1. Introduction
From 2016 to now, the Ministry of Education (MOE) of Taiwan conducted a digital application promotion plan in rural schools to improve digital skills, enrich digital competency (e-commerce and e-marketing), and promote enjoyment of mobile services as well as applications. The MOE e-learning partner program was one part of the plan [
1]. Based on companionship, learning, and improving the motivation and interest of rural students, undergraduates used the internet to overcome barriers between urban and rural areas. The purpose was to promote equal learning opportunities for elementary and secondary students everywhere. In addition, the core values of this program included life accompanying life and living teaching living. There are two main targets: one is the learning of the children and the other is the teaching of the undergraduates. There was a total of ten weeks in one semester, with classes twice a week and two lessons each time (Chinese, English, or Mathematics). Each lesson lasted for 45 min with one-on-one online interaction. Regarding the partners, there were five primary and secondary schools that cooperated with the case university from 2017 until now. Many studies reported good performance of the e-learning partner program [
2,
3,
4,
5]. In the case university, the e-learning partner program has been promoted for four semesters. However, the outcome report of this program stated that the retention rate of the undergraduates was below 30% over three semesters [
6]. At the beginning of each new semester, the program hosts had to invite lecturers to announce and impart the same knowledge and skills as those of the e-leaning partner program. The problem, as per program hosts, was considered an economic loss for the universities and talent loss for the team members of the program.
This present study aims to collect specific data of the undergraduates’ inner characteristics and the outer environmental support to examine the significant factors that influence the persistence of the program. A sound and complete model can predict the user performance with technology by the analysis of personal learning behavior [
7]. We use an advanced procedure to set up the predictive model of undergraduates’ self-evaluation in an online e-learning partner program to improve the high talent turnover problem and decrease the cost of human-resource training.
The evaluation of self-efficacy will directly affect a person’s motivation for the job. Yesilyurt et al. concluded that teacher self-efficacy, academic self-efficacy, and computer self-efficacy are important predictors of prospective teachers’ attitudes toward applying computer-supported education [
8]. Also, personalization features can match the user’s attitude and behavior; therefore, they could enhance the user’s performance and persistency with technology [
9]. Some studies reported that there was a positive relationship between self-efficacy and positive attitude and predicted the attitude [
10,
11,
12]. The variables of teaching efficacy and teaching satisfaction were also discussed [
13,
14]. In order to explore the results of the undergraduates’ retention during one semester in the case university, collecting data on teaching, counseling, administration, and equipment evaluation was necessary.
Furthermore, the retention factor and designing of a questionnaire and its implementation during the mid-term and the end of the semester were important. For this reason, it is essential to detect which variables affect prospective undergraduates’ attitudes toward applying for the e-learning partner program and to devise a predictive model in this program. In the e-learning partner program, there were 94 teaching undergraduates who completed the survey at the case university. When using a small sample, the partial least squares structural equation modeling (PLS-SEM) can be used to estimate the relationships between latent variables with several indicators [
15,
16]. Moreover, the method was applied in various research disciplines, such as marketing [
17,
18,
19], management [
20,
21,
22,
23], finance [
24,
25,
26], hospitality and tourism [
27,
28,
29], information [
30,
31,
32], social media [
33,
34], psychology [
35], and science [
36], and it was shown as capable of solving real-world problems. Many studies stated the predictive estimation capabilities of PLS-SEM [
37,
38,
39]. Exploratory factor analyses of scales were analyzed via SPSS. For the confirmatory factor analyses of scales and the structural equation modeling, partial least squares (PLS) was used to predict the latent variable of retention in this program.
The aims of this study are (1) diagnosing the outcome of self-evaluation of undergraduates and determining which factors are significantly different, (2) evaluating the correlation among scores of self-evaluation, and (3) predicting the willingness of retention by carrying out PLS modelling. From the suggestions for improvement, a decrease in the gap of undergraduates’ expectations and enhanced sustainability in the e-learning partner project can be achieved.
2. Literature Review
An interactive environment, administrative support, and perceived institutional resources influenced the learning completion intention of undergraduates [
40]. Thus, teaching, counseling, administration, and equipment were included in our evaluation. Related to inner assessment, self-efficacy was described in many studies. The prediction method of PLS was also applied in many researches.
2.1. Self-Efficacy
Self-efficacy, as developed by Bandura, indicated that people had confidence in a relevant area [
41]. Effectiveness and self-efficacy of teaching were discussed in many studies [
42,
43,
44]. Self-efficacy was effective in different aspects such as classroom management, teaching methods and techniques, and the use of computers and instructional tools [
45]. In addition, it also affected students’ learning performance. Sun found that environmental transformation, teaching innovation, class management, parent-teacher communication, teaching execution, and teaching evaluation influenced teaching quality [
44]. Xu found that teacher self-efficacy was positively related to the teacher’s perceived pressure, class management style, teaching-thinking style, commitment to teaching, willingness to carry out teaching innovation, and students’ learning performance [
46]. In this study, considering teaching self-efficacy and retention (passion and innovation), teaching indicators that included teaching preparation, teaching methods, and teaching attitude were adapted from Danielson, Keller, Pan et al. and the Ministry of Education [
47,
48,
49].
Secondly, counseling self-efficacy is defined as a belief in the ability of individuals to perceive counseling cases effectively [
50]. According to Larson and Daniels, studies on counseling self-efficacy are mostly conducted to understand the self-efficacy of counselors, including trainees or graduate students, school counselors, psychologists, or mental health-related personnel. Regarding the factors that affect the effectiveness of counseling, Larson and Daniels found that counselor characteristics, sexual orientation, age, training background, work experience, and other variables were important factors in predicting the self-efficacy of counselors in multiple studies. Case studies based on self-assessment by trained counselors have shown that those with higher self-efficacy in counseling have higher self-evaluations [
51,
52]. There are three sections related to counseling training: knowledge, skill, and belief [
53]. In this study, counseling self-efficacy is divided into two parts. One is the attitude when tutoring the child (this includes understanding the child’s position, accepting the child, trusting the child’s ability, and respecting the child). The other is the positive perception of counseling knowledge and attitude.
Finally, the outer environment support and institutional resources, as well as administrative support and IT equipment were two important factors in the questionnaire that were designed by the MOE of Taiwan. The program hosts and assistants shared their experiences with the administrative support. With regard to the IT section, the platform of MOE, the e-learning interactive equipment (such as the software and the writing board) and the computers’ network flow speed were employed to investigate the self-evaluation of the undergraduates during the semester at the case university.
2.2. Retention
According to [
54], individual characteristics such as background and attitude affect students’ dropout rates. Vianden and Barlow studied student loyalty with the assumption that students who develop positive attitudes toward their institutions are more likely to continue [
55]. Furthermore, passion for long-term goals predicted retention among novice teachers [
56,
57]. Some researchers showed a strong link between talent retention and innovation [
58,
59]. Therefore, innovation and passion were two key factors related to retention in this paper. The control variables for the background used in this research were college, gender, subject, e-learning partner experiences, and teacher education program background.
2.3. PLS
PLS-SEM is a method used to estimate path models with latent variables and extend the principal component and canonical correlation analysis in statistics [
60,
61]. PLS can cope with smaller sample sizes than structural equation modeling (SEM) for the same effect size and model complexity, and it can more easily specify formative constructs [
62,
63,
64]. In the formative model, it was necessary to assess the indicator weights and loads, and perform redundancy analyses. Chin provided redundancy analysis, in which each formatively specified construct correlated with its alternative measure [
65]. The SmartPLS 3 software with a graphical user-interface was used to estimate the PLS-SEM models [
66]. The first PLS software was published nearly ten years after LISREL III [
67]. A relational predictive model is a research model that aims at determining the presence and the extent of the retention among the four self-evaluation variables.
3. Methods
In this context, the effect of the undergraduate retention toward using teaching, counseling, administration, and equipment evaluation on each other and their mutual relation are considered in this study.
3.1. Samples and Procedure
This research was conducted in two stages. The first was the factor analysis and the reliability and validity analyses of all items of the questionnaire to reduce the number of questions. ANOVA and post hoc analyses were conducted to determine the significant factors. The demographic features of the participants were gender, year level, college, teaching subjects, experience, and teacher education program background. Through ANOVA, we detected which of the background variables were significantly different in this study. Through post hoc analysis, multiple comparison analysis (over two levels) was explored. Furthermore, the PLS modelling in the final semester questionnaire was created to detect the correlation between retention (passion and innovation) and the four factors (teaching, counseling, administration, and IT equipment). Thereafter, it was extended to predict the future trend in this e-learning partner program to enhance the perseverance in the tutoring and teaching side.
The students completed the first questionnaire in the fifth week of the semester and the second one in the tenth week. In the 2019 academic year (from 21 October 2019 to 20 December 2019), 94 valid self-evaluation records were obtained twice. There were 10 weeks in total, with classes twice a week, covering two lessons each time (Chinese, English, or Mathematics). Each lesson lasted for 45 min with one-on-one online companion and learning. The ethical rule was stated as a declaration in the beginning of the questionnaire, and it was answered online. All the participants voluntarily attended the scale implementation process.
The demographic features of the participants were as follows. In terms of gender, 80.9% of the participants were female, and 19.1% were male participants. In terms of year level, 17% of the participants were in the first year, 38% in the second year, 24% in the third year, 13% in the fourth year, and 7% were masters students. In terms of college, 30% of the participants studied in the College of Humanities and Social Sciences, 6% in the College of Science and Engineering, 10% in the College of Design, 24% in the College of Management, and 30% in the College of Informatics. Considering the teaching subjects, 16% of the participants taught Chinese, 41% taught English, and 43% taught Mathematics. Considering experience, 76% were novices and 24% were experienced. Concerning teacher education program background, 34% studied in the program, and 66% did not.
3.2. Measures and Variables
The main variables of the teacher were college, gender, subject, e-learning partner experience, and teacher education program background. There were teaching, counseling, administrative, and equipment questions in the self-evaluation survey. With regard to retention, passion and innovation were two main variables. Students answered the questionnaire, using a five-point scale. The options were “strongly agree”, “agree”, “neutral”, “disagree”, and “strongly disagree”. The scoring order was 5–1 points, respectively. The items on teaching, counseling, administration, and IT equipment are shown in
Table 1.
3.3. The PLS Algorithm Procedure
A participant group with smaller numbers was necessary for complex models in partial least square structural equation modelling. The bias-corrected and accelerated (BCa) bootstrapping procedure was used to assess the control variables’ significance and interaction effects. The assessment of the significant control variables should use the f2 effect size to consider their relevance. In the beginning, the reliability and validity (first stage) were examined. Then, the degree of collinearity of the indicators and the significance and relevance of the indicator weights were analyzed. Finally, redundancy analysis was conducted.
4. Results
There were 94 valid self-evaluation records in the 2019 academic year (from 21 October 2019 to 20 December 2019). The post scores were significant (p-value < 0.05) and higher in T11, E5, and Innovation than those in the first survey. Teaching self-efficacy, administration, and equipment evaluations increased with time. Counseling self-efficacy decreased in the second stage; however, it was not significant.
4.1. Factor Analysis
According to the factor analysis used to reduce the items, there are four main functions in this study. The items that did not score higher than 0.5 in the loading weight would be deleted from the study. Initially, in the teaching items, the 17 items were classified into four main factors (TP = teaching preparation, TC = teaching conference, TA = teaching attitude, and TW = teaching method). The factor of TP = {T1-T6, T11-T12, T15}, TW = {T7-T10, T13-T14}, TA = T16, and TC = T17. Moreover, E1 was deleted because the weight was lower than 0.5. Through the principal analysis, a total of four functions were created in this study as shown in
Table 2, and the significance was smaller than 0.05 in the KMO and Bartlett test. Related to the rotation sum of the squared loading, the cumulative percentage was 76.52 among the four main factors.
After the factor analysis, we classified TP, TC, TA, and TW as Teaching; C1–C5 as Counseling; A1–A6 as Administration; and E2–E5 as Equipment. In the basic statistics in
Figure 1, the score of Administration is the highest (mean = 4.761) and the score of Teaching is the lowest (mean = 4.276). It showed a high level of inner self-efficacy and outer support. Besides, the reliability and validity of the questionnaire were explored in
Table 3. The Cronbach’s alpha value of the scale was detected to be 0.886 for Teaching, 0.942 for Counseling, 0.912 for Administration, and 0.881 for Equipment. The Cronbach alphas of all four functions are greater than 0.8. This indicated that the result had a good reliability. In convergence validity, the average variance extracted (AVE) was greater than 0.5. In addition, all the correlation coefficients were better than 0.3. In the discriminate validity, the coefficient was better than 0.7. Furthermore, the four weights were higher than the others. This indicated good validity.
4.2. ANOVA and Post Hoc Analysis
The demographic features of the participants were gender, year level, college, teaching subjects, experience, and teacher education program background. The freshman and sophomores were coded as 1, the juniors and seniors were coded as 2, and the graduates were coded as 3. Through ANOVA and post hoc analysis (Scheffe method), year level, experience, and teacher education program background were significantly different in this study as demonstrated in
Table 4. The scores of third- and fourth-year participants were significantly higher than those of the first- and second-year undergraduates. In the e-learning partner program experience, the scores of the experienced participants was higher (4.47) than those of the novices (4.21) in self-efficacy of teaching. The participants who were enrolled in the teacher education program concurrently scored significantly higher in teaching (4.48 vs. 4.17) and counseling (4.81 vs. 4.60) than those who were not enrolled in the teacher education program at the case university.
4.3. PLS Analysis
In this study, SmartPLS is used to process the original data from SPSS to CSV file. SmartPLS is advantageous for samples smaller than 100. There are a total of 94 participants in the survey, which is an appropriate sample size. Reasons for analyzing the data with this tool include small sample size, non-normal data, formative measures, focus on prediction, model complexity, exploratory research, and theory development. There are reflective, formative, and redundancy models in the PLS. From previous studies, a model that shows the effect level of the latent variables of teaching, counseling, administration, equipment, and retention on each other and their ratios to each other was put forward. In
Figure 2, all the outer loadings are higher than 0.7. In collinearity statistics, the inner variance inflation factor was smaller than 10. The four factors were not collinear in this research.
The path coefficients of the prediction model were positive in teaching, counseling, and administration; however, they were negative in equipment to the latent variable of retention. The model also presented innovation (0.915) and passion (0.921), which had positive path coefficients to retention.
In path coefficient analysis, the factor of teaching coefficient was higher than 0.7, the total effect
p-value was smaller than 0.05, and t > 3.29 (
Table 5). This indicates that the formative construct in teaching can explain at least 50% of the variance of retention.
First, from
Table 6, all the composite reliability values are around 0.8–0.9, supporting the internal consistency and reliability of the measures. Second, the values of convergence validity (AVE) were all higher than 0.5, and the values of composite reliability (CR and Cronbach’s alpha CA) were all higher than 0.7. In discriminate validity, the coefficient was smaller than 0.7. Although the teaching coefficient was 0.873, it is smaller than 0.918. Therefore, it is still in the reasonable range. Considering redundancy, if it is higher, the model is a good fit (goodness of fit = Sqrt (redundancy)). This is similar to the model with blindfolding redundancy > 0, indicating that the variable has a predictive relevance in the model. The construct cross redundancy Q
2 (=1-SSE/SSO) = 0.63 in retention was calculated, and it was a good fit in this model. Third, the R
2 (0.767) was higher than 0.67, indicating that it was valuable in real application. Fourth, the effect size f
2 (0.79) was higher than 0.35. This indicated that the outer variable (Teaching) significantly influenced the inner variable (Retention) and the R
2 would increase.
6. Implications and Recommendations
It was necessary for the online e-learning partner program to improve the high talent turnover and decrease the cost of human resource training. The purpose of this study was to analyze factors that could contribute to this. Although the participants in this research were limited to one case university, the model of predicting teachers’ retention can be implemented at any university, especially novices in instruction in the world. If the institutional manager offered sufficient resources for teachers and learners, their interaction and communication could be stable and free of worry for a long-period in the e-learning process. The knowledge and subject area during each learning stage were consistent and systematic; therefore, the improvement of inner self-efficacy (teaching and counseling) and outer support could contribute to real or virtual classrooms anywhere Through workshops, professional seminars, instructors’ communities, retention rewards, or apprenticeships held by the MOE in Taiwan, the e-learning partner program could be more successful and beneficial to the accompaniment of children in happiness and harmony.