It is still hard to define academic persistence, as consensus has not been made concerning the definition of persistence. There are two major controversies: (1) trait vs. state; (2) help-seeking behavior. Early research defined academic persistence as the consistent investment in learning despite obstacles, difficulties, failures, and situations (Zimmerman and Risemberg, 1997). However, recent research argued that persistence is more complex (Roland et al., 2018). It can vary from person to person, depending on the situation and one’s personal preference. The early definition took the trait-dependent view and suggested that persistence is a stable trait, which means that a persistent person will struggle through hardship to achieve their goals across various domains and settings (Sellman et al., 1997; Sommer and Baumeister, 2002). However, the later definition took the state-dependent view that persistence is no more than a state, so people who persist in one context might not persist in a different situation (Baker et al., 2008; Hershkovitz and Nachmias, 2009).

The second controversy is around help-seeking behavior. Even though help-seeking behaviors could help students continue on a challenge or difficult task (Jackson et al., 2003; Terrell et al., 2015), most of the time, persistence was seen as a self-reliance physical and mental process. Some researchers viewed help-seeking behavior as a sign of weakness (Tyssen et al., 2004).

This research would like to propose a new definition for academic persistence to address the conceptual issues stated above. As discussed, previous research had debates around trait vs. state and whether help-seeking is a form of persistence. This research would like to move away from trait vs. state debate and trade help-seeking as a form of persistence because student overthrows their psychological barrier of being seen weak and tries to achieve their academic goals by asking for help. The current study defined academic persistence as an individual’s self-perceived and actual willpower and behaviors (may vary in different situations) to overcome obstacles, difficulties, and failures by oneself or by seeking help from others to achieve learning goals. This definition contains three types of persistence: self-perceived (trait), self-reliance (state), and help-seeking persistence (state). Self-perceived persistence is how a person thinks he or she will behave when facing difficulties and obstacles. The latter two types of persistence are individual’s actual reactions when facing difficulties or challenges in learning. Self-reliance persistence means a person continues to work hard on the problem by oneself. Help-seeking persistence is defined as a person seeks help from others to overcome obstacles and complete an arduous task.

Persistence is not only hard to define, it is also extremely hard to measure and quantify because of its multidimensionality and dynamic nature. Scholars used self-report scales (Renaud-Dubé et al., 2015) and the amount of time invested (Pelletier et al., 2001; Jõesaar et al., 2011) measure persistence. However, experimental studies of persistence are limited and often do not reflect real-world problem-solving scenarios. More importantly, few experimental studies have captured the multidimensionality of persistence. In this study, a self-reported scale and experimental design are implemented in order to capture the three aspects of persistence. The measurement of academic persistence will be specified in the “Materials and Methods” section.

Besides the calling for a better understanding of academic persistence, it is also essential to understand what supports students’ academic persistence. Whether persistence could depend on (a) students’ emotions (Tulis and Ainley, 2011); (b) autonomy support (Pelletier et al., 2001). The following section will review the literature on emotion and autonomy to explain how these two factors would influence academic persistence.

Emotions in this study were defined as the various emotions directly induced by learning activities and learning outcomes. Traditionally, researchers study cognition and emotion separately, and emotions have not been studied intensively in education before the 1990s (Pekrun, 2019). The recent 20 years sees a rise in studies of emotion in education, as emotions were discovered to activate and deactivate cognition and metacognition processes (e.g., persistence) related to learning (Artino and Jones, 2012; King and Areepattamannil, 2014; Ramirez-Arellano et al., 2019).

Emerging literature addresses the importance of emotions in the online learning context (e.g., Feidakis et al., 2014; D’Errico et al., 2016). O’Regan (2013) concluded that emotions played a central role in students’ lived online learning experience through interviews with eleven students. Both anxiety and excitement were stood out in students’ discussion of the online learning experience. D’Errico et al. (2018) also detected and classified 11 cognitive emotions students showed in video-lecture and chat with teachers. Parlangeli et al. (2012) argued that, within online learning, cognitive emotions were crucial, but social emotions also needed attention. Most of the studies concerning emotions in an online context aimed to address the importance of emotions or identify the types of emotions students demonstrated, but not much research explored the association of these emotions and students’ academic persistence.

However, in the traditional face-to-face learning context, there was burgeoning consciousness of the significant role of emotions in students’ academic performance (Tulis and Fulmer, 2013). Anxiety, especially test and math anxiety, was studied massively as a predictor for academic performance (Cassady and Johnson, 2002; Zeidner, 2014; Putwain et al., 2016). Anxiety is the outcome of negative (unpleasant) emotions like anger and frustration, and most of the time associated with academic performance negatively (Chapell et al., 2005; Karatas et al., 2013). There is also joy, which is seen as the outcome of enjoyment. Such joy of learning deepens the learning process and promotes academic achievement (Goetz et al., 2008; Villavicencio and Bernardo, 2013; Putwain et al., 2016).

Not only associated with academic performance, joy and anxiety also potentially related to academic persistence. Students who experience positive emotion (joy) would perceive themselves have enough ability or resources to achieve their goal. On the other hand, students with unpleasant emotions (anxiety) would be frustrated by the current situation and avoiding continuing their goals (Linnenbrink and Pintrich, 2002; Ainley et al., 2005; Tulis and Ainley, 2011). Aforementioned research suggested a potential association between these two emotions and students’ persistence, so the joy of solving specific problems or the anxiety activated by failures or challenging tasks would be this study’s focused emotions.

Most previous studies used questionnaires to measure emotions, for instance, the Achievement Emotions Questionnaire (AEQ; Pekrun et al., 2011) and the Epistemically-Related Emotion Scale (EES; Pekrun et al., 2017). However, the research mentioned above, and the self-reported questionnaires have several limitations. Firstly, they did not provide the real-time emotional status during students’ problem-solving process. Secondly, self-reported data has widely acknowledged drawbacks: for instance, participants would conceal their real opinion, and discrepancies might exist between how people behave and how people think they would behave (Sallis and Saelens, 2000; Subar et al., 2015). Thirdly, self-reported data does not provide observed valence and activation of synchronous emotions during task completion, but these characteristic of synchronous emotions are essential for learning (Pekrun, 2006).

To address these limitations, this study incorporated the Contain Intelligent Facial Expression Recognition System (CIFERS) to measure two dimensions of emotions: valence (positive or negative) and activation (activating or deactivating) (Pekrun, 2000, 2006). CIFERS could track students’ macro- and micro-facial expressions as indicators for different emotions (both positive and negative). It could also provide information on real-time emotion change, emotion intensity (activating and deactivating), as well as the specific time an emotion occurs. Before elucidating more on CIFERS in the “Materials and Methods” section, certain suspicion must be squelched: why use a facial expression as an emotion indicator and whether this approach is accurate? The implication of facial expression in emotion studies was presented below to answer these two questions.

Facial expressions have long been used to indicate emotions and stayed central in emotion studies (Tomkins and McCarter, 1964; Russell, 1994; Ruba and Repacholi, 2020). The accuracy of using facial expressions to identify emotions has been justified through many ways, for instance, self-report instruments (Matsumoto, 1987; Matsumoto et al., 2000) and facial coding systems (Ekman et al., 1980; Clark et al., 2020; Rosenberg and Ekman, 2020). There is debate around the universality of facial expression. Early research discovered that even people in an isolated tribe in New Guinea shared the same emotional interpretation of facial expression (Izard, 1992). This finding was later replicated by Matsumoto (1992) and Ekman (1994). Other researchers questioned such findings. For instance, Jack et al. (2009, 2012) argued that facial expressions are not universal. However, they can only prove that the intensity of emotions and degree of the movement of people’s faces are different. More importantly, the differences they identified did not exist in facial expression but in how people use their own cultural understandings to interpret facial expressions. In this paper, we believed that facial expressions, both macro and micro facial expressions, are shared by different cultures; only the intensity and interpretation might be different from culture to culture (Ekman and Friesen, 2003; Cowen and Keltner, 2020).

The CIFERS equipment adopted in this research was established based on Ekman and Rosenberg (1997) and Rosenberg and Ekman (2020) facial expression theory and facial coding systems, which divided the face into 47 units. With CIFERS, the facial movement could be obtained within 50 ms. CIFERS’s basic mechanism is out of the scope of this research, but more information could be found in Supplementary Appendix A and previous studies (Scherer and Scherer, 2011; Krumhuber et al., 2012). The CIFERS has one more advantage: its artificial intelligent feature allows it to improve its own accuracy through data collection. It has already been trained and improved by more than 100,000 people’s emotional data before this study (see Supplementary Appendix A for more information). The CIFERS collects macro-facial expressions and micro-facial expressions, which means even when students try to conceal their emotions, the machine could still identify that emotion.

Besides emotions, another factor that would affect academic persistence is autonomy support. According to self-determination theory, autonomy is the basic psychological need to make choices without pressure, external control, or compulsions (Deci and Ryan, 2000). It has been primarily acknowledged that having autonomy would support learning persistence (Pelletier et al., 2001; Vansteenkiste et al., 2004; Bonneville-Roussy et al., 2013). Specifically, in Pelletier et al. (2001) study, student-athletes who perceived more autonomy support were persistent longer in the sports that they play. Bonneville-Roussy et al. (2013) also found in a longitudinal study that college students are more persistent in learning within an autonomy-supportive environment.

Autonomy support not only associates with academic persistence but also impacts an individual’s emotional function. In Wang et al. (2007) longitudinal investigation, if under an autonomy support parenting style, children had functioned better emotionally, while a constraining parenting style would dampen children’s emotional functioning for both the United States and Chinese seventh-grade students. Another research also supported such findings. If parents and teachers showed more support for children’s autonomous behavior, children’s emotions would be more positive, and they were better at emotion regulation (Liew et al., 2011). Most of the time, autonomy support was manifested as providing choices to students (Benita et al., 2014; Lewthwaite et al., 2015). In this study, we adopted the same approach of conveying choice as a way of autonomy support.

To determine the required sample size, we conducted the Power Analysis based on the root mean squared error of approximation (RMSEA) in Structural Equation Modeling. The results showed that to achieve the power for acceptable RMSEA, the lower bound of sample sizes is 66 in each group. There were 177 college students randomly sampled from a university to ensure sufficient power for statistical inference. An experiment related to persistence was performed with participants being randomly assigned to either an experiment group (n = 88) or a control group (n = 89). In the control group, participants were allowed to choose which type of task (either math or literacy) they prefer to complete. Participants will be given no choice in the experimental group. The demographic information was presented in Table 1.

Before starting the experiment, participants were asked to answer an online survey. The survey includes demographic information, a persistence scale, and an anxiety scale.

The experiment aims to determine whether students exhibit different persistence levels under the choice or non-choice scenarios and different emotions. Participants had the chance to solve either math or literacy problems. Each participant was asked to answer 25 questions in both the experiment and control group. The questions were taken from the Cultural Fair Intelligence Test (CFIT) from the Genius Tests.¹ The CFIT provides types of tasks suited to the various task conditions in this experiment.

Participants would sign up for the study through a university’s data collection system, and then the online system assigned a four-digit research ID to the student. A Zoom meeting link would be provided to the participant. The instruction would notify the participant to temporarily change their Zoom ID to their four-digit research ID temporarily before the experiment Zoom meeting. When a participant joined the research Zoom Meeting, the investigator would send him/her a consent form via the Zoom chat function. The participant would E-sign a consent form (concealed the emotion tracking information for the test’s accuracy), which indicated all the experiment information and clarified that participants could drop out of the study anytime if they feel uncomfortable. After signing the consent form, the investigator would send out an online survey (specified above) link via the Zoom chat function. When the participant finished the survey, he/she would be randomly assigned to a treatment condition (choice vs. no choice).

Participants in the experiment group could choose freely from the two groups of tasks (i.e., math and literacy tasks). Participants in the control group were presented with the math and literacy tasks, but the investigator will assign only one task type randomly to participants without providing any option for choice. Participants were informed that the task has 25 questions, and there was no time limit. Whenever they answer a question wrong, they can work on the questions by themselves more, or click the hint button, or skip the question. After assigning the group, if the participants did not have questions about the task, the investigator would be on mute and turn off the video to give the participant time to solve the 25 questions. The emotion tracking machine (specified in the next section) would be started at this point to capture participants’ facial expressions. Participants were asked to show their faces and try to face their camera the whole time.

The instructions in the assigned task stated that participants could answer as many items as possible correctly with no time limit. When participants answered a question incorrectly, they were given three options. (1) They can skip the question, in which case their answer will be considered wrong. (2) They can request a hint and then continue solving the question; if they come to the correct answer after receiving the hint, their answer will be considered correct. (3) They can continue to try to solve the question by themselves without a hint until they get the right answer. For example, if the question is, Bruce likes 324 but not 325. He likes 2,500 but not 2,400. He likes 121, but not 122. Which does he like? (a) 900; (b) 800; (c) 700; (d) 600. If the participant’s answer is (a), then he/she gets it correct and will automatically move to the next question. If the student’s answer is not (a), then he/she can choose to skip the question by clicking the skip button and move to the next question. The participant can also choose the hint option by clicking the hint button, and then the system will show the hint: James likes square numbers, then the student can continue to solve the problem. Persistence would be calculated according to students’ actual behaviors, and the specific method will be discussed in the “Plan of Analysis” section.

After the participant finished the task, he/she would be asked to sign a post-experiment consent form which indicated that their facial expression data were captured. If they allowed the research team to use the data, they would sign the form. After E-sign the post consent form, the participant could leave the Zoom session.

A facial tracking system was running through the duration of the experiment as participants complete their assigned tasks to track their emotional state. The Contain Intelligent Facial Expression Recognition System (CIFERS) camera would capture participants’ real-time facial expressions from the Zoom window. CIFERS is a software program that uses macro- and micro-expression data modeling to track individuals’ facial movements and determine participants’ cognitive and psychological states. The software recorded participants’ 11 basic emotions and emotional changes over time (more information about the equipment appears in Supplementary Appendix A). This study used joy and anxiety as the targeted emotions.

Data analysis was conducted following two steps. In the first step, we categorized each item responses for each participant into three types of persistence behaviors (self-reliance behavior, help-seeking behavior, and low persistence) according to three criteria: (1) the number of times participants click the “submit” button (participants have to try at least one); (2) whether they clicked “hint” button, and (3) whether they skipped an item. Moreover, the total response time for each group of items was recorded (T_sr, T_hs, T_l) which represented the time length they showed self-reliance behavior, help-seeking behavior, and low persistence accordingly during the experiment. In the second step, we analyzed five regression models in which the three types of persistence were considered as outcomes.

As shown in Table 2, each participant’s persistence was defined using participants’ behaviors. Specifically, the criteria for item-level persistence were as follows: items with self-reliance behavior – the individual responded to an item more than once, did not ask for hints, and did not skip the item; items with help-seeking behavior – similar to self-reliance persistence except the individual asked for hints; items without persistent behavior – the individual answer incorrectly and skipped the items, so they did not demonstrate high or moderate persistence. The persistence index Yp in this study was defined as the total time spent on the items which show specific persistence behavior for participant p. We set J as the total number of items, p as the person index, and i as the item index. For example, each participant’s time interval showing self-reliance behavior can be computed as follow:

where Y_p is a P × 3 test-level persistence score vector representing three types of persistence scores for total participants. I_j is an indicator matrix for item j suggesting whether this item shows specific persistence behavior, whose values were either 0 or 1 and I is J × 3 matrix; for example, I_1p = [1, 0, 0] indicated that Item 1 shows self-reliance behavior for person p. T_pj represents the time response vector for item j answered by person p and then T is a P × J × 3 matrix.

We use the regression model below to examine Hypothesis 1:

where Y_persis is one of two types of persistence indices of interest (help-seeking persistence or self-reliance persistence), SS (sum scores) represents the achievement scores which were computed with the number of items each participant answer correctly, RT (response time) represents the total time each participant used, choice indicates whether the participant was allowed to select the type of task.

To examine Hypothesis 2, we estimated the following regression models:

where Anxiety and Joy are the maximum level of that emotion of participants during the experiment; gender indicates whether the participants are females or males; trait_of_anxiety indicates the state of anxiety level measured by State-Trait Anxiety Inventory. There were two major reasons for using the maximum level of anxiety and joy. Firstly, according to previous literature, only a higher level of emotional arousal would induce behavior and influence the decision-making process (Kaufman, 1999; Hu et al., 2015). Secondly, using other indices, for instance, mean (total emotion show/time), would be inaccurate since students were not showing joy or anxiety all the time during the experiment.

To examine Hypothesis 3, we used the regression model as follow:

Finally, to examine Hypothesis 4, a structural equation model containing all variables was fitted to test the effects of emotion and choice condition (see Figure 2). In SEM, the dependent variable was either the emotion index (Anxiety or Joy) or persistence index (Y_persis); independent variables were the choice condition, anxiety, and joy. Other controlling variables included anxiety traits, achievement scores (SS), and average response time (RT). To examine the goodness-of-fit of SEM, root mean square error of approximation (RMSEA), and maximum likelihood (ML)-based standardized root mean squared residual (SRMR) were reported to evaluate the adequacy of the model. RMSEA and SRMR values close to or lower than 0.08 are acceptable, although values approaching 0.05 are preferable (Hu and Bentler, 1999).