Evidence from studies observing a correlation between negative AGS and subsequent declines in performance cannot be considered causal evidence of the phenomenon or of its underlying mechanisms. Experimental studies allowing the deliberate manipulation of stereotypes are therefore needed. For this purpose, several experimental paradigms have been developed to explore competing explanations.

An early approach to activating positive or negative self-relevant AGS used a verbal priming technique, where age-related adjectives (e.g., “confused”) were briefly flashed on a screen, resulting in cognitive activation of the semantic networks related to the concept. Negative AGS priming was shown to cause temporary declines in walking speed (Bargh et al., 1996) and memory performance (Stein et al., 2002); however, these experiments caused controversy regarding their replicability (Rivers and Sherman, 2018), and the approach does not give participants the perspective of an older person; thus, the activation of self-relevant AGS lacks external validity.

Haslam et al. (2012) used two parallel manipulations with age-related self-categorizations by assigning middle-aged participants to either “old” or “young” groups (manipulating self-relevant affiliation with the stereotyped group) and by shaping deficit expectations with newspaper articles on either age-related memory decline or other age-related problems (manipulating the stereotype content). The study found that when participants were labelled “old,” they performed worse on memory tests, particularly when they had been primed with a newspaper article on memory decline. These results support the link between stereotype content and behavioral outcomes. However, the effects were short lived, and the introduction of a social comparison group possibly diluted the self-relevant nature of the stereotype activation, as it added a competitive component.

A more lifelike approach was used by Varkey et al. (2006), where medical students partaking in the ageing game were asked to perform simple everyday tasks while wearing technical appliances to reduce their general abilities. With this artificial simulation of living in a frail body, Varkey et al. were able to improve significantly the participants’ empathy and attitudes toward older adult patients. This manipulation offered the continuous embodiment of a physical and perceptual condition commonly associated with older age, even though the design was non-experimental and only allowed an indirect activation of self-related AGS. Still, it demonstrates how complex embodiment manipulations create effective multi-level experiences, leading to promising results.

Eibach et al. (2010) chose an experimental approach with a more complex age-related phenomenology and stronger ecological validity. The manipulation of vision decline and induction of a generation gap experience led to the alteration of subjective age in an adult sample. Furthermore, the successful manipulation of subjective age moderated the influence of AGS on the subjects’ self-evaluation. In a comparable approach, Stephan et al. (2013) presented participants with manipulated performance feedback for a handgrip strength task. Participants in the experimental condition were given positive feedback regarding their performance in comparison to same-aged peers, leading to a reduction in subjective age and an increase in handgrip strength scores for a second measurement. Thus, an experimental manipulation of AGS’ self-relevance is indeed possible by inducing stereotypically old age-related experiences on a perceptual or social level.

Even more complex and visually realistic manipulations are possible whenever stimulus content is created and presented in gaming scenarios in which participants control an avatar whose appearance can be modified. The resulting Proteus effect (Yee et al., 2009; named after the shape-changing sea god in Greek mythology) is based on high standardization and multi-sensory stimulus manipulation, implying that self-relevant stereotypes can be activated by visual identity cues, affecting people’s behavior. Earlier studies successfully manipulated social behavior and aggression by increasing body height (Peña et al., 2009) and influenced negotiation interactions by changing the avatar’s attractiveness (Yee et al., 2009). Unfortunately, most studies used a desktop setting, which cannot offer an intuitive immersive experience or multi-level sensory stimulation, and results can only be partially generalized to self-relevant stereotypes.

While the above-mentioned longitudinal relationships offer a view of possible health-related outcome variables, many short-term experimental findings are more difficult to transfer beyond the laboratory and fail to show lasting effects and external validity; hence, proving causality remains complicated. Meanwhile, combining a realistic aging experience with a small-step experimental variation has a continuous impact on perceptions or behaviors and seems the next step necessary to push the limits of our understanding of this psychological mechanism and, furthermore, to develop interventional techniques for the known medical and social implications of strong self-relevant negative AS.

Together with experts in the field, such as, Slater (2018), Rizzo et al. (2019) or Tuena et al. (2020), we argue that virtual reality (VR) technology enables a realistic immersive experience and hence allows the combination of a high level of standardization and the presentation of true-to-life environments, thus maximizing both internal and external (or ecological) validity. Particularly, users of VR technology can interact with their surroundings more naturally (at least more so than in an interaction with a computer mouse in front of a computer screen), and researchers can manipulate the characteristics of these surroundings in a highly flexible and cost-effective way (at least more flexibly and cost-effectively than setting up, furnishing, or driving to a desired setting). Numerous clinical applications (e.g., Brown et al., 2020) have demonstrated the benefits of an effective immersive experience when simulating general counselling settings (Slater et al., 2019), as well as of applying established therapy approaches to specific conditions, such as obesity and diabetes (Rizzo et al., 2011), eating disorders (Riva et al., 2021), and post-traumatic stress disorder (Rizzo and Shilling, 2017). Earlier studies successfully applied VR to non-clinical topics with regard to interactional or cognitive parameters. Seeing oneself as Sigmund Freud resulted in improved interpersonal problem-solving skills (Osimo et al., 2015), while the virtual embodiment of Albert Einstein improved cognitive performance when compared to a control group (Banakou et al., 2018). Hence, VR scenarios could give participants a realistic aging experience when embodying an older avatar by activating self-relevant AGS in a standardized environment. One implementation of this approach by Reinhard et al. (2020) demonstrated a replication of priming studies using a VR embodiment technique instead of verbal priming; young participants who first embodied an older avatar in a VR scenario showed a decreased walking speed. Furthermore, the approach offers an actual shift in perspective toward an older self, as demonstrated by Sims et al. (2020), who showed age-progressed images to influence young participants’ attitudes toward financial security topics in later life.

Aside from the various applications of VR, some technical details seem to alter the interventional value of the approaches. Slater (2018) argues that VR applications allowing detailed head tracking and the ability to look around at objects provide a higher degree of realism, making it more likely for participants to experience the simulation as if it was really happening. Daher et al. (2017) showed that this even applied to interacting with a virtual agent instead of a real person. If certain technical criteria are met, a strong body ownership illusion is ensured as one possible mechanism behind VR. The body ownership illusion theory (e.g., Kilteni et al., 2015) focuses on the plasticity of the processing of perceptual stimuli. Based on the classic rubber-hand illusion findings (Botvinick and Cohen, 1998), it is argued that VR creates an intersensory bias to perceive the displayed body parts as one’s own (Banakou et al., 2018). The combination of tactile, visual, or auditory stimuli must be coordinated well enough, together with the experience of actually controlling the virtual avatar by moving its virtual head and arms around. By doing so, a realistic body ownership experience is provided, thus activating expectations and stereotypes directly and intuitively that are connected to the virtual avatar. In contrast, the Proteus effect suggests that people in general (Frank and Gilovich, 1988) and computer gamers in particular (Yee et al., 2009) adjust their behavior to their own expectations and stereotypes related to the visual identity cues of their digital avatar. The activation of self-relevant stereotypes takes place indirectly from perceiving one’s own appearance, and it thus depends on appropriate perceptions and semantic interpretations of visual cues.

Based on previous findings and the characteristics of our design and methods, we present three hypotheses concerning the age avatar embodiments’ effects on the subsamples’ performance scores and their associated moderating parameters.

A priori power analysis using G*Power (Faul et al., 2007) was conducted to determine the sample size necessary. Parameters for the calculations were obtained from the experimental design of the study by Reinhard et al. (2020), which was closest to the current methodological approach and where a comparison of two avatar groups’ walking speed scores before and after applying avatar embodiment led to a Cohen’s f of 0.43. The power analysis included a repeated measures analysis of variance (ANOVA) with three measures in total for two groups to compare: an error rate of α = 0.05, a power level of 1—β = 0.95, and a correlation of r = 0.50 between measures, and it resulted in a minimum total sample size of N = 50.

To account for adequate randomization effects concerning possible parameters influencing the main results, the data analysis of each study started with a t-test comparison of the two experimental conditions based on relevant baseline parameters: immersion intensity, performance domain-related baseline measure (physical fitness or memory self-efficacy), positive and negative affect, and implicit and explicit AGS. The main analysis for each study was carried out using ANOVA and repeated-measures ANOVA to compare the experimental conditions (with and without age manipulation in VR) concerning their physical performance and cognitive performance accordingly. The statistical assumptions for the inferential tests were checked prior to the analysis and did fulfill the criteria necessary in accordance with the relevant literature (Kaur and Kumar, 2015).

To provide evidence of the robustness of our results against alternative explanations and possible undisclosed associations and to provide evidence of the validity of our approach, an ANCOVA analysis was carried out including the aforementioned moderators and controls as covariates and looking at 1) the main effects of this covariate on performance, 2) the interaction effects with our experimental condition on performance, and 3) whether the ANOVA’s main effect results remained identical.

In total, N = 68 participants (54 female) aged 18–35 (M = 22.46, SD = 3.10) were recruited using both online and offline university blackboards and randomly assigned to the experimental (n = 35) or control group (n = 33). There were more female subjects in the total sample and an almost equal distribution of gender (79 vs. 80% female) and age (M = 22.12, SD = 2.27 vs. M = 22.77, SD = 3.71) in both the control and experimental conditions.

The 75-min procedure (see Figure 3) started with information and consent forms, followed by a demographic questionnaire, an assessment of affective state, measures for implicit and explicit AGS, and a physical fitness questionnaire. The VR headset was then mounted, and motion trackers inside the headset and hand controllers allowed synchronization of the participants’ movements. The avatar conditions’ random assignment was based on a randomization list and was not announced or discussed by the laboratory members. The young participants were assigned to the younger (YA_y) or older (OA_y) age avatar group. Participants were presented with a 90-s audio instruction directing their attention to their avatar’s appearance and movements to facilitate adequate immersion. After this introduction, a repeated measure of the affective state was followed by physical performance measures for handgrip strength and endurance. After leaving the 60-min VR scenario, participants completed a presence and body ownership questionnaire.

The virtual age embodiment did not directly affect handgrip strength. The significant interaction between condition and the repeated measurement, however, indicates that physical performance was affected in a more complex and subtle way than hypothesized. Thus, our first hypothesis is only partially confirmed. Compared to both the pattern shown in the control group and that known from other research (Innes, 1999), participants in the experimental group exhibited a conspicuously low performance in the first measurement; consequently, there was almost no fatigue-related decline in performance, typically found after a major effort. Experiencing being older appears to have resulted in a strategic attempt to save efforts and thus avoid premature resource depletion. This interpretation would be consistent with notions of selective optimization with compensation (Freund and Baltes, 1998) and reminiscent of an anticipated loss-based selection vis-à-vis a potentially challenging task or, as Ebner et al. (2006) would say, a goal orientation guided by loss prevention. The drawback of this interpretation is its assumption that younger people have at least tacit knowledge of the developmental regulation strategies of older individuals. A more simplistic explanation would be that participants approached the unknown task more cautiously, considering the possibly excessive demands they stereotypically expected for older people in unknown performance situations. Our data are compatible with both interpretations and invite further exploration of this issue in future research.

An intriguing finding of Study 1 was the significant interaction of the experimental condition with domain-matched (but not domain-general) explicit AGS, suggesting that AGS activation is indeed the effective mechanism behind our findings. In line with our expectations, participants only behaved stereotypically old in the aspects they believed related to old age. This aligns with research by Levy and Leifheit-Limson (2009), who showed stronger effects when the stereotype content matched the outcome performance domain.

Surprisingly, we did not find significant main effects of avatar age on endurance. A more strenuous variant with a heavier weight might have produced clearer effects, as the subjects’ physical resources would then be depleted more strongly. It could also be that the task did not match the content of the self-relevant stereotypes people hold. As our findings from the handgrip strength task suggest, changes in performance seem quite sensitive to the specific domain and hence to the specific content.

Study 2 comprised 45 (34 female) participants aged 19–30 years (M = 23.04, SD = 1.99) that were assigned to the control (n = 24) and experimental conditions (n = 21).

The design, procedure, and time spent in VR remained almost identical to Study 1, except the dependent variables included cognitive performance measures instead of physical ones (see Figure 3) and the respective baseline control measure was replaced with a memory self-efficacy questionnaire. The young participants again were assigned to the younger (YA_y) or older (OA_y) avatar group.

In Study 2, a multifaceted selection of cognitive tasks was used. Even though the required sample size was not reached and the study was possibly underpowered, we found a strong main effect of avatar age on VMR. With strong connections to the participants’ motivation and persistence, this finding is in line with our expectations, and the effect size of η ² = 0.18 is large. For Study 2, the first hypothesis was confirmed. However, not all cognitive measures applied in this study showed a meaningful difference concerning avatar age condition. Most importantly, it is unclear why this experimental approach affects cognitive performance more strongly than physical performance. A certainly obvious difference lies within the details of test application. The cognitive measure of VMR was carried out right inside the virtual scenario with the relevant words displayed on the virtual monitor. Subjects were forced to direct their attention to the general direction of the monitor and mirror. Meanwhile AWM and handgrip strength tests were both administered without any visual reference within the VR scenario. Perhaps stereotypes related to cognitive performance decline in old age are more widespread than those targeting physical performance (Levy, 2003). However, it must be noted that the main effect of avatar age on VMR was no longer significant when the implicit AGS score was included as a covariate. Unfortunately, a post-hoc inspection of the correlations between implicit AGS and the experimental condition, as well as memory performance did not provide a clear picture to explain this finding. More research is needed to understand the role of implicit AGS in this context.

Interesting results were also obtained when including domain-matched explicit AGS as a moderator in the analyses. We found a substantial offset in the learning slopes in the experimental older avatar group. A descriptive inspection of the data could indicate that strong domain-matched negative AGS were associated with low performance expectations for the older avatar group prior to the experiment, leading to a steeper learning curve once the task started and once participants gained confidence. Examining the data more closely, however, leads to another possible interpretation. The largest difference in performance occurred at the first measurement, where participants differing in their domain-matched AGS also showed a considerable offset in their performance. This finding resembles the handgrip strength effect from Study 1, and it might indicate the use of strategies to prevent the depletion of cognitive resources by using limited resources economically.

Meanwhile, the validity of these findings stands in question, as a significant baseline difference in this moderator was found between the experimental conditions, and future research would require higher sample sizes with careful randomization or matching of participants.

The effect we found for VMR, as well as its large size, cannot be generalized across all cognitive tasks. Data from the AWM task did not show significant effects following avatar age manipulation. One possible interpretation considers the limited capacity of people’s working memory (Baddeley & Hitch, 1974). With limited resources on a neuro-functional level, variations in self-relevant AGS activation and motivational aspects may not account for these performance differences.

Study 3 comprised N = 117 (66 female) participants aged 50–83 years (M = 61.23, SD = 7.5) that were assigned to the control (n = 59) and experimental conditions (n = 58). The recruitment for Study 3 resulted in a much larger sample size compared to the earlier studies, whose recruitment was considerably limited by, for instance, the availability of student participants during certain times in the semester. An age range starting at 50 was selected for two reasons. First, other studies on aging and on AGS sometimes (e.g., Wurm et al., 2007, 2010) include participants who do not yet fall into the category of being old to include this transition phase and to study precursors of aging. To cover both groups in the data analysis, chronological age will be included as a covariate. Second, we were guided by more pragmatic reasons given the demanding recruitment situation. The parents and relatives of the university students were in their 50 s or 60 s and could thus be recruited more easily than the general public during the COVID-19 pandemic.

Recruitment was carried out using a university blackboard, leaflets in a campus-attached GP practice’s waiting room, and newspaper bulletins. Prior to the VR assessment, the participants received a bundle of documents by mail and were asked to complete them prior to visiting the VR lab. This included information and consent forms on the study, together with questionnaires on demographic variables, explicit AGS, and memory self-efficacy that were identical to aforementioned studies but carried out in a paper-pencil mode at home. By transferring these questionnaires to a paper-pencil assessment prior to the laboratory appointment, the procedure for Study 3 was deliberately reduced in comparison to Studies 1 and 2 (see Figure 3). As the earlier procedure had reportedly been demanding on the physical and cognitive levels, a trade off had to be reached with a set of sufficiently valid and reliable performance measures and a minimization of possible fatigue effects toward the end of the laboratory appointment. The resulting 75-min procedure started with the assessment of implicit AGS. Participants were then introduced to the lab room, and the VR headset was mounted. Participants were assigned to the younger (YA_o) or older (OA_o) avatar group, and the introduction procedure remained identical to Studies 1 and 2. The following assessment of performance measures included verbal memory and alpha-numerical working memory, followed by the handgrip strength task. After a break of 2 minutes, the arm-holding endurance task was administered.

The results of Study 3 suggest that what we have obtained for the young sample cannot be simply generalized to the older group. Most importantly, we did not find any evidence of the idea that embodying a younger avatar leads to improved performance parameters, so our second hypothesis was not confirmed. Thus, this part of the project was unable to provide the expected cross-validation. Possibly, the most prominent explanation is that the virtual avatars’ incongruence with the subject’s own appearance itself might have caused the observed performance decline due to an increased cognitive load (Sweller, 2011) caused by distraction or excitement. For the age-different avatar in both the young and old samples, this visual incongruence was possibly stronger than for the avatar that was closer to the subject’s own chronological age. For future research, an additional control condition with, for instance, a gender/age-neutral mannequin or a personalized avatar appears a reasonable step to provide better insight into the relevant mechanism of our virtual reality approach. Furthermore, Study 3 was expected to provide evidence of the development of negative AGS interventions aimed at improving long-term health parameters known from previous research and thus minimizing negative effects during the aging process. Such effects were not found, even though the sample size exceeded what was necessary based on our previous power analysis and the selected performance measures were proven reactive to the intervention among younger participants. The roles of several possible influential parameters remain unclear and must be explored in future studies: a strong variation in chronological age within the two age groups was matched with fixed-age avatars, which poses the risk of strong intra-group variations in self-relevant age perceptions compared to the avatar. Interactional findings from Study 3 support this argument, as participants with a chronological age closer to that of the virtual avatar’s age performed better in the endurance task in both the YA_o and OA_o groups. Again, a baseline difference between experimental conditions in explicit negative AGS was found, which poses a limitation in the results and requires careful consideration in future experimental designs and randomization or matching procedures. Furthermore, the application of our performance tasks might have caused a stereotype threat (e.g., Hess et al., 2003), inducing higher stress among the older sample, thus diluting our results.

Our research proposes a novel approach to studying the short-term effects of activating self-relevant AGS in the lab, and this approach has obvious strengths when compared to more conventional paradigms (Varkey et al., 2006). Our studies, however, also have several weaknesses that might be responsible for the lack of significant findings or might limit the generalizability of our results. The conditions for sample recruitment led to a highly selective sample, as psychology students from our university might have been aware of the content or intention of our study or might have recognized some assessment procedures. The older sample was recruited from a university blackboard, a nearby GP’s waiting room, and newspaper bulletins, which might have caused a somehow selective sample with higher characteristics on positive health behavior, openness to science in general, or interest in VR applications in particular. Such selectivity effects are known from cross-sectional age-comparative studies, even if all relevant measures are taken to prevent them (e.g., Lüdtke et al., 2003). In addition, the sample size was relatively small, especially in Study 2, and our careful randomization procedure nevertheless resulted in partial baseline differences in the experimental conditions concerning the relevant covariate of explicit AGS. A possible lack of identification with the virtual avatar might have resulted from the lack of personalization of the avatars (see Liu et al., 2019). The set of four avatars was chosen to ensure a maximum coverage of our participants’ visual features but, as the avatars were not further individualized, a possible inter-individual variation in avatar identification must be mentioned as a limitation of this piece of research. In addition, more movement trackers could have improved the body illusion effect (Kim et al., 2020). These latter two limitations of the immersion, and hence the validity of our effects, were due to economic and technical factors. In Study 3, several questionnaires were moved from a laboratory assessment to paper-pencil questionnaires prior to the laboratory appointment to make the latter less time-consuming and less exhausting for the older sample. This might limit the comparability of our questionnaire assessments in an age group comparison.

We found that the average endorsement of items measuring self-reported immersion was not the highest possible. Instead of lamenting this fact, we want to draw two other conclusions. First, it seems the VR approach is quite effective in activating self-relevant stereotypes, even when the experience of immersion is not perfect. This is a promising result for future applications of the paradigm. Second, there is room for improvement, and higher immersion could have been obtained with a higher-resolution VR headset, more individualized avatars, sensory feedback on more levels, or a virtual room matching the subjects’ home environments. As stronger effects on cognition and behavior with higher immersion are expected, the full potential of the VR approach is yet to be explored.

Another possible weakness is the possibly invalid outcome variables that were not directly affected by AGS activation. Especially in Study 3, we reduced the overall testing time for pragmatic reasons but meanwhile combined two demanding performance measures in one session. We needed to focus on certain variables and deliberately decided to have the widest possible range of performance domains vis-à-vis the given constraints rather than assessing only one aspect in full. Future research can investigate more diverse outcome variables with the VR paradigm.

It should be noted that concerning the moderating role of pre-existing AGS, they were not manipulated experimentally. Hence, our interpretation that this finding identified the actual mechanism behind the performance declines lacks causal evidence. Future research must manipulate both virtual age and the valence of stereotypes to prove the entire causal chain. Furthermore, the moderating role of pre-existing AGS in both handgrip strength and VMR performance appears similar only at first glance. Upon closer inspection, the initial performance is impaired in both tasks, but only for VMR do we find a compensation of this impairment, as indicated by a steeper learning curve. It would be premature to draw substantial conclusions from this difference in performance patterns, as it would generalize this moderating effect too broadly.

One further limitation is that the test administrators were completely aware of both the assignment to one of the two conditions and to the research question, which might have led to a Rosenthal effect (e.g., Dumke, 1978). Their own expectations or AGS might have resulted in differences in behavior toward the control and experimental condition participants. Blinding the test investigators via automatic selection of and assignment to the conditions appears a feasible adjustment in the future.

As Marilyn Ferguson and Toms, (1982) stated, “Of all the self-fulfilling prophecies in our culture, the assumption that aging means decline and poor health is probably the deadliest” (p. 272). For this reason, research on self-relevant AGS deserves more attention. At the same time, this research faces the fundamental challenge of manipulating age experimentally. We can therefore come closer to an understanding of the causal processes that link stereotypical attitudes to biological and psychological outcomes. Fortunately, current technology allows us to close the gap between what is possible and what would be needed from a methodological perspective. With the advent of VR applications, we might overcome at least some of the obstacles related to lifespan experimental research.

The data and materials for all experiments are available at https://osf.io/rnz62/?view_only=d5a447aec74a4aeeab73e8eb38f0a8a1.

All of the experiments were preregistered using OSF-Framework:

1) Physical performance: https://osf.io/dbns9