The final study of this dissertation takes a more in-depth approach to examine one making stage. The context for this data collection is organizational decision-making, and more specifically, information search on a company website (Sipilä et al., 2016). It has been proposed that research on the time-related aspects of information search calls for “non-traditional” methodologies (Einhorn and Hogarth, 1981). Accordingly, continuous facial expression and clickstream data were collected for the final study included in this dissertation to study ambivalence during the information search stage in more detail than in the other studies. Facial expressions were used because they enabled the continuous measurement of emotions while the participants were searching for information on a website. Continuous measurement was beneficial especially because this dissertation’s focus on processes, as not only the end state, but also the unfolding of evaluations as they form is observable in overt motor behavior (Schneider et al., 2015), such as facial expressions. Additionally, the facial expression approach enabled ambivalence to be measured on an implicit level, therefore complementing the verbal
measures of ambivalence used in the other data collections. This was beneficial because in this study, the affective component of ambivalence was the focus area, and in terms of emotions, verbal expressions have their limitations, as emotions are not only a verbally conveyable experience; they are also embodied (Larsen et al., 2008). Therefore, one school of emotion researchers have approached emotions using physiological signals to understand psychological processes (Larsen et al., 2008). This approach was also adopted in this study.
3.4.1 Sampling
Because the study investigates organizational buyers’ decision-making, it was necessary to recruit participants who either had some experience with organizational buying tasks or could be conducting such tasks in organizations (Sipilä et al., 2016). The participants were therefore recruited from professional events, email lists, and social media forums targeting university alumni or students who were close to graduation (Sipilä et al., 2016).
A total of 29 individuals participated in the study (Sipilä et al., 2016). One participant had to be excluded for not following the instructions, and three additional participants were excluded because their faces were not properly identified by the Noldus FaceReader software, which was used to identify emotions form facial expressions (Sipilä et al., 2016). Thus, the final sample consisted of 25 participants (Sipilä et al., 2016). The participants’ mean age was 34 years, most had a higher education degree (n=20), and they had a business (n=16) or engineering (n=9) background (Sipilä et al., 2016). Thus, the sample was comparable to young professionals who might be assigned the task of searching for information and evaluating a software service for their company. In addition, it mainly represents novice B2B decision-makers who are accustomed to searching for information in an online environment (Marshall et al., 2012). However, there is an oversampling of doctoral degree holders, which is a limitation of the study.
3.4.2 Data collection
Upon arrival, the participants were led to a controlled laboratory setting, where they were briefed and signed an informed consent form (Sipilä et al., 2016). They were then seated in front of a computer screen, and were asked to fill in a pre-questionnaire with questions about their demographics and work experience (Sipilä et al., 2016). Next, they were asked to imagine that they had been assigned to purchase a professional services automation (PSA) software for a small- or medium-sized (SME) company at which they were currently employed (Sipilä et al., 2016). They were then asked to search for information and evaluate the suitability of an actual PSA software presented on an existing website for the company’s PSA needs (Sipilä et al., 2016). The participants were allowed to chat with sales representatives via an automatically opening chat window, but were not allowed to use any other websites during the task (Sipilä et al., 2016). The participants were given a maximum time of 20 minutes to spend on the website (Sipilä et al., 2016).
While the participants were on the website, their clickstream was recorded using eye tracking software (Tobii Studio), and their facial expressions were recorded with a video camera (Microsoft LifeCam) (Sipilä et al., 2016). After the search task, the participants received a questionnaire, in which their behavioral intentions towards the company and the control variables were measured (Sipilä et al., 2016). After the questionnaire, they were debriefed and left the laboratory (Sipilä et al., 2016). The entire session took 45–60 minutes (Sipilä et al., 2016). The process is described in more detail in Publication IV.
The dependent variable and control variables were measured in the post-questionnaire (Sipilä et al., 2016). The dependent variable was constructed of three behavioral intention items, which indicated how the participants would have proceeded after the task (Sipilä et al., 2016). The items were “I would send a request for an offer,” “I would schedule a demo with a sales person,” and “I would start trial,” measured on a sliding scale (0 = very unlikely; 100 = very likely) (Sipilä et al., 2016). A summated scale was calculated based on these items (Sipilä et al., 2016). Perceived risk related to the decision was used as a control variable and was measured on a scale from 1 to 7, using three items adopted from Sitkin and Weingart (1995). The question was: “How would you characterize the decision at hand?” The items were: “significant opportunity - significant threat,” “potential for loss - potential for gain,” and “positive situation - negative situation.” Product evaluation was measured with items adapted from Mukherjee and Hoyer (2001) and measured on a sliding scale (0 = totally disagree; 100 = totally agree). The items were: “This software seems good,” “I like this software,” “This software is useful,” and “This software is high quality.”
The ambivalence measure was constructed from the emotion measurements exported from face reading software (the software is described in more detail in the next section).
In line with the existing research (e.g., McGraw and Larsen, 2008), happiness and sadness were used for operationalizing the positive and negative emotion dimensions (Sipilä et al., 2016). In the raw data exported from the FaceReader, each emotion has a value between 0 and 1, 0 meaning that the emotion is not expressed on the face, and 1 meaning a very strong expression (Sipilä et al., 2016). For both happiness and sadness, there was thus a value between 0 and 1, measured 30 times per second, and these values were used to calculate an ambivalence score using the Griffin calculation, which was outlined in section 3.3.2 (Sipilä et al., 2016).
3.4.3 Analysis
After the data collection, the videos of the participants’ facial expressions were run through the Noldus FaceReader software, which uses the Facial Action Coding System (FACS; Ekman and Friesen, 1976) to identify seven basic emotions (happiness, sadness, anger, fear, disgust, surprise, and neutral) (Sipilä et al., 2016). The FACS is the most commonly used, comprehensive, and rigorous observer-based system of facial expression recognition, and uses action units (AUs), which are the smallest visually discriminable
facial movements (Cohn et al., 2007). From the AUs, it is then possible to interpret emotions according to different combination rules (Cohn et al., 2007). This step was automatically conducted by the FaceReader.
The website was divided into two-page categories based on the contents of each page, namely the landing and product information pages (Sipilä et al., 2016). The participants started the task on the landing page, and subsequently moved to the product information pages, including a page with an overview of the product and more detailed product information pages (Sipilä et al., 2016). The time periods during which the participants were on the landing page or the product information pages were acquired from the clickstream data and consequently event marked into the FaceReader data (Sipilä et al., 2016). For data-reduction purposes, the mean values of ambivalence on the landing and product pages were calculated for each participant (Sipilä et al., 2016). These values were imported into SPSS 22, and hierarchical multiple regression was performed with ambivalence on the landing and product information pages as independent variables and behavioral intentions as dependent variables, with perceived risk and product evaluations as control variables (Sipilä et al., 2016). A more detailed description of the analysis is included in Publication IV.
3.4.4 Validity and reliability
The content validity of the questionnaire measures was ensured using constructs from the existing research, which had been accurately defined and validated in previous studies.
Any modifications to the wording of the items was made only to adapt the measures to the context of the study. In addition, all materials were examined by three researchers prior to the data collection. The validity of the facial expression measures was assessed based on the existing literature, which has provided evidence on the validity of measures of facial muscle activity as indicators of valence (Bolls et al., 2001; Wang and Minor, 2008). These studies have been conducted with electromyography (EMG), which involves placing electrodes on the participant’s skin, and is therefore slightly different from the facial expression measures used in this dissertation. Correspondingly, the validity of Noldus FaceReader has been confirmed by comparing the output of the FaceReader with EMG (D’Arcey, 2013). Additional efforts to validate the FACS method are based on the performed action criterion, in which people were trained to perform certain actions on request, and records of the performances were coded without knowing the actions requested. FACS can recognize the performed behaviors (Kanade et al., 2000).
Furthermore, the stability of FACS action units has been found to be good over a four-month interval (Cohn et al., 2002). These studies provide evidence of the validity of the FACS coding, based on which the FaceReader interprets emotions.
The reliability of the questionnaire measures was established prior to data collection by using multi-item scales and ensuring that the instructions were clear and the items were unambiguous (Peter 1979). The scales consisted of multiple items, and the questionnaire was pretested, which prevented ambiguity. The reliability of the survey measures was
assessed with Cronbach’s alphas, which ranged from 0.713 (perceived risk related to the decision) to 0.915 (product evaluation), indicating good to very good reliability (Sipilä et al., 2016). The reliability of the facial expression analysis was ensured using a controlled laboratory setting with a frontal camera orientation. The camera was placed at an optimal distance from the participants to ensure that the FaceReader software would analyze the expressions correctly, and the participants were asked to remain as still as possible during the recording. These steps were important for ensuring reliability at the data-collection stage, because the camera orientation, head motion, and size of the head in an analyzed figure are reasons for poor reliability in facial expression studies (Sayette et al., 2001).
Additionally, using a computer software for FACS coding instead of human coders may increase reliability, as the software makes fewer coding errors than human coders (Cohn et al., 2007). Finally, the assumptions of regression analysis (i.e., linearity, homoscedasticity, normality of residuals, and multicollinearity assumptions) were met on a satisfactory level, which lends further support for the trustworthiness of the survey results (Appendix C).
4 Summary of the publications and review of the findings
This section summarizes the results and outlines the publications included in the dissertation. First, a summary of the results of the entire dissertation is outlined. Next, the objectives, main results, and contributions of each paper are discussed.
4.1
Review of the findingsFigure 8 presents an overview of the findings of this dissertation regarding the role of ambivalence in the decision-making process, organized into a classical model of the decision-making process (e.g., Puccinelli et al., 2009). As visualized in Figure 8, the findings suggest that in the need recognition stage, ambivalence arises from the preliminary information of different product attributes, but does not have many consequences. During the information search stage, information is actively searched, and therefore different types of information about products, brands, and their aspects become relevant antecedents. Additionally, ambivalence has more consequences than in the need recognition stage. Furthermore, different types of ambivalence are coped with differently.
While cognitive ambivalence is coped with through an increased information search, both cognitive and affective ambivalence are coped with through emphasizing the valence of the most important product attribute. Further, intercomponent ambivalence is coped with through the rejection of symbolically unacceptable product options. In the information search stage, ambivalence carryover and the intention to continue in the decision-making process are also found to be consequences of ambivalence. In the evaluation stage, information about different product attributes continues to influence ambivalence.
However, potentially because of the temporal proximity of the choice stage, ambivalence in the evaluation stage does not have numerous consequences. For example, in the video diaries, many participants combined the evaluation and choice stages, doing their final evaluation only moments before the final choice. In the choice stage, concrete information reduced ambivalence, as theorized based on CLT. Information about some product attributes still continued to elicit different types of ambivalence, which were coped with by emphasizing the most important attribute. Finally, any remaining ambivalence had a negative influence on choice.
Figure 8. Summary of the results