Persuasive appeals frequently prove ineffective or produce unintended outcomes, due to the presence motivated reasoning. This research delves into the impact of emotional content levels, message valence, and the coherence of pre-existing attitudes on the process of biased information evaluation. By conducting a factorial survey (N = 480) and incorporating a computational model of attitude formation, we aim to gain a deeper insight into the cognitive-affective mechanisms driving motivated reasoning. Our experimental findings reveal that motivated reasoning is most pronounced when persuasive appeals employ a combination of emotional and rational elements within a negatively valenced argumentation. Furthermore, our computational model, which estimates belief and affect adjustments underlying attitude changes, elucidates how message framing influences cognitive-affective processes through emotional coherence. The results provide support for a negative correlation between shifts in coherence in response to new information and the propensity for motivated reasoning. The research contributes to computational models of opinion dynamics and social influence, offering a psychologically realistic framework for exploring individual reasoning's impact on population-level dynamics, particularly in policy contexts, where it can enhance communication and informed policy discussions.
Article
The critical role of emotional communication for motivated reasoning
https://doi.org/10.21203/rs.3.rs-3575719/v1
This work is licensed under a CC BY 4.0 License
You are reading this latest preprint version
In our daily lives, emotions, cognitions, and behaviors of others significantly influence our attitudes and reasoning 1,2. Our perspectives on new technologies 3, political matters 4 and scientific evidence 5 often hinge on the opinions of those around us. However, not every attempt to change our attitudes, such as persuasive communication, produces the intended impact. Beyond social factors like cultural and identity-affiliating group membership 6, cognitive elements like attitudinal alignment 7, attitude strength 8,9 and affective coherence 10 on issues have a substantial influence on the strength of and response to social influence. When individuals encounter opposing arguments and emotions, they tend to reinterpret and modify information to support their existing beliefs, goals, and motives 11. This inclination toward motivated reasoning underscores a commitment to prior mental representations and may even reinforce existing biases, possibly resulting in backfire effects 12,13. While understanding and modeling these effects on an individual, group, or societal level is crucial, little is known about the extent and direction of belief and affect adjustments underlying observable changes in attitudes in response to persuasive appeals with varying levels of congruence.
Numerous empirical studies on motivated reasoning, also known as biased assimilation 14 or selective perception 15,16, provides substantial evidence across diverse domains that human information processing is both influenced and constrained by preexisting beliefs, attitudes 17, culturally ingrained values and ideologies 18,19 or religious worldviews 20. People employ a range of strategies to achieve desired outcomes and resist persuasive appeals, including selectively avoiding information, discounting the source, and experiencing negative emotional reactions 21. Furthermore, prolonged exposure to counterattitudinal messages may exacerbate attitudinal polarization among social groups with opposing views in both offline and online communication 22–24. Scholars in this field have identified individual-level moderators related to the internal structure of attitudes, such as attitude strength 25, factual knowledge 26 and affective evaluation of concepts 27 which influence the propensity for engaging in motivated reasoning.
Theoretical research on information processing and attitude formation has attributed motivated reasoning to individuals' inclination to reduce inconsistencies in their beliefs and emotions while maintaining attitudes that align with their goals 11,28. The theory of emotional coherence suggests that devaluing conflicting information while favoring or endorsing supporting data can enhance the coherence of existing mental representations, including goals, beliefs, and emotions 29. Localist connectionist models of mental reasoning indicate that coherence-driven mental processing of information can be computationally described as a parallel constraint satisfaction process (PCS) 30–35. Recent PCS models of attitudes offer a practical, formalized framework for understanding the complex structures and processes involved in attitude change, along with a quantitative measure of coherence 36. However, existing approaches, with some notable exceptions such as Simon et al. 37, have mainly relied on aggregated empirical or simulated data, limiting their ability to provide detailed empirical insights at a cognitive-affective level. These models lack the capacity to generate testable predictions about specific mental structures that either encourage or discourage motivated reasoning based on the direction of advocacy (inconsistency or consistency of beliefs and affects) and the nature of the argumentation (rational or emotional) in persuasive communication. Furthermore, most coherence models have not explicitly considered the influence of affective processes associated with motivated reasoning, as documented in empirical studies 25.
In this study, we integrate experimental data with computational modeling to explore the cognitive and affective processes that underlie motivated reasoning in response to persuasive appeals. First, we conducted an online factorial survey experiment, allowing us to analyze whether attitudinal responses to persuasive appeals differ across various levels and valences of emotional message frames. We aimed to test the hypothesis that motivated reasoning contributes to a negative evaluation of information sources and leads to attitude polarization. Second, we simulated the process of information exchange and integration by creating a parallel constraint satisfaction network (PCS) model for each sender (vignette statement) and receiver (participants) in the experiment. PCS models represent issue-specific mental structures as nodes (goals and attitude objects) and connected links (beliefs and affective meanings). For each of the three experimental conditions, we estimated the adjustments made by respondents to their pre-existing network structures based on empirically reported attitude changes. This allowed us to identify specific mental structures and conditions that are most susceptible to motivated reasoning. Furthermore, our approach allowed us to quantitatively assess the theoretically predicted increase in overall coherence resulting from motivated reasoning.
Methods
We conducted a factorial-survey experiment, utilizing statements both advocating for and opposing the use of internal combustion engine (ICE-cars) and electric cars (E-cars). By using a vignette design we aimed to investigate the interplay between pro- and counterattitudinal information, emotional content, and the presence of motivated reasoning. The vignettes described the attitudes and experiences of a hypothetical person advocating for or against ICE- or E-cars under three experimental conditions: rational, emotional, and a combination of emotional and rational argumentation. Example vignettes of the three conditions are provided in the Supplemental Material. The vignettes varied independently along four key attributes: the type of vehicle powertrain (i.e., internal combustion engine or electric engine), beliefs about the contribution of these vehicle types to domain-specific goals (e.g., independence, comfort, eco-friendliness, driving experience, good conscience), and the valence of the attitude object (positive or negative). We created a total of 120 different vignettes by systematically combining these attributes, resulting in 40 vignettes for each experimental condition.
Pilot study
To ensure the specificity and relevance of the characteristics of the three message types, all vignettes underwent pilot testing. Participants in the pilot study (n = 180) were recruited and incentivized through a commercial online research panel in Germany, namely Bilendi & respondi Group. They were randomly assigned to one of three groups, each consisting of 40 vignettes of a specific type (i.e., rational, emotional, or combined). We excluded five individuals due to incomplete data, resulting in a final sample of n = 175 respondents (with 56 in the rational messages group, 59 in the emotional messages group, and 60 in the combined messages group). In the study, participants rated the emotionality and information content of each vignette on a six-point scale, ranging from 'not at all' (1) to 'completely' (6). As anticipated, participants considered the rational messages significantly more informative than the combined statements and the emotional ones (rational: M = 3.58, SD = 0.09; combination: M = 3.19, SD = 0.08; emotional: M = 3.06, SD = 0.08; F(2,172) = 9.79, p < .0001). Additionally, they evaluated emotionally framed messages as having significantly more emotional content than combined and rational statements (emotional: M = 4.23, SD = 0.10; combined: M = 4.13, SD = 0.10; rational: M = 3.80, SD = 0.1; F(2,172) = 5.36, p < 0.005).
Main study
In the main study, we employed a between-subjects design to investigate how individuals respond to various types of persuasive messages, both in support of and in opposition to the use of ICE-cars and E-cars. Our study included a total of 480 participants, with a gender distribution of 47.1%, aged between 18 and 65 years, and a mean age of 46.63 (SD = 13.90). These participants were also recruited by Bilendi & respondi Group. All participants held a valid driver's license, and the majority (89.8%) indicated having at least moderate knowledge of electric vehicles. Participants were randomly allocated to one of three experimental conditions, with each group consisting of 160 participants: i) rational persuasion, which featured vignettes presenting factual and goal-oriented arguments; ii) emotional persuasion, where vignettes conveyed emotional experiences with the different car types; and iii) combined persuasion, involving vignettes that integrated both goal-oriented and emotional arguments. Each participant received one set of vignettes, which was randomly selected from four sets available for each condition. Each set contained 10 vignettes, evenly split between statements about conventional internal combustion engine (ICE) cars and electric cars, maintaining a balance in attitudinal valence.
Before presenting the vignettes, we assessed participants' pre-existing attitudes toward conventional internal combustion engine (ICE-cars) and electric cars (E-cars). For ICE-cars, we used 13 items, and for E-cars, we used 15 items. Participants were asked to rate their agreement or disagreement on a six-point Likert scale in response to statements related to various goal-related aspects of transport mode use (e.g., cost efficiency goal: ‘the cost of driving an E-Car is too expensive for me’; independence goal: ‘Driving an ICE-car means freedom to me’). The items for each transport mode were aggregated into corresponding ICE-car and E-car attitude scales and rescaled to a range of -1 to + 1. To gather the necessary empirical parameters for the PCS model described later, participants were requested to indicate the importance of five goals related to their choices of transport modes (independence, comfort, eco-friendliness, driving experience, good conscience) on a six-point Likert scale. They were also asked to rate the emotional valence of these goals and the transport mode options (ICE-cars and E-Cars) on a semantic differential scale, which ranged from − 4 to + 4. Furthermore, participants provided their beliefs regarding the extent to which these goals could be satisfied by using ICE-cars and E-Cars on a six-point Likert scale.
In the main segment of the factorial survey, the set of 10 vignettes were displayed on separate screens in a random order. After each vignette, four dependent variables were assessed. Participants were required to provide responses on the following: a) Their level of agreement with the statements, rated on a scale from 1 (strongly disagree) to 6 (strongly agree), to measure their cognitive dissonance with the messages, b) the likability of the author, evaluated on a scale from 1 (very unpleasant) to 6 (very likable), aimed at quantifying the extent to which they discounted the source of the messages, c) the competence of the author, rated on a scale from 1 (very low) to 6 (very high), also employed to gauge the discounting of the message source and d) whether they perceived an attitude change in a more positive or negative direction after evaluating the persuasive message, assessed on a seven-point scale ranging from 'much more negative' (1) to 'much more positive' (7).
All procedures conducted in this study adhere to the ethical principles outlined in the 1964 Helsinki Declaration and its subsequent revisions, or equivalent ethical standards. The data collection followed the institutional guidelines of ESOMAR21, as well as the international quality standard ISO 20252:2019. All study participants provided their informed consent to participate, with no impact on their physical or psychological well-being, privacy, or personal rights. The compliance with national regulations governing the protection of human subjects was reviewed and approved by the Ethics Committee of the Research Institute for Sustainability – Helmholtz Centre Potsdam.
We initiated our analysis by examining whether the pre-treatment attitudes toward internal combustion engine cars (ICE-cars) and electric cars (E-cars) varied among the experimental groups. A one-way ANOVA revealed no significant differences between the groups for both attitude objects (ICE-cars: F(1,318) = 0.0, p = .965; E-cars: F(1,318) = 0.28, p = .600). Therefore, we can confidently assume that the subsequent results are not influenced by disparities in prior attitudes across the groups.
To investigate the influence of prior attitudes on the assessment of information sources and subjectively perceived attitude changes, we utilized Pearson correlations. In Fig. 1 (A-D), we present separate correlations between prior attitudes toward conventional and electric cars and each vignette's ratings, categorized into positively and negatively framed messages. It's important to note that these analyses are aggregated across the three message types. The level of agreement with the statements displayed a strong association with participants' prior attitudes (positive: ICE-car: r = .58, p = .001; E-car: r = .66, p = .001; negative: ICE-car: r = .49, p = .001; E-car: r = .43, p = .001). These results reveal robust instances of motivated reasoning for both types of transport modes and message framings (positive/negative). Authors of messages that align closely with participants' pre-existing attitudes are evaluated as more likable and competent compared to authors of information that diverges significantly from those pre-existing attitudes. Interestingly, we observed evidence of a backfire effect. Attitude polarization (Fig. 1D), which signifies the shift and reinforcement of attitudes among individuals with strong pre-existing attitudes in the opposite direction of message valence, was notably stronger for positive information (ICE-car: r = .30, p = .001; E-car: r = .47, p = .001). In contrast, negative messages about cars did not yield significant results (ICE-car: r = .07, p = n.s.) and showed a weaker relationship (E-car: r = 0.15, p = 0.01). Overall, the correlations between prior attitudes and message evaluations were more pronounced for positive statements than for negative ones. This pattern was consistent for positively valenced messages about E-cars compared to ICE-cars, with a contrasting tendency observed for negative statements about the two vehicle types.
Subsequently, we conducted a separate analysis of the effects of positive and negative persuasive messages. In Fig. 2 (A-D), we present the average responses to all vignette-related measures, categorized by condition and valence. Consistent with the notion that positive messages tend to be more appealing and persuasive than negatively valenced ones 38, participants' ratings, on average, were significantly higher for positive information. Separate dependent t-tests revealed substantial differences across all three conditions concerning: agreement with the statements (rational: t(159) = 6.50, p < .001; emotional: t(159) = 4.56, p < .001; combined: t(159) = 6.88, p < .001), perceived competence of the sender (rational: t(159) = 5.88, p < .001; emotional: t(159) = 4.83, p < .001; combined: t(159) = 7.76, p < .001), likability of the sender (rational: t(159) = 6.06, p < .001; emotional: t(159) = 6.96, p < .001; combined: t(159) = 8.69, p < .001), and perceived attitude shift in response to the information (rational: t(159) = 5.69, p < .001; emotional: t(159) = 2.60, p < .01; combined: t(159) = 4.24, p < .001).
When comparing the impact of different types of persuasive appeals on perceived attitude changes (Fig. 2D), we employed one-way ANOVA followed by Tukey's HSD post-hoc tests (p < .05). Surprisingly, we observed no significant differences either among the groups or in any pairwise comparisons between groups, whether for positive or negative messages (F(2,477) = 0.15, p = .8590, ηp2 = .0003), nor for negative messages alone (F(2,477) = 1.80, p = .166, ηp2 = .0006). Conversely, when assessing evaluations regarding sender characteristics, we detected notable effects across all types of persuasive appeals. For rational statements, participants indicated the highest levels of agreement (positive: F(2,477) = 8.85, p = .001; negative: F(2,477) = 13.45, p = .001), sender competence (positive: F(2,477) = 7.23, p = .0001, ηp2 = .029; negative: F(2,477) = 19.40, p = .001, ηp2 = .075), and likability (positive: F(2,477) = 2.36, p = .10, ηp2 = .009; negative: F(2,477) = 20.04, p = .001, ηp2 = .077), regardless of whether the message was positively or negatively framed. Responses to emotional information were slightly, though not significantly, more positive than those to the combined condition, as indicated by the post-hoc test.
However, a more detailed analysis of covariance (ANCOVA), with control variables for perceived dissonance (i.e., agreement), prior attitudes, and their two-way interactions with the treatment, painted a nuanced picture for all dependent variables. The main effect of message type on perceived attitude change showed slight significance (positive: F(2,477) = 3.29, p = .0379, ηp2 = 0.014; negative: F(2,477) = 3.80, p = .0231, ηp2 = .016). Notably, covariates, such as prior attitude towards E-cars (positive: F(1,478) = 9.32, p = .0024, ηp2 = .019; negative: F(1,478) = 11.34, p = .0008, ηp2 = .019), and agreement with the statement, had a significant effect on perceived attitude change (positive: F(1,478) = 215.83, p = .0001, ηp2 = 0.312; negative: F(1,478) = 45.23, p = .0001, ηp2 = .087). However, prior attitudes toward ICE-cars showed no significant influence on perceived attitude change (positive: F(1,478) = 0.76, p = .7829, ηp2 = .018; negative: F(1,478) = 0.93, p = .3352, ηp2 = .017). Furthermore, the results revealed a small yet significant interaction between message type and agreement for negatively framed statements (F(2,477) = 3.80, p = .0023, ηp2 = .018), and a non-significant interaction for positively valenced statements (F(2,477) = 2.15, p = .117, ηp2 = .009). When participants exhibited more dissonance with the information, they perceived combined messages as the most persuasive. In contrast, when respondents showed greater agreement with the statements, rationally and emotionally framed messages led to stronger perceived attitude changes.
A second ANCOVA examining the perceived competence of the message sender, with prior attitudes, agreement, and their interactions as covariates, showed that the effect of message type became non-significant for positively framed statements (F(2,477) = 0.58, p = .5595, ηp2 = .002). However, it remained significant for negatively framed statements (F(2,477) = 5.12, p = .0063, ηp2 = .021), with rational messages leading to the highest perceived competence. No significant interaction between message type and agreement was found, but the main effect of agreement was strongly associated with attributed competence (positive: F(1,478) = 680.07, p = .0001, ηp2 = .588; negative: F(1,478) = 807.27, p = .0001, ηp2 = .629). Additionally, prior attitudes towards E-cars significantly influenced competence attribution for negatively framed messages (positive: F(1,478) = 0.39, p = .5278, ηp2 = .001; negative: F(1,478) = 25.12, p = .0001, ηp2 = .050). This finding aligns with previous research indicating the impact of motivated reasoning on expertise perception and suggests that competence attribution to the sender was primarily driven by perceived dissonance with persuasive appeals rather than the type of argumentation or prior attitudes.
In summary, the empirical results demonstrate the influence of prior attitudes and dissonance on information evaluation, providing support for the proposed motivated reasoning processes. This effect was most pronounced for negatively valenced combined statements, resulting in more negative judgments about senders' characteristics and lower persuasiveness compared to positive rational or emotional messages. The impact of different message types was primarily moderated by the level of dissonance or disagreement with the statements.
Model structure
In our experiment, we observed the process of motivated reasoning, wherein respondents tended to evaluate both the messages and the sender of the information in a way that aligned with their existing attitudes. However, these results beg the question of the underlying cognitive-affective structures and mechanisms that facilitate this process and the extent to which these structures are modified when attitudes change. To address this, we conducted an in-depth analysis of the observed effects by developing a computational model of attitude formation and change based on the parallel constraint satisfaction (PCS) theory of emotional coherence 29. PCS, in general, is a connectionist model that conceptualizes cognitive consistency, drawing from established theories like cognitive dissonance 28 and congruity theory 40. Connectionist models posit that information processing is guided by a network of pre-existing beliefs and goals. When new evidence is presented, it is integrated through an automatic parallel activation spread in this network. The process is geared toward maximizing the satisfaction of constraints, reflecting the innate desire to minimize inconsistencies and uphold prior beliefs and emotional responses. Prior simulation studies have demonstrated that motivated reasoning can be likened to a process of parallel constraint satisfaction (e.g. 36).
We designed the experimental framework as a one-way communication between two PCS models, each serving as a sender (i.e., the vignette statement) and a receiver (i.e., the respondent). Both models had identical network structures. The process involved four key steps, visually depicted in Fig. 3B. Our starting point is a PCS network that represents the attitudes of the respondents. Drawing from the general PCS model structure known as HOTCO (an abbreviation for "HOT COherence") as proposed by Thagard 29, our computational model encompassed three interconnected components of attitudes (cf. 41): cognitive (pertaining to beliefs about the relationship between goals and action options), affective (related to feelings about goals and action options), and conative (concerning behavioral tendencies regarding action options). In localist models like PCS networks, psychological concepts (e.g., goals or actions) are represented as individual units, and the links connecting these units denote the relationships between the concepts (e.g., beliefs or feelings). Our model, illustrated in Fig. 3A, featured a central activation node (on the left) and three layers of nodes (units) representing valence (the second layer from the left, signifying the pleasantness or unpleasantness of concepts), goals (the third layer from the left, indicating the motives behind choices in transport modes), and options for action (on the right, representing various transport mode choice options). Mutual constraints between these concepts were expressed through bidirectional links between nodes within each layer. Connections between goal and option nodes could either be excitatory (reflecting supportive beliefs, e.g., the goal of personal comfort aligns with the use of an E-car) or inhibitory (indicating conflicting beliefs, e.g., the goal of environmentally responsible travel behavior doesn't align with the use of an ICE-car). Links between the valence node and goal or option units signaled positive (excitatory) or negative (inhibitory) affective associations with concepts. The strength of beliefs or affect was quantified by connection weights ranging from − 1.0 to 1.0. Lines connecting the general activation node and the goal units represented the personal importance of a goal, with weights ranging from 0 to 1.0. In this structure, each PCS network of the participants consisted of five goal nodes (independence, comfort, eco-friendliness, driving experience, good conscience) and two option nodes, representing attitudes towards the use of ICE- and E-cars. Based on this framework, we modeled the individual attitudes of each participant (n = 480) as a unique PCS model. These models were parameterized using the participants' survey responses, which included information about the importance of goals, beliefs, and affective meanings of concepts (as described in the Factorial Survey section). All other model parameters were consistent with our prior studies (cf. 42).
An iterative updating algorithm simulates the process of attitude formation by spreading activations through the network, aiming to maximize coherence while considering all constraints within the given network structure 43. At the initial time, t = 0, all nodes start with an activation close to zero, except for the general activation node, which is set to 1.0. The units' activations are updated iteratively and in parallel until they reach a stable pattern of activation, with the degree of change falling below a predefined threshold. This updating process follows the equation:
where aj(t) is the activation of the unit j at iteration t. The constant parameter d (0.05) is the rate of decay of activation for each unit at every cycle, min is the minimum activation (-1.0) and max is the maximum activation (+ 1.0). The incoming activation from other nodes is computed by the connection weight wij between each unit i and j, the net input netj to a node
Updating continues until a stable level of activation is achieved, typically requiring fewer than 100 iterations. The ultimate activation of option nodes reflects the evaluation of an object, with higher positive values indicating a more positive attitude and higher negative values suggesting a more negative attitude. At this stage, the network's overall coherence, represented as ‘Energy’ 44,45 is calculated by:
Where wij is the weight between each unit i and j, and ai (t) is the activation of unit i at time (t). This equation operates as follows: when the product of activation for two units aligns with the constraint between them, the network's energy decreases; conversely, when the product of activation for two nodes contradicts the constraint between them, the network's energy increases (see Fig. 2c). The network's overall coherence is high when all cognitive and affective elements associated with one option align, while they do not align with the second option. Conversely, when beliefs and affect linked to both options are equally inconsistent, meaning that both option units are connected to a similar number of excitatory and inhibitory weights, attitudinal ambivalence arises, making it challenging to favor one option and subsequently reducing coherence.
In a second step, we applied the same PCS model, depicted in Fig. 2a, to represent persuasive messages. Here, we incorporated the unique attributes of each individual vignette statement as weighted constraints in the network. For instance, in a persuasive appeal promoting the use of E-cars (o1) and emphasizing compatibility with the goals of environmentally friendly (g1) and comfortable (g2) travel behaviour, the connection weights were set as follows: links between goals and the E-car option nodes (wg1−o1, wg2−o1) and between valence, goals, and E-car option nodes (wv−g1, wv−g2, wv−o1) were initialized with + 1.0, while all other weights were set to 0.01. In contrast to the PCS model of respondents described above, the activation of nodes is not relevant in this case because it doesn't factor into the subsequent process of information exchange. In the third step, we implemented the three modes of persuasive communication, corresponding to the experimental conditions, as a transfer of different types of information or network elements between the sender and receiver (see Fig. 2b). In the rational persuasion mode, specific link weights between goal and option nodes (i.e. beliefs) corresponding to the argumentation of the respective vignette text were communicated. In the rational persuasion mode, specific link weights between goal and option nodes (i.e., beliefs) aligned with the arguments presented in the respective vignette text were transmitted. In the emotional persuasion mode, connection weights between valence and goal as well as valence and option nodes (i.e., affective meanings) were conveyed. The information transfer in the third mode, combined persuasion, included both types of weights (i.e., beliefs and affective meanings) from the preceding mechanisms. In line with the experimental procedure, we generated simulations for 10 unique persuasive attempts per receiving agent, resulting in a total of 4800 (1600 per condition) simulated vignette treatments.
In the last step, we modeled the adaptation of attitudes resulting from persuasive communication attempts as modifications of connection weights within the receivers' network structure. Unlike previous learning processes in PCS networks 46, these weight adjustments were specific, meaning that only links transmitted by the sender in the respective conversation and condition were altered in the receivers' networks. For instance, in a rational persuasive communication regarding the use of electric cars, which featured positive arguments about comfort and eco-friendliness, only the connections between the goal units (comfort and eco-friendliness) and the option unit (E-car) were modified. Drawing from the literature on network learning 47,48, we employed a variant of error-correcting learning rules to adjust link weights based on the degree of error between simulated and empirically reported attitude changes. These adjustments were calculated using a standard backpropagation algorithm aimed at minimizing the total sum squared error function D across all vignette treatments, defined as:
where Asim represents the activation of the respective option unit in response to new information i.e. receivers simulated attitude towards a transport mode. Aemp is the empirically reported attitude after vignette treatment, calculated as by respondents’ original attitude plus the self-reported attitude change, which has been rescaled to values ranging between − 1 and + 1. N is the total number of simulated vignette treatments.
The minimization of the error function is achieved by fitting free parameters to the following equation.
∆w = b 1 ∆ws,r + b2 wr + b3 ∆ws,r wr + b4 wr 2 + b5 ∆ws,r2 (5)
where bi are the parameters systematically iterated in in given upper and lower limits of -1 and + 1. ∆ws,r is the difference (i.e. dissonance) between the corresponding link weights of sender and receiver, i.e. the difference between existing beliefs and affects and new information. wr is the connection weight in the PCS network of the receiver before communication. The function also includes the interaction between ∆ws,r and wr and is approximated up to quadratic order. Best fitting parameters (b1 − 5) were estimated by iterating systematically through the set free parameters in the permitted range between − 1.0 and + 1.0.
Using the parameterized PCS models, we conducted a simulation of each respondent's attitude prior to the influence of persuasive messages. To validate whether our model accurately reflects the empirically observed attitude patterns before persuasion attempts, we correlated the activations of the option nodes for ICE cars and E-cars with the corresponding self-reported attitudes of the respondents. These results demonstrated a strong resemblance between the simulations and the empirical findings (ICE car: r = 0.93, p = 0.001; E-car: r = 0.93, p = 0.001). Furthermore, we investigated the relationship between the simulated coherence of networks, quantified by the energy of PCS networks, and the respondents' simulated prior attitudes across different conditions. Our findings revealed an inverted U-shaped pattern (Fig. 4A), which was also supported by the results of quadratic regression models for ICE cars (R2 = .50, F(1, 478) = 476.4, p < .001) and for E-cars (R2 = .57, F(1, 478) = 631.7, p < .001). As predicted by the PCS approach, both strongly negative and strongly positive prior attitudes were associated with low levels of energy, indicating that these attitudes were consistent, stable, and resistant to change. In Fig. 4B, we present the average empirically measured agreement with statements (i.e., dissonance) for each condition alongside the corresponding coherence obtained from the simulations. These results exhibit a similar inverted U-shaped pattern to the one observed for prior attitudes. (rational pos.: R2 = .14, F(1, 318) = 50.59, p < .001; rational neg.: R2 = .05, F(1, 318) = 16.48, p < .001; emotional pos.: R2 = .10, F(1, 318) = 36.52, p < .001; emotional neg.: R2 = .05, F(1, 318) = 16.86, p < .001; combination pos.: R2 = .23, F(1, 318) = 93.79, p < .001; combination neg.: R2 = .08, F(1, 318) = 30.23, p < .001). Across experimental conditions and for both positively and, to a somewhat lesser degree, negatively valenced messages, high and low agreement was associated with low levels of energy, while moderate agreement corresponded to the highest energy levels. In other words, individuals with low attitudinal ambivalence, indicating greater coherence, tended to exhibit stronger (dis)agreements with (counter)pro-attitudinal positions, as one might expect.
Figure 4. Legend see below.
To what extent are the cognitive-affective structures that underlie attitudes influenced by persuasive messages with different levels of emotionality? Fig. 4C displays the average weight changes (∆w of beliefs and affect in each experimental condition that minimize the respective error functions D, which represent the gap between simulated and measured attitudes following persuasive communication. The results indicate across conditions varying patterns of belief and affect convergence and divergence. The results reveal diverse patterns of belief and affect convergence and divergence across conditions. These patterns depend on the strength and valence of the prior beliefs and affect held by receivers, as well as their disparity from those of the sender. In the rational condition, for instance, when the sender and receiver maintain strong opposing prior beliefs (sender: ws = + 1.0/-1.0 and receiver wr = − 1.0/+1.0), receivers tend to adjust their beliefs in a direction contrary to the new information, reinforcing the strength of their initial beliefs. However, the extent of this backfire effect differs between receiving agents with prior positive and negative beliefs. Interestingly, in response to opposing positive information, this backfire effect leads to more significant negative weight adjustments than belief changes induced by negative messages that align with pre-existing negative beliefs. Conversely, when the sender supports the receivers' prior beliefs, the weights are updated in the direction of the sender.
The most pronounced effects of belief polarization/convergence and valence asymmetry are observed when two agents exclusively exchange affect, particularly in the emotional condition and, to a slightly lesser extent, in the combined condition. Intriguingly, in the emotional condition, opposing negative messages (i.e., when the receiver holds strong positive affect) do not lead to polarization but instead prompt a convergence of posterior affect towards the message.
To ascertain the impact of weight adjustments in response to persuasive messages on the coherence of PCS networks, we conducted a comparison of the average energy levels among all agents before and after the treatment for each condition (Fig. 4D). Prior to receiving persuasive messages, no significant differences in coherence levels were observed between conditions (F(2,477) = 0.19, p = .8824, η2 = .001). However, exposure to persuasive information resulted in a statistically significant decrease in energy in the emotional and combined conditions, but not in the rational condition (F(1,479) = 410.96, p = .0001, η2 = .0462). These findings align with the expectations of the coherence maximization process, indicating that beliefs and affect are adjusted in a manner that enhances the coherence among existing mental representations. Notably, this effect is most pronounced when both rational and emotional communication strategies are employed simultaneously.
Our research had two primary goals. First, we aimed to investigate the impact of different levels of emotional content in persuasive messages on motivated reasoning. Our findings revealed that regardless of message framing, motivated reasoning influenced the persuasiveness of statements, resulting in dissonance with the message and a tendency to discredit the information source, particularly when individuals encountered opposing viewpoints. Moreover, we identified backfire effects, where individuals with strongly held views and high dissonance actually reinforced their existing attitudes, consistent with previous research 49. A novel aspect of our study was that the extent of this process was moderated by message type and the valence of the advocated positions. Biased reasoning was most prevalent when participants encountered negatively framed appeals and when messages employed a combined rational and emotional tone, rather than exclusively rational or emotional approaches. The persuasiveness of messages was inversely related to the degree of motivated reasoning. Additionally, our results suggested that the impact of motivated reasoning on perceptions of the source was primarily determined by dissonance (agreement) with persuasive appeals, rather than by prior attitudes.
Our second goal was to elucidate the observed effects using computational mechanisms related to emotional coherence, achieved by extending an existing parallel constraint satisfaction (PCS) model 29 with a learning algorithm. Our simulation results indicated that modifications of the existing mental representations underlying attitude changes exhibited significant variations across valence and different message types. Consistent with prior experiments 50, we found that negative affect and beliefs were more resistant to counterattitudinal persuasive attempts. Furthermore, we demonstrated that negative mental structures had a greater tendency to lead to backfire effects, which were most pronounced when emotionally positive appeals were encountered. These findings suggested that mental structures respond to and integrate information with varying emotional qualities and intensities. Recent neuroimaging studies supported the idea that positively and negatively framed information is processed differently in the brain and is differentially used to update prior affect and beliefs 51.
Previous research on PCS models52, which assumed coherence maximization as the core mechanism of information integration, argued that this process naturally operates to increase the coherence of existing mental representations. By employing a more realistic learning PCS network model, we complemented these earlier studies by providing evidence that the level of coherence shift associated with attitude changes depended on the type of mental structure involved in information integration. Information that solely influenced the cognitive structures of attitudes resulted in no increase in overall coherence after updating beliefs. In contrast, when information involved affective and, particularly, both affective and cognitive elements of attitudes, global coherence significantly increased. These findings emphasized the role of affect in reducing dissonance and ambivalence toward options for action and supported the notion that decision and attitude formation are driven by maximizing the emotional coherence of mental structures, combining cognition and affect, relevant to the decision at hand 29. Moreover, the pattern of coherence shifts that is inversely related to the observed tendency of motivated reasoning suggests an effortful process underlying the integration of new information, in particular when both strong cognitive and affective components of attitudes are involved. Furthermore, the pattern of coherence shifts, inversely related to the observed tendency of motivated reasoning, suggested an effortful process underpinning the integration of new information, especially when strong cognitive and affective components of attitudes were involved. The greater the effort required for a person to arrive at a coherent conclusion, the lower the persuasiveness and credibility of the message source. Future research could integrate additional elements such as specific identity concepts into the model, further investigating the interaction between the goal of attitude protection and the goal of identity protection, both of which moderate responses to new information.
Our findings make a valuable contribution to the growing body of literature on the role of motivated reasoning in information processing and attitude formation. While earlier studies focused on the influence of belief strength, values, political or group identity on attitude responses and group polarization in the context of biased processing of evidence and persuasive appeals 53, we extended this work by demonstrating that positively valanced and cognitively framed messages can mitigate reasoning bias and prevent polarization between communication partners. Our results suggested that reducing the emotional content of messages that contradict individuals' beliefs and affect, while increasing emotional intensity in persuasive appeals that align with respondents' representations, may offer an effective strategy to enhance the persuasiveness of communication and mitigate backfire effects.
Additionally, our research contributes to the field of computational models of opinion dynamics and social influence 54. Our findings underscore the importance of studying and comprehending the microfoundations of processes underlying motivated reasoning. Our modeling framework exemplifies how existing theoretical models of attitude and opinion formation can be extended to incorporate psychologically realistic processes related to information processing, with empirical support as a basis. We anticipate that our approach paves the way for future research aimed at understanding how individual reasoning influences dynamics at a population level. This approach has practical implications, such as in policy contexts related to the transition of mobility, where it can enhance the understanding of how policymakers can effectively communicate policy issues, accelerate public comprehension of available evidence, and foster informed policy discussions 42,55.
.
Acknowledgements
We thank Jakob Dordevic for his assistance in data collection. This study was partly funded by the German federal government (grant # BMBF 16j1610), the funding programme “Open Access Publicationskosten” of the German Research Foundation (#491075472) and the Research Institute for Sustainability – Helmholtz Centre Potsdam (RIFS).
Author Contributions
I.W. and T.S designed the study and discussed the results. I. W. performed the data analysis and programmed and executed the simulations. I. W. drafted the manuscript, and T. Schröder provided critical revisions. Both authors approved the final version of the manuscript for submission.
Data availability
Data is available from the corresponding author upon request.
Competing interests
The authors declare no competing interests.
- Albarracin, D. & Shavitt, S. Attitudes and Attitude Change. Annu Rev Psychol 69, 299–327 (2018).
- Falk, E. & Scholz, C. Persuasion, Influence, and Value: Perspectives from Communication and Social Neuroscience. Annu Rev Psychol 69, 329–56 (2018).
- Pettifor, H., Wilson, C., Axsen, J., Abrahamse, W. & Anable, J. Social influence in the global diffusion of alternative fuel vehicles – A meta-analysis. J Transp Geogr 62, 247–261 (2017).
- Brady, W. J., Wills, J. A., Jost, J. T., Tucker, J. A. & Van Bavel, J. J. Emotion shapes the diffusion of moralized content in social networks. Proceedings of the National Academy of Sciences 114, 7313–7318 (2017).
- Guilbeaulta, D., Beckera, J. & Centola, D. Social learning and partisan bias in the interpretation of climate trends. Proceeding of the national academy of scienes 115, 9714–9719 (2018).
- Kahan, D. M., Braman, D., Gastil, J., Slovic, P. & Mertz, C. K. Culture and identity-protective cognition: Explaining the white male effect in risk perception. Journal of Empirical Law Studies 4, 465–505 (2007).
- Hart, W. et al. Feeling Validated Versus Being Correct:A Meta-Analysis of Selective Exposure to Information. Psychologcial Bulletin 135, 555–588 (2009).
- Taber, C. S., Cann, D. & Kucsova, S. The motivated processing of political arguments. Polit Behav 31, 137–155 (2009).
- Zumofen, G., Stadelmann-Steffen, I. & Bühlmann, M. No, It Is Not All About Selective Exposure: Information Selection Strategies in Referendums. Polit Behav (2023) doi:10.1007/S11109-023-09893-8.
- Huntsinger, J. R. Incidental Experiences of Affective Coherence and Incoherence Influence Persuasion. Pers Soc Psychol Bull 39, 792–802 (2013).
- Kunda, Z. The case for motivated reasoning. Psychol Bull 108, 480–498 (1990).
- Zhou, J. Boomerangs versus Javelins: How Polarization Constrains Communication on Climate Change. Env Polit 25, 788–811 (2016).
- Chinn, S. & Hart, P. S. Climate Change Consensus Messages Cause Reactance. Environ Commun 17, 51–59 (2023).
- Schmittat, S. M. & Englich, B. If you judge, investigate! responsibility reduces confirmatory information processing in legal experts. Psychology, Public Policy, and Law 22, 386–400 (2016).
- Lord, C. G., Ross, L. & Lepper, M. R. Biased assimilation and attitude polarization: The effects of prior theories on subsequently considered evidence. J Pers Soc Psychol 37, 2098–2109 (1979).
- Sherrod, D. R. Selective Perception of Political Candidates. Public Opin Q 35, 554 (1971).
- Koehler, J. J. The Influence of Prior Beliefs on Scientific Judgments of Evidence Quality. Organizational Behavior and Human Decision Processes vol. 56 28–55 Preprint at https://doi.org/10.1006/obhd.1993.1044 (1993).
- Badas, A. Motivated Reasoning and Attitudes Towards Supreme Court Confirmation Hearings: Evidence from Five Nominations and an Experiment. Polit Res Q 76, 540–552 (2023).
- Kahan, D. M. et al. The polarizing impact of science literacy and numeracy on perceived climate change risks. Nat Clim Chang 2, 732–735 (2012).
- Ecklund, E. H., Scheitle, C. P., Peifer, J. & Bolger, D. Examining Links Between Religion, Evolution Views, and Climate Change Skepticism. Environ Behav 49, 985–1006 (2017).
- Jacks, J. Z., Cameron, K. A., Jacks, J. Z. & Cameron, K. A. Strategies for Resisting Persuasion. Basic Appl Soc Psych 25, 145–161 (2003).
- Bail, C. et al. Exposure to opposing views can increase political polarization: evidence from a large-scale field experiment on social media. Proceeding of the national academy of scienes 117, 9216–9221 (2018).
- Druckman, J. N., Peterson, E. & Slothuus, R. How elite partisan polarization affects public opinion formation. American Political Science Review 107, 57–79 (2013).
- Merkley, E. & Stecula, D. A. Party Cues in the News: Democratic Elites, Republican Backlash, and the Dynamics of Climate Skepticism. Br J Polit Sci 51, 1439–1456 (2021).
- Taber, C. S. & Lodge, M. Motivated Skepticism in the Evaluation of Political Beliefs. Am J Pol Sci 50, 755–769 (2006).
- Druckman, J. N. & Bolsen, T. Framing, motivated reasoning, and opinions about emergent technologies. Journal of Communication 61, 659–688 (2011).
- Nisbet, E. C., Cooper, K. E. & Garrett, R. K. The Partisan Brain: How Dissonant Science Messages Lead Conservatives and Liberals to (Dis)Trust Science. Annals of the American Academy of Political and Social Science 658, 36–66 (2015).
- Festinger, L. The theory of cognitive dissonance. (Stanford University Press, 1957).
- Thagard, P. Hot thought: Mechanisms and applications of emotional cognition. (MIT Press, 2006).
- Thagard, P. & Verbeurgt, K. Coherence as Constraint Satisfaction. Cogn Sci 22, 1–24 (1998).
- Read, S. J. & Simon, D. Parallel constraint satisfaction as a mechanism for cognitive consistency. in Cognitive consistency: A fundamental principle in social cognition. 66–86 (Guilford Press, 2012).
- Glöckner, A. & Betsch, T. Decisions beyond boundaries: when more information is processed faster than less. Acta Psychol (Amst) 139, 532–42 (2012).
- Jekel, M., Glöckner, A. & Bröder, A. A new and unique prediction for cue-search in a parallel-constraint satisfaction network model: The attraction search effect. Psychol Rev 125, 744–768 (2018).
- Kunda, Z. & Thagard, P. Forming impressions from stereotypes, traits, and behaviors: A parallel- constraint- satisfaction theory. Psychol Rev 103, 284–308 (1996).
- Heck, D. W. & Erdfelder, E. Linking process and measurement models of recognition-based decisions. Psychol Rev 124, 442–471 (2017).
- Monroe, B. M. & Read, S. J. A general connectionist model of attitude structure and change: the ACS (Attitudes as Constraint Satisfaction) model. Psychol Rev 115, 733–59 (2008).
- Simon, D., Stenstrom, D. M. & Read, S. J. The coherence effect: Blending cold and hot cognitions. J Pers Soc Psychol 109, 369–394 (2015).
- Nordmo, M. & Selart, M. The asymmetrical force of persuasive knowledge across the positive–negative divide. Front Psychol 6, 1324 (2015).
- Kahan, D. M., Jenkins‐Smith, H. & Braman, D. Cultural cognition of scientific consensus. J Risk Res 14, 147–174 (2011).
- Osgood, C. E. & Tannenbaum, P. H. The principle of congruity in the prediction of attitude change. Psychol Rev 62, 42–55 (1955).
- Ajzen. Attitudes structure and behavior. in Attitude Structure and Function (eds. Pratkanis, A. R., Breckler, S. J. & Greenwald, A. G.) 241–274 (Erlbaum, 1989).
- Wolf, I., Schröder, T., Neumann, J. & de Haan, G. Changing minds about electric cars: An empirically grounded agent-based modeling approach. Technol Forecast Soc Change 94, 269–285 (2015).
- McClelland, J. L. & Rumelhart, D. E. An interactive activation model of context effects in letter perception: Part I. An account of basic findings. Psychol Rev 88, 365–405 (1981).
- Hopfield, J. J. Neural networks and physical systems with emergent collective computational abilities. Proceeding of the national academy of scienes 79, 2554–2558 (1982).
- Hopfield, J. J. Neurons with graded response have collective computational properties like those of two-state neurons. Proceedings of the National Academy of Sciences 81, 3088–3092 (1984).
- Read, S. J. & Monroe, B. M. Modeling cognitive dissonance as a parallel constraint satisfaction network with learning. Cognitive dissonance: Reexamining a pivotal theory in psychology (2nd ed.). 197–226 (2019) doi:10.1037/0000135-010.
- Rescorla, R. A. & Wagner, A. R. A theory of Pavlovian conditioning: Variations in the effec- tiveness of reinforcement and nonreinforcement. in Classical conditioning II: Current research and theory (eds. Black, A. H. & Prokasy, W. F.) 65–99 (Appleton-Century-Crofts, 1972).
- Read, S. J. & Urada, D. I. A neural network simulation of the outgroup homogeneity effect. Personality and social psychology review 7, 146–69 (2003).
- Nyhan, B. & Reifler, J. When corrections fail: The persistence of political misperceptions. Polit Behav 32, 303–330 (2010).
- Bizer, G. Y. & Petty, R. E. How we conceptualize our attitudes matters: The effects of valence framing on the resistance of political attitudes. Polit Psychol 26, 553–568 (2005).
- Sharot, T. & Garrett, N. Forming Beliefs: Why Valence Matters. Trends Cogn Sci 20, 25–33 (2016).
- Glöckner, A. & Betsch, T. Modeling option and strategy choices with connectionist networks : Towards an integrative model of automatic and deliberate decision making. Judgm Decis Mak 3, 215–228 (2008).
- Druckman, J. N. & McGrath, M. C. The evidence for motivated reasoning in climate change preference formation. Nature Climate Change 2019 9, 111–119 (2019).
- Flache, A. et al. Models of Social Influence: Towards the Next Frontiers. Journal of Artificial Societies and Social Simulation 20, 2 (2017).
- Schröder, T. & Wolf, I. Modeling multi-level mechanisms of environmental attitudes and behaviours: The example of carsharing in Berlin. J Environ Psychol 52, 136–148 (2017).
No competing interests reported.