1 Introduction

Owing to the limited information on how to correctly perform exercise behavior in the fitness domain, behavior modeling has become a popular persuasive technique used to motivate behavior change. Behavior modeling is a persuasive strategy employed by domain experts and coaches in a physical or virtual setting to demonstrate how to correctly perform a given behavior to an observer [1, 2]. The behavior modeling strategy can be considered successful when the behavior of the observer is visibly changed in the future in line with the observed behavior [3]. In a systematic review of persuasive strategies used in top-ranked fitness apps in the marketplace, Conroy et al. [4] found that exercise behavior modeling (such as instructions, videos and simulations) is the most commonly used persuasive technique in motivating behavior change in the health domain. For example, in virtual (video- or simulation-based) exercise behavior modeling, users can visualize the steps involved in performing a given exercise behavior, just as computer-based learners in the education domain can visualize abstract cognitive processes or concepts of a given subject simulated by the computer [3]. Moreover, video-based behavior models are employed to visualize the positive outcome expectations of certain health behaviors, for example, in exercise behavior, specific muscles that are targeted by specific body movements. Virtual behavior models have some advantages over real-life human models such as gym personal trainers and coaches. They include: (1) they are more affordable; and (2) they are more convenient, i.e., they enable the target users to observe and perform the modeled behavior at any time and wherever that is most convenient for them, e.g., at home, in the hotel, etc. [3, 5].

However, in the fitness domain, there is limited research on the moderating effect of personalization of exercise behavior models on their target users. Personalization is defined as the alignment of the features of a persuasive system to the characteristics of the end-user by the designer of the system [7]. It can be implemented at various levels—individual and group—with user characteristics being defined based on demographic variables such as age, gender, race, ethnicity, etc. In the context of this paper, we focused on group-based personalization based on the race of the end-user, also known as tailoring [6]. Research in the persuasive technology domain has shown that persuasive systems are more likely to be successful if tailored to the user by the designer. Thus, there have been several calls in the persuasive technology domain on designers to tailor or personalize persuasive systems aimed at behavior change to the characteristics of the users, as they are more likely to be successful [8, 9]. However, there is limited research in the area of race-based personalization of persuasive systems to the target user, especially in the health and fitness domain. Although, personalization has been applied in persuasive technology design in different health contexts, e.g., substance use [10], medication adherence [11], etc., there is limited evidence of its moderating effect in the fitness domain in the context of behavior modeling [7].

In this paper, in the context of user perception, we investigated the effect of race-based tailoring of exercise behavior model design on the target users using path analysis. Specifically, we examined the relationship between the perceived persuasiveness of bodyweight-exercise behavior model design and three key social-cognitive determinants of behavior (self-efficacy beliefs, self-regulation beliefs and outcome expectations) [5, 12] in a tailored, contra-tailored and untailored context. Primarily, our investigation is based on two same-race (tailored) observer-model dyads and two opposite-race (contra-tailored) observer-model dyads. They include: (1) black/brown observers who evaluated black/brown behavior models (BB); (2) black/brown observers who evaluated white behavior models (BW); (3) white observers who evaluated white behavior models (WW); and (4) white observers who evaluated black/brown behavior models (WB). In addition, we investigated four other (untailored) dyads which resulted from the combination (summation) of two of the aforementioned dyads: (1) the group of random observers (BB and WB) that evaluated black/brown models (BM); (2) the group of random observers (WW and BW) that evaluated white models (WM); (3) the group of black/brown observers that evaluated random models (BR); and (4) the group of white observers that evaluated random models (WR).

The results of our study showed that the perceived persuasiveness of behavior model design has the potential to influence social-cognitive determinants of behavior, with the relationship between the persuasive design and social-cognitive constructs being moderated by race. Our path models showed that exercise behavior models are more likely to be effective in influencing outcome expectations and self-regulation beliefs than self-efficacy beliefs. In tailored (BB vs. WW) and untailored (BR vs. BR) contexts, behavior models are more likely to be effective in influencing outcome expectations and self-regulation beliefs of black/brown observers than those of white observers. For example, in the untailored context (BR vs. WR), the influence of perceived persuasiveness on self-regulation beliefs and outcome expectations is significantly stronger for black/brown users [(β = 0.52, p < 0.001) and (β = 0.59, p < 0.001), respectively] than for white users [(β = 0.38, p < 0.001) and (β = 0.39, p < 0.001), respectively]. Similarly, in the tailored context (BB vs. WW), the influence of perceived persuasiveness on self-regulation beliefs and outcome expectations is significantly stronger for black/brown users [(β = 0.59, p < 0.001) and (β = 0.62, p < 0.001), respectively] than for white users [(β = 0.41, p < 0.001) and (β = 0.35, p < 0.001), respectively]. However, in the contra-tailored (BW vs. WB) context, though the path coefficients for the three investigated relationships turned out to be higher for the black/brown observers than for the white observers, we found no significant differences between both user groups (BW vs. WB).

In sum, our findings revealed that behavior models, tailored and untailored, have the potential of being more effective among the black/brown user population than among the white user population, especially with respect to self-regulation beliefs and outcome expectations. Moreover, we found that behavior models, among black/brown users, are more likely to be effective, especially with respect to influencing self-efficacy beliefs, if tailored than otherwise. Therefore, we argue that, given that most existing fitness apps on the market are targeted at the western population (especially white users), there is a need for designers to begin tailoring their apps to the racial characteristics of the target populations. Tailoring behavior models to the target users based on race will allow designers to deliver more effective fitness apps that users can identify with and use to support their behavior change.

The rest of this paper is organized as follows. Section 2 focuses on background information. Section 3 dwells on related work, while Sect. 4 on research method. Finally, Section 5 focuses on the result, while Sects. 6 and 7 on the discussion and conclusion, respectively.

2 Background

In this section, we provide an overview of the Social Cognitive Theory (SCT) and behavior modeling in the context of observational learning.

2.1 Social cognitive theory

The SCT is one of the most established behavior theories in the field of Psychology. It was put forward by Bandura [13] to explain human behaviors. As shown in Fig. 1, the SCT holds that personal factors, environmental factors, and the target behavior itself reciprocally influence one another. The bidirectional interplay among the three SCT factors is known as the triad of reciprocal determinism. In this paper, we are particularly concerned with the influence of the environmental factors on three social-cognitive factors of behavior: self-efficacy beliefs, self-regulation beliefs, and outcome expectations. Table 1 shows the definitions of the three factors. The environmental factor includes friends, family or other influential people in face-to-face conversations, mass communication media (e.g., newspaper, radio, television, etc.), electronic communication systems (e.g., mobile phone, computer, etc.) or social networks (e.g., Facebook, Twitter, etc.) where people engage online. Given the critical role mobile technology plays in our lives today, to the extent that mobile devices such as smartphones have become an extension of the human body and are becoming ubiquitous and invisible, it has become important to investigate how mobile technology affects our cognition in the context of beneficial behavior change. Thus, in our study, we are focused on investigating, at the perception level, the effect of mobile-technology-based artifacts on three key social-cognitive factors using bodyweight-exercise behavior models featured in a mobile fitness app as our case study. More specifically, we set out to investigate how the perceived persuasiveness of behavior model design influences the three SCT factors. All of the three SCT factors (self-efficacy beliefs, self-regulation beliefs, and outcome expectations) have been shown by prior research [14,15,16,17,18,19] to be significant determinants of physical activity behavior.

Fig. 1
figure 1

SCT framework depicting the triad of reciprocal determinism [2]

Full size image
Table 1 Study’s constructs and definition
Full size table

2.2 Behavior modeling

Research [22] shows that most behavior changes require new knowledge, though knowledge alone may not be sufficient to bring about the desired behavior change. According to Jimison et al. [22], human cognition (what people know and think) and behavior (people’s action) go hand in hand. Behavior modeling, in the context of SCT, is a behavior change technique through which a domain expert or coach teaches an observer (by way of demonstration) how to perform a given behavior correctly, in either physical setting (e.g., gym) or virtual setting (e.g., video). In this paper, we are concerned with the latter. Behavior modeling stems from the concept of observational learning, which is based on Bandura’s [23] Social Learning Theory (SLT). The SLT, over time, evolved into the SCT. According to Wouters et al. [3], “the social cognitive model of sequential skill acquisition [sic] describes how learners initially start with observing a model, but then start practicing and gradually learn how to self-regulate their own performance” (p. 328). During the observation of the behavior model, Wouters et al. explain [3], the learners constructs a mental representation of the modeled behavior without actually performing it immediately. Eventually, the observed behavior is performed by the observer in different situations and contexts. Wouters et al. [3] argue that the transition from observing the model to self-regulated practice is “accompanied by different cognitive processes and that the interaction takes different forms that all facilitate different cognitive processes” (p. 328). Research has shown that behavior modeling—also regarded as vicarious modeling—can influence behaviors through mediating social-cognitive factors such as self-efficacy, self-regulation, outcome expectation, etc. [24, 25]. In behavior modeling, individuals observe the actions and consequences of the behavior of others and then determine through social-cognitive processes whether to imitate and/or adopt the newly learned behavior. Three reasons have been put forward to explain why behavior modeling has the potential to be effective. They include the following [3]:

  1. 1.

    By observing an expert performing a complex behavior, the observer can construct a sufficient cognitive representation, which enables him/her to rehearse and fine-tune the observed behavior mentally and physically.

  2. 2.

    Observational learning might be more effective than other forms of learning because the expert modeling the behavior does not only demonstrate to the observer what is happening but explains why it is happening as well.

  3. 3.

    While the performance of a complex behavior by novices might place high demands on memory resources such that vital information can no longer be processed, observing an expert performing the complex behavior, on the other hand, tends to free up cognitive resources that can be channeled to processing the vital information.

3 Related work

In this section, we present a review of the literature related to behavior modeling in the health domain. Jimison et al. [22] investigated the effectiveness of using interactive exercise videos for remote coaching of older adults in their homes. They developed an online-based interactive video exercise system (composed of a human behavior model and Kinect camera’s skeleton representation), which was used to demonstrate the performance of certain types of chair exercises to older adults residing in their homes. They found that the online interactive video exercise system can be used to foster adherence to exercise goals despite the challenge in finding adequate room in participants’ homes. However, the authors did not investigate the perceived effect the interactive video exercise system has on participants’ cognition such as self-efficacy beliefs. Oyibo et al. investigated the influence behavior model design has on three SCT factors. They found that the perceived persuasiveness of behavior model design affects self-regulation beliefs, outcome expectations and self-efficacy beliefs (the least affected). However, they did not investigate the moderating effect of race-based personalization of the behavior model design.

Moreover, Devi et al. [26] conducted a study to investigate the effect of breast-feeding behavior modeling and other media influences on postnatal women. The study used a video in which correct breastfeeding techniques and safe practices were taught to postnatal women by a behavior model. They found that the video-based model of breastfeeding—a cost-effective method of disseminating breastfeeding information—yielded positive results, including providing the target women with the right knowledge and fostering a prolonged breastfeeding period. However, apart from the study being in the breast-feeding subdomain of health, it is difficult to tell how much influence the video-based behavior model had on the target audience due to their exposure to other media (e.g., television, newspapers), health personnel, friends and family, etc.) as well during the same time period. Torres et al. [26] carried out an experimental study to evaluate the effects of a video-enhanced activity schedule on exercise behavior among adolescents with autism spectrum disorder (ASD) using an iPad®. They found that the intervention, consisting of a video-based model and graduated guidance, increased the independent schedule-following behavior and on-task behavior of adolescents with ASD, with participants’ skills generalizing to a novel exercise and novel setting (such as a new fitness center). Specifically, they found that the intervention was socially acceptable to behavior analysts, instructors, and paraprofessionals. However, their study, though experimental, did not investigate the effect the video modeling had on cognitive factors. Our current study aims to bridge this gap in a large-scale study. Specifically, at the cognitive level, we set out to investigate the influence of exercise behavior model design on three important social-cognitive factors (outcome expectations, self-efficacy and self-regulation beliefs), which mediate behavior change, and the potential effectiveness of race-based personalization in fitness app design.

4 Method

In this section, we present the research objective, empirical instruments used in measuring the constructs in our research model and demographics of participants who took part in the study.

4.1 Research objective

In this paper, we aimed to investigate the moderating effect of race-based tailoring of exercise behavior models by uncovering the differences that exist between two different groups of observers in tailored, contra-tailored and untailored contexts. This means, in all three contexts, the two comparative groups of interest evaluated behavior models of the same, opposite and random race, respectively. In a nutshell, we aimed to answer this overarching research question—“is race-based tailoring of behavior model design more likely to be effective for both racial target groups than otherwise?” To answer this question, we adopted a prior research model [2] shown in Fig. 2 to study the relationships between the perceived persuasiveness of behavior model design and the three social-cognitive factors of interest.

Fig. 2
figure 2

Research model based on the SCT triad of reciprocal determinism [2]

Full size image

Using path analysis, we investigated whether behavior model designs are more likely to be effective if tailored to the target users or not. Owing to the limited studies in the context of tailoring exercise behavior models to the race of the target users in the fitness domain, we decided to adopt an exploratory approach using Partial Least Square Path Modeling (PLS-PM) [27]. We chose PLS-PM because it is a technique used for predicting response and not for explaining a phenomenon, for which covariance modeling is a better suited alternative modeling technique. While explanatory models may be “perceived to be primarily concerned with testing the faithful representation of causal mechanisms by the statistical model and the estimation of true population parameter values from samples” (p. 2) [28], predictive models, though possibly based on causal mechanisms, are developed in a more exploratory and data-driven way. In other words, a PLS-based predictive model emphasizes the observed data over the underlying theory, which makes it preferable to covariance analysis for the prediction of target constructsFootnote 1 [28]. According to Tobias [29], PLS is a method for building predictive models which lays emphasis on “predicting the responses and not necessarily on trying to understand the underlying relationship between the variables” (p. 1) in the causal model [28]. In our case, we aim at investigating the ability of the perceived persuasiveness of behavior model design in fitness apps to predict SCT factors (perceived self-efficacy, perceived self-regulation and outcome expectation) in the context of tailoring, contra-tailoring and non-tailoring and not the interrelationships among the SCT variables in our path model. In a nutshell, our research design is based on the research model shown in Fig. 2 and the following four observer-model dyads:

  1. 1.

    Black/brown observers who evaluated black/brown behavior models (BB).

  2. 2.

    Black/brown observers who evaluated white behavior models (BW).

  3. 3.

    White observers who evaluated white behavior models (WW).

  4. 4.

    White observers who evaluated black/brown behavior models (WB).

Furthermore, using exploratory analysis, we analyzed the path models for the following four observer-model dyads:

  1. 1.

    The group of random observers (BB and WB) that evaluated black/brown models (BM).

  2. 2.

    The group of random observers (WW and BW) that evaluated white models (WM).

  3. 3.

    The group of black/brown observers that evaluated random models (BR).

  4. 4.

    The group of white observers that evaluated random models (WR).

We examined in a multigroup analysis the statistical significant differences that exist in the respective path coefficients between pairs of path models, e.g., BB versus. WW, BB versus BW, WW versus WB, BM versus WM, BR versus WR, etc. We also examined the numerical differences in the respective path coefficients between a tailored dyad and an untailored dyad (BB versus BR and WW versus WR) in order to see whether perceived persuasiveness is more likely to predict the three SCT factors in a tailored context than in an untailored context.

For the sake of proper organization and easy reference, we present all of our research questions in Table 2, in which the third column represents their symbolic representations. For example, the symbol, BM > WM? or WM > BM?, means: are the relationships between the perceived persuasiveness of the behavior model design and the SCT factors stronger in the BM path model than those in the WM path model or vice versa? Overall, our findings will provide empirical evidence on the moderating effect of tailoring exercise behavior models to the race of the target users.

Table 2 Research questions in words and observer-model dyadic symbols
Full size table

4.2 Research design and measurement instruments

To answer our research questions, we designed a fitness app prototype we called “Homex App,” aimed at encouraging the performance of bodyweight exercise on the home front. The app features eight versions of exercise behavior model designs (see Fig. 6 in Appendix 1), each of which demonstrates to the target user (in text and graphics) how to correctly and successfully perform a given bodyweight exercise. The textual representation of the modeled exercise behavior (see the top-left corner of the application screen) is similar to the Tunneling strategy in the Persuasive System Design (PSD) model [30]. Tunneling is a persuasive strategy in which a user is guided in a step-by-step fashion to perform the target behavior [10, 31]. For example, the tunneling-based instructions for the squat exercise include: (1) Stand upright, with your arms stretched forward; (2) Push your butt backward until hips are lower than knees; (3) Return to the starting position; and (4) Repeat as many times as possible. Apart from focusing on two exercise-types (push-up and squat), in designing the behavior models, we considered gender (male vs. female) and race (black/brown vs. white). However, in this paper, we are focused on investigating the moderating effect of race-based tailoring of behavior model designs to the target users (black/brown observers vs. white observers). In our questionnaire, to put our study in proper context, we described to participants the functionality of the fitness app prototype as shown in Appendix 1. Afterwards, we presented each participant with one of the randomized versions of the eight behavior model designs (a GIF-based video). Next, we requested the participant to answer questions relating to the behavior model design and SCT factors (see Table 3).

Table 3 Study’s constructs and measurement items [20, 21]
Full size table

4.3 Participants

The survey questionnaire was submitted to and approved by the Behavior Research Ethics Board of our university. Afterward, we posted it on Amazon Mechanical Turk (AMT) to recruit participants. Prior to answering the questionnaire, participants were presented with an informed consent form, which explained what the study was about, the potential risks, potential benefits, compensation, rights to withdraws, etc. Participants agreed to the information contained in the consent form before completing the survey. In appreciation of participants’ time, each was compensated with US $0.6. Table 4 shows the demographic information of the valid participants after data cleaning.

Table 4 Demographics of participants (n = 669)
Full size table

5 Results

In this section, we present the results of our path analysis and multigroup analysis (MGA). The analyses were conducted using PLS-PM library (“plspm”) in R programming language [27]. Statistical significance tests were done using the bootstrapping method with 5000 resamples [33].

5.1 Measurement models

Before analyzing the structural models, we evaluated the measurement models and ensured the preconditions for the former were met [33, 34]. The definitions of the preconditions (criteria) and the result of the evaluation of the measurement models are shown in Table 5.

Table 5 Evaluation of measurement models [33,34,35,36]
Full size table

5.2 Structural models

We began our path analysis by building structural models that will help in addressing our first research question: “Overall, are the relationships between perceived persuasiveness of behavior model design and the SCT factors more likely to be stronger when observers evaluate black models or white models?” Figure 3 shows the path models for the evaluation of the two different race-based behavior model designs. The path models represent the relationships between the perceived persuasiveness of behavior model design and the three SCT factors for the two groups of random observers that evaluated the behavior models in an untailored context. In the survey, the first group evaluated black/brown models (BM) and the second group evaluated white models (WM). In the path models, the goodness of fit (GOF) represents how well the respective models fit their data, while the coefficient of determination (R2) stands for the amount of variance of an SCT factor explained by perceived persuasiveness. Finally, the path coefficient (β) represents the strength of the relationship between perceived persuasiveness and an SCT factor. The GOF for BM and WM is 49% and 47%, respectively (roughly moderate), while the R2-value for self-regulation and outcome expectation is about 17%. In particular, the R2-value for self-efficacy is quite low (4% and 1%, respectively). Moreover, the β-value for self-regulation and outcome expectations is over (β = 0.4, p < 0.001). However, with respect to self-efficacy, while the relationship for the black/brown models is significant (β = 0.21, p < 0.01), that for the white models is non-significant (β = 0.10, p = n.s). The MGA will reveal whether the numerical differences between the corresponding path coefficients in the BM and WM path models are statistically significant at p < 0.05 or not.

Fig. 3
figure 3

Path models for the evaluation of black/brown and white behavior models

Full size image

Furthermore, to be able to answer the second to sixth research questions, we built the path models for the untailored (BR and WR), tailored (BB and WW) and contra-tailored (BW and WB) dyads we introduced earlier on (see Fig. 4). The left column represents the black/brown group of observers, while the right column represents the white group of observers. The GOF for the six dyads ranges from 46 to 56%—almost similar to those for BM versus WM path models. However, unlike in the BM versus WM path models, where the β- and R2-values, especially for self-regulation and outcome expectations, are approximately equal for both path models, in the six dyads shown in Fig. 4, both parameters are higher for the black/brown observers than for the white observers. For example, regarding BB versus WW dyads, the β-values for self-efficacy, self-regulation and outcome expectations are higher for the black/brown observers (β = 0.30, p < 0.5; 0.59, p < 0.001; and 0.62, p < 0.001, respectively) than for the white observers (β = 0.11, p = n.s; 0.41, p < 0.001; and 0.35, p < 0.001, respectively). The same applies to the R2-values for the black/brown observers (9%, 35%, and 39%, respectively) compared with those for the white observers (1%, 17%, and 12%, respectively). Again, the MGA will reveal whether the numerical differences between the corresponding path coefficients in the respective comparative path models are statistically significant at p < 0.05 or not.

Fig. 4
figure 4

Race-based tailored, contra-tailored and untailored path models

Full size image

5.3 Multigroup analysis

To investigate the moderating effect of the race of the behavior models and that of the observers in the evaluation of the former, we conducted MGAs as shown in Table 6. The MGA is based on the following six dyadic pairs: (1) BM versus WM, (2) BR versus WR, (3) BB versus WB, (4) WW versus BW, (5) BB versus BW, (6) WW versus WB, and (7) BB versus WW. The results of the MGAs showed that, regarding BR versus WR, there are significant differences in two relationships: PERS → SR (p = 0.059—marginal) and PERS → OE (p < 0.05). Secondly, regarding BB versus WB, there are significant differences in two relationships: PERS → SR (p < 0.05) and PERS → OE (p < 0.05). Finally, regarding BW versus WW, there are significant differences in only one relationship: PERS → OE (p < 0.05). However, regarding BR versus WR, BB versus BW and WW versus WB, there are no significant differences in the three investigated relationships. We discuss the implications of the significant differences between the respective dyadic pairs in Sect. 6.

Table 6 Multigroup analysis for untailored, contra-tailored and tailored dyadic pairs
Full size table

6 Discussion

We have presented a path model of the relationships between the perceived persuasiveness of behavior model design and three key social-cognitive factors (perceived self-efficacy, perceive self-regulation and outcome expectations) in tailored, contra-tailored and untailored contexts. Moreover, we showed the significant differences that exist between key observer-model dyadic pairs in a number of MGAs. In this section, we discuss our key findings and recommend persuasive technology design guidelines accordingly.

6.1 Perceived effect of black/brown and white behavior model design on SCT factors: BM versus WM

Overall, as shown in Fig. 3, the perceived persuasiveness of behavior model design, irrespective of its race characteristic, has a positive influence on the three social- cognitive factors, especially on perceived self-regulation and outcome expectations. In the WM path model, for example, the influence of perceived persuasiveness on perceived self-regulation (β = 0.42, p < 0.001) and outcome expectations (β = 0.40, p < 0.001) is stronger than that on perceived self-efficacy (β = 0.10, p = n.s). The numerical difference between the first two path coefficients and the third is also evident in the tailored, contra-tailored, and untailored path models shown in Fig. 4. These results suggest that irrespective of the race characteristic of the behavior model, the higher users perceive behavior model design to be persuasive, the higher will be their outcome expectations and their perceived ability to regulate themselves to engage in the modeled behavior. Based on these findings, we can conclude that, in the context of perception, irrespective of tailoring, exercise behavior models are more likely to be effective in increasing outcome expectations and self-regulation beliefs and less likely to be effective in increasing self-efficacy beliefs. As a result, we recommend that fitness apps designers should leverage exercise behavior models to promote regular exercise, as their perceived persuasiveness has the potential of impacting outcome expectations and self-regulation, both of which mediate behavior change. For example, as seen in Fig. 6 (Appendix 1), positive outcome expectations (e.g., physical benefits of exercise) can be fostered by highlighting the body parts or groups of muscles affected when users perform a given bodyweight exercise. Moreover, users should be given the opportunity, as a way of self-regulation, to set goals, plan their workouts, track their activities, performance and progress.

Moreover, we found that it is only the perceived persuasiveness of the black/brown behavior model design that has a significant influence on users’ perceived self-efficacy (β = 0.21, p < 0.01). That with regard to white behavior model design is non-significant (β = 0.10, p = n.s). While the numerical difference between both path coefficients is not significant, in the light of our first research question, this result seems to suggest that the black/brown behavior models are more likely to influence users’ perceived self-efficacy than white behavior models. This also seems to be evident in the BB dyad (β = 0.30, p < 0.05) versus WB dyad (β = 0.19, p < 0.01), in which the relationship PERS → SE—unlike that of WW (β = 0.11, p = n.s) and BW (β = 0.23, p = n.s)—is statistically significant. However, as a result of the non-statistically significant difference between BM and WM path models with respect to PERS → SE, there is a need for further research to answer our first research question, which, in other words, reads: “in a one-size-fits-all fitness app, which of the behavior model designs (black/brown vs. white) is more likely to influence the social-cognitive factors of observers?” Meanwhile, a possible answer to this question could be found in the qualitative data we collected (i.e., comments provided by participants on the respective race-based behavior model designs). Thus, we look forward, in future studies, to conducting a sentiment analysis on participants’ comments to uncover any possible explanations as to why the perceived persuasiveness of the black/brown models seems to have a stronger significant effect on observers’ perceived self-efficacy,  in particular, and social-cognitive factors, in general.

6.2 Perceived effect of untailored behavior model design on SCT factors: BR verus WR

Figure 4 (first row) shows the path model for the untailored dyads (BR and WR). Overall, the three model parameters (β, R2, and GOF) are similar to those of BM versus WM in Fig. 3. In the BR and WR path models, the influence of perceived persuasiveness of behavior model design on outcome expectations and perceived self-regulation is stronger than on perceived self-efficacy. Moreover, the three relationships tend to be numerically stronger for the BR dyad than for the WR dyad. It is worthy of note that the PERS → SE relationship (β = 0.15, p = n.s) for BR might have been non-significant because of the relatively small sample size (n = 152) compared with the larger sample size of WR (n = 517) for which the relationship is significant (β = 0.13, p < 0.05). Despite that the PERS → SE relationship is significant for WR and non-significant for BR, the MGA shows that there is no significant difference between both dyads regarding this relationship. However, for PERS → OE and PERS → SR, the MGA results, which are fully significant (p < 0.05) and marginally significant (p = 0.059), respectively, indicate the relationships are stronger in the BR path model than in the WR path model. These findings suggest that, in the light of our second research question, the influence of perceived persuasiveness on outcome expectations and self-regulation is more likely to be stronger among black/brown observers than among white observers. Moreover, the stronger influence of the untailored behavior models on black/brown observers than on white observers is reflected in the amount of variance of the target social-cognitive factors explained by perceived persuasiveness. For example, the variances of outcome expectations and perceived self-regulation (see Fig. 4) explained by perceived persuasiveness for the black/brown observers (BR) are 27% and 34%, respectively, while those for the white observers (WR) are 15% and 13%, respectively. These findings, coupled with those regarding BB versus WW and BW versus WB (see Fig. 4), suggest that the use of behavior models in fitness apps, whether in a tailored, contra-tailored or untailored context, is more likely to be effective among black/brown users than among white users, especially regarding the effect of their perceived persuasiveness on outcome expectations and self-regulation beliefs.

6.3 Race-based tailoring and contra-tailoring of behavior models

In general, the results of our path analyses (Fig. 4) showed that exercise behavior model designs are more likely to be effective among black/brown users than among white users, especially in a tailored context in which the PERS → SE relationship is significant for the former group. For easy visualization of the pairwise comparisons among the tailored and contra-tailored dyadic path models, we summarize the results of the MGAs (from Table 6) in Fig. 5. The results of the MGAs showed that there are significant differences between BB and WB, between BW and WW, and between BB and WW, at p < 0.05. However, there are no significant differences between BB and BW, and between WW and WB, and between WB and BW.

Fig. 5
figure 5

Abstraction of pairwise comparisons of tailored and contra-tailored path models

Full size image

First, regarding BB versus WB, the MGA showed that the perceived persuasiveness of the black/brown model design has a significantly stronger effect (p < 0.05) on the self-regulation beliefs (β = 0.59, p < 0.001) and outcome expectations (β = 0.62, p < 0.001) of black/brown observers than those of white observers: (β = 0.37, p < 0.001) and (β = 0.36, p < 0.001), respectively. Overall, this suggests that, in a real-life fitness app, black/brown models are more likely to be effective for black/brown users (tailored) than for white users (contra-tailored) in impacting outcome expectations and self-regulation.

Second, with respect to BB versus WW, the MGA showed that, in a tailored context, the perceived persuasiveness of behavior model design is more likely to influence the self-regulation beliefs (β = 0.59, p < 0.001) and outcome expectations (β = 0.62, p < 0.001) of black/brown observers than white observers: (β = 0.41, p < 0.001) and (β = 0.35, p < 0.001), respectively. (Note: the significant difference between the two groups with respect to self-regulation beliefs is marginal: p = 0.059.) Overall, this finding suggests that, in a real-life fitness app, tailored behavior models are more likely to be effective among black/brown users than among white users in influencing outcome expectations and self-regulation.

Third, with respect to BW versus WW, the MGA showed that the perceived persuasiveness of the white model design has a significantly stronger effect (p < 0.05) on the outcome expectations (β = 0.56, p < 0.001) of black/brown observers than those of white observers (β = 0.35, p < 0.001). The result suggests that, in a real-life fitness app, regarding outcome expectations, white models are more likely to be effective among black/brown users (contra-tailored) than among white users (tailored). This finding, coupled with that in the prior paragraph, suggests that, regarding outcome expectations, behavior models—whether tailored or contra-tailored—are more likely to be effective among black/brown users than among white users. This confirms the prior finding in Sect. 6.2, in which we saw that, in an untailored context (BR vs. WR), behavior models are more likely to influence the outcome expectations of black/brown users (β = 0.59, p < 0.001) than those of white users (β = 0.36, p < 0.001).

6.4 Summary of main findings

All of our findings based on the MGA results in Table 6 are summarized in Table 7 to provide answers to the research questions we posed earlier in Table 2.

Table 7 Summary of answers to research questions
Full size table

As shown in Tables 6, 7 and Fig. 4, we see that, regardless of the race of the observers, behavior model design is more likely to be effective in enhancing outcome expectations and self-regulation beliefs than self-efficacy beliefs. Secondly, we see that, in general, behavior models, whether tailored (BB vs. WB/WW) or contra-tailored (BW vs. WW) or untailored (BR vs. WR), are more likely to have a greater effect on outcome expectations for black/brown users than for white users (see RQ2, RQ7, RQ8, and RQ9). Thirdly, as seen in RQ2, RQ7 and RQ9, behavior models, whether tailored (BB vs. WB/WW) or untailored (BR vs. WR), are more likely to have a greater effect on self-regulation beliefs among black/brown users than among white users.

To wrap up the discussion, we would like to revisit the overarching research question of this study—“is race-based tailoring of behavior model design more likely to be effective for both racial target groups than otherwise?”—in the light of our results. Going by the MGA results for BB versus BW (tailored vs. contra-tailored) and for WW versus WB (tailored vs. contra-tailored), the answer to our question seems to be “No,” as there are no significant differences between both comparative path models in either case regarding the three SCT relationships. However, for black/brown users, given the following:

  1. 1.

    the three relationships in the path models are numerically stronger when the behavior models are race-tailored (BB) than otherwise (Fig. 4);

  2. 2.

    the relationship between perceived persuasiveness and self-efficacy beliefs is only significant (β = 0.30, p < 0.05) when the behavior models are tailored (BB); and

  3. 3.

    there are significant differences between BB and WB (tailored vs. contra-tailored) but none between BW and WB (contra-tailored vs. contra-tailored).

we may conclude that behavior model design in fitness apps are more likely to be effective if tailored to the race of black/brown users than otherwise. However, this conclusion needs further investigation in future research efforts to confirm it.

On the other hand, with respective to white users, we found no evidence that tailored behavior models are more likely to be effective than contra-tailored or untailored behavior models. Specifically, for the three relationships, we found the following:

  1. 1.

    There is no substantial numerical difference between WW and WR (see Fig. 4).

  2. 2.

    There is no statistically significant difference between WW and WB regarding all three relationships (see Table 6).

However, unexpectedly, we found that the relationship between perceived persuasiveness and self-efficacy beliefs for WB (β = 0.19, p < 0.01) is significant, but for WW, it is non-significant (β = 0.11, p = n.s). These findings seem to suggest that white users are very likely to be indifferent to the race characteristic of the behavior model design. However, more investigation still needs to be carried out in the future to verify this hypothesis.

6.5 Contributions

Our main contribution to the body of knowledge is that our study provides empirical evidence on the potential effectiveness of the race-based personalization of behavior models in fitness apps aimed at exercise behavior change, especially for black/brown users. We showed that, in general—in an untailored context—behavior models, in fitness apps, are more likely to be effective among black/brown users than among white users. Secondly, we showed that race-based personalization is more likely to be effective among black/brown users than among white users. Finally, for black/brown users, we demonstrated that tailoring of behavior models is more likely to be effective than contra-tailoring and non-tailoring. In the context of fitness apps aimed at promoting regular bodyweight exercise, our study is one of the first to contribute these findings to the body of knowledge.

6.6 Limitations

There are several limitations to our study. The first limitation is that our study is based on user perceptions and not the actual use of fitness apps featuring exercise behavior models. Thus, our findings may not generalize to the context of actual use, in which users have to interact with the behavior models through monitored (logged) usage over a given period of time. For this reason, we recommend that, in future research efforts in the area of race-based personalization, researchers investigate our findings in an actual use context to uncover how far our findings generalize to or differ from a real-life setting.

The second limitation of our findings is that we treated black/brown users as a monolithic group given that both subgroups belong to the collectivist culture based on Hofstede’s framework of cultural classification [37, 38]. Similarly, in our study, we did not differentiate between black and brown behavior models in terms of their physical characteristics, e.g., skin color, hairstyle, etc. Rather, our black/brown behavior models seem to be more tailored to the physical characteristics of black users than brown users, e.g., based on skin color, hairstyle, and physique. As a result of these limitations, we recommend that future research efforts in this area should differentiate between black and brown users to uncover the influence of the acknowledged limitation on our study and provide more nuanced findings related to race-based personalization of exercise behavior model design in the fitness domain.

The third limitation of our study is that our findings may neither generalize to the entire white race nor black/brown race, as our study mainly focuses on both groups of users from North America (Canada and United States). As a result of this limitation, we recommend that, in addition to investigating our findings in an actual application setting, future research efforts should focus on users from other countries and continents to uncover how our current findings generalize to them.

7 Conclusion

In this paper, we presented the path model of the relationships between perceived persuasiveness of behavior model design and three social-cognitive factors (outcome expectations, self-regulation beliefs and self-efficacy beliefs) in a tailored, contra-tailored and untailored contexts. In addition, we presented the results of our MGAs aimed at uncovering the potential effectiveness of tailoring exercise behavior model design in fitness apps to the target users. Overall, our path models showed that exercise behavior models are more likely to be effective in influencing users’ outcome expectations and self-regulation beliefs than self-efficacy beliefs. Moreover, we found that, in a tailored context (BB vs. WW) and untailored context (BR vs. BR), behavior models are more likely to be effective in influencing the outcome expectations and self-regulation beliefs of black/brown observers than those of white observers. However, in the contra-tailored context (BW vs. WB), we found no significant differences between both user groups, though the path coefficients for the three relationships turned out to be higher for the black/brown observers than for the white observers. Overall, we did not find a statistically significant difference between tailored and contra-tailored behavior models for the black/brown observers (BB vs. BW) and the white observers (WW vs. WB). However, our path models indicated that race-based personalization is more likely to be effective for black/brown users as the path coefficients for the three relationships turned out to be greater in the tailored than in the contra-tailored and untailored context. Based on this finding and the finding that the relationship between perceived persuasiveness and self-efficacy beliefs is statistically significant when behavior models are tailored to black/brown users, we recommend that fitness apps featuring behavior models should be specifically tailored to this group of users. In future work, we look forward to conducting sentiment analysis on participants’ comments on the behavior model designs and studies among other populations to determine the generalizability of our findings.