Study 1 examined how the type of physician AI support used in hybrid medical decision-making influences potential patients’ trust, perceptions, and behavioral intentions. To this end, we used a between-subjects design, in which participants were randomly assigned to one of three experimental conditions: (1) physician without AI support (No AI), (2) physician with descriptive AI support, which provides only visualizations or additional analyses (Descriptive AI), or (3) physician with diagnostic AI support, where the AI system also suggests a preliminary diagnosis to the physician (Diagnostic AI).

We expected that trust in the medical decisions (hypothesis [H]1a), perceived provider trustworthiness (H1b), perceived uniqueness neglect (H1c), and willingness to follow the advice (H1d) would not be rated differently in the conditions No AI and Descriptive AI but less favorably in the condition Diagnostic AI.

Participants were recruited via a subject pool comprising employees who were enrolled as part-time students at a German university alongside their regular employment. Data collection took place between March and April 2025 (median participation time: 26 minutes). Prior to data collection, we conducted a power analysis using data simulations in R (R Core Team) with the package simr [62], based on small to medium effects observed in the pilot study (approximately Cohen d=0.10‐0.52). The simulations indicated that a total sample size of 450 would provide adequate power (approximately 0.80 across outcomes) at a significance level of α=.05.

Each participant read 4 vignettes in randomized order, each describing a different medical consultation (Figure 3). The consultations described in the vignettes varied in symptoms, clinical settings, diagnoses, and medical recommendations, as detailed in the subsequent section. Presenting multiple scenarios allowed us to assess participants’ perceptions across a broader range of realistic medical contexts and AI use cases. This approach increases reliability by reducing the influence of idiosyncratic features of any single vignette.

The experimental condition remained consistent across all vignettes for each participant to make the core manipulation less transparent. This design avoided direct comparisons between the experimental conditions and reduced the risk that participants artificially align their responses. Participants were asked to imagine themselves as patients in the described situation and then responded to the outcome variables. After reading all vignettes and answering the outcome measures for each, participants completed a manipulation check in which they indicated whether the physician (1) did not use an AI system, (2) used an AI system for visual support only, or (3) used an AI system also for diagnostic support. Manipulation checks were administered without allowing participants to revisit the vignette text. To supplement the quantitative data, participants then answered open-ended questions to provide deeper insight into their thoughts and feelings about hybrid medical decision-making. Finally, participant characteristics were collected.

Our vignettes described ophthalmology consultations—a field in which AI-CDSS hold great potential and offer diverse use cases [44]. The vignettes were codeveloped with ophthalmologists to ensure clinical plausibility. Together, we identified areas in which AI tools are already being used or hold potential for future application. We selected the use cases to reflect different clinical stakes and settings. The 4 selected use cases were screening (diabetic retinopathy), emergency (retinal detachment), risk prediction (oculomics), and follow-up (glaucoma). Two ophthalmologists reviewed and refined the written vignettes, with particular attention to the realism of the diagnostic procedures and clinical interactions described.

The vignettes were written in the first-person perspective to facilitate participant immersion and because prior research has shown that perceptions of AI in medical decision-making differ when individuals evaluate care for themselves compared with care for others [27,33]. Each vignette detailed the reason for the visit, diagnostic procedures (including retinal imaging), and how the physician derived the medical decisions, varying by experimental condition. In the condition No AI, the physician analyzed the retinal images and determined the diagnosis and medical advice without AI support. In the condition Descriptive AI, the AI support system highlighted features in the retinal images to support physician interpretation. The physician then determined the diagnosis and medical advice. In the condition Diagnostic AI, the AI support system also suggested a preliminary diagnosis, which the physician reviewed. The physician then determined the final diagnosis and medical advice.

Note that the manipulations were implemented in a relatively subtle manner to reduce demand characteristics and avoid participants deducing the overall study goal. For example, no condition labels or bolded text were used. As it was not the aim of this study to examine the effect of matching versus contradicting assessments by the physician and AI, the vignettes in the hybrid conditions did not contain information about the specific results of the AI analyses nor whether they confirmed or contradicted the physician’s assessment. Furthermore, the consultations were set in unfamiliar medical settings (eg, a new clinic or regular physician on leave) to avoid that participants’ associations with their own, real doctors influenced the results. Perceived vignette realism was captured by 4 items (eg, “The situation was believable.” “I was able to imagine the situation.”) on 7-point Likert scales, with higher values indicating higher realism. Mean ratings, ranging between 5.50 (risk prediction) and 5.62 (emergency), indicated that participants were able to imagine the situations sufficiently well (Table S1 in Multimedia Appendix 2). Figure 4 displays an example vignette including the different experimental versions. Multimedia Appendix 3 contains all vignettes.

The following four outcome variables were assessed, capturing key components of trust [63] and aligning with outcome measures used in prior research on AI-supported medical decision-making [22,29,33,40]:

All items were rated on 7-point Likert scales, with higher scores reflecting higher levels of the measured constructs. English scales were back-translated into German. Multimedia Appendix 4 provides an overview of all items. Participants also answered open-ended questions addressing influencing factors on their quantitative responses (Q1), their appraisal of the physician’s use of AI in the situations (Q2), and suggestions for increasing trust (Q3).

We fitted linear mixed-effects models (LMMs) in R (version 4.4.0) using the lme4 package [65], with Condition as a fixed effect and Participant as a random intercept to account for the nested data structure. Significant effects were followed by pairwise comparisons of estimated marginal means (EMMs) using Bonferroni-Holm correction for multiple testing (emmeans [66]). As robustness checks, we additionally corrected the family-wise error rate across the 4 outcomes using Bonferroni-Holm correction on the omnibus tests, repeated the main analyses in intention-to-treat samples including participants who failed the manipulation checks, and conducted exploratory analyses examining the specific vignette as an additional main effect and its potential interaction with the AI support condition. Note that the vignettes differed on several dimensions at once, including clinical setting, symptom severity, diagnostic procedures, and invasiveness of recommended treatments. Therefore, these exploratory analyses of specific vignette effects should be interpreted as tests of broad scenario–level differences rather than as isolating the specific effects of stakes, acuity, or any other single vignette feature.

Analysis of the open-ended responses followed a 3-step process. First, in line with prior studies that used AI to analyze large qualitative datasets [67,68], we used ChatGPT-4 (OpenAI) to provide an initial list of candidate recurring themes across the 489 responses for each of the 3 open-ended questions. The model was prompted in German using the original response language, and its role was limited to preliminary theme surfacing. To enhance transparency, the prompts and suggested themes are included in Multimedia Appendix 5. The use of a large language model at this stage was motivated by the size of the qualitative dataset. The output provided an initial overview of potential themes but served only as a starting point. Human coders carefully reviewed and refined the themes in the subsequent steps, ensuring that themes and interpretations are grounded in the original data. Second, initial codebooks for Q1-Q3 were developed by grouping themes and removing redundancies. Third, a human rater applied the initial coding scheme, coding which themes were present in each response. During this process, the coding scheme was iteratively refined to capture all relevant content and reapplied. For example, when new themes emerged inductively, these were added to the codebook. To mitigate potential hallucinations or model bias, ChatGPT was not used for final coding, interpretation, or quote selection. All reported themes resulted from human coding. A second human rater independently coded 15% of the responses for each question to assess the reliability of the final coding scheme. Interrater agreement was substantial for all 3 questions, with Cohen κ=0.71 for Q1, κ=0.71 for Q2, and κ=0.71 for Q3 [69].

Additionally, we subjected responses to Q2 to exploratory sentiment analysis using the Hugging Face checkpoint cardiffnlp/twitter-xlm-roberta-base-sentiment, a multilingual XLM-RoBERTa-base model [70], in Python (version 3.11.11; Python Software Foundation). The underlying XLM-T model was trained on approximately 198 million tweets and fine-tuned for 3-class sentiment classification on an 8-language benchmark that includes German. In the source paper [70], German performance reached macro-F₁ values of 76.13 in monolingual evaluation and 77.35 in multilingual evaluation on held-out test data. For each response, the model produced 3 class logits, which were transformed using the softmax function into normalized scores for the negative, neutral, and positive sentiment classes (ranging from 0 to 1). These continuous class scores were then compared across conditions using 1-way ANOVAs, followed by 2-tailed Bonferroni-Holm–adjusted t tests when significant effects were observed. Results of the sentiment analysis are provided in Tables S2 and S3 in Multimedia Appendix 2. Because the model was not trained and validated on health care–related text specifically, its validity for the present task is unclear. Therefore, the sentiment analyses were treated as exploratory, and results should be interpreted with caution.

Study 2 sought to replicate findings from study 1 and to extend them by varying not just the type but also the timing of AI support. The 2 hybrid conditions in study 1 exemplified a concurrent workflow, in which AI support was reviewed alongside the primary clinical information. Study 2 therefore crossed AI Support Type (descriptive vs diagnostic) with AI Support Timing (concurrent vs sequential), yielding 4 hybrid configurations in addition to the no-AI baseline (Figure 2).

We hypothesized that evaluations would not be rated differently between No AI and Concurrent Descriptive AI but lower for Concurrent Diagnostic AI across all 4 outcome variables (H2a-H2d). Additionally, we explored whether AI Support Type and AI Support Timing interactively influence participants’ evaluations of AI-assisted medical decisions.

Participants were recruited via subject pools of 4 German universities, with the majority sampled again from the pool of employees enrolled as part-time students. Data collection for study 2 took place between June and October 2025 (median participation time: 25 minutes). We estimated the required sample size for study 2 based on parameters from study 1. We preregistered a target sample size of 500 to achieve a power of 0.80, given α=.05.

Participants were randomly assigned to 1 of 5 between-subjects conditions: No AI, Concurrent Descriptive AI, Concurrent Diagnostic AI, Sequential Descriptive AI, or Sequential Diagnostic AI (Figure 3). We updated the vignettes from study 1 to distinguish between concurrent and sequential use of physician AI support. Mean ratings of vignette realism, ranging between 5.52 (screening) and 5.72 (emergency), indicated that participants were able to imagine the situations sufficiently well (Table S1 in Multimedia Appendix 2).

Two manipulation checks were applied. First, participants indicated whether the physician (1) did not use an AI system, (2) used an AI system for visual support only, or (3) used an AI system also for diagnostic support. Second, participants in the hybrid conditions additionally indicated whether the physician communicated their own assessment before reviewing the AI output (sequential) or not (concurrent). Otherwise, design and procedure mirrored those of study 1. We used the same quantitative measures as in study 1. Internal consistency per vignette was again high for all composite scales (trust in the medical decisions: α=.91-.96, perceived provider trustworthiness: α=.97, and perceived uniqueness neglect: α=.84-.85). Participants were also asked the open-ended questions Q2 and Q3 from study 1.

Given that the condition No AI cannot be categorized by AI Support Timing, our design was partially factorial. Therefore, we preregistered 2 complementary analyses using LMMs for each dependent variable. To examine main and interaction effects of AI Support Type and AI Support Timing, we used a 2x2 factorial model, which included only the 4 hybrid conditions. To compare all 5 conditions, we additionally used an all-conditions model (as in study 1). In both analyses, Participant was entered as random intercept. The analyses were followed up with Bonferroni-Holm–adjusted comparisons of EMMs. We conducted the same robustness checks as in study 1.

Open-ended responses to Q2 and Q3 were qualitatively analyzed using a deductive-inductive approach. A coder, who was experienced in qualitative analyses and who had not been involved in the analyses for study 1, applied the coding scheme developed in study 1 to the responses from study 2 while remaining open to modifications if new themes emerged. Additionally, open-ended responses to Q2 were subjected to sentiment analysis following the same procedure as in study 1. These results are reported in Tables S16 and S17 in Multimedia Appendix 2.

Of 625 respondents, 8 (1%) were excluded because they completed the survey more than 2 times faster than the median duration (as recommended by Leiner [71]), 33 (5%) were excluded because they failed at least 1 of 2 attention checks, and 95 (15%) were excluded because they failed the manipulation check. The final sample for study 1 consisted of 489 participants (n=336 females, 69%) with a mean age of 25.6 (SD 4.6) years, ranging from 18 to 49 years. Detailed demographics are reported in Table 1.

EMMs and results from the model comparisons (likelihood ratio tests) are reported in Table 2. Including Condition as a fixed effect significantly improved model fit compared with a null model for all outcomes except willingness to follow the advice, indicating that the type of AI involvement affected all outcomes but this one. When controlling for family-wise error across the omnibus tests using Bonferroni-Holm correction, the pattern of results remained unchanged (Table S4 in Multimedia Appendix 2). Results from all (including nonsignificant) pairwise comparisons can be found in Tables S5-S8 in Multimedia Appendix 2.

In the condition Diagnostic AI, trust in the medical decisions was rated significantly lower than in the conditions Descriptive AI (t₄₈₆=3.51; P=.002) and No AI (t₄₈₆=2.67; P=.02). A similar pattern was observed for perceived provider trustworthiness, with Diagnostic AI rated significantly lower than Descriptive AI (t₄₈₆=4.46; P<.001) and No AI (t₄₈₆=3.81; P<.001). For perceived uniqueness neglect, where higher scores indicate greater concerns about lack of individualized consideration, Diagnostic AI received significantly higher ratings than

Descriptive AI (t₄₈₆=2.72; P=.02) and No AI (t₄₈₆=2.47; P=.03). Descriptive AI and No AI, however, did not differ significantly from each other on any of these outcomes (all P≥.39). No significant differences emerged between any of the conditions regarding willingness to follow the advice (all P≥.25). These results therefore support H1a-H1c but not H1d.

Intention-to-treat analyses including participants who failed the manipulation checks showed a largely similar overall pattern. However, effects for perceived uniqueness neglect failed to reach significance, suggesting that this effect depends more strongly on successful processing of the manipulation and should therefore be interpreted more cautiously. Full results are reported in Tables S9-S13 in Multimedia Appendix 2.

In exploratory analyses, adding Vignette as a main effect significantly improved model fit for all outcomes, and these effects remained significant after correction for family-wise error across outcomes. The vignette follow-up (glaucoma) generally elicited less favorable evaluations, whereas the vignettes screening (diabetic retinopathy) and emergency (retinal detachment) were evaluated somewhat more favorably across outcomes. By contrast, adding the interaction term Condition × Vignette did not significantly improve model fit for any outcomes, suggesting that the experimental effects were stable across scenarios (Table S14 in Multimedia Appendix 2).

Table S15 in Multimedia Appendix 2 summarizes recurring themes for influencing factors (Q1), appraisal of AI support (Q2), and approaches to increasing trust (Q3). Given the aim of this work, we focus particularly on AI-related aspects among these. The open-ended responses suggest that several aspects shape potential patients’ trust in hybrid medical decision-making. These can be roughly categorized as relating to the medical situation, the physician, the participants themselves, and the AI system and dynamics of the physician-AI interaction.

Overall, several participants acknowledged the potential of AI in medical decision-making, for example, to help prevent human oversights. However, many participants were concerned that physicians would overrely on AI recommendations.

As physicians were considered more capable of holistically assessing patients’ unique cases than AI, participants preferred AI-CDSS that play only a smaller role in the diagnostic process over AI-CDSS that provide diagnostic support—consistent with the quantitative results.

Ultimately, the physician’s interaction with the AI system played an important role. For example, participants considered the physicians’ assumed AI competence and whether they appeared to critically examine the AI output.

Moreover, several participants expressed a need for more information about the AI system itself, its results, and its usage.

Another recurring theme was the wish for the use of AI-CDSS as an independent “second opinion” for the physician. Participants described a preferred decision-making workflow where physicians first review all diagnostic information and only look at the AI output after having formed their own assessment, so that potentially incorrect AI output does not bias their clinical judgment prematurely.

Overall, the qualitative data suggest that many participants were open to physicians using AI support but only under certain conditions. For example, they valued AI only as an augmentation of the physician’s expertise and voiced concerns about overreliance on AI results. Participants expressed that transparency and the way physicians use AI-CDSS are crucial. The latter includes both the type of AI support (ie, for which tasks) and when that support is used in the clinical decision-making workflow. From their perspective, careful choices about the support type and timing can help safeguard against AI output that is incorrect or overlooks patients’ individual circumstances.

Of 850 respondents, 53 (6%) were excluded because they completed the survey more than 2 times faster than the median duration, 38 (4%) were excluded because they failed at least 1 attention check, 81 (10%) were excluded because they failed the first manipulation check, and 108 (13%) were excluded because they failed the second manipulation check. The final sample consisted of 570 participants (n=410 females, 72%) with a mean age of 25.1 (SD 5.3) years, ranging from 18 to 58 years. Detailed demographics are reported in Table 1.

For trust in the medical decisions, perceived provider trustworthiness, and perceived uniqueness neglect, 2x2 LMM analyses found a significant main effect of AI Support Type, which was qualified by a significant interaction effect (Table 3). For willingness to follow the advice, a significant main effect of AI Support Type emerged, but the interaction effect failed to reach statistical significance (P=.07). The interaction plots are shown in Figure 5. When controlling for family-wise error across the omnibus tests using Bonferroni-Holm correction, the pattern of results remained unchanged (Table S4 in Multimedia Appendix 2).

Intention-to-treat analyses including participants who failed the manipulation checks produced the same pattern of results (Table S18 in Multimedia Appendix 2). In exploratory analyses investigating main and interaction effects of vignette, adding Vignette as a main effect significantly improved model fit for all outcomes, also after correcting for family-wise error. Generally, the vignette follow-up (glaucoma) elicited less favorable evaluations, whereas screening (diabetic retinopathy) and emergency (retinal detachment) tended to be evaluated more favorably. By contrast, adding the higher-order interaction terms involving Vignette did not improve model fit for any outcome, suggesting that the effects of AI support type and timing were stable across scenarios (Table S19 in Multimedia Appendix 2).

EMMs and results of the likelihood ratio tests are displayed in Table 4. Results from LMMs including all 5 conditions showed that including Condition as a fixed effect significantly improved model fit compared with a null model for all outcomes (all P values <.001). When controlling for family-wise error across the omnibus tests using Bonferroni-Holm correction, the pattern of results remained unchanged (Table S4 in Multimedia Appendix 2). Results from all (including nonsignificant) pairwise comparisons can be found in Tables S20-S23 in Multimedia Appendix 2.

Regarding trust in the medical decisions, Concurrent Diagnostic AI received significantly lower ratings than Concurrent Descriptive AI (t₅₆₅=5.09; P<.001), Sequential Descriptive AI (t₅₆₅=5.57; P<.001), Sequential Diagnostic AI (t₅₆₅=4.13; P<.001), and No AI (t₅₆₅=2.65; P=.04). Compared with the baseline condition No AI, both Concurrent Descriptive AI (t₅₆₅=2.76; P=.04) and Sequential Descriptive AI (t₅₆₅=3.14; P=.01) received significantly higher ratings of trust in the medical decision. Together, these results partially support H2a.

Also regarding perceived trustworthiness of the medical provider, Concurrent Diagnostic AI evoked less favorable ratings than all other conditions. Specifically, Concurrent Diagnostic AI was rated significantly lower than Concurrent Descriptive AI (t₅₆₅=4.95; P<.001), Sequential Descriptive AI (t₅₆₅=5.22; P<.001), Sequential Diagnostic AI (t₅₆₅=4.12; P<.001), and No AI (t₅₆₅=3.52; P=.003). These results support H2b.

The reverse pattern emerged for perceived uniqueness neglect. Here, Concurrent Diagnostic AI was rated significantly higher than Concurrent Descriptive AI (t₅₆₅=4.46; P<.001), Sequential Descriptive AI (t₅₆₅=3.99; P<.001), Sequential Diagnostic AI (t₅₆₅=2.79; P=.04), and No AI (t₅₆₅=3.84; P=.001). These results support H2c.

Regarding willingness to follow the medical advice, Concurrent Diagnostic AI also received significantly lower ratings than Concurrent Descriptive AI (t₅₆₅=3.74; P=.002) and Sequential Descriptive AI (t₅₆₅=3.55; P=.004), whereas no other contrasts for this outcome variable reached statistical significance. These results partially support H2d.

Overall, the results mirror those of the 2x2 LMMs: When the timing of AI use in the decision-making workflow was framed as concurrent, Diagnostic AI received significantly less favorable evaluations than Descriptive AI on all outcomes, replicating results from study 1. When the timing of AI use in the decision-making workflow was framed as sequential, the 2 types of AI support did not differ significantly from each other anymore. This was driven by improved ratings of diagnostic AI support in sequential decision-making compared with concurrent decision-making.

Intention-to-treat analyses including participants who failed the manipulation checks yielded a largely unchanged pattern of results, with the differences between the conditions Diagnostic AI and Descriptive AI remaining significant in concurrent workflows and not significant in sequential workflows for all outcomes (Tables S24-S28 in Multimedia Appendix 2). Exploratory analyses investigating main and interaction effects of Vignette showed significant main effects for all outcomes. Again, the vignette follow-up (glaucoma) tended to elicit lower trust-related evaluations. However, no interaction effects were significant after correction for family-wise error. This suggests that participants responded differently to the clinical scenarios overall, whereas the relative pattern of condition effects was stable across vignettes (Table S29 in Multimedia Appendix 2).

The qualitative findings derived from Q2 and Q3 in study 2 largely mirrored those of study 1, with broadly similar proportions of mentions across themes (Table S30 in Multimedia Appendix 2). No additional themes emerged. Participants again expressed ambivalent views toward AI-supported medical decision-making, recognizing both the potential of AI to support physicians as well as concerns about physician overreliance and the possibility that incorrect AI output could bias the final decision.

Several responses further suggested that comfort with physician-facing AI support depends on how it is used. For example, one participant in the Concurrent Descriptive AI condition noted positively that “It is also very good that no diagnosis is suggested and that the doctor is not actively influenced in their diagnosis.” More specifically, several responses reflected the central quantitative finding of study 2, namely, that sequential workflows were experienced as more reassuring than concurrent ones.

More specifically, participants seemed to value not only the sequential process itself but also the communication of the physician’s independent assessment before consulting AI.

Overall, these qualitative results reinforce the interpretation that evaluations of hybrid medical decisions depend not only on whether AI is used but also on perceptions of the physician-AI collaboration.

Our results suggest that when the physician uses AI support during (concurrent) medical decision-making, patients trust the resulting decision less if the AI system had already suggested a preliminary diagnosis to the physician than if the AI system had provided only supplementary analyses or visualizations. The same applies to the extent to which patients perceive the medical provider to be trustworthy and considerate of their unique circumstances.

Qualitative insights from participants’ open-ended responses complement these quantitative results. For example, participants expressed that deriving a diagnosis requires a holistic consideration of the patients’ case, which encompasses more than the interpretation of medical imaging. Therefore, they preferred AI rather for smaller, supportive tasks but wanted the diagnosis to remain with human expertise. Additionally, participants expressed concerns that incorrect diagnoses suggested by AI could bias physicians’ judgments and thereby compromise the final clinical decision. Together, these qualitative insights suggest that diagnostic AI support is disadvantaged by a “double curse.” First, it is perceived as more likely to produce inadequate recommendations than descriptive AI support, with its diagnostic quality expected to fall short compared with a physician. Second, when such shortcomings occur, diagnostic AI support is seen as more harmful to the final decision than descriptive AI support, because its suggestions may influence physicians’ judgments more directly.

Study 2 revealed that aversion toward diagnostic AI support was eliminated when physicians reviewed AI output only after forming their own independent clinical assessment (ie, in sequential workflows). Participants’ open-ended responses suggest that they view the timing of AI support as a way to reduce the risk of overreliance on AI advice and safeguard the final decision from potentially incorrect AI output, namely, through sequential decision-making.

In contrast to reported trust and perceptions, participants’ willingness to follow the medical advice was comparably less affected by the use of AI support, remaining stable across conditions in study 1 and showing only modest variation in study 2. While the use of diagnostic AI in a concurrent workflow still lowered participants’ willingness to follow the medical advice compared with some other types of AI involvement in study 2, behavioral intentions seem to be less sensitive to the use of AI support than subjective perceptions and trust. A possible explanation might be that when deciding whether to follow medical advice, other factors are more relevant to patients than how the medical decision was reached.

For example, in our vignettes, the diagnostic procedure always yielded a clinically concerning health result, which required medical action. Consequently, patients may be inclined to follow recommendations for active intervention “to be on the safe side,” even when they feel uneasy about how the decision was reached. This could partly explain the rather high willingness to follow the medical advice across conditions. Such a tendency has also been observed in previous studies [30,72,73]. For example, Detjen et al [30] found that although AI interpretations of mammograms were trusted less than those of physicians, participants were still willing to follow AI recommendations for a biopsy—presumably due to risk aversion. In our study, condition effects might have been more pronounced if the vignettes had described scenarios with favorable health results recommending no immediate action, as suggested by the findings of previous studies [72,74].

At the same time, the possibility of ceiling effects limiting sensitivity to detect differences between conditions should be noted. Our vignette framing may have contributed to strong baseline intentions to follow the medical advice and therefore restricted upper-end variability on this variable. Although we still observed some variance on this variable (SD_{Study 1}=1.50, SD_{Study 2}=1.49), responses in both studies were skewed toward the upper end of the scale. Accordingly, the comparatively weak condition effects for willingness to follow the medical advice should be interpreted cautiously.