspective of four variables: gender, age, residence, vehicle type, vehicle age. Regarding the gender of the respondents, there are no significant differences, the number of female respondents being close to the number of male respondents. A description of the defining characteristics of the respondents is provided in
Section 3.3.1.
3.3. Examination and Validation of Data
Missing values may be random, when respondents skip a question due to inattention, or non-random, when they choose not to answer a certain question for different reasons [
51]. This fact is also confirmed by the analysis of missing values conducted in the statistical program WARP PLS version 16. By pretesting the data, the program shows the existence or non-existence of missing values within each variable.
Extreme multivariate values represent the extreme scores for two or more variables or an atypical score pattern [
52]. The extreme multivariate values in WarpPLS can be observed in the analysis of the path coefficients between two latent variables, when we encounter absurd values of the significance threshold, under the conditions in which the sample is representative [
53]. Extreme multivariate values can be identified through the diagrams of linear or non-linear relationships between the latent variables.
Techniques for modelling structural equations based on variance analysis (PLS) do not imply the existence of normally distributed data [
53,
54,
55,
56]. PLS uses the non-parametric bootstrapping technique to obtain the standard errors required for hypothesis testing and assumes that the sample distribution is a reasonable representation of the target population distribution [
53,
54,
55]. Due to the fact that the collected data does not respect the principle of normality, PLS is the best method of data analysis.
3.3.1. Respondents Profile and Representativeness
Drivers have different socio-demographic characteristics and buying habits. For a better knowledge of their defining characteristics, the profile of the respondents was established.
The complete profile of potential consumers, who have a driver’s license and who have answered yes to the request to participate in the survey, is presented in
Table 3. It can be noticed that there are no significant differences in the gender of the respondents, 45.42% being women and 54.57% being men.
Very large differences appear within the demographic variable “age”. Most of the respondents are between 16 and 30 years old: 41.21% are between 16 and 30, 30.33% between 31 and 40, 14.8% between 41 and 50, 10.59% are between the ages of 51 and 60, and 3.04% are over 60 years old.
Regarding their place of residence, 80.11% of the respondents live in the urban environment, and 19.88% live in rural areas.
The vast majority of respondents, namely 80.11%, reside in the city; only 19.88% declared that they reside in the rural environment. Regarding their vehicle types, as expected, the majority of respondents drive passenger vehicles (87.22%), described as cars or taxis. 5.51% drive a bus or coach, followed by motorcycles at 4.2% and HGVs at 1.59%. Respondents are more evenly distributed according to the age of their vehicles. Regarding vehicle age, most of the respondents, namely 35.12% of the total respondents, declared that their vehicle is less than 5 years old, followed by less than 10-year-old vehicles (32.36%) and more than 10 but less than 15 years old (27.57%). At the opposite pole are those who own vehicles that are over 15 years old (4.93%).
With regards to driving frequency, most of the respondents, 57.91%, drive daily, followed by those who drive several times a week, representing 21.62%.
The main area of driving is in the city/village (76.77%), with only 22.35% declaring that they most often drive outside the city/village.
3.3.2. Measurement Accuracy
In any data set there will be a certain amount of error that must be minimized for better accuracy of the investigated phenomenon [
57]. Moreover, no measuring instrument is completely safe, and there is always the possibility of a measurement error [
57]. The measurement error represents a potential threat to the validity of the empirical results, which is why it is necessary to validate the measuring instrument before presenting the research results [
58].
The most commonly used method for evaluating the accuracy of measurements is internal consistency [
59]. The popularity of internal consistency as a method for determining accuracy is given by its ability to measure the accuracy of measurements from a single administration of the questionnaire [
60].
It starts from the idea that the scales formed of several items are effective only if they behave in a homogeneous manner, measuring the same variable [
61]. Thus, internal consistency ensures that all items of a variable consistently contribute to determining the measurement of a particular variable [
62].
Usually, internal consistency is measured using Cronbach Alpha coefficients [
59,
61,
62], the technique of calculating coefficient estimates medium accuracy based on all the ways of dividing the set of items into two parts.
The Cronbach Alpha technique correlates each individual item with the other items of the same construct and with the overall score [
62]. The Cronbach’s Alpha coefficient (α) is calculated based on the formula:
where
K represents the number of items,
represents the variation of the total scores observed, and
represents the variation of the item “
i” [
63].
The Cronbach Alpha coefficient can take values between 0 and 1. There are disputes over the value of Cronbach Alpha coefficients indicating acceptable or adequate accuracy of measurements [
64] (p. 221).
Most researchers cite Nunnally’s recommendation for an acceptable threshold of over 0.7 [
64], (p. 221). Nunnally recommends a Cronbach coefficient value of over 0.7 for preliminary research and over 0.8 for applied research [
65] (p. 245).
Internal consistency can be regarded as an inappropriate term, since it is not based only on the average correlations between the items, but also on the number of items that make up that variable [
65], which is why a variable made of 10 items will have a better internal consistency than a variable consisting of three items. Thus, determining the accuracy of measurements based on Cronbach Alpha coefficients will be done if the number of indicators is limited [
64] (p. 222).
To test the accuracy of each variable used in the present research, we used the WARP = PLS software, version 6.0. From the output generated by the program, we compiled
Table 4.
In the case of all variables, the Cronbach Alpha coefficient indicates a good internal consistency of this scale [
65]. The total correlation of each item with the cumulative score of the other items supports this assertion.
Furthermore, by eliminating one of the items in the construct, we obtain a Cronbach Alpha coefficient lower than that generated by all items. Thus, it is not necessary to give up any item of the construction [
66].
Another way to check the internal consistency of items is represented by measurement of composite reliability [
53], which represents a measure of the overall accuracy of a set of items that measure the same construct [
67].
Composite reliability coefficients can take values from 0 to 1, recommending a value of composite reliability coefficients of over 0.6 [
58,
67] and 0.7 [
65], respectively, which indicate acceptable internal consistency.
To test the accuracy of the measurements by the composite reliability coefficients, we used the statistical software WarpPLS, version 6.0. The results obtained are presented in
Table 5.
All coefficients of the composite reliability exceed the recommended minimum threshold of 0.6 [
67], having values between 0.915, in the case of the variable “perceived usefulness”, and 0.938 in the case of “attitude”. As with the Cronbach’s Alpha coefficients, the composite reliability coefficients indicate a good consistency of measurements.
3.3.3. Measurement Validity
Validity refers to the extent to which the measurements adequately reflect the real meaning of the investigated concept [
68]. Unlike exploratory factorial analysis, in the case of structural equation modelling (SEM) based on variance analysis, namely PLS, a confirmatory factorial analysis is performed, starting with certain theoretically anticipated factors in an attempt to confirm that the set of items correspond to them [
69] (p. 92).
The accuracy of this pre-specified model is examined to determine convergent and discriminant validity [
69] (p. 92). Convergent validity evaluates the relationship between the scores of the proposed indicators to measure a concept or construct. If all indicators of a construct show similar results [
70] and significantly load on the construct for which they were defined [
69], then the measuring instrument is assumed to have good convergent validity.
Convergent validity was tested with the statistical software WarpPLS, version 6.0, which generates factor loadings and cross-loadings based on structure matrix, pattern matrix, and a combination of the two types of matrices, the result of which is three methods being generated in separate tables.
Rotation has been viewed as a set of methods in factorial analysis, whereby the researcher tries to associate the calculated factors with theoretical entities, regardless of whether the researcher expects them to correlate or not.
The oblique rotation methods are most appropriate in a SEM (structural equation modelling) analysis, since it starts from the existence of certain correlations between the latent variables, in the absence of which the link coefficients would be insignificant [
53].
Thus, the first table analyzed will be the one of the factor loadings and the cross loadings from the pattern matrix. They were obtained from the transformation of a structure matrix (structure matrix) by the oblique rotation method, also called Promax [
53].
In
Table 6, we observe that each set of items loads very well within its defined construct. For example, within the INT construct, representing consumers’ intention to purchase/use the eCall IVS, factor loadings take values from 0.864 to 1.01, while cross-loadings are very small. These values indicate that the elements of the INT construct, namely INT1, INT2, INT3 load efficiently inside the construct and very weak outside of it.
Table 7 displays the factor loadings and the cross loadings from the structure matrix. As the matrix contains Pearson correlations between indicators (items) and latent variables, this is not significant before rotation in the context of validation of the measuring instrument [
53].
Comparing both the factor loadings and the cross loadings of the template matrix and the structure matrix, we observe that they tend to be higher within the structure matrix. This is because within the template matrix, the loadings reflect both the unique relationship between item and factor, as well as the relationship between item and common variance between factors [
54].
In order for the measurements to be well defined, that is, to respect the principle of convergent validity, it is necessary that the
p values associated with factor loadings be less than 0.5 and the loadings equal to or greater than 0.5 [
71].
Indicators for which these criteria are not met are excluded from the model. A more permissive case is that of the formative variables, for which this rule does not apply [
53].
Table 6 shows factor loadings above 0.5 and cross loadings below 0.5, confirming the divergent validity of the measurements.
Table 8 shows the combined loadings and cross-loadings in the pattern matrix [
53]. Due to the fact that the loadings come from an unstructured matrix, they will have values between −1 and 1. In order to allow the convergent validity of the measurements, the loadings must be equal to or greater than 0.5, and the
p values associated with the loadings must be more less than 0.05 [
71,
72].
From
Table 8, we can see that each item loads very well inside the construct, with loading values over 0.5, at
p < 0.001, thus meeting the convergent validity criteria.
Discriminant validity determines the extent to which one latent variable differs from the other latent variables in the model [
56]. Discriminant validity confirms that the latent variable, or construct, is unique, not just a reflection of the other variables [
73].
Discriminant validity is confirmed in the structural equation modelling (SEM) based on the PLS technique, by analyzing the mean variance extracted (AVE) [
69]. The average extracted variance measures the variance captured by the lantern construct, or the variance explained [
69]. This is calculated using the formula:
where
λi is the loading of each item within the corresponding construct [
69].
The extracted mean variance values have a recommended minimum threshold of 0.5 [
67]. As we can see in
Table 9, the average variance extracted for each construct exceeds the critical threshold of 0.5.
Good discriminant validity demonstrates that the question-statements (items) associated with each latent variable are not confused by the questionnaire respondents with the items of the other latent variables in terms of the meaning of each question-statement [
53].
Based on
Table 10 data, the values on the diagonal (corresponding to the square root of AVE) must be larger than any other value above or below it, from the same column (corresponding to the correlations between the latent variables) [
53]. Thus, it is observed that the square roots of AVE are much larger than any correlation between two variables, fulfilling the criterion of validity.