Trust and the Compliance–Reliance Paradigm: The Effects of Risk,Error Bias,and Reliability on Trust and Dependence

Abstract

Objective:

This study provides a theoretical link between trust and the compliance–reliance paradigm. We propose that for trust mediation to occur, the operator must be presented with a salient choice, and there must be an element of risk for dependence.

Background:

Research suggests that false alarms and misses affect dependence via two independent processes, hypothesized as trust in signals and trust in nonsignals. These two trust types manifest in categorically different behaviors: compliance and reliance.

Method:

Eighty-eight participants completed a primary flight task and a secondary signaling system task. Participants evaluated their trust according to the informational bases of trust: performance, process, and purpose. Participants were in a high- or low-risk group. Signaling systems varied by reliability (90%, 60%) within subjects and error bias (false alarm prone, miss prone) between subjects.

Results:

False-alarm rate affected compliance but not reliance. Miss rate affected reliance but not compliance. Mediation analyses indicated that trust mediated the relationship between false-alarm rate and compliance. Bayesian mediation analyses favored evidence indicating trust did not mediate miss rate and reliance. Conditional indirect effects indicated that factors of trust mediated the relationship between false-alarm rate and compliance (i.e., purpose) and reliance (i.e., process) but only in the high-risk group.

Conclusion:

The compliance–reliance paradigm is not the reflection of two types of trust.

Application:

This research could be used to update training and design recommendations that are based upon the assumption that trust causes operator responses regardless of error bias.

Keywords

warning systems human–automation interaction automation decision making dual task time sharing task switching

Introduction

The complexity and ubiquity of automation have grown, which has relegated the human to a monitor of automated systems. To help the human manage numerous complex systems, sensor-based signaling systems have also flourished in many domains, for example, aviation (Pritchett, Vándor, & Edwards, 2002), hospitals (Cvach, 2012), and ground transportation (Lees & Lee, 2007). Yet, because signaling systems are not always reliable, humans do not always depend upon associated signals. One factor that guides operator dependence is trust (Bliss, Gilson, & Deaton, 1995). Research suggests system error type (i.e., false alarm or miss) leads to two independent types of trust that induce two categorically different responses: compliance and reliance (Meyer, 2001; Rice, 2009). In the current work, we address the role of trust in the compliance–reliance paradigm.

Compliance–Reliance

The operator responding when a signal is issued is referred to as compliance. The operator refraining from a response when the system is silent, or indicating normal operation, is referred to as reliance. Together, compliance and reliance are referred to as dependence. Meyer’s (2001, 2004) initial work implies that compliance should be affected when a signal is issued and therefore degraded by the error associated with an issued signal (i.e., false alarm); reliance should be affected when the system indicates normal operation and therefore degraded by the error associated with no signal (i.e., miss).

Rice (2009) suggested that an extreme version of Meyer’s (2001) conceptualization takes the form of Figure 1a: Signal trust mediates the false alarm–compliance relationship and trust in nonsignals mediates the miss–reliance relationship independently (i.e., two types of trust). Supporting this theory, Dixon and Wickens (2006) reported a false-alarm-prone (FP) system degraded compliance, whereas a miss-prone (MP) system degraded reliance. Moreover, the MP system did not affect compliance. However, the FP system affected both compliance and, to a lesser degree, reliance. Based on these results, it is unclear whether false alarms and misses exclusively affect two types of trust that lead to independent responses (i.e., Figure 1a) or if trust in signals affects both compliance and reliance (i.e., Figure 1b).

Figure 1.

Signaling system errors on dependence: (a) selective two-process model and (b) Mandler’s two-process model. Adapted from Rice (2009).

Supporting evidence for Figure 1b, Dixon, Wickens, and McCarley (2007) reported an FP system affected both compliance and reliance. Rice (2009) proposed that the nonselective effect of false alarms on both responses implies trust in signals also affects reliance, whereas trust in nonsignals affects only reliance. Dixon et al. (2007) suggested that false alarms may affect both responses because false alarms are accompanied by perceptually salient events (e.g., alarms) and are more noticeable errors than misses (p. 571). In an investigation of this saliency hypothesis, although Rice and McCarley (2011) matched misses and false alarms for perceptual salience, the FP systems still led to lower dependence than the MP systems. The authors suggested that false alarms might also be more cognitively salient. In a second experiment, results indicated that when false alarms were framed as neutral messages, the asymmetry was reduced.

Rice (2009), however, described a study (Rice & McCarley, 2008) in which misses, in addition to false alarms, affected both compliance and reliance. Rice suggested that these results indicate that either there is a singular trust affecting both responses (Figure 2a) or there are two types of trust that affect both compliance and reliance (Figure 2b). Rice’s analysis indicated two forms of trust (i.e., Figure 2b). Specifically, false alarms strongly affect compliance and weakly affect reliance. Alternatively, misses strongly affect reliance and weakly affect compliance.

Figure 2.

Signaling system errors on dependence: (a) single-process model and (b) nonselective two-process model. Adapted from Rice (2009).

Although these studies show that false alarms and misses affect compliance and reliance differentially, the two types of trust are inferred from responses alone. By assuming that two types of trust are the sole determinants of these behaviors, researchers could be oversimplifying the processes. Alternative explanations may be that, depending upon error bias, a singular trust is formed in qualitatively different ways or even absent. To investigate these alternative explanations, a theoretical discussion linking trust to the compliance–reliance paradigm is necessary.

Trust and the Compliance–Reliance Paradigm

Lee and See (2004) provide arguably one of the most comprehensive and integrative perspectives of human–automation trust, highlighting components critical for trust development. A common theme among trust models is vulnerability, where the trustor willingly assumes risk by delegating responsibility to the trustee (cf. Mayer, Davis, & Schoorman, 1995). This responsibility implies that the trustee is advancing the goal of the trustor. Additionally, trust is conceptualized as an attitude, or an affective evaluation of a belief. Beliefs are the informational bases influenced by experience and information availability. Lee and See propose bases of human–automation trust: Performance reflects observable behaviors and describes what the automation does, according to current and historical operation; process describes how the automation operates, reflecting the appropriateness of the automation’s algorithms for achieving the operator’s goals; purpose describes why the automation was developed, reflecting the match with the designer’s intent, and is not based on observed behaviors.

An operator’s affective evaluations of beliefs about automation form the basis for adopting a particular level of trust (Lee & See, 2004). This level of trust leads a person to adopt an intention that leads to a behavior. Lee and See (2004) suggested that considering trust as a behavior or intention could confound its effects with other variables (e.g., workload, self-confidence). Here, there is a clear distinction between trust as an attitude and its behavioral effects (compliance and reliance). Consequently, Lee and See (2004) define trust as “an attitude that an agent will help achieve an individual’s goals in a situation characterized by uncertainty and vulnerability” (p. 54).

Unreliable systems should degrade trust, leading to reduced dependence (Chancey, Bliss, Proaps, & Madhavan, 2015). However, the stability of trust depends upon the degree to which the goal-oriented information (i.e., performance, process, purpose) about the automation provides the basis to form that trust. The availability of these bases determines the appropriateness of trust, affecting behavioral responses. The way in which this information is conveyed, however, depends on its error characteristics.

Salient choice

FP systems are qualitatively different from MP systems in relation to error saliency (Rice & McCarley, 2011). False alarms present a salient, explicit choice to comply or not. By comparison, misses offer a nonsalient, implicit choice to rely or not. The absence of a cue represents the key compliance–reliance distinction. Intervening without a signal requires an operator to notice cue absence, a difficult task (Hearst, 1991).

One characteristic of trust that distinguishes it from other constructs, such as confidence, is that the individual must choose one action in preference to another (Luhmann, 1988; Mayer et al., 1995). Lee and See (2004) proposed that trust develops from observing automation performance, which provides evidence for trust to then reflect greater attributional abstraction (i.e., process, purpose). Thus, false alarms are more likely to act causally through operator trust because of their salience.

Risk

Additionally, if operators do not feel risk when ignoring signals, trust has less influence upon behaviors. Here, risk is a characteristic of decisions that is defined as “the extent to which there is uncertainty about whether potentially significant and/or disappointing outcomes of decisions will be realized” (Sitkin & Pablo, 1992, p. 10). Mayer et al. (1995) argued that risk is an essential component of trust, where trust was characterized as a willingness to be vulnerable to another party. Simply having a willingness to be vulnerable, however, does not require an individual to take on any risk. Risk is integral in the behavioral demonstration of trust (Lyons & Stokes, 2012; see Figure 3).

Figure 3.

Model of human–automation trust on compliance/reliance behaviors according to Lee and See’s (2004) bases of trust. Adapted from Mayer, Davis, and Schoorman (1995).

Conditions for dependence

Two key conditions allow trust to mediate the relationship between system errors and dependence: First, the operator faces a salient choice to depend on the system. Second, the operator recognizes the risk of nondependence. Following these conditions, one of two outcomes is possible: (a) If system trust is higher than the perceived risk of nondependence, then the operator will depend upon the system, or (b) if system trust is lower than the perceived risk of nondependence, then the operator will not depend upon the system.

An example of the first outcome would be a pilot who engages in a violent course correction following a collision-avoidance alarm. The pilot, in this case, trusts the alarm (Condition 1) more than the risk associated with not complying with it, potentially leading to a midair collision (Condition 2). Moreover, by taking evasive action, the pilot receives performance feedback. Trust should then grow if the pilot can confirm loss of separation (i.e., hit) or decline if the loss of separation cannot be confirmed (i.e., false alarm).

If the operator does not recognize a choice (Condition 1) or, therefore, the risk of deviating from the current state (Condition 2), trust is less likely to act as a causative factor. Because misses involve no discernable signal, there is no salient choice and the status quo should continue. Although trust could mediate the miss–reliance relationship, the challenge is for the operator to recognize that it is missing signal events (i.e., performance). Indeed, misses may severely degrade trust when the system misses events that are easily detectable by the operator (Madhavan, Wiegmann, & Lacson, 2006). Misses, in that case, would be more salient and trust should play a larger role in determining reliance.

Even when presented with a salient choice (Condition 1), if the operator risks nothing by depending upon the system (Condition 2), then trust is also less relevant. Consequently, trust can vary but not determine behaviors (Lyons & Stokes, 2012). The purpose of this work is to investigate the proposed conditions for trust mediation as they relate to the compliance–reliance paradigm.

Hypotheses

Hypothesis 1: Trust will mediate the relationship between reliability and compliance for the FP systems (Chancey, Bliss, Liechty, & Proaps, 2015; Lee & See, 2004; Rice, 2009; Figure 4).

Hypothesis 2: Risk will moderate the mediating effect of trust (Mayer et al., 1995; Parasuraman & Riley, 1997; Figure 4).

Hypothesis 3: Trust will not mediate the relationship between reliability and reliance for the MP systems (Chancey, Bliss, Liechty, et al., 2015; Figure 4).

Hypothesis 4: Higher reliability will lead to higher trust (Chancey, Bliss, Proaps, et al., 2015; Lee & See, 2004).

Hypothesis 5: An interaction will indicate false alarms affect compliance but not reliance (Chancey, Bliss, Liechty, et al., 2015; Meyer, 2001).

Hypothesis 6: An interaction will indicate misses affect reliance but not compliance (Chancey, Bliss, Liechty, et al., 2015; Meyer, 2001).

Hypothesis 7: Participants in the high-risk group will report higher perceived risk than those in the low-risk group.

Figure 4.

Hypothesized model reflecting signaling system errors on dependence (similar to the model in Figure 1a, with the addition of risk moderating the effect of trust; i.e., model in Figure 3). Solid arrows indicate strong effect, whereas dashed arrows indicate weak effect.

Method

Participants

Eighty-eight undergraduate students (56 females; mean age = 19.28 years, SD = 2.13 years) participated for research credit. All participants reported having normal (or corrected-to-normal) visual acuity at the time of participation. No participant indicated color deficiency or hearing impairment. This research complied with the American Psychological Association Code of Ethics and was approved by the Institutional Review Board at Old Dominion University. Informed consent was obtained from each participant.

Experimental Tasks

Experimental tasks were housed on two desktop computers with 12-inch monitors. RadioShack® PRO-100 Communications Headset headphones presented auditory alarms.

Primary tasks

The Multi-Attribute Task Battery (MATB II) is a set of programmable tasks that simulate pilot responsibilities (Santiago-Espada, Myer, Latorella, & Comstock, 2011). Participants interacted with the compensatory tracking task, which simulates maintaining stable flight, and the resource management task, which simulates maintaining adequate fuel levels. For the compensatory tracking task, participants used a joystick to keep a continuously drifting reticle at the center of a pair of crosshairs. For the resource management task, participants used a 10-key number pad to transfer fuel among continuously depleting tanks.

Secondary task

Participants viewed aerial pictures for tanks. Participants were exposed to the same 30 pictures with and without an imbedded tank (60 pictures). Participants could use the “tank-spotting aid,” which signaled the presence of a tank in one of four quadrants of the image by surrounding it in red. Alarms were accompanied by a tone that increased in frequency from 700 to 1700 Hz in 0.85 s with an interruption interval of 0.12 s. If the aid diagnosed the absence of a tank, no alarm occurred. A point bank was provided in the task window to indicate performance, where a point was added or subtracted for correct or incorrect responses respectively (Figure 5).

Figure 5.

Signaling system task graphical depiction. (1) Delay between images randomized at 10, 14, and 18 s. (2) Aerial image presented for 3 s with visual and auditory alarm. (3) Screen presented after aerial image indicating the tank-spotting aid’s diagnosis and requesting the participant’s response.

Design

We employed a 2 (error bias: FP, MP) × 2 (reliability: 90%, 60%) × 2 (risk: high, low) split-plot design, where all variables were fixed effects. Signaling system error bias was manipulated between subjects, and reliability was manipulated within subjects. Reliability indicated the percentage of images the signaling system correctly indicated tank present or absent (see Table 1). Participants were told the 90%-reliable aid “tends to be pretty reliable, so it probably won’t make a lot of mistakes” and the 60%-reliable aid “tends to be pretty unreliable, so it probably will make a lot of mistakes.” Risk was manipulated between subjects. We instructed the high-risk group that poor task performance would cause additional noncompensated time in the experiment, yet this consequence was not enforced. We gave no instructions on the consequences of task performance to the low-risk group.

Table 1:

Detection Response Matrix for the False-Alarm-Prone (FP) and Miss-Prone (MP) Systems According to Reliability for the Signaling System Task

Detection Response	90% FP	60% FP	90% MP	60% MP
Hits	30 (.50)	30 (.50)	24 (.40)	6 (.10)
False alarms	6 (.10)	24 (.40)	0 (.00)	0 (.00)
Misses	0 (.00)	0 (.00)	6 (.10)	24 (.40)
Correct rejections	24 (.40)	6 (.10)	30 (.50)	30 (.50)

Note. Numbers outside of parentheses represent the raw number of responses per category that will occur during the session. Numbers in parentheses represent the proportions of responses out of the total number of responses during each session.

We used three modified factors from the Human–Computer Trust Questionnaire (Madsen & Gregor, 2000; see appendix). The questionnaire showed adequate internal consistency for trust (α = .97) and its factors: performance (α = .96), process (α = .91), and purpose (α = .93). We used a modified version of the risk questionnaire by Simon, Houghton, and Aquino (1999) to measure perceived risk (α = .85).

Compliance was the number of tank found responses when the system issued an alarm, out of the total number of alarms. Reliance was the number of no tank responses when the system remained silent, out of the total number of times the system remained silent. The performance metrics for the primary tasks were the root mean square deviation of a reticle from a center target for the tracking task and the average fuel level deviation from a prespecified amount for the resource management task.

Mediation and Moderated Mediation Analyses

To test for mediation, we used ordinary least squares path analysis and consulted bias-corrected bootstrap confidence intervals (CI) for the indirect effects of trust based on 10,000 bootstrap samples (Hayes, 2013; Preacher & Hayes, 2004, 2008). Preacher and Hayes (2008) recommend the use of bootstrapping over other methods (e.g., Sobel test, causal-steps approach) on the grounds that this approach has higher power while maintaining reasonable control over Type I error rate (p. 880). Moreover, an indirect effect does not require a significant total effect (Hayes, 2013). For example, if a model has two mediators working in opposite directions, this could result in two significant indirect effects in the absence of a total effect. Therefore, based on the pervasive theoretical presumption that trust mediates the relationships between error characteristics and dependence, we interpret significant indirect effects as evidence of mediation. For the simple mediation analyses, we accounted for the within-participant manipulation of reliability by using the technique outlined by Montoya and Hayes (2016).

To examine if risk modifies the degree to which trust mediates the tested relationships, we consulted conditional indirect effects and indices of moderated mediation. Preacher, Rucker, and Hayes (2007) define a conditional indirect effect as “the magnitude of an indirect effect at a particular value of a moderator” (p. 186). Alternatively, Hayes (2015) proposes that moderated mediation can be tested by whether the moderator has a nonzero weight in the function linking the indirect effect to values of the moderator, called the index of moderated mediation (p. 7). It should be noted, however, we did not account for the within-subjects reliability manipulation, as we were able to do with the simple mediation analyses. Therefore, it is possible that we violated the assumption of independence and our standard errors are underestimated, thus causing the CIs to be too narrow.

Finally, to test Hypothesis 3, we consulted Bayes factors (BFs), commonly labled as BF₁₀, as a measure for the data supporting evidence for either a model where trust mediates the relationship between miss rate and reliance (full model) or a model that indicates no mediating effect of trust (reduced model). A BF₁₀ greater than 1 indicates evidence for the full model, and a BF₁₀ less than 1 indicates evidence for the reduced model. We used the R package BayesMed, which provides BFs for indirect effects (Nuijten, Wetzels, Matzke, Dolan, & Wagenmakers, 2015). To describe the strength of effects, we adopt the terms from Nuijten et al. (2015): anecdotal, moderate, strong, very strong, and extreme.

Procedure

Participants first completed an informed-consent and demographics form. Participants received instructions indicating either “high risk” or “low risk” associated with poor performance (randomized). Following instructions, participants practiced the primary tasks and then searched through 10 images containing a tank. Participants were required to find the tank in each image. Participants then completed the perceived-risk questionnaire and a 10-min practice session (system 100% reliable).

Following practice, participants completed two 20-min experimental sessions, where the system varied in reliability (90%, 60%; counterbalanced). The system was either MP or FP (randomized). We measured trust halfway through each session and analyzed only the responses obtained after the questionnaire. The proportions of system errors were equal for pre- and postquestionnaire administration. Participants were then debriefed and received research credit.

Results

Compliance, reliance, and questionnaire data were transformed to scale from 0 (min) to 1 (max). All data were inspected for normality and that no data were missing. We applied Levene’s tests to inspect for homogeneity of variance. We used multiple 2 (90%, 60%) × 2 (FP, MP) × 2 (high risk, low risk) split-plot analyses of variance (ANOVAs) to test main effects and interactions among dependent measures. The moderated-mediation and mediation analyses employed the heteroscedasticity-consistent HC3 standard error estimator to address the assumption of homoscedasticity (Hayes & Cai, 2007). We applied an outlier-labeling rule (multiplier 2.2) to identify outliers among all measures (Hoaglin & Iglewicz, 1987). We identified two lower-limit outliers for reliance (reliance < .04), where reliance was 0 for both. These data were adjusted to .32 to be .01 below the identified next-lowest value of .33. We established alpha level p < .05 to indicate statistical significance. Nonhypothesized interactions were interpreted by alpha-corrected simple effects.

Main Effects and Interactions

Primary task performance

There was a significant main effect of risk on tracking task performance, F(1, 84) = 10.42, p = .002, partial η² = .11. The high-risk group (M = 39.89, 95% CI [37.16, 42.62]) kept the reticle more stable than the low-risk group (M = 46.16, 95% CI [43.43, 48.89]). No other significant main effects or interactions occurred.

Perceived risk

Supporting Hypothesis 7, there was a significant main effect of risk on perceived risk, F(1, 84) = 12.46, p = .001, partial η² = .13, where the high-risk group (M = .58, 95% CI [.52, .65]) assigned higher perceived risk than the low-risk group (M = .43, 95% CI [.36, .49]). No other significant main effects or interactions occurred.

Trust

Supporting Hypothesis 4, there was a significant main effect of reliability on trust, F(1, 84) = 185.79, p < .001, partial η² = .69, where the 90% condition (M = .72, 95% CI [.69, .76]) rated the signaling system more trustworthy than the 60% condition (M = .47, 95% CI [.43, .50]). No other significant main effects or interactions occurred.

Performance

There was a significant main effect of reliability on the performance factor of trust, F(1, 84) = 176.14, p < .001, partial η² = .68, where participants in the 90% condition (M = .70, 95% CI [.66, .74]) rated the performance factor higher than those in the 60% condition (M = .39, 95% CI [.36, .44]). There was also a significant main effect of error bias, F(1, 84) = 4.09, p = .046, partial η² = .05, where those in the FP group (M = .58, 95% CI [.54, .63]) rated the performance factor higher than the MP group (M = .52, 95% CI [.47, .56]). No other significant main effects or interactions occurred.

Process

There was a significant main effect of reliability on the process factor of trust, F(1, 84) = 103.69, p < .001, partial η² = .55, where participants in the 90% condition (M = .78, 95% CI [.74, .81]) rated the process factor higher than those in the 60% condition (M = .58, 95% CI [.54, .63]). No other significant main effects or interactions occurred.

Purpose

There was a significant main effect of reliability on the purpose factor of trust, F(1, 84) = 195.87, p < .001, partial η² = .70, where participants in the 90% condition (M = .69, 95% CI [.66, .73]) rated the purpose factor higher than those in the 60% condition (M = .47, 95% CI [.38, .46]). No other significant main effects or interactions occurred.

Compliance

Supporting Hypothesis 5, there was a significant interaction between reliability and error bias on compliance, F(1, 84) = 77.45, p < .001, partial η² = .48. Simple effects indicated an effect of reliability on compliance but only for the FP group, Wilks’ λ = .346, F(1, 84) = 158.94, p < .001, partial η² = .65 (Figure 6). No other significant main effects or interactions occurred.

Figure 6.

Mean compliance as a function of reliability and error bias. Error bars represent 95% within-subject confidence intervals based on the main effect of reliability (Loftus & Masson, 1994).

Reliance

Supporting Hypothesis 6, there was a significant interaction between reliability and error bias, F(1, 84) = 15.93, p < .001, partial η² = .15. Simple effects indicated an effect of reliability on reliance but only for the MP group, Wilks’ λ = .58, F(1, 84) = 60.18, p < .001, partial η² = .42. The effect of false-alarm rate on reliance approached significance, Wilks’ λ = .95, F(1, 84) = 4.88, p = .03 (greater than the alpha-corrected .025), partial η² = .06 (Figure 7). No other significant main effects or interactions occurred.

Figure 7.

Mean reliance as a function of reliability and error bias. Error bars represent 95% within-subject confidence intervals based on the main effect of reliability (Loftus & Masson, 1994).

FP Systems Mediation Analyses

Trust

Supporting Hypothesis 1, simple mediation analyses indicated that trust mediated the relationship between reliability and compliance, indirect effect = .07, 95% CI [.02, .15], but not between reliability and reliance. No other significant effects were observed.

Performance

Simple mediation analyses indicated that the performance factor mediated the relationship between reliability and compliance, indirect effect = .07, 95% CI [.03, .12], but not between reliability and reliance. No other significant effects were observed.

Process

Although not hypothesized, moderated mediation showed a significant conditional indirect effect between reliability and reliance for the high-risk group, indirect effect = .06, 95% CI [.01, .14], but not in the low-risk group, indirect effect = –.06, 95% CI [–.05, .22]. However, these conditional indirect effects were not significantly different from each other, index of moderated mediation = .01, 95% CI [–.012, .16]. No other significant effects were observed.

Purpose

Simple mediation analyses indicated that the purpose factor mediated the relationship between reliability and compliance, indirect effect = .07, 95% CI [.02, .13], but not between reliability and reliance. Moderated mediation analyses indicated that the purpose factor mediated the relationship between reliability and compliance but only for the high-risk group. Partially supporting Hypothesis 2, a significant conditional indirect effect occurred in the high-risk group, indirect effect = .09, 95% CI [.002, .17], but not in the low-risk group, indirect effect = .04, 95% CI [–.08, .17] (Preacher et al., 2007; see Figure 8). However, according to Hayes (2015), Hypothesis 2 was not supported because these conditional indirect effects were not significantly different from each other, index of moderated mediation = .04, 95% CI [–.10, .19]. No other significant effects were observed.

Figure 8.

Path coefficients for the moderated mediation model depicting the relationship between reliability, purpose, risk, and compliance.

MP Systems Mediation Analyses

Supporting Hypothesis 3, the data showed anecdotal evidence indicating trust did not mediate the relationship between reliability and reliance for the MP system (BF₁₀ = .44). Moreover, the data showed moderate evidence that the performance factor did not mediate miss rate and reliance (BF₁₀ = .27) and anecdotal evidence that neither the process factor (BF₁₀ = .43) nor the purpose factor (BF₁₀ = .35) mediated miss rate and reliance. According to simple mediation and moderated mediation analyses, neither trust nor any of its individual factors mediated the relationships between miss rate and compliance or reliance.

Discussion

Similar to other compliance–reliance studies, predicted main effects and interactions were observed (Dixon & Wickens, 2006; Meyer, 2001). Yet, supporting our hypotheses, trust mediated the FP–compliance relationships, as indicated by the significant indirect effect, whereas the Bayesian mediation analyses provided evidence suggesting trust did not mediate the MP–reliance relationship. For the MP system, although trust was related to reliance (i.e., main effects of miss rate on trust and reliance), the results indicated that the effect on trust was a by-product of miss rate rather than a causal mechanism affecting reliance (cf. Bustamante, 2009).

Theoretical Implications

If (as shown here) trust mediates the FP–compliance relationship but not the MP–reliance relationship, Meyer’s (2001) initial proposal of two independent cognitive processes linking each relationship is accurate. However, it may be more correct to conceptualize trust as strongly mediating one relationship and weakly mediating the other. From this perspective, there are not two types of trust per se, but instead, trust links false alarms and compliance, whereas other constructs may link misses and reliance.

FP–dependence relationship

The faulty behavior associated with false alarms was expected to act causally through operator trust because of their salience. Trust, as well as the trust factors of performance and purpose, mediated the relationship between false-alarm rate and compliance. Moreover, the moderated mediation analysis indicated that the process basis of trust mediated the relationship between false-alarm rate and reliance but only for the high-risk group. Therefore, for FP systems, compliance and reliance likely manifest from different bases of trust (i.e., two manifestations of the same construct).

Risk of dependence

It is unclear from our results whether risk moderated the mediating effect of trust on compliance and reliance (i.e., conditional indirect effects versus index of moderated mediation results). Regardless, risk is a critical element of most human–automation and interpersonal theories of trust (cf. Lyons, Stokes, Eschleman, Alacron, & Barelka, 2011). Because the sample participants in the current study were college students, additional time investment was suspected to be personally negative to participants. As expected, this risk manipulation did lead to conditionally significant indirect effects of some trust factors according to risk group assignment. Yet if automation fails in operational settings, the human faces real and sometimes detrimental consequences (Sheridan & Parasuraman, 2005). In these high-risk real-world situations, trust in automation likely plays a much more important role in determining automation dependence.

MP–reliance relationship

Trust did not mediate the MP–reliance relationship (cf. Chancey, Bliss, Liechty, et al., 2015). Theoretically, other constructs might be better candidates. For example, state-level suspicion may offer an alternative explanation to describe the MP–reliance relationship (Bobko, Barelka, & Hirshfield, 2014; Lyons et al., 2011). Missing information is a component associated with increased suspicion, which can degrade performance as operators experience greater workload when searching for information. In fact, Dixon and Wickens (2006) reported increased workload in participants, who divided attention among tasks to offset errors produced by an MP system. However, further research is needed to substantiate any potential mediating role of suspicion in MP–reliance relationships.

Another candidate construct is confidence (Chancey, Bliss, Liechty, et al., 2015). Luhman (1988) proposed that trust requires one choice in preference to another, but confidence does not. Systems that are MP do not offer salient choices to deviate from the status quo. Smith (2005) proposed that confidence allows people to “bracket” life’s contingencies so they can avoid continuous uncertainty and anxiety (p. 307). This scenario could be easily used to describe the concept of complacency, where individuals assume the “expert system” will work properly and engage in other activities without worrying the system will make an undetected error.

Trust requires the acceptance and acknowledgement of risk, whereas confidence does not (Luhman, 1988; Mayer et al., 1995). In the case of misses, if an operator fails to acknowledge risk, confidence is at work.

Conclusion

The results from the current work suggest that false alarms have a stronger impact upon trust than misses. Commonly, sensor thresholds are set to minimize the chance of missing an abnormal event (Sorkin & Woods, 1985). One reason for adopting this setting is legalistic policies associated with manufacturers’ “obligation to warn” (Bliss & Gilson, 1998). Moreover, the costs associated with a signaling system’s missing a critical event are potentially disastrous. Yet overly liberal thresholds, which generate frequent false alarms, will likely adversely affect trust in the long run. The consequence is that either operators will not use the FP system, indicating the system represents a waste of resource allocation, or the operator will not comply with the system when it issues a true alarm. Although the conclusion that false alarms affect trust and subsequent compliance is not a novel concept (i.e., cry-wolf effect; Bliss et al., 1995), the idea that misses might not have an equal impact upon trust is.

These findings are particularly relevant to physiological-based alarming thresholds (e.g., pulse oximetry), in which some hospitals have altered criteria to be more conservative (i.e., MP) to combat excessive false alarms (Whalen et al., 2014). Similarly, the error bias of in-vehicle collision-avoidance systems depends upon driving style, where drivers with shorter headways experience more misses and drivers with longer headways experience more false alarms (Lees, 2010, p. 38; Ben-Yaacov, Maltz, & Shinar, 2002; Maltz & Shinar, 2004). From this perspective, systems that are more likely to either false alarm or miss should be studied differently, as not only are the behavioral consequences different (i.e., compliance or reliance), but so too are the psychological mechanisms driving those behaviors.

Key Points

The compliance–reliance paradigm was evaluated in the context of Lee and See’s (2004) human–automation trust model. Trust was conceptualized as an affective evaluation of beliefs and measured in terms of representing trust bases of performance, process, and purpose.

False alarms affected compliance but not reliance. Alternatively, misses affected reliance but not compliance.

Trust mediated the relationship between false-alarm rate and compliance but not the relationship between miss rate and reliance.

This work indicates that there are not two forms of trust in the compliance–reliance paradigm. Instead, trust likely plays a stronger mediating role in false alarm–dependence relationships and a weaker role in miss–dependence relationships.

Footnotes

Appendix

Acknowledgements

This manuscript was greatly improved by the comments provided by two anonymous reviewers. The authors received no financial support for the research, authorship, and/or publication of this article.

Eric T. Chancey is a senior behavioral scientist in the Multi-Sensor Exploitation and Countermeasures Division at Leidos. He earned his PhD in human factors psychology from Old Dominion University in 2016.

James P. Bliss is a professor of the Psychology Department at Old Dominion University. He earned his PhD in human factors psychology from the University of Central Florida in 1993.

Yusuke Yamani is an assistant professor of the Department of Psychology at Old Dominion University. He received his PhD in psychology (visual cognition and human performance) from the University of Illinois at Urbana-Champaign in 2013.

Holly A. H. Handley is an associate professor in the Engineering Management and Systems Engineering Department at Old Dominion University. She earned her PhD in information technology and engineering from George Mason University in 1999.

References

Ben-Yaacov

Maltz

Shinar

(2002). Effects of an in-vehicle collision avoidance warning system on short- and long-term driving performance. Human Factors, 44, 335–342.

Bliss

J. P.

Gilson

R. D.

(1998). Emergency signal failure: Implications and recommendations. Ergonomics, 41, 57–72.

Bliss

J. P.

Gilson

R. D.

Deaton

J. E.

(1995). Human probability matching behaviour in response to alarms of varying reliability. Ergonomics, 38, 2300–3212.

Bobko

Barelka

A. J.

Hirshfield

L. M.

(2014). The construct of state-level suspicion: A model and research agenda for automated and information technology (IT) contexts. Human Factors, 56, 489–508.

Bustamante

E. A.

(2009). A reexamination of the mediating effect of trust among alarm systems’ characteristics and human compliance and reliance. In Proceedings of the Human Factors and Ergonomics Society 53rd Annual Meeting (pp. 249–253). Santa Monica, CA: Human Factors and Ergonomics Society.

Chancey

E. T.

Bliss

J. P.

Liechty

Proaps

A. B.

(2015). False alarms vs. misses: Subjective trust as a mediator between reliability and alarm reaction measures. In Proceedings of the Human Factors and Ergonomics Society 58th Annual Meeting (pp. 647–651). Santa Monica, CA: Human Factors and Ergonomics Society.

Chancey

E. T.

Bliss

J. P.

Proaps

A. B.

Madhavan

(2015). The role of trust as a mediator between system characteristics and response behaviors. Human Factors, 57, 947–958.

Cvach

(2012). Monitor alarm fatigue: An integrative review. Biomedical Instrumentation & Technology, 46, 268–277.

Dixon

S. R.

Wickens

C. D.

(2006). Automation reliability in unmanned aerial vehicle control: A reliance-compliance model of automation dependence in high workload. Human Factors, 48, 474–486.

10.

Dixon

S. R.

Wickens

C. D.

McCarley

J. S.

(2007). On the independence of compliance and reliance: Are automation false alarms worse than misses? Human Factors, 49, 564–572.

11.

Hayes

A. F.

(2013). Introduction to mediation, moderation, and conditional process analysis: A regression-based approach. New York, NY: Guilford.

12.

Hayes

A. F.

(2015). An index and test of linear moderated mediation. Multivariate Behavioral Research, 50, 1–22.

13.

Hayes

A. F.

Cai

(2007). Using heteroskedasticity-consistent standard error estimators in OLS regression: An introduction and software implementation. Behavioral Research Methods, 39, 709–722.

14.

Hearst

(1991). Psychology and nothing. American Scientist, 79, 432–443.

15.

Hoaglin

D. C.

Iglewicz

(1987). Fine tuning some resistant rules for outlier labeling. Journal of American Statistical Association, 82, 1147–1149.

16.

Lee

J. D.

See

K. A.

(2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46, 50–80.

17.

Lees

(2010). Context conditions drivers’ disposition towards alarms (Doctoral dissertation). Retrieved from Iowa Research Online: http://ir.uiowa.edu/etd/845

18.

Lees

M. N.

Lee

J. D.

(2007). The influence of distraction and driving context on driver response to imperfect collision warning systems. Ergonomics, 50, 1264–1286.

19.

Loftus

G. R.

Masson

M. E. J.

(1994). Using confidence intervals in within-subjects designs. Psychonomic Bulletin & Review, 1, 476–490.

20.

Luhmann

(1988). Familiarity, confidence, trust: Problems and alternatives. In Gambetta

D. G.

(Ed.), Trust: Making and breaking cooperative relations (pp. 94–107). New York, NY: Basil Blackwell.

21.

Lyons

J. B.

Stokes

C. K.

(2012). Human–human reliance in the context of automation. Human Factors, 54, 112–121.

22.

Lyons

J. B.

Stokes

C. K.

Eschleman

K. J.

Alacron

G. M.

Barelka

A. J.

(2011). Trustworthiness and IT suspicion: An evaluation of the nomological network. Human Factors, 53, 219–229.

23.

Madhavan

Wiegmann

D. A.

Lacson

F. C.

(2006). Automation failures on tasks easily performed by operators undermine trust in automated aids. Human Factors, 48, 241–256.

24.

Madsen

Gregor

(2000, November/December). Measuring human–computer trust. Paper presented at the Eleventh Australasian Conference on Information Systems, Sydney, Australia.

25.

Maltz

Shinar

(2004). Imperfect in-vehicle collision avoidance warning systems can aid drivers. Human Factors, 46, 357–366.

26.

Mayer

R. C.

Davis

J. H.

Schoorman

F. D.

(1995). An integrative model of organizational trust. Academy of Management Review, 20 (3), 709-734.

27.

Meyer

(2001). Effects of warning validity and proximity on responses to warnings. Human Factors, 43, 563–572.

28.

Meyer

(2004). Conceptual issues in the study of dynamic hazard warnings. Human Factors, 46, 196–204.

29.

Montoya

A. K.

Hayes

A. F.

(2016, June 30). Two-condition within-participant statistical mediation analysis: A path-analytic framework. Psychological Methods. Advance online publication. https://dx-doi-org.web.bisu.edu.cn/10.1037/met0000086

30.

Nuijten

M. B.

Wetzels

Matzke

Dolan

C. V.

Wagenmakers

(2015). A default Bayesian hypothesis test for mediation. Behavioral Research Methods, 47, 85–97.

31.

Parasuraman

Riley

(1997). Humans and automation: Use, misuse, disuse, abuse. Human Factors, 39, 230–253.

32.

Preacher

K. J.

Hayes

A. F.

(2004). SPSS and SAS procedures for estimating indirect effects in simple mediation models. Behavior Research Methods, Instruments, & Computers, 36, 717–731.

33.

Preacher

K. J.

Hayes

A. F.

(2008). Asymptotic and resampling strategies for assessing and comparing indirect effects in multiple mediator models. Behavior Research Methods, 40, 879–891.

34.

Preacher

K. J.

Rucker

D. D.

Hayes

A. F.

(2007). Addressing moderated mediation hypotheses: Theory, methods, and prescriptions. Multivariate Behavioral Research, 42, 185–227.

35.

Pritchett

A. R.

Vándor

Edwards

(2002). Testing and implementing cockpit alerting systems. Reliability Engineering and System Safety, 75, 193–206.

36.

Rice

(2009). Examining single- and multiple-process theories of trust in automation. Journal of General Psychology, 13, 303–319.

37.

Rice

McCarley

(2008). The effects of automation bias and saliency on operator trust. International Congress of Psychology, 43, 747–747.

38.

Rice

McCarley

J. S.

(2011). Effects of response bias and judgment framing of operator use of an automated aid in a target detection task. Journal of Experimental Psychology: Applied, 17, 320–331.

39.

Santiago-Espada

Myer

R. R.

Latorella

K. A.

Comstock

J. R.

(2011). The Multi-Attribute Task Battery II (MATB-II) software for human performance and workload research: A user’s guide (NASA/TM-2011-217164). Hampton, VA: National Aeronautics and Space Administration, Langley Research Center.

40.

Sheridan

T. B.

Parasuraman

(2005). Human–automation interaction. Reviews of Human Factors and Ergonomics, 1, 89–129.

41.

Simon

Houghton

S. M.

Aquino

(1999). Cognitive biases, risk perception, and venture formation: How individuals decide to start companies. Journal of Business Venturing, 15, 113–134.

42.

Sitkin

S. B.

Pablo

A. M.

(1992). Reconceptualizing the determinants of risk behavior. Academy of Management Review, 17, 9–38.

43.

Smith

(2005). Understanding trust and confidence: Two paradigms and their significance for health and social care. Journal of Applied Philosophy, 22, 299–316.

44.

Sorkin

R. D.

Woods

D. D.

(1985). Systems with human monitors: A signal detection analysis. Human–Computer Interaction, 1, 49–75.

45.

Whalen

D. A.

Covelle

Piepenbrink

J. C.

Villanova

K. L.

Cuneo

Awtry

E. H.

(2014). Novel approach to cardiac alarm management on telemetry units. Journal of Cardiovascular Nursing, 29, E13–E22.