Fear Responses to Threat Appeals: Functional Form,Methodological Considerations,and Correspondence Between Static and Dynamic Data

Abstract

The study tested (a) the extent to which an inverted-U pattern of fear response predicted persuasion, (b) the degree to which the fear curve mediated the effects of the four components of threat appeals on persuasion, (c) the correspondence between the static measures of fear used in between-subjects designs and the dynamic indices required by the within-subject approach, and (d) the methodological threats inherent to dynamic designs. Participants (N = 418) read a message that advocated for colorectal cancer screening. Results showed that the inverted-U fear curve predicted intention to obtain a colonoscopy, and that susceptibility and response efficacy exerted their influence on persuasion via the fear curve while severity and self-efficacy did not. The static measure of fear showed poor absolute correspondence with the peak and end indices of dynamic fear, but strong pattern correspondence. Hazards to inference posed by dynamic designs of the type used in this study appear negligible.

Keywords

persuasion fear fear appeals threat appeals peak-and-end rule colorectal cancer

Studies of threat appeals typically have employed experimental designs in which research participants are exposed first to either a strong or weak threat appeal, then their agreement with the advocacy is measured. Investigations of this sort are plentiful. Collectively, they have provided the basis for several meta-analyses, most of which conclude that individuals who experience higher levels of fear are more likely to be persuaded by the message (e.g., Boster & Mongeau, 1984; Sutton, 1982; Witte & Allen, 2000; but see Peters, Ruiter, & Kok, 2013). In other words, the relationship between fear and persuasion appears to be linear when assessed across persons. Notably, this result is distinct from the prediction made by one of the earliest theories of threat appeals. Drive theory anticipated that persuasion was the product of messages that both induced fear and alleviated it (Hovland, Janis, & Kelley, 1953). Testing this prediction requires measuring fear at multiple points in time (Rossiter & Thornton, 2004; Shen & Dillard, 2014; Sutton, 1982), something that only a few studies have done. But, in every project in which fear was assessed before, during, and after the message, the form of the fear-persuasion relationship was curvilinear, not linear (Meczkowski, Dillard, & Shen, 2016; Shen & Dillard, 2014; Shen, Wang, & Seung, 2015).

Evidence of a curvilinear, within-persons effect raises significant questions about how to interpret previous research. If the true underlying process is one in which fear must be evoked and then assuaged, then what part of that process has been captured by between-person designs? One possibility is that individuals report only part of their emotional response, such as their highest or lowest feeling of fear. But, perhaps they give some average or weighted average of their overall experience? In either case, the between-subjects data might under- or overestimate the effectiveness of threat appeals. Mis-estimation could have serious ramifications for applied persuasion projects that depend on threat appeals to promote behavior change. Yet, at present, even the direction of the bias is unknown, a condition that makes post-hoc adjustments impossible.

We designed the current project to examine emotional responses to threat appeals. The first part of our effort focused on the question of functional form. That is, we sought to replicate the results of studies that show a curvilinear association between fear and persuasion when fear is assessed within persons. Our second motivation was to better understand the relationship between the earlier between-subjects studies and the current within-subjects study. Toward that end, we considered the correspondence between the static measures of fear that have been used in between-persons designs and the dynamic indices that define within-persons studies. Finally, we recognized that it would be impossible to conduct these comparisons without creating methodological hazards to causal inference. Thus, we designed a study that would allow for explicit examination of the most prominent threats to internal validity. To put all of these issues in context, we turn next to a closer consideration of research on threat appeals.

Content, Structure, and Terminology of Threat Appeals

Over the many years that threat appeals have been studied, consensus has emerged regarding the content of the messages (Mongeau, 2013; Rogers, 1983). There are two major parts to a threat appeal, each of which possesses two subcomponents. The threat element encompasses the severity of the issue—the degree to which the consequences are negative and undesirable—along with the audience member’s susceptibility—the likelihood that the consequences will befall the message recipient. The action component consists of response efficacy, which refers to the effectiveness of the recommendation in reducing or avoiding the threat, and self-efficacy, which indexes one’s perceived ability to perform the recommended behavior. In terms of arrangement, the threat component precedes the action element either by convention (Rogers, 1983) or by definition (Witte, 1994).

Linear and Curvilinear Effects of Fear on Persuasion

The drive or drive-reduction model supposes that humans are subject to unpleasant motivational states, such as hunger. These drive states energize trial-and-error behavior (Dollard & Miller, 1950). Any behavior that successfully reduces the motivational state is reinforced, and therefore likely to become the habitual response. The application of learning theory concepts to threat appeals begins with the idea that emotional states, including fear, have the properties of drives. From this starting point, Hovland et al. (1953) describe the conditions necessary for persuasion:

It is assumed that a threat appeal is mostly likely to induce an audience to accept the communicator’s conclusion if a) the emotional tension aroused during the communication is sufficiently intense to constitute a drive state; and b) silent rehearsal of the recommended belief or attitude is immediately followed by reduction of tension. (p. 62)

The excerpt makes clear the proposed mediating process and the central mechanism of drive theory. As shown in Panel A of Figure 1, a message will be persuasive to the extent that it first arouses emotional tension, then alleviates it. This is the original curvilinear or inverted-U hypothesis. To fully appreciate the prediction, it is helpful to focus on the axes of the graph. The X axis represents time with measures of fear from one person taken at three points: prior to message exposure (t₀), just following presentation of the threat component (t₁), and then again after exposure to the action component of the appeal (t₂). Panel B of Figure 1 displays a linear pattern between measurement occasion and fear response. Drive theory predicts no persuasion for the person in Panel B because fear is not reduced.

Figure 1.

The association between message content and fear response at three points in time.

Some later theories also proposed a curvilinear relationship between fear and persuasion. For example, Janis (1967) and McGuire (1968) developed frameworks that explained persuasion in terms of individuals’ fear responses. Both theories propose an optimal level of fear arousal for any given individual which, if matched with the corresponding level of threat, will maximize persuasion. Too little fear or too much will result in decreased agreement with the message. This is the prediction tested in many studies of fear appeals and in subsequent meta-analyses (Boster & Mongeau, 1984; Sutton, 1982; Witte & Allen, 2000). It is illustrated in Figure 2 and, again, close attention to the graph’s axes is instructive. In this case, the X axis is composed of scale values for multiple research participants. Each person has one value that describes his or her overall fear response to the message. In other words, where the X axis of Figure 1 shows the fear response of a single individual over time, Figure 2 shows the fear score for multiple individuals at a single point in time (i.e., after the message). The Y axis in Figure 2 is persuasion. This is distinct from Figure 1 where the Y axis represents fear intensity.

Figure 2.

The association between fear response and persuasion across persons at a single point in time.

Collectively, Figures 1 and 2 enable two conclusions. First, both linear and nonlinear effects are possible in between-persons studies or in within-persons studies. But, second, the associations displayed in Figures 1 and 2 are not interchangeable. Given appropriate data, one might infer that the fear-persuasion relationship is linear, not curvilinear, in between-persons investigations. Indeed, this is precisely the conclusion reached by several of the meta-analyses (Boster & Mongeau, 1984; Witte & Allen, 2000). It accurately describes the pattern of data in those papers. However, this conclusion has no bearing on the question of functional form of the fear-persuasion relationship in within-persons data. As comparison between Figures 1 and 2 makes plain, there are different variables involved in the two accounts. The upshot is that we know a good deal about the fear-persuasion relationship in between-persons data. However, very little is understood about the original drive theory prediction simply because it has not been tested.

Several conditions are necessary to an appropriate test of the inverted-U hypothesis. For one, the research design must be capable of generating a curve for each message recipient. Operationally, this means that the design must measure fear at three or more times. As suggested by Figure 1, a minimal design would assess fear prior to message exposure (t₀), again just following presentation of the threat component (t₁), and then again after exposure to the action component of the appeal (t₂). A second necessary condition is that at the level of means, the data array themselves as a curve of a particular sort: an inverted U. If the fear data do not curve, then it is impossible to test for the effects of a curve. Third, there must be some inter-individual variation in the person-specific curves. Different individuals must display different curves or the independent variable would have no variance. One indication that this condition has been satisfied is if the standard deviations at t₀, t₁, or t₂ are non-zero values. Finally, an analytic technique is needed that is capable of capturing the information in the curves and using it as a predictor of persuasion. Latent growth curve modeling is one such technique (Bollen & Curran, 2006; Hancock & Lawrence, 2006).

To date, only a few studies have been conducted that satisfy all four requirements (Meczkowski et al., 2016; Shen & Dillard, 2014; Shen et al., 2015). The results, however, are consistent across projects. Each of the investigations showed that fear was related to persuasion in the manner predicted by drive theory (cf. Dolinski & Nawrat, 1998). In the current study, we sought to replicate those results with a different advocacy and in a new topic domain. The main expectation was that

Hypothesis 1 (H1): To the extent that an inverted-U response is present in the fear response data, it will predict persuasion.

Drive theory underwrites another prediction that is substantially at odds with contemporary theories of threat appeals. That is with regard to the location and the function of fear in the persuasive process. Consider that the revised Protection Motivation Theory (PMT) in which fear is a sort of appendage to the main theoretical process (Rogers, 1983, Figure 6-2). In this conception, fear affects and is affected by severity and vulnerability, but it has no direct effect on either threat or on coping appraisal.¹ The Extended Parallel Processing Model (EPPM) posits two roles for fear (Witte, 1992, 1994, Figure 1). One is as a proximal cause of defensive avoidance and reactance: Fear is predicted to prompt maladaptive responses. The other role is as reciprocal cause of threat and efficacy appraisals. As with PMT, fear appears to have a minor influence on the main theoretical process. Witte (1994) summarizes these predictions succinctly when she writes that “the emotion of fear is [directly] associated with message rejection and is not directly related to message acceptance” (p. 118).

In contrast to PMT and EPPM, drive theory locates fear as central to the persuasion process. It is the rise and fall of fear that proximally determines agreement with the message. More precisely, drive theory suggests that message judgments such as severity and susceptibility will influence the fear curve, but that they will not directly lead to persuasion. This prediction is conceptually plausible with regard to all four of the message judgments, although previous research suggests that self-efficacy may have only a direct effect on intention (Shen & Dillard, 2014). Thus, there is reason to suspect that the expectation of full mediation by fear response may be inaccurate. Nonetheless, for the sake of simplicity, we advanced the strong form of the hypothesis that follows the logic of drive theory:

Hypothesis 2 (H2): Message judgments of perceived severity, susceptibility, response efficacy, and self-efficacy will directly affect the pattern of fear responses and only via fear will they affect persuasion.

Correspondence Between Static and Dynamic Measures of Emotion

If we assume that the fear-persuasion relationship is best understood as a dynamic process, we then confront the question of how to interpret the results of investigations that have utilized between-subjects designs. In other words, if the relevant emotional process unfolds over time during message processing, what aspect of that process is captured by a post-message survey item?

Several studies already exist that are relevant to the question. For example, Rossiter and Thornton (2004, Study 1) had subjects in one group provide static, post-message data on their emotional responses to seven anti-speeding advertisements. A separate group made continuous judgments on a dial mechanism throughout each the advertisement. Correlation of the peak dynamic scores with the static judgments was τ = .71, p = .02 (where message was the unit of analysis). The researchers also note a correlation of r = −.43, p = .18 between fear reduction and the static measure, “which indicates that the normal static rating of fear completely misses the drive reduction process when it is present” (p. 953).

Another line of research, on a phenomenon known as duration neglect, yielded similar findings. This work has demonstrated that static, retrospective judgments of affect are generally not influenced by the length of the pleasant or unpleasant event: hence, duration neglect (Fredrickson, 2000). Instead, overall judgments are well predicted by the peak and end values of the dynamic measure. Of special relevance to the study of messages, Baumgartner, Sujan, and Padgett (1997) provide evidence of peak and end effects using commercial advertisements as stimuli.

While remaining studies focused on threat appeals, we aimed to extend previous work by utilizing a more stringent method of assessing correspondence. The correlation statistics, used in previous research, are limited in that they are sensitive to the overall pattern of data, but they do not demand category-to-category concordance. Thus, high coefficients could arise from different patterns of static-dynamic correspondence. To better appreciate this point, assume that research participants make judgments on five-point scales where 0 = none of this emotion and 4 = a great deal of this emotion. Further assume that we possess data from three participants whose peak dynamic responses are 1, 2, and 3 and whose corresponding static judgments are 1, 2, and 3. The association between these two vectors is r = 1.00. A different group of three subjects might generate peak dynamic data of 1, 2, and 3 along with static data of 2, 3, and 4. The correlation is again 1.00, but the exact match is 0 because none of the participants produced the same response at different points in time. Thus, we can distinguish between pattern correspondence, indexed by the correlation coefficient, and absolute correspondence, which tests for the degree to which one set of values matches exactly with another set of values. The latter may be assessed by a statistic such as Krippendorff’s alpha (α_K; Hayes & Krippendorff, 2007), which would estimate the number of exact matches between dynamic and static values. We asked,

Research Question 1 (RQ1): To what extent do peak and end values of dynamic emotion judgments correspond with static judgments of the same emotions?

Methodological Threats to Internal Validity

All research designs suffer from certain weaknesses (Campbell & Stanley, 1966). Given the nature of our research, two possible threats to internal validity were testing and order effects. Testing refers to the potential for act of measurement early in the research to influence responses on measures later in the research. In our case, the concern was that the heightened attention to emotion required by the dynamic measures might alter responses to the subsequent static measures or to the persuasion indices. Accordingly, we included the presence versus absence of the dynamic measures as a factor in our experiment and asked,

Research Question 2 (RQ2): Do the data show evidence of testing effects?

Order effects create a threat to internal validity when the arrangement of experimental procedures causes unwanted effects on the outcomes of interest. In our investigation, one such risk was the placement of the static emotions measures relative to the persuasion measure. Although we judged the risk as small, it seemed prudent to directly address this inferential danger via empirical test. Hence, we created two conditions that alternated the order of the static fear and persuasion measures.

Research Question 3 (RQ3): Do the data show evidence of order effects?

Context and Advocacy: Screening for Colorectal Cancer

Because all behavior occurs in context, it was necessary to choose a domain in which to examine the hypotheses. Colorectal cancer was chosen, in part, because of the frequency and severity of the disease. Worldwide, colorectal cancers are the third most common type of cancer (Ferlay, Bray, Pisani, & Parkin, 2004; Parkin, Whelan, Ferlay, Teppo, & Thomas, 2002). The National Cancer Institute (n.d.) estimates that over 50,000 people will die from colorectal cancer in 2014. Colorectal cancer affects men and women in equal numbers (Centers for Disease Control, n.d.-c).

Another reason to consider colorectal cancer is its preventability. In a study of the United States and Europe, the survival rate was 66% during the period 2000-2002 (Verdecchia et al., 2007). Likelihood of survival is increased by early detection (Ransohoff, 2009) and, as a consequence, screening is recommended for persons 50 years of age and older (Centers for Disease Control, n.d.-a). Several different methods of screening are available. However, colonoscopy is considered the most effective of the available options because it enables inspection of the entire length of the colon.

Method

Sampling and Participants

Participants were a geographically diverse group of adults recruited by Qualtrics via a national paid opt-in online survey panel that consisted of equal numbers of males and females. Individuals were asked three screening questions to ensure their suitability for the study. First, because colonoscopy is not recommended until age 50, participants were screened by age. Persons who were younger than 39 or older than 49 were filtered out. Next, we eliminated individuals who had been diagnosed with colorectal cancer prior to the survey. Finally, respondents who had previously sought screening for colorectal cancer were eliminated. Following these screens, participants were randomly assigned to experimental condition.

In the final sample (N = 418), 209 (50.0%) participants were male and 209 (50.0%) were female, ranging in age from 39 to 49 years old (M = 44.06, SD = 3.19). Three hundred twenty-five (77.8%) of the participants identified as White, 46 (11.0%) as Black or African American, 26 (6.2%) as Asian or Pacific Islander, 22 (5.3%) as Hispanic, 7 (1.7%) as Native American or American Indian, and 7 (1.7%) as “other” without specifying their ethnicity. Fourteen (3.3%) participants identified with multiple ethnicities. Participants provided data in a single online session. Data collection took place over a six-day period.

Statistical Power

In the full sample of 418, power to detect small (.10), medium (.30), and large (.50) bivariate effects at α = .05, two-tailed was .86, .99, and .99, respectively. Only 209 participants were available in the conditions that included the dynamic measures. The corresponding power values in this subsample were .29, .99, and .99.

Message

Participants read the message online in PowerPoint format that intermingled text and images. The message was divided into two components: threat and action. In the threat component, participants read information about their susceptibility to colorectal cancer, the progression of the disease, and the potential consequences of colorectal cancer in various stages. In the action component, the message identified colonoscopy as an effective means of detecting colorectal cancer in an early stage, and encouraged participants to talk to their doctor about the screening procedure. The message was the same across all conditions; however, individuals in the dynamic fear condition were asked to report their emotional reaction to the message following the threat component. The entire message is available at the first author’s website (see Dillard, 2015).

Research Design

The experiment was a 2 × 2 factorial design with varying numbers of repeated measures for fear. The first between-persons factor was presence versus absence of the dynamic measures of fear (i.e., testing). When the measures were present, data on emotional responses were gathered just prior to message exposure, immediately following presentation of the threat component, and immediately following exposure to the action component. This feature of the design allowed a test of the degree to which dynamic measures might induce measurement reactivity.

The second between-persons factor was timing of the persuasion measures (i.e., order). In half of the cells, behavioral intention was measured immediately after the message and prior to the static measure of fear. In the remaining conditions, the intention measure appeared at the end of the survey, following the static measure of fear.

Measurement

Message judgments

Four aspects of the message were evaluated with multi-item scales. For perceived severity, the items were: I believe colorectal cancer is severe; colorectal cancer has serious negative consequences; and colorectal cancer is extremely harmful. For perceived susceptibility, the items were: There’s a chance that I could contract colorectal cancer; it is possible that I could get colorectal cancer; and I could be at risk for colorectal cancer. For response efficacy: Getting a colonoscopy is a good way to protect myself against colorectal cancer; getting a screening is helpful if I want to avoid colorectal cancer; and getting a screening can help me prevent colorectal cancer. And, for self-efficacy, I am able to get screened for colorectal cancer; I can easily get a colonoscopy if I want to; and I am confident in my ability to get screened for colorectal cancer. Judgments were rendered on 7-point Likert-type scales where 1 = strongly disagree and 7 = strongly agree. Means, standard deviations, and alpha reliability estimates are given in Table 1.

Table 1.

Zero-Order Correlations and Descriptive Statistics for the Measurement Model.

	Severity	Susceptibility	Response efficacy	Self-efficacy	Intention	Family history %	Fear t₀	Fear t₃
Severity	.80
Susceptibility	.37**	.88
Response efficacy	.44**	.27**	.87
Self-efficacy	.18*	.23**	.47**	.90
Intention	.21**	.31**	.36**	.45**	.93
Family history %	−.06	−.12**	−.01	.09*	−.07	—
Fear t₀	.03	.10*	.03	−.04	−.05	−.01	.93
Fear t₃	.20**	.27**	.20**	.05	.22**	−.01	.35**	.91
M	6.20	5.14	6.27	5.70	8.56	7.2	1.52	2.36
SD	0.88	1.33	0.90	1.32	2.72	—	0.92	1.17

Note. N = 418. Cronbach’s alpha values are shown in diagonal. Fear at t1 and t2 do not appear in this table because they were measured in only half of the sample.

p < .05. **p < .01.

Emotion

Fear was assessed with two items, afraid and scared. These were treated as separate single-item measures in the correspondence analyses, but otherwise combined to form a scale (r = .83 in the static data, designated at t₃ in Table 1). All participants provided emotion data at t₀ as well as after reading the message in its entirety. The item stems were, respectively, How much of this emotion are you feeling right now? and Overall, how did the message make you feel? Participants in the dynamic measures conditions also reported on their level of fear following the threat component of the message and the action component of the message. The corresponding item stems were How did Part I (II) of the message make you feel? The 5-point response scale was anchored at 0 = none of this emotion and 4 = a great deal of this emotion.

Behavioral intention

Intention to comply with the advocacy was assessed by asking respondents to assess the likelihood that they would engage in two actions: talk to my doctor about receiving a colonoscopy and get screened for colorectal cancer. The 11-point response scale ran from 0% to 100% with anchors representing 10-unit intervals.

Control variables

We gathered data on two indirect indices of involvement with the message topic. Family history of colorectal cancer was assessed by asking participants whether or not any of their family members had been diagnosed with colorectal cancer. Seven percent replied yes. Even within the relatively narrow age range of our sample, we considered the possibility that respondents closer to 50 might be more involved than younger participants. Given our screening procedures, values for age ranged from 39 to 49.

Results

Measurement Model

A confirmatory factor analysis was conducted to evaluate the multi-item scales in the presence of the control variables. Cross-loadings were not permitted nor were error terms allowed to correlate with one another. Consequently, the model had six latent variables and 58 parameters to be estimated. The analysis returned the following: χ²(126) = 186.18, p = .00, comparative fit index (CFI) = .98, Tucker-Lewis index (TLI) = .98, and root mean square error approximation (RMSEA) = .034 (90% confidence interval [CI]: [.023, .044]). In terms of the criteria suggested by Hu and Bentler (1998, 1999), the model fit the data well. Attention to the factor loadings revealed that they were uniformly high, ranging from .72 to .95. We concluded that the measurement model was acceptable. Descriptive statistics and reliability estimates are given in Table 1.

Effects of Control Variables

When respondent age and family history were correlated with message judgments, fear, and intention, in the dynamic data subsample three small relationships emerged for family history and none for age: for severity, −.12, p = .06; for susceptibility, −.18, p = .009, and fear at t₁ = −.12, p = .07. These results suggested that it would be prudent to control for family history in subsequent analyses, but that controlling for age was unnecessary.

Methodological Threats to Internal Validity

Analysis of variance was used to assess possible effects of testing (RQ2) and timing (RQ3) on static emotions and on persuasion. This procedure showed that the presence versus absence of the dynamic measures of fear had no discernible impact on intention to obtain a colonoscopy, F(1, 416) = .58, p = .446, η² = .001, but that there was a significant association with static fear, F(1, 416) = 4.63, p = .032, η² = .011. Inspection of the means revealed that the dynamic conditions produced slightly depressed responses for static fear (2.24 vs. 2.49). However, given the very high statistical power of the test and the rather small size of the effect, we concluded that testing did not pose a substantial threat to internal validity of the study.

Analysis of the effects for order (persuasion measured before versus after static fear) showed several small, but significant effects. The order of measurements significantly influenced static fear, F(1, 416) = 6.41, p = .012, η² = .015, as well as behavioral intention, F(1, 416) = 4.05, p = .045, η² = .009. Specifically, the conditions in which persuasion was measured after static emotions led to relatively lower level of reported static fear and intention to obtain a colonoscopy. Given strong statistical power and the small magnitude of the effect, we inferred that order did not pose a substantial threat to internal validity of the study. Nonetheless, we controlled for it statistically in analyses where possible.

Input Data

To control for the effects of family history and order, we created a matrix of partial covariances. After subtracting the degrees of freedom associated with the covariates, the effective n was 205. This matrix and the means of each variable were used in the input data for subsequent analyses (Table 2).

Table 2.

Partial Covariances Controlling for Family History and Order.

	Severity	Susceptibility	Response efficacy	Self-efficacy	Intention	Fear t₀	Fear t₁	Fear t₂	Fear t₃
Severity	.81
Susceptibility	.42	.88
Response efficacy	.53	.24	.88
Self-efficacy	.23	.17	.44	.91
Intention	.25	.20	.36	.40	.94
Fear t₀	.01	.15	.03	−.04	−.02	.93
Fear t₁	.17	.27	.20	.12	.24	.39	.96
Fear t₂	.04	.09	.06	−.08	.04	.49	.56	.93
Fear t₃	.16	.26	.25	.01	.18	.34	.63	.64	.91
M	6.13	5.13	6.29	5.70	8.46	0.46	1.65	0.77	1.24
SD	0.92	1.31	0.93	1.28	2.90	0.89	1.31	0.97	1.12

Note. N = 205. Cronbach’s alpha values are shown in diagonal.

Evaluating the Growth Trajectory

Among the conditions necessary to testing for a curvilinear association between fear and persuasion are (a) curvilinearity in the data and (b) inter-individual variation in the form of the curves. Evidence of the first condition can be seen in Figure 3, which plots the mean fear responses at t₀, t₁, and t_2. The means of .46, 1.65, and .77 show the anticipated inverted-U pattern. The second necessary condition can be assessed by attending to the variance statistics at each time point. Because the standard deviations for the three conditions were non-zero (i.e., .89, 1.30, and .97), we concluded that the data showed inter-individual variation.

Figure 3.

Mean fear response to the colonoscopy message at three points in time.

More formal assessments were also possible using Latent Growth Modeling. The first step was to create a growth curve model from the dynamic indices of fear. Hence, we created three latent variables that represented different aspects of the curve: (a) the intercept with paths to the indicators fixed at 1, 1, and 1; (b) the linear component with paths fixed at 0, −1, and −2; and (c) the quadratic component with paths fixed at 0, −2, and −4. These values were chosen for the linear and quadratic components such that the latter modeled the inverted-U relationship predicted by the hypothesis. The intercept and the slope were allowed to covary, but the error terms for the observed variables were not. The linear model did not fit the data χ²(1) = 131.25, p < .001, CFI = .08, Goodness of Fit Index (GFI) = 1.00, RMSEA = .69.

Because the three-component growth curve model was just-identified (df = 0), it could not be tested for fit. However, it is possible to estimate the parameters of a just-identified model. When we did this, it produced

Fear (y) = 0.46 + 2.23 x - 1.04 x^{2},

(1)

where x = time of measurement. All three estimates were more than twice the size of their respective standard errors. The positive linear component and negative quadratic component confirm that the data take the form of an inverted-U.

Effects of the Inverted-U on Persuasion

The next step was to assess the effect of the inverted-U on persuasion. Perceived severity and susceptibility were cast as predictors of the linear and quadratic components of emotional response, which were predictors of behavioral intention. Following previous work, response efficacy and self-efficacy predicted intention directly (Shen & Dillard, 2014). The message judgment and intention measures were specified as single indicator latent constructs, with their error terms fixed as (1-α) multiplied by their respective variance. After deleting non-significant paths and removing variables that had no significant cause or effect, we were left with the model that appears in Figure 4. Initially, this model yielded an improper estimate of the intercept (−.13). Consequently, we fixed the intercept at the mean value of fear at t₀ (i.e., .46) and recomputed the model. The resulting fit indices were χ²(105) = 19.32, p = .05, CFI = .97, GFI = .97, RMSEA = .07 (90% CI [.008, .11]). Thus, the theoretical model fit the data.

Figure 4.

The obtained model for the colonoscopy message.

There are two features of the obtained model that bear on the hypotheses. First, there is a significant relationship between the quadratic component of the fear curve and persuasion (standardized coefficient of .22, p = .004). This result indicated that the inverted-U predicted intention to obtain a colonoscopy. H1, which predicted this relationship, was supported.

The second important feature concerns the behavior of the message judgment scales. The pattern of relationships has implications for H2, which predicted that message judgments would exert their influence on persuasion via the fear curve. Although perceived severity showed significant bivariate relationships with fear and intention (see Table 1), these associations vanished in the full model. Perceived susceptibility and response efficacy both emerged as significant predictors of the fear curve. Self-efficacy showed a direct effect on intention, but no discernible impact on fear response. Thus, support for H2 was mixed: Susceptibility and response efficacy behaved as predicted; self-efficacy gave results that were counter to expectations.

Correspondence Between Static and Dynamic Measures of Fear

A strong test of correspondence between the static and dynamic measures of fear was conducted via α_K. Because the version of α_K that we planned to use tests for exact match between the two measures, we conducted separate analyses for each fear item. For the scared item, correspondence between the t₁ measure and the static measure was .25. For t₂, the value was .33. For the afraid item, the t₁ and t₂ values were .30 and .29, respectively. These coefficients suggest that the static measure of fear, common to between-persons research designs, shows poor correspondence with either the peak or end indices of dynamic fear.

There are two methodological issues that might bear on the low α_K values. One is the effect of order. However, our analysis of order effects showed them to be quite small. Thus, while order effects may be present, they should not influence the α_K by more than 1% to 2%. An alternative is that the coefficients were depressed due to imbalanced marginal frequencies, a condition that is known to attenuate the reliability coefficient (Feinstein & Cichetti, 1990). This was true of all the fear response distributions, which were positively skewed. To partially address this situation, we collapsed the top two emotion response categories (i.e., 3 and 4, where 4 = a great deal of this emotion), then recomputed α_K. The results for scared were .27 and .35, and for afraid .32 and .32. These results differed only trivially from the results for the original distributions. Thus, we rejected the possibility that imbalanced marginals were responsible for attenuation of the coefficients and moved to a simple answer to RQ1: The static measure of fear showed poor absolute correspondence with both the post-threat and post-action indices of fear. Conceivably, data that are gathered via the static method represent some combination of dynamic influences, as suggested by the duration neglect literature. To evaluate this possibility, we used correlation and regression analyses to assess pattern correspondence in the data set that controlled for order and family history. Specifically, the three dynamic measures were treated as predictors of the static measure (the t₀ measure was included as a control). The results, given in Table 3, show that the t₁ and t₂ measures are both substantial and significant predictors of static fear. The regression analyses indicate that the combination of the two is superior to either one alone. That is, each predictor contains variance that contributes uniquely to the overall prediction. Attention to the means (Table 2) suggested that the static measure captured something close to the average of the peak and end values. In general, these results give a more nuanced answer to the question of correspondence (RQ1): Despite the high bivariate correlations, the regression analyses demonstrate that even in terms of pattern correspondence, the static measure is an incomplete index of the dynamic properties of the fear response.

Table 3.

Association Between Measures of Static Fear and Three Dynamic Measures.

Predictor	Scared				Afraid
Predictor	r _p	B	SE B	β	r _p	B	SE B	β
Fear t₀	.35***	.08	.07	.06	.32***	−.03	.07	−.03
Fear t₁	.57***	.27***	.05	.31	.62***	.38***	.05	.44
Fear t₂	.63***	.51***	.07	.43	.58***	.39***	.07	.35
R				.70***				.69***

Note. N = 205. These results control for family history and order.

p < .05. **p < .01.***p < .001.

Discussion

Processing Threat Appeals

Several issues emerge from the model in Figure 4 that bear on the question of how individuals process threat appeals. First, there is evidence that fear and persuasion are related in a curvilinear fashion. Not only did the overtime fear measures show an inverted-U pattern, but that pattern was a significant predictor of persuasion. The linear component of the fear response showed no discernible influence on persuasion. The current study is not the first investigation to show this result, but it adds to existing evidence by replicating the effect in a novel message context using a diverse sample of adults. Thus, contrary to conventional wisdom, we conclude that the relationship between fear and persuasion is curvilinear when assessed in an appropriate research design.

The results challenge current orthodoxy in another way. The obtained model located fear as a central mechanism in message processing. The results were complex in that some of the message judgment variables exerted their influence on persuasion through fear whereas others did not. Specifically, the effects of susceptibility and response efficacy were mediated by fear. This pattern runs counter to the basic tenets of PMT and EPPM, but is perfectly consistent with drive theory, which presents fear as the central mechanism in message processing. Still, efforts to draw a simple, uniform conclusion about the operation of message judgments are hampered by two other findings: (a) severity showed no significant effect in the obtained model and (b) self-efficacy demonstrated a direct influence on intention. The absence of a significant association for severity may be the result of a ceiling effect, it may reflect something unique to the topic of colonoscopy, or it may be nothing more than the result of sampling error. Further testing is needed to evaluate the degree to which this outcome is robust across samples and message domains.

The same may be true for self-efficacy, but prior theory suggests one other alternative. Notably, the Theory of Planned Behavior makes the prediction that self-efficacy will be directly correlated with intention (Ajzen, 1991). Though the prediction is phrased with appropriate scientific caveats, it echoes the commonplace belief that people do things because they can. But, it is unlikely that this commonplace notion is true without qualification. Rather, there is evidence that self-efficacy has a positive effect on intention only when attitude or subjective norms toward the behavior are favorable (Dillard, 2011; Yzer, 2007). Although attitude and subjective norms were not measured in the current study, it is likely that these important side conditions are in place regarding intention to obtain a colonoscopy and in the threat appeal literature as a whole. It would be more accurate to say that, generally, when actions are seen as desirable, people perform those actions if they are able.

It also is important to compare the findings for self- and response efficacy. From the standpoint of message design, these ideas refer to different aspects of the threat appeal. But, from contemporary theory, it is less evident that they are discrete judgments. Consider that PMT treats the two ideas as if they are parts of a greater whole (Rogers, 1983, p. 168). The EPPM position is similar, but even more specific: Researchers are explicitly advised to create a single perceived efficacy score by combining the two judgments (Maloney, Lapinski, & Witte, 2011; Witte, Cameron, McKeon, & Berkowitz, 1996). The motivation for doing so is justifiable. Integrating the two variables produces more parsimonious theory and it reflects the fact that the two concepts have shown parallel empirical associations with persuasion in some previous studies. However, the current results suggest that it would be unwise to follow this advice: Self- and response efficacy play different roles during message processing in longitudinal data. Response efficacy influences the shape of the fear response, which, in turn, guides the formation of intention. Self-efficacy affects intention directly. Of course, the findings must be viewed with caution in that they derive from a single study. If, however, they are replicated, it may be necessary to partition message processing concepts such as appraisal and coping more finely. This could, in turn, require rethinking message design such that response efficacy components should appear early in the appeal whereas self-efficacy material should be placed near the end of the message. One conclusion appears clear from our data: contra PMT and EPPM, self- and response efficacy are distinct and should be treated separately.

Correspondence Between Static and Dynamic Measures of Fear

If the true relationship between fear and persuasion is one that unfolds over time, then what sort of data do respondents provide when forced to characterize a dynamic process in static terms? We noted earlier that this is actually two questions, one having to do with absolute correspondence and the other with pattern correspondence. With regard to absolute correspondence, our data showed that static, retrospective fear is a poor proxy for either peak fear or end fear. The static measure underestimates peak fear and overestimates end fear. To the extent that an accurate representation of the fear response is of theoretical interest, researchers will be poorly served by the use of static indices (see also Algie & Rossiter, 2010; Rossiter & Thornton, 2004). This result should be of signal importance to duration neglect researchers, who thus far have ignored questions of absolute correspondence.

In terms of pattern correspondence, peak and end values were strongly related to static fear. Our results replicate the basic findings of the duration neglect studies (which all use correlational methods), but with several noteworthy differences (Fredrickson, 2000). One is a limitation: Because our study did not vary duration of the affective experience, we can make no claims concerning duration neglect (nor did we plan our study to do so). But, second, by focusing on fear in particular, we demonstrated peak and end effects for this specific, discrete emotion rather than the typical diffuse judgments of positive and negative affect (see also Rossiter & Thornton, 2004). For emotion researchers, it will be of interest to know whether the peak and end effects hold for other discrete emotions, as messages are often capable of arousing multiple affective responses (Dillard, Plotnick, Godbold, Freimuth, & Edgar, 1996). Third and finally, it is important to note that, even though highly correlated, peak and end scores made unique contributions to the prediction of static fear. It appears that the static, retrospective measure of fear is roughly the average of the peak and end measures. Hence, it could be said that the static measure represents the worst of both worlds: It underestimates peak fear and overestimates end fear. In any case, the assumption that a static measure of fear is substitutable for a peak or end measure is untenable. With regard to the threat appeals literature more generally, it is yet more evidence that the results of between- and within-persons studies cannot be easily reconciled.

Methodological Hazards of Repeated Measures Designs

Assessing change in fear over time requires repeated measurement of that emotion. And, repeated assessment introduces the possibility that measurement itself may influence the outcome of research (Campbell & Stanley, 1966). Participants might become hyper-cognizant of their feelings, which could inflate the fear-persuasion correlation, or repeated measurement might produce desensitization and artificial attenuation of the results. Our design allowed a test of this hypothesis. The static conditions measured fear at t₀ and again after the message, whereas the dynamic conditions included both of these measures as well as an assessment of fear at t₁ and t₂. Comparison of these conditions revealed no discernible effect of testing on intention coupled with a small, but significant, effect on static fear. Given that the statistical power of our design was in excess of .99 for small effects, it is not surprising that a significant difference was observed. To our knowledge, only one previous study has assessed testing effects in a threat appeals study with three measures of fear (Dillard & Anderson, 2004). That investigation showed no observable effects of measurement. Considered jointly with the current results, we conclude that the threat to internal validity is small to none in designs of the type employed here.

We are less certain about alternative methods. At this point, it is unclear whether studies that employ continuous measures of emotional response are similarly robust against the effects of testing.

Encouraging Colorectal Cancer Screening

The data also have implications for colorectal cancer screening. Recall that the two intention items measured intention to talk to a doctor and intention to be screened. For both items, the mean score was a self-reported 75% likelihood. These values are substantially larger than the roughly 2/3s of people who currently receive screening (Centers for Disease Control, n.d.-b). Though we recognize that intentions do not translate perfectly into behavior (Kim & Hunter, 1993; Webb & Sheeran, 2006), the data suggest that threat appeals may be a viable method for improving screening rates for colorectal cancer. How, when, and where to disseminate the messages are separate problems.

Conclusions

This project contributes to the literature on threat appeals in three important ways. First, it demonstrates the viability of understanding message processing as an overtime process in which emotional onset and offset are key to persuasion. It also shows that the question of correspondence between static and dynamic measures of fear is more nuanced than previously thought. Finally, the study provides evidence that threats to internal validity arising from longitudinal designs of the sort used here are negligible. This is good news on the methodological front. In contrast, the two main theoretical findings—the inverted-U and the incommensurability of static and dynamic measures—lead to a different conclusion: Building meaningful connections between past and current research findings will not be a simple process.

Footnotes

Authors’ Note

Lijiang Shen is now at The Pennsylvania State University. Meczkowski et al. (2016) was conducted prior to the current paper. They are meant to be read in chronological order.

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Notes

Author Biographies

James Price Dillard (PhD, Michigan State University, 1983) is liberal arts research professor of Communication Arts & Sciences at The Pennsylvania State University. His research focuses on the role of emotion in persuasion.

Ruobing Li (MA, Pennsylvania State University, 2014) is a doctoral student in the College of Communications at The Pennsylvania State University. Her primary research interest focuses on the persuasive effects of media technology and message appeals.

Eric Meczkowski (MA, Pennsylvania State University, 2014) is a doctoral candidate at The Pennsylvania State University. He conducts research at the intersection of persuasion and health.

Chun Yang (MA, Washington State University, 2013) is a doctoral candidate at The Pennsylvania State University. He is interested in the cognitive and emotional effects of media.

Lijiang Shen (PhD, UW-Madison, 2005) is Associate Professor of Communication Arts & Sciences at The Pennsylvania State University. His primary area of research considers the impact of message features and audience characteristics in persuasive health communication, message processing, and the process of persuasion/resistance to persuasion as well as quantitative research methods in communication.

References

Ajzen

(1991). The theory of planned behavior. Organization Behavior and Human Decision Processes, 50, 179-211. doi:10.1016/0749-5978(91)90020-t

Algie

Rossiter

J. R.

(2010). Fear patterns: A new approach to designing road safety advertisements. Journal of Prevention & Intervention in the Community, 38, 264-279. doi:10.1080/10852352.2010.509019

Baumgartner

Sujan

Padgett

(1997). Patterns of affective reactions to advertisements: The integration of moment-to-moment responses into overall judgments. Journal of Marketing Research, 34, 219-232. doi:10.2307/3151860

Bollen

K. A.

Curran

P. J.

(2006). Latent curve models: A structural equation perspective. Hoboken, NJ: John Wiley.

Boster

F. J.

Mongeau

(1984). Fear-arousing persuasive messages. In Bostrom

R. N.

(Ed.), Communication yearbook 8 (pp. 330-375). Newbury Park, CA: Sage.

Campbell

D. T.

Stanley

J. C.

(1966). Experimental and quasi-experimental designs for research. Boston, MA: Houghton Mifflin.

Centers for Disease Control. (n.d.-a). Colorectal cancer screening guidelines. Retrieved from http://www.cdc.gov/cancer/colorectal/basic_info/screening/guidelines.htm

Centers for Disease Control. (n.d.-b). Colorectal cancer screening rates remain low. Retrieved from http://www.cdc.gov/media/releases/2013/p1105-colorectal-cancer-screening.html

Centers for Disease Control. (n.d.-c). Colorectal cancer statistics. Retrieved from http://www.cdc.gov/cancer/colorectal/statistics/

10.

Dillard

J. P.

(2011). An application of the integrative model to women’s intention to be vaccinated against HPV: Implications for message design. Health Communication, 26, 479-486. doi:10.1080/10410236.2011.554170

11.

Dillard

J. P.

(2015). Message advocating obtaining a colonoscopy. Retrieved from https://www.researchgate.net/profile/James_Dillard2/publications?pubType=dataset

12.

Dillard

J. P.

Anderson

J. W.

(2004). The role of fear in persuasion. Psychology & Marketing, 21, 909-926. doi:10.1002/mar.20041

13.

Dillard

J. P.

Plotnick

C. A.

Godbold

L. C.

Freimuth

V. S.

Edgar

(1996). The multiple affective outcomes of AIDS PSAs: Fear appeals do more than scare people. Communication Research, 23, 44-72. doi:10.1177/009365096023001002

14.

Dolinski

Nawrat

(1998). “Fear-then-relief” procedure for producing compliance: Beware when the danger is over. Journal of Experimental Social Psychology, 34, 27-50. doi:10.1006/jesp.1997.1341

15.

Dollard

Miller

N. E.

(1950). Personality and psychotherapy. New York, NY: McGraw-Hill.

16.

Feinstein

A. R.

Cichetti

D. V.

(1990). High agreement but low kappa: I. The problems of two paradoxes. Journal of Clinical Epidemiology, 43, 543-549. doi:10.1016/0895-4356(90)90158-L

17.

Ferlay

Bray

Pisani

Parkin

D. M.

(2004). GLOBOCAN 2002: Cancer incidence, mortality and prevalence worldwide (IARC CancerBase No. 5, Version 2.0). Lyon, France: IARC Press.

18.

Fredrickson

B. L.

(2000). Extracting meaning from past affective experiences: The importance of peaks, ends, and specific emotions. Cognition and Emotion, 14, 577-606. doi:10.1080/026999300402808

19.

Hancock

G. R.

Lawrence

F. R.

(2006). Using latent growth curve models to evaluate longitudinal change. In Hancock

G. R.

Mueller

R. O.

(Eds.), Structural equation modeling: A second course (pp. 171-186). Greenwich, CT: Information Age.

20.

Hayes

A. F.

Krippendorff

(2007). Answering the call for a standard reliability measure for coding data. Communication Methods and Measures, 1, 77-89. doi:10.1080/19312450709336664

21.

Hovland

Janis

Kelley

(1953). Communication and persuasion. New Haven, CT: Yale University Press.

22.

Bentler

P. M.

(1998). Fit indices in covariance structure modeling: Sensitivity to underparameterized model misspecification. Psychological Methods, 3, 424-453. doi:10.1037/1082-989X.3.4.424

23.

Bentler

P. M.

(1999). Cutoff criteria for fit indexes in covariance structure analysis: Conventional criteria versus new alternatives. Structural Equation Modeling, 6, 1-55. doi:10.1080/10705519909540118

24.

Janis

I. L.

(1967). Effects of fear arousal on attitude change: Recent developments in theory and research. In Berkowitz

(Ed.), Advances in experimental social psychology (Vol. 3, pp. 167-222). New York, NY: Academic Press.

25.

Kim

M. S.

Hunter

J. E.

(1993). Relationships among attitudes, behavioral intentions, and behavior. Communication Research, 20, 331-364. doi:10.1177/009365093020003001

26.

Maloney

E. K.

Lapinski

M. K.

Witte

(2011). Fear appeals and persuasion: A review and update of the extended parallel process model. Social & Psychology Personality Compass, 5/4, 206-219. doi:10.1111/j.1751-9004.2011.00341.x

27.

McGuire

W. J.

(1968). Personality and susceptibility to social influence. In Borgatta

E. F.

Lambert

W. W.

(Eds.), Handbook of personality theory and research (pp. 1130-1180). Chicago, IL: Rand McNally.

28.

Meczkowski

E. J.

Dillard

J. P.

Shen

(2016). Threat appeals and persuasion: Seeking and finding the elusive curvilinear effect. Communication Monographs, 83, 373-395.

29.

Mongeau

P. A.

(2013). Fear appeals. In Dillard

J. P.

Shen

(Eds.), The handbook of persuasion (2nd ed., pp. 184-199). Thousand Oaks, CA: Sage.

30.

National Cancer Institute. (n.d.). Colon and rectal cancer. Retrieved from http://www.cancer.gov/cancertopics/types/colon-and-rectal

31.

Parkin

D. M.

Whelan

S. L.

Ferlay

Teppo

Thomas

D. B.

(2002). Cancer incidence in five continents (Vol. 8, no. 155). Lyon, France: International Agency for Research on Cancer.

32.

Peters

G. Y.

Ruiter

R. A. C.

Kok

(2013). Threatening communication: A critical re-analysis and a revised meta-analytic test of fear appeal theory. Health Psychology Review, 7, s8-s31. doi:10.1080/17437199.2012.703527

33.

Ransohoff

D. F.

(2009). How much does colonoscopy reduce colon cancer mortality? Annals of Internal Medicine, 150, 50-52. doi:10.7326/0003-4819-150-1-200901060-00308

34.

Rogers

R. W.

(1983). Cognitive and physiological processes in fear appeals and attitude change: A revised theory of protection motivation. In Cacioppo

J. T.

Petty

R. E.

(Eds.), Social psychophysiology: A source book (pp. 153-176). New York, NY: The Guilford Press.

35.

Rogers

R. W.

Prentice-Dunn

(1997). Protection motivation theory. In Gochman

D. S.

(Ed.), Handbook of health behavior research I: Personal and social determinants (pp. 113-132). New York, NY: Plenum.

36.

Rossiter

J. R.

Thornton

(2004). Fear-pattern analysis supports the fear-drive model for road safety TV-ads. Psychology & Marketing, 21, 945-960. doi:10.1002/mar.20042

37.

Shen

Dillard

J. P.

(2014). Threat, fear, and persuasion: Review and critique of questions about functional form. Review of Communication Research, 2, 94-114. doi:10.12840/issn.2255-4165.2014.02.01.004

38.

Shen

Wang

Seung

(2015, May). Putting the fear back again: Revisiting the role of fear in fear appeal. Paper presented at ICA conference, San Juan, Puerto Rico.

39.

Sutton

S. R.

(1982). Fear-arousing communications: A critical examination of theory and research. In Eisner

J. R.

(Ed.), Social psychology and behavioral medicine (pp. 303-337). New York, NY: John Wiley.

40.

Verdecchia

Francisci

Brenner

Gatta

Micheli

Mangone

Kunkler

(2007). Recent cancer survival in Europe: A 2000-2002 period analysis of EUROCARE-4 data. Lancet Oncology, 8, 784-796. doi:10.1016/s1470-2045(07)70246-2

41.

Webb

T. L.

Sheeran

(2006). Does changing behavioral intentions engender behavior change? A meta-analysis of the experimental data. Psychological Bulletin, 132, 249-268. doi:10.1037/0033-2909.132.2.249

42.

Witte

(1992). Putting the fear back into fear appeals: The extended parallel process model. Communication Monographs, 59, 329-349. doi:10.1080/03637759209376276

43.

Witte

(1994). Fear control and danger control: A test of the extended parallel process model (EPPM). Communication Monographs, 61, 113-134. doi:10.1080/03637759409376328

44.

Witte

Allen

(2000). A meta-analysis of fear appeals: Implications for effective public health campaigns. Health Education & Behavior, 27, 591-615. doi:10.1177/109019810002700506

45.

Witte

Cameron

K. A.

McKeon

Berkowitz

(1996). Predicting risk behaviors: Development and validation of a diagnostic scale. Journal of Health Communication, 1, 317-341. doi:10.1080/108107396127988

46.

Yzer

(2007). Does perceived control moderate attitudinal and normative effects on intention: A review of conceptual and methodological issues. In Ajzen

Albarracin

Hornik

(Eds.), Prediction and change of health behavior: Applying the reasoned action approach (pp. 111-127). Mahwah, NJ: Lawrence Erlbaum.