It’s complicated: explaining the relationship between trust, distrust, and ambivalence in online transaction relationships using polynomial regression analysis and response surface analysis

Abstract

Trust and distrust are considered crucial elements affecting online relationships – particularly those involving electronic transactions. Although some studies propose that they are distinct, others claim that they are merely opposite ends of one continuum. Further adding to the debate is the possibility of ambivalence, a topic that has not been examined in electronic transaction relationships. Unfortunately, current models of trust and distrust have limitations that impede explanations of how – or even if – ambivalence is generated by feelings of trust and distrust and how these two constructs can best coexist. We thus propose a hybrid model which considers the limitations and strengths of previous models. Namely, we posit that trust and distrust can coexist as separate components with related continua. We use polynomial regression analysis (PRA) and response surface analysis (RSA) to test these complex relationships. Using an empirical study of online consumer behaviour with 521 experienced online consumers, strong empirical validation is found for the model. We examine the effects of ambivalence on the truster’s intentions towards a website and find a small positive effect which increases such intentions. PRA and RSA confirm that trust and distrust are most likely separate components – not opposite ends of a continuum – with related continua. The continua within the subconstructs of trust and distrust likely have more complex and interesting relationships than have been considered previously. These findings lead to interesting future research opportunities on trust, distrust and ambivalence using advanced techniques such as PRA and RSA.

Appendices

Appendix 1

See Table 14.

Table 14 Instrument detail

Appendix 2

See online Appendix 2.

Appendix 3: factorial validity and analysis support

Assessing construct validity

In this section, we assess several key elements of construct validity, including determining which constructs are formative and which are reflective (Diamantopoulos and Winklhofer, 2001); assessing factorial validity as determined by discriminant validity and convergent validity (Straub et al, 2004); evaluating multicollinearity (Cenfetelli and Bassellier, 2009); and checking for common methods bias (Podsakoff et al, 2003). We used partial least squares (PLS), employing SmartPLS version 2.0 (Ringle et al, 2005) for model validation and analysis because PLS is especially adept at validation of mixed models of formative and reflective indicators (Chin et al, 1996; Diamantopoulos and Winklhofer, 2001; Chin et al, 2003; Gefen and Straub, 2005; Lowry and Gaskin, 2014).

Determining which constructs are formative and which are reflective

A key step of preparation for assessing factorial validity is to determine which constructs are formative and which are reflective (Diamantopoulos and Winklhofer, 2001). The basic difference is that items within formative constructs are theoretically distinct, and thus cannot be replaced with other items in the same construct; meanwhile, the items in reflective constructs are theoretically similar, and thus are replaceable (Diamantopoulos and Winklhofer, 2001). This theoretical and methodological distinction has recently become a serious issue in IS research, where it has been discovered that many previous IS studies were mis-specified because they did not distinguish between reflective and formative constructs (Petter et al, 2007). Such mis-specification can lead to problems in empirical results and theoretical interpretations, including a potential increase in both Type I and Type II errors (Petter et al, 2007).

We used the latest leading recommendations (Diamantopoulos and Winklhofer, 2001; Petter et al, 2007; Cenfetelli and Bassellier, 2009; Lowry and Gaskin, 2014) to determine how to analyse our constructs. Notably, constructs are not inherently reflective or formative; it is up to the researcher and the literature to make that determination and to choose appropriate items for the conceptualisation (MacKenzie et al, 2011). Our conceptualisation and measurement, along with formal model specification, followed the latest conventions (e.g., MacKenzie et al, 2011) Trust and distrust beliefs and the dispositions to trust and distrust have previously been theorised, modelled and validated as second-order constructs, composed of first-order reflective subconstructs (McKnight et al, 1998; McKnight et al, 2004; McKnight and Choudhury, 2006). We have neither theoretical nor methodological reasons to contradict these previous construct conceptualisations, and thus we have validated and modelled our constructs accordingly.

Establishing factorial validity

Factorial validity is established through both convergent and discriminant validity, which are two highly interrelated concepts which must coexist. Convergent validity is the basic idea that measurement items which should be related are related. It is established ‘when items thought to reflect a construct converge, or show significant, high correlations with one another, particularly when compared to the convergence of items relevant to other constructs, irrespective of method’ (Straub et al, 2004, p. 391). Discriminant validity is the basic idea that items that should not be related are in fact not related. Thus, it can be established when items thought to diverge show insignificant, low correlations with one another, particularly compared to items in other constructs (Straub et al, 2004). Importantly, factorial validity is established in different ways for reflective and formative constructs; thus, we address these analyses separately.

Factorial validity of reflective constructs

To establish the factorial validity of our reflective constructs, we followed procedures by Gefen and Straub (2005) and Lowry et al (2014), and further demonstrated in (Lowry et al, 2008; Lowry et al, 2009). For an especially conservative analysis, we used two established techniques to establish convergent validity and two established techniques to establish discriminant validity.

Convergent validity of reflective constructs

First, we examined the outer model loadings. Convergent validity can be established when the t-values of the outer model loadings are significant. In all cases but one (one ambivalence item), each latent variable’s indicators strongly converged on the latent variable and was highly significant, as summarised in Table 15. As a second check, we correlated the latent variable scores against the indicators as a form of factor loadings, and then examined the indicator loadings and cross-loadings to establish convergent validity. Although this approach is typically used to establish discriminant validity, convergent validity and discriminant validity are interdependent and help to establish each other. Convergent validity is also established when each loading for a latent variable is substantially higher than those for other latent variables. This is done by correlating the latent variable scores against the indicators as a form of factor loadings. Table 16 illustrates the loadings in italics. Based on this analysis, only the same indicator in ambivalence showed poor loading on its intended construct, in comparison to all other constructs.

Table 15 Outer-model weights t values of reflective items to test convergent validity

Table 16 Items in latent variable analysis for discriminant validity

Discriminant validity of reflective constructs

We also used two approaches to establish discriminant validity. First, like with convergent validity, we examined the factor loadings, but this time, wanted to ensure significant that overlap did not exist between the constructs (see Table 16). All loadings, excluding an item from following advice, were appropriate, given the dropped ambivalence indicator in the previous step.

Second, we used the approach of examining the square roots of the average variances extracted (AVEs), as summarised in Table 17. The basic standard followed here is that the square root of the AVE for any given construct (latent variable) should be higher than any of the correlations involving the construct. The numbers are shown in the diagonal for constructs (bolded). Strong discriminant validity was shown for all constructs.

Table 17 AVE analysis to establish discriminant validity

Factorial validity of formative constructs

Establishing factorial validity for formative indicators is more challenging than validating reflective indicators, because the established procedures which exist to determine the validity of reflective measures do not apply to formative measures (Gefen and Straub, 2004; Petter et al, 2007). Moreover, the procedures for validating formative measures are less known or well-established (Diamantopoulos and Winklhofer, 2001), although standards are beginning to emerge in IS research (Cenfetelli and Bassellier, 2009).

Validating items within formative measures is particularly challenging because these items can move in different directions from one another. Whereas reflective indicators must demonstrate considerably high correlations with each other (i.e., high conceptual overlap) to be valid internally, the indicators of a formative construct need not meet this criterion, and instead need to represent distinct facets of the overall construct being modelled (Bollen and Lennox, 1991; Diamantopoulos and Winklhofer, 2001; Petter et al, 2007). Reflective items are interchangeable, but formative items are not; hence, reliability measurements are not appropriate for formative constructs (Diamantopoulos and Winklhofer, 2001). Specifically, internal consistency examinations of formative constructs with Cronbach’s α and AVE calculations are not methodologically appropriate (Bagozzi, 1994; Marakas et al, 2007; Petter et al, 2007; Cenfetelli and Bassellier, 2009).

Researchers have traditionally used theoretical reasoning alone to support the validity of formative constructs (Diamantopoulos and Winklhofer, 2001). Over time, methodological approaches have emerged to improve the validation of formative constructs, such as the modified multitrait–multimethod approach and assessment of multicollinearity (Straub et al, 2004; Marakas et al, 2007; Petter et al, 2007). This foundation has been improved on in work by Cenfetelli and Bassellier (2009), which we follow for our validation process.

As an initial step, we assessed the absolute indicator contributions (i.e., zero-order correlations) of the individual items for service quality against the overall average of service quality. The idea is to improve internal validity by removing items not exhibiting a significant association with the overall construct (Diamantopoulos and Winklhofer, 2001; Cenfetelli and Bassellier, 2009). All of the items exhibited significant associations with the overall measure at the 0.05 level of significance. We also performed inter-item correlational diagnostics to assess whether there were high correlations among the formative indicators, as these can significantly weaken formative measures (Diamantopoulos and Siguaw, 2006). However, the biggest potential issue that must be addressed is multicollinearity (Cenfetelli and Bassellier, 2009). We thus we assessed the possibility of multicollinearity amongst all of the indicators (reflective and formative) in the model. Variance inflation factors (VIFs) less than 10 are traditionally viewed as justification for a model’s lack of multicollinearity, with 5.0 being ideal, but formative methodologists have recently called for a more stringent cut-off of less than 3.3 (Diamantopoulos and Siguaw, 2006; Petter et al, 2007; Cenfetelli and Bassellier, 2009).

A number of concerns emerged from this analysis. While all of the reflective indicators had VIFs of 5.0 or less, a few of the formative indicators were above the more stringent 3.3 cut-off. All such instances were found in trusting beliefs and trusting intentions, but these were second-order formative constructs made up of reflective subconstructs. However, given the extensive use and theory backing the second-order formative nature of these two constructs, we retained them in the model as specified by research, noting that small amounts of multicollinearity may obfuscate the results.

Establishing a lack of common-methods bias

To diminish the likelihood of common methods bias occurring in our data collection, we randomised items within the instrument so that participants would be less apt to detect underlying constructs, another potential source of common method bias (Cook and Campbell, 1979; Straub et al, 2004). However, all data were collected using a similar-looking online survey; thus, we still needed to test for common methods bias to establish that it was not a likely negative factor in the data remaining for our analysis. We used two approaches to increase validity and rigor.

The traditional approach to establishing lack of common methods bias is to conduct a Harman’s single factor test; however, the validity of this approach is increasingly under attack, and thus, we used a couple of stronger methods instead (Podsakoff et al, 2003; Pavlou et al, 2007).

The first approach was suggested by Podsakoff et al (2003), and adapted for PLS by Liang et al (2007). The objective of this technique is to measure the influence of a common latent method factor on each individual indicator in the model versus the influence of each indicator’s corresponding construct. To perform this technique in PLS, constructs of the theoretical model and their relationships are modelled as is normally conducted with two major additions: (1) A single-indicator construct is created for each indicator in the measurement model. Each subconstruct is then linked to each of the single-indicator constructs which comprise the subconstruct. This effectively makes each subconstruct in the model a second-order reflective construct. (2) A construct representing the method is created, reflectively composed of all indicators of the instrument. The method construct (the latent method factor) is then linked to each single-item construct. Based on this analysis the average substantively explained variance of the items is .833, while the average method-based variance is −.001. This results in a ratio of 637:1. In addition, most of the relationships between the items and the method-based construct were insignificant—indicating a lack of common-methods bias.

However, this approach by Liang et al (2007) is now under increase dispute as to its effectiveness. Thus, we used a second approach, which to simply examine a correlation matrix of the constructs and determine whether any of the correlations were above 0.90, which would be evidence that common method bias may exist (Pavlou et al, 2007). To be conservative, we conducted this analysis for the constructs and for the subconstructs. All construct correlations were below this threshold.

Manipulation checks

Two approaches were used for manipulation checks, as follows: (1) asking the participants if they noticed the manipulation and (2) statistical manipulation checks to see whether the treatments provided the desired manipulations. To assess the manipulation validity of the experiment, questions were added to the post-test to determine whether participants perceived their treatment manipulations. This elucidated whether the participants had noticed the process abnormalities, website design abnormalities and informational abnormalities. Table 18 shows the results of these manipulation questions. As can be seen, when asked whether participants perceived the manipulations, a majority was aware of it. However, a substantial portion of the manipulations were not perceived/remembered by the participants (see below). Nevertheless, these data were retained because they provide a more realistic test of our data. Straub et al suggest that although unmanipulated participants add additional variance to results, data for these participants may profitably be retained in the dataset to provide ‘a more robust testing of the hypotheses’ (2004, p. 408).

Table 18 Summary of manipulation checks – qualitative assessment

Several of the manipulations were relatively weak in comparison to the majority that was correctly perceived (the highlights in italics indicate the manipulations that were perceived with less than 50% accuracy). The most frequently under-perceived manipulation was the request for sexual orientation and mother’s maiden name (information abnormalities for treatments #4, 7 and 8). This manipulation was only accurately recalled once (treatment #3, 59%). Interestingly, while participants largely did not recall this manipulation, later analysis revealed that this type of abnormality did produce changes in overall trust. This shows that several of our manipulations were subtle, but in the end effective, providing all the more reason to retain all data, not just data which were properly perceived. Perhaps other manipulations of this type of abnormality would be more blatant and produce stronger results (e.g., Everard and Galletta, 2005). The downside of more blatant manipulations is giving up realism.

Two other treatments had less than expected perceived manipulations for informational abnormalities (#5 and 6). For the fifth treatment group, participants did not perceived the subtle shift from an Amazon shopping cart to that of Google, which is akin to a finding in the literature on change blindness (e.g., Levin and Simons, 1997; Silverman and Mack, 2006). Change blindness refers to individuals’ inability to notice changes in their current settings. Perhaps the change between two of the most major e-commerce shopping carts was too subtle for participants to perceive; again, later analysis indicates an effect from this manipulation despite the participants’ inability to perceive it. Likewise, participants did not correctly recall that no informational abnormalities existed, potentially because various other abnormalities were present. It is possible that the mixed signals in other areas resulted in a faulty recall of this one area that was not anomalous.

Higher than market prices were also incorrectly perceived in two treatment groups (#5 and 7), while they were correctly perceived in two other treatments (#2 and 8). Perhaps since the manipulation of high price was only marginal in comparison to low price, participants incorrectly perceived this manipulation (50% of the total), whereas all low price manipulations were correctly perceived. In the instances where the high price manipulation was not perceived, it is possible that such a manipulation may be due to other abnormalities present in the process which may have interfered with participants’ memories regarding price.

Finally, half of the treatment groups incorrectly recalled whether production-related information was being manipulated (treatments #2, 3, 4 and 6). Treatments #2, 3 and 4 incorrectly recalled that information was present about their products, despite the absence of such information (or the inclusion of information focussing deliberately on the wrong product– a car battery). Such inattention to detail may be attributed to the nature of the product being ‘purchased’ by the participants (i.e., rechargeable AA batteries). As participants are expected to be highly familiar with such items, it is possible that they largely ignored this information, as it would not factor into their buying decision.

Given that large portions of the participant sample were not aware of specific manipulations, this study also relies upon mean comparisons between treatment groups to assess the effectiveness of the manipulations. We conducted several rounds of comparative analysis, summarised in Tables 19, 20 and 21, which establish that the treatments mainly worked in the directions as intended.

Table 19 Summary of situational abnormality manipulation tests

Table 20 Summary of trust manipulations

Table 21 Summary of distrust manipulations

The means of the relevant constructs that were manipulated by the treatment groups are shown in Table 19. Each of the treatments significantly altered the levels of situational abnormality, which followed the study design. Specifically, the abnormality treatments (#2–5) reported even higher scores for situational abnormalities than the ambivalence treatments (#6–8). Table 19 indicates that all abnormality manipulations were significant and in the correct direction.

The same procedure was used to verify the trust manipulations found in treatments #3–5. These results are shown in Tables 20 and 21. Tables 20 and 21 indicate that all trust manipulations were significant and in the intended direction. For clarity, the trust dimension is highlighted with the corresponding manipulation, which is expected to be the lowest mean in the given column.

These results indicate that, with the notable exception of the deceit manipulation on integrity, the manipulations tended to produce the most pronounced results in their intended subdimensions of both trust and distrust. However, we note that all manipulations that contained some distrusting or negative cue (i.e., treatments #2–8) resulted in higher levels of distrust when compared with the control treatment. This indicates that the effects of the manipulation for a specific subdimension of distrust tend to bleed over to other subdimensions. This supports the assumption that intra-attribute ambivalence is likely not to be present in such relationships, as trusters do not distinguish between the subdimensions in great detail required for such ambivalence.

A multivariate analysis of variance of our manipulations onto trust, distrust and ambivalence shows that the manipulations significantly affected each of the dimensions of trust, distrust and ambivalence at the p = 0.000 level with the exception of price. Price showed significant results only with Roy’s largest root on these constructs (R = 0.032, F = 6.0, p = 0.016), whereas all other estimates were insignificant (Wilks’ lambda = 0.966, F = 12.0, p = 0.148; Lawley-Hoteling trace = 0.0354, F = 12.0, p = 0.146; Pillal’s trace = 0.034, F = 12.0, p = 0.151). Although price is not clearly shown to affect the results of trust, distrust or ambivalence, we retain it in our model due to the results shown above, and the significant effects it is found to have on the various dimensions in isolation.

We also regressed the effects of each manipulation of the study on the variables of interest to show the partial effects that each manipulation has on each dimension. The summarised results of these regressions are shown in Table 22. We note that the majority of the effects are highly significant.

Table 22 Regression results of manipulations on trust and distrust dimensions and ambivalence

Manipulations

The manipulations targeted process, information and website design abnormalities (i.e., mistakes). Each was manipulated as present – containing the listed errors, or absent– having no such errors. It was not the purpose of this study to explore what specific error causes what changes in a trust or distrust subdimension, but rather to ascertain that such a line of inquiry would be beneficial for future research. Based on previous literature (Everard and Galletta, 2005; Ou and Sia, 2010) and the results of our first pilot study, we identified the three general types of abnormalities that can be present on websites.

• 1. Process abnormality: an aspect of the typical buying process is disrupted.

  • Present: The buying process involves providing additional information that is not usually collected (e.g., mother’s maiden name, sexual orientation)

  • Absent: Shopping cart with credit card payment option as typically offered through most sites.

• 2. Information abnormality: information regarding the desired item/service is abnormal.

  • Present: Extremely low or high price in comparison to listed other sellers, missing product description, highly negative review and rating score, product description and name do not match displayed picture, or no sales history for the given seller

  • Absent: Comparable price to other listed sellers, commonly available description, expected customer reviews and ratings (average for sellers of this product)

• 3. Website design abnormality: can include extremely poor website design, errors and/or broken links that are not specifically relevant to the product/service information.

  • Present: Frequent and blatant misspellings, look-and-feel of webpage changes during the process

  • Absent: Consistent appearance throughout the entire process

Distrust manipulations consists of three levels: malevolence, incompetence and deceit. These were manipulated by providing customer feedback on the feedback page of the experiment which specifically manipulated that dimension only.

More details on experimental description

Participants were recruited from the two readily available subject pools at a large, public eastern US university. Initially, they were asked to complete a pre-experiment survey to gather stable personality characteristics (e.g., demographics, Internet experience and the dispositions to trust and distrust). Once participants completed the initial survey, they proceeded to an online survey containing the experimental manipulations, manipulation checks and post-manipulation survey.

Participants were told to imagine that they were buyers of a given product (i.e., battery pack) and that a given search provided the following scenario. They were asked to review the indicated screenshots and to respond to several questions concerning the attitudes and intentions that they would have if they had been making such a purchase. Each webpage was listed and described in the order that it appeared (screen shots of the webpages are given in Appendix 2).

First, participants viewed the main product page for the item that he or she was purchasing. This page contained an item picture, price, description and so on normally found on a product page. An initial view of the page was presented; then, additional zoomed-in portions of the page were presented to ensure that subjects became familiar with the information there (i.e., product description, price and seller information).

Second, customer reviews and ratings were displayed along with several comments from previous customers, such as those commonly found on Amazon.com. Like the product information page, portions of the customer ratings were zoomed into increase the likelihood of subjects being familiar with those portions of that page.

Third, subjects were shown a buyer’s information page, which requested personal and shipping information.

Fourth, subjects were shown a page where buyers would enter credit card and billing information. Finally, they were then shown a product confirmation page, which summarised the item, price, shipping and billing information.

To increase the likelihood of coexisting rival attitudes and potential ambivalence, several different yet important product attributes and dimensions were manipulated to be either normal or abnormal. Following research in ambivalence, several versions of the purchase process were utilised to focus on an overall attempted manipulation for normality, abnormality and ambivalence rather than focussing on specific manipulations of website factors. The abnormality manipulation groupings are summarised in Table 23.

Table 23 Summary of experimental manipulations

Finally, subjects proceeded to the instruments to respond to questions about distrusting and trusting beliefs, intentions and ambivalence in regards to this situation if they imagined themselves being buyers in this situation. They were also asked to provide their intentions concerning the seller and the website.