Participants were recruited through flyers displayed in public locations such as university buildings, supermarkets, general practitioner practices, pharmacies, and hospitals, as well as through social media groups, the SONA system (Research Participation System of the Philipps-University of Marburg; https://www.sona-systems.com/), and mailing lists at both Philipps-University of Marburg (students and employees) and other universities in German-speaking countries (i.e. Luxemburg and Austria). In addition, support groups for individuals with depression and anxiety disorders were contacted. Recruitment occurred from April 2021 to February 2022. For each participant, one Euro was donated to an organization working to reduce the stigma surrounding mental illness. Psychology students could receive credit points for their bachelor’s degree.

The target sample included individuals with mild depressive and/ or anxiety symptoms. A score between 5-9 on the German version of the Patient Health Questionnaire-9 (PHQ-9; Gräfe et al., 2004) was considered indicative of mild depression. Alternatively, a score between 8-25 on the Beck Anxiety Inventory (BAI; Margraf & Ehlers, 2007), indicating mild to moderate anxiety not clinically relevant, was accepted. These cutoffs were selected for ethical reasons, as the planned behavioral experiments by participants were not controlled by a licensed therapist. Participants were required to be at least 18 years old, have sufficient proficiency in the German language, and have access to a personal email account.

The literature suggests that online interventions tend to have small to medium effect sizes (Lintvedt et al., 2013; Sander et al., 2016; Zhou et al., 2016). Prior power analyses were calculated for a MANOVA with repeated measures on within- and between-subject factors, using an effect size of f = 0.2 and a power of .80, indicated that 152 participants would be needed. This power analysis was conducted as part of the study’s preregistration (Ewen et al., 2021S) to ensure that the sample size was adequate for detecting meaningful effects given the expected variance in cognitive immunization processes. The selected parameters were based on previous research on online psychological interventions with comparable effect sizes.

The study was conducted entirely online using the survey platform formr (Arslan et al., 2020). At each measurement point, the questionnaires were automatically sent via email to participants through formr. Email addresses were collected at the beginning of the study and stored pseudonymously by the program. Upon clicking on the study link, participants were directed to the study information and informed consent page, which they had to agree to before proceeding. Participants were then automatically randomized by the program (Arslan et al., 2020) into one of three groups: the EFPI group, the ACG, or the CG (see below). Additionally, all the participants were asked to complete baseline questionnaires (see variables for details). After the baseline assessment, the procedures varied among the groups (see Figure 1). For the EFPI group, a psychoeducation video was presented immediately after the completion of the baseline questionnaires. This video covered the role of expectations in shaping behavior and emotions, as well as the impact of dysfunctional expectations, those that contribute to dysfunctional behaviors and long-term negative consequences. Three days later, participants were for the first time introduced to behavioral experiments, designed to test their burdensome situation-specific expectations (see Instructions, Table A1, Appendix). 'Situation-specific expectations' refer to beliefs about particular situations or events that may be unrealistic or rigid, contributing to maladaptive responses to those situations. The chosen expectation had to be very specific and testable over the following three days. Participants documented the planned experiment, their specific expectation regarding the situation, and their level of belief in this expectation (rated from 0-100%). They also recorded the associated emotion. Furthermore, they were asked to reflect on their future behavior in the given situation by considering how their actions could either confirm or disconfirm the expectation (How can I confirm/disconfirm the expectation through my behavior?). After three days, participants evaluated the experiment and planned a new one. This process was repeated twice a week for four weeks.

For the ACG, the same psychoeducation video was shown. Participants received a standardized psychoeducational intervention aimed at promoting cognitive restructuring, but without the active expectation testing component. The intervention focused on providing participants with knowledge about cognitive biases and common distortions, aiming to reduce maladaptive thinking patterns. This group was chosen to account for the effects of general psychoeducation, and we expected that any changes in expectations in this group would be less pronounced compared to the EFPI group, given the lack of behavioral experimentation. They were asked every three days, over four weeks, to document their burdensome expectations in the days between the questionnaires and to evaluate them as helpful or unhelpful, as well as noting the associated emotions and behaviors.

The CG did not receive any psychoeducation or interim questionnaires.

After four weeks, all groups completed the post-questionnaires. A follow-up measurement was conducted four weeks later.

The animated video using the program Powtoon (Puspitarini et al., 2019) guided participants through the five key questions (Anne-Catherine Ewen, 2021; https://www.youtube.com/watch?v=q_L56tmckos): What are expectations? Why do humans have expectations? How do expectations arise? Why do certain expectations remain stable? When are expectations causing problems? The questions were addressed using the basic principles of the ViolEx model and psychoeducational content based on Rief and Glombiewski (2016). The video explained that expectations were future-directed thoughts that could be neutral, positive, or negative and interact with one’s environment. They were arising automatically and were not always consciously recognized. Expectations were shaped by experience, social influences, and individual differences. The evolutionary advantage of planning behavior to prevent harm was highlighted. While the stability of functional expectations would be adaptive, the persistence of dysfunctional expectations could lead to distress by reinforcing avoidance behaviors. Avoidance and immunization processes contribute to the rigidity of expectations, making them resistant to change. Avoidance reduces exposure to corrective experiences, leading to a distorted perception of reality. Even when an expectation is violated, the brain may dismiss the event as insignificant, preventing the correction of distorted perceptions of reality. The video emphasized the importance of consciously observing negative, burdensome, dysfunctional expectations and periodically testing them.

The cognitive intervention was implemented in both the EFPI and ACG groups. It was based on the SORC model, commonly used in cognitive-behavioral therapy (Borg-Laufs, 2020), which facilitates a structured analysis of situations by identifying thoughts, physical sensations, emotions, behaviors, and consequences. Participants selected a burdensome expectation they had in the past three days. They were asked to identify the emotions and the physical sensations triggered by this specific expectation, describe their behavioral response, and reflect on its consequences. The cognitive intervention aimed to help participants become aware of their expectations, understand their influence on emotions and behavior, and encourage a more flexible adaptation of expectations over time. In the EFPI group, this intervention was combined with behavioral experiments to actively test and modify expectations, while in the ACG group, participants engaged only in cognitive reflection without direct behavioral testing. The intervention was designed to reduce cognitive immunization and promote more adaptive expectation processes.

Behavioral experiments were conducted exclusively by the EFPI group. The instructions for the behavioral experiments were adapted from Rief and Glombiewski (2016). Participants selected a specific negative or burdensome expectation and formulated it in a precise, situation-specific manner to ensure testability over the next three days. They rated their belief in the expectation (0-100%) and identified the associated emotion. Additionally, they reflected on their potential behavior in the given situation, considering how they could act in ways that either confirmed or disconfirmed their expectation (How can I behave to confirm my expectation? How can I behave to disconfirm my expectation?). A key aim of the behavioral experiments was to counter cognitive immunization by actively engaging participants in testing their expectations. By requiring them to plan and execute experiments that challenge their expectations, the intervention sought to make it more difficult for participants to dismiss disconfirming experiences as mere exceptions. This structured approach was intended to promote a more flexible adaptation of expectations. After three days, they reassessed their belief in the expectation, documented any perceived evidence supporting or contradicting it, and considered alternative interpretations of the situation.

Age, gender, nationality, education, profession, current and past diagnosed mental disorder, and current and past psychotherapeutic treatment were assessed.

The Immunization Scale (IMS; Ewen et al., 2022) was chosen as the primary outcome measure because it assesses key mechanisms involved in expectation persistence, making it a relevant indicator of cognitive immunization. The IMS consists of three subscales: (1) Cognitive Immunization, which measures reappraisal strategies used to maintain expectations despite disconfirming experiences, (2) Negative Expectations, capturing the persistence of negative beliefs that often accompany cognitive immunization, and (3) Assimilation, referring to the integration of expectation-violating information without fundamentally altering the expectation. While the Cognitive Immunization subscale directly measures immunization processes, the other two subscales assess related mechanisms that sustain rigid expectations, aligning with the ViolEx model. The IMS was validated through factor analyses and demonstrated excellent internal consistency (Cronbach’s α = .94). Significant correlations with depression, anxiety, and psychological flexibility further support its construct validity. Cognitive immunization was selected as the primary outcome instead of direct expectation change because expectations are highly specific and individual, making it difficult to systematically assess whether a particular expectation was tested and updated. Instead, we focused on cognitive immunization as the assumed underlying mechanism, as it represents a generalized process by which individuals maintain their expectations despite disconfirming evidence. Depressive and anxiety symptoms were considered moderators rather than primary outcomes because the study aimed to investigate mechanisms of expectation maintenance and change rather than direct symptom reduction.

Depressive symptoms were assessed using the Patient Health Questionnaire – German Version, section PHQ-9 (Gräfe et al., 2004). This 9-item scale primarily evaluates depression criteria as defined by the Diagnostic and Statistical Manual of Mental Disorders DSM-IV. Severity levels are classified as follows: a sum score of 1 to 4 indicates minimal depressive symptoms, 5 to 9 mild, 10 to 14 moderate, and 15 to 27 severe depressive symptoms. Anxiety levels were measured using the Beck Anxiety Inventory BAI (Margraf & Ehlers, 2007), a 21-item scale. A sum score of 0 to 7 indicates minimal anxiety, 8 to 15 mild, 16 to 25 moderate, and 26 to 63 severe anxiety symptoms. A sum score of 26 or higher is considered clinically relevant anxiety.

An ANOVA of the mixed-effects model revealed no significant main (timepoint: F(2,121) = 0.18, p = .311; group: F(2,122) = 0.20, p = .820) or interaction (F(4,121) = 1.23, p = .300) effects (see Figure 4).

The results of the linear mixed-effects model revealed no significant effects of timepoint or group, as well as no significant interactions could be found (see Table 3). The PHQ-9 score did not significantly vary between the groups or measurement timepoints.

Random effects analysis showed that the intercept variance (τ00) was 0.77. The variance in the slopes (τ11) for the post timepoint was 4.41, and for the follow-up timepoint, 5.54. The ICC was 0.77, suggesting that most variability in the PHQ-9 scores is due to between-group differences. The model's conditional R² was 0.769, suggesting that the model explains a substantial proportion of the variability in the outcome, while the marginal R² was 0.014, indicating that the fixed effects alone explain only a small proportion of the variability.

For the calculated ANOVA, a significant main effect for timepoint could be found (F(2,125) = 14.46, p < .001). The main effect group (F(2,122) = 0.20, p = .819) was non-significant, the interaction effect (F(4,125) = 2.89, p = .025) showed a significant result. Contrast analyses looking at the change scores per group, only the EFPI group showed a significant difference between post and follow-up measurement (t(119) = 5.03, p < .001; see interaction plot Figure 4).

The results of the linear mixed-effects model further revealed the main significant effect of the follow-up timepoint (β = - 2.55, 95% CI [-4.75, -0.36], p = .023). No significant effect was found for the group variable (see Table 4), nor were any significant interactions observed between the timepoint and group.

The random intercept variance (τ00) was 18.28, the random slope variance (τ11) at the post timepoint and the follow-up timepoint was 30.49 and 24.84, respectively. The total variability was attributable to 74% of the differences between the groups. The model's conditional R² was 0.744, the marginal R² was 0.034.

The model integrating the depressive symptoms as a moderator (AIC 3013.2) did not explain significantly more variance than the model without a moderator (AIC 3016.5; χ²(9) = 16.28, p = .061).

In the triple interaction model, ANOVA calculations showed a significant main effect of timepoint (F(2,123) = 4.68, p = .011) and a significant interaction of timepoint and PHQ-9 (F(2,123) = 3.48, p = .034). The triple interaction was not significant (F(4,123) = 3.48, p = .417). The results of the linear mixed-effects model revealed no significant main effects or interactions (see Table 5).

Random effects analysis showed that the variance in the intercepts across sessions (τ00) was 131.26, and the variance in the slopes (τ11) for the post and the follow-up timepoint was 54.11 and 102.20, respectively. The ICC was 0.71. The model's conditional R² was 0.749, and the marginal R² was 0.123.

Integrating anxiety as a moderator in the mixed model (interaction model: AIC 2981.6, triple interaction model: AIC 2974.1) explained the data significantly better (χ²(9) = 25.45, p = .003). The ANOVA revealed a significant main effect in group (F(2,119) = 6.23, p = .002). The interaction between timepoint x BAI (F(2,122) = 6.43, p = .002) and group x BAI (F(2,119) = 4.27, p = .016) were significant. The triple interaction was non-significant (F(4,122) = 0.40, p = .806).

The results of the linear mixed effects model further revealed the main significant effect of the CG, indicating an increase in the IMS score in the third group compared to the EFPI group (CG: β = 14.52, 95% CI [3.47, 35.57], p = .010). Higher BAI scores were significantly associated with higher IMS scores (β = 1.05, 95% CI [0.35, 1.75], p = .003). A significant interaction between the follow-up timepoint and the BAI score was found (β = -0.87, 95% CI [-1.61, -0.13], p = .021), indicating that the relationship between the BAI score and the IMS score differs significantly at the follow-up timepoint compared to the other timepoints. Another significant interaction was found between the third group, CG, and the BAI score (β = -1.30, 95% CI [-2.26, -0.33], p = .008; see Table 6), indicating that the relationship of the BAI score and the IMS score differs significantly in the CG compared to the other groups.

The random intercept variance (τ00) was 113.50, and the random slope variance (τ11) was for the post timepoint and the follow-up timepoint 54.34 and 85.34, respectively. The ICC was 0.71, indicating that 71% of the total variability in the outcome could be explained through the grouping factors and not individual variance. The model's conditional R² was 0.752, suggesting that the model explains a substantial proportion of the variability in the outcome, while the marginal R² was 0.148.