Central to the present study is Pintrich’s (2004) model of self-regulation, which includes metacognitive, motivational, and behavioral components of self-regulation. Students’ metacognition is regarded as the core component of SRL (Veenman, 2017; Usher and Schunk, 2018). It encompasses a range of key skills and strategies for regulating and controlling learning behaviors, such as task-orientation, monitoring, and self-evaluation (Pintrich, 2004; Vandevelde et al., 2013). The regulation of motivation involves the motivational beliefs that students hold regarding a task, including their self-reinforcement, positive self-talk, and interest enhancement (Pintrich, 2004). Last, the regulation of behavior entails the regulation of effort, persistence, time management, and help seeking (Skinner et al., 2008; Usher and Schunk, 2018). The inclusion of the motivational and behavioral components make Pintrich’s model one of the most comprehensive models available and particularly relevant for the elementary school context (Panadero, 2017). These components are considered to be crucial for elementary students’ academic performance (Dignath and Büttner, 2008; Perry et al., 2019).

Two possible factors may promote students’ SRL-skills, including teacher self-efficacy in fostering SRL-skills (Dignath, 2016; Karlen et al., 2023). Teachers’ self-efficacy in SRL-skills, in short TSE for SRL-skills, reflects teachers’ beliefs about their ability to affect students’ SRL-skills (De Smul et al., 2018). Based on social cognitive theory it is commonly believed that TSE for SRL may play a role in students’ SRL-skills by its influence on behaviors and competencies, such as goals and aspirations, outcome expectations, and metacognition (Bandura, 1997). Empirical studies indeed suggest a link between TSE and various SRL-skills, showing that feelings about their own abilities and their own level of self-regulation may be related to students’ self-regulated learning (see Moos and Ringdal, 2012, for an overview).

Despite considerable progress in this literature, current findings seem to call for some nuance. First, it should be noted that general TSE can fluctuate according to the circumstances in the classroom or the tasks at hand (Bandura, 1997; Zee and Koomen, 2016). Unfortunately, however, only a few studies have considered the task-and domain specific nature of TSE in studies on SRL-skills (e.g., Dignath, 2016, 2021; De Smul et al., 2018; Lawson et al., 2019; Karlen et al., 2020). This may be problematic, as the outcome domain of interest specifies the type of TSE that weighs most heavily on a given SRL-skill.

Second, from these studies, the study by Dignath (2021) is the only study that included direct measures of students’ SRL-skills. This study, which involved 191 third-to fourth-grade teachers and their students, assessed TSE for SRL-skills and the level of students’ self-regulatory skills. Results showed that students did not report an increase in their levels of SRL-skills when their teachers reported higher levels of TSE for SRL-skills. These results deviate from the commonly assumed belief that TSE leads to higher SRL-skills in students. It may be that the informant of student SRL-skills (teacher or student reported) might play an important role in whether associations between TSE and students’ SRL-skills appear to be positive. Hence we used student reports to measure SRL-skills in the current study.

Third, in Dignath’s (2021) study the relation of TSE and students’ SRL-skills has only been considered at the class level (Zee and Koomen, 2016). Little is known about the relation at the student level, that is between a teacher’s self-efficacy toward individual students (student-specific TSE) and student’s SRL-skills. Possibly, similar to studies in other teaching domains, the student-specific self-efficacy beliefs teachers hold are a better predictor of individual differences can be removed students’ SRL-skills than teachers’ general feelings of TSE (Zee et al., 2016a, 2018). Given these previous findings, we expect that a large portion of variance in the components of SRL is explained by student-specific TSE and that student-specific TSE has a stronger relationship with individual students outcomes (i.e., the student level), than general forms of TSE (i.e., teachers’ judgments at the class level, based on a subsample of students; Zee et al., 2016a, 2018). Our aim is therefore to examine associations between TSE for SRL-skills and students’ SRL-skills both within and between classes.

A second factor that may promote students’ SRL-skills is the quality of teachers’ interactions with their students (Dignath, 2016; de Ruig et al., 2023; Hofkens et al., 2023). The quality of daily interactions between teachers and students is considered a driving force behind learning and development, including the development of SRL-skills (Bronfenbrenner and Morris, 1998; Hamre et al., 2013). These interactions are conceptualized around three broad domains: instructional support, classroom organization, and emotional support (Hamre et al., 2013; Fauth et al., 2021).

Instructional support consists of the degree to which teachers are able to promote inquiry skills, use a variety of learning formats, provide high-quality feedback, involve students in instructional dialogue, and help them understand content at a deeper level (Pianta et al., 2012). It is commonly assumed that if instructional support is well-tailored to students’ prior knowledge and includes real-life examples it will provide opportunities to show current skills, broaden knowledge, and develop new metacognitive SRL-skills (Hamre et al., 2013; Veenman, 2017). Moreover, direct forms of instruction by teachers are believed to help students obtain knowledge about strategy application and its benefits, which might also be related to their self-regulation (Veenman, 2017; Dignath and Veenman, 2021). Different reviews have demonstrated that instructional support can be an effective way through which teachers support students’ metacognitive SRL-skills. However, in these reviews its positive relationship with the motivational and behavioral SRL-skills appears to be less strong (Dignath and Büttner, 2008; Dignath and Veenman, 2021; de Ruig et al., 2023).

The domain of classroom organization represents teachers’ organization of and managerial tasks within the classroom, such as behavioral management, daily routines, and maximization of learning time (Pianta et al., 2012; Fauth et al., 2021; Hofkens et al., 2023). Organized classrooms and daily routines are believed to help students focus on learning tasks such that they can achieve their learning goals (Hamre et al., 2013). In addition, teachers in well-organized classrooms are expected to provide students with enough instructional time so that students encounter sufficient challenges and receive enough strategy instruction. Findings of a recent review by de Ruig et al. (2023) showed that well-organized classrooms also had students with higher metacognitive SRL-skills. The association appeared to be less strong compared to the overall relation for instructional support (de Ruig et al., 2023). Findings were mixed for the relationships between teachers’ classroom organization and motivational and behavioral SRL-skills.

Lastly, teachers who provide high levels of emotional support create a warm classroom climate, are sensitive, and have regard for students’ perspectives by being flexible and providing autonomy (Pianta et al., 2012). Emotionally supportive teachers create a secure and safe learning climate, in which students may feel self-reliant, competent, and willing to take risks to explore the classroom environment (Hamre et al., 2013). Moreover, emotionally supportive teachers are expected to help students feel engaged, autonomous, and motivated to work on a task. It has been suggested that these feelings support students’ SRL-skills (Rieser et al., 2013). Previous research has shown that when teachers were emotionally supportive they were also more likely to promote metacognitive SRL-skills in their students (de Ruig et al., 2023). However, emotionally supportive teachers did not always have students with higher levels of motivational or behavioral SRL-skills (Rimm-Kaufman et al., 2014; Havik and Westergård, 2020).

Two features of current evidence on the relation between teacher–student interactions and SRL-skills are of note. First, most studies in this field predominantly rely on teacher and student reports. de Ruig et al. (2023) found only four studies which included observations of the quality of interactions, all using the Classroom Assessment Scoring System (CLASS; Pianta et al., 2012). This is noteworthy as observations can provide valuable insights about the quality of teacher-student interactions, including real-time classroom dynamics, nonverbal behaviors, and other contextual factors (Pianta et al., 2012; Bell et al., 2019; Wang et al., 2020). It seems therefore likely that observations can provide additional information on how the quality of interactions within a classroom can promote students’ SRL-skills.

A second feature of current evidence is that the four studies that did use classroom observations of the quality of teacher–student interactions were all focused on motivational and behavioral SRL-skills (de Ruig et al., 2023). Metacognitive SRL-skills were not included in these studies. Moreover, findings about the relationship between the quality of interactions and students’ motivational and behavioral SRL-skills appeared to be mixed. Two cross-sectional studies revealed that when high levels of classroom organization were observed, lower and upper elementary school students were also more likely to show higher levels of motivational and behavioral SRL-skills (Cadima et al., 2015; van Dijk et al., 2019). Yet, two other studies did not find association of instructional support, classroom organization, and emotional support with students’ ability to regulate their motivation and behavior (Rimm-Kaufman et al., 2014; Schenke, 2018). Therefore, we go a step further than previous studies by examining whether observed interactions may be associated with all three SRL-skills, thus including metacognitive self-regulation skills.

The role of interaction quality in promotion of SRL-skills seems rather complex, and may generate interaction effects in addition to the expected direct effects. These interaction have, to our knowledge, not been previously researched. This is notable, given the commonly held believe that the potential relationships between student-specific TSE and students’ SRL-skills at different levels of classroom ecology cannot be examined in isolation from the context in which this relationship occurs (Bandura, 1997; Eccles and Wigfield, 2002).

It seems likely that the quality of teacher–student interactions provides teachers with positive and negative information regarding their own ability to teach SRL-skills to students and about their own performance in providing sufficient support. Lower-quality classroom interactions may therefore negatively effect the relationship between TSE and students’ SRL-skills, whereas in classrooms with high-quality interactions a much stronger effect of TSE on SRL-skills may be expected. Hence, our third and final aim to explore the potential moderating effect of the quality of teacher–student interactions on the relationship between TSE and SRL-skills.

This study addresses three research questions that extend prior research on SRL-skills. The first question is how TSE for SRL-skills and SRL-skills are related at the student and the class level. Rather than taking a class-level only perspective, as an extension to previous research, we also examined these relationships at the student level to gain a deeper insight into how TSE might promote students’ SRL-skills at different levels of classroom ecology. The second question concerns how teacher–student interactions are associated to students’ SRL-skills. We used observations instead of teacher or student reports to obtain information about the quality of teacher–student interactions in class. The third question is whether teacher–student interactions moderate the relationship between TSE and SRL-skills.

Based on prior research, several hypotheses were tested. First, we hypothesized that associations between TSE for SRL-skills and SRL-skills may not be positive (Dignath, 2021). Second, we expected that student-specific TSE may have a positive relationship with individual students’ outcomes (e.g., Zee et al., 2016a). Third, we hypothesized that associations between the three domains of teacher–student interactions and metacognitive SRL-skills may be positive (e.g., Rieser et al., 2013). Given that relationships with motivational and behavioral SRL-skills were mixed in prior research (Rimm-Kaufman et al., 2014; Cadima et al., 2015; Schenke, 2018; van Dijk et al., 2019), we assumed weaker relationships with motivational and behavioral SRL-skills than with metacognitive SRL-skills. Finally, our fourth hypothesis was that a possible relationship between TSE for SRL-skills and SRL-skills was moderated by the quality of teacher–student interactions. We expected that high-quality interactions may positively moderate a potential relationship between TSE and SRL-skills, whereas lower-quality interactions may negatively effect such a association.

The sample consisted of 1,278 third-to sixth grade students (46.9% boys) from 63 classrooms of 33 Dutch regular elementary schools in both urban and rural areas across the Netherlands. Of these students, 126 were in grade 3 (9.9%), 264 in grade 4 (20.7%), 401 in grade 5 (31.4%), and 452 in grade 6 (35.4%). Another 35 students (2.7%) did not report their grade. Students’ age ranged from 7 to 13 years (M = 10.00, SD = 1.19). Of the participating students, 75.7% identified themselves as ethnically Dutch. Teachers’ age ranged from 22 to 59 years old (M = 34.72, SD = 10.75). Teachers (N = 63) generally had 9.23 years of teaching experience (SD = 7.54) and class sizes ranged from 12 to 37 students per classroom (M = 24.88, SD = 3.18).

Teachers were asked to fill out a questionnaire on student-specific TSE for SRL-skills for eight students who were randomly selected from their classrooms. Completing questionnaires for eight students is considered to be a manageable workload for teachers and a good indication of the class mean for TSE (Zee et al., 2018, 2024; Zee and Rudasill, 2021). Due to illness or time constraints, not all teachers were able to complete the questionnaire for those eight students, resulting in completed questionnaires on teachers’ TSE for SRL-skills for an average seven students per class. This resulted in a subset of 449 students for which information was available on student-specific TSE for SRL-skills.

We reached out to school principals and teachers through various channels, including telephone, email, and LinkedIn. Invited teachers received an information brochure and were asked to sign an informed consent form. Of the teachers that were contacted, 13.2% agreed to participate. Moreover, parents were asked to provide consent for their child to participate. The average teacher consent rate per school was 73.3% and the parental consent rate for student participation per class was 84.9%.

Data was gathered in three ways, through: (1) classroom videos, (2) student-report questionnaires, and (3) teacher-report questionnaires. Trained research assistants or the first author made a 45-min video within each participating classroom to capture teacher and student behavior as well as classroom-level interactions (see Instruments). Videos were recorded at the beginning of a regular school day. After the video was recorded, on the same day, teachers and students were asked to complete a questionnaire. Completing the questionnaires took around 30 min. Data was collected during the fall of schoolyear 2019–2020. Teachers and students were allowed to stop at any given moment during the project. The study was approved by the Ethics Review Board of the Faculty of Social and Behavioral Sciences of the University of Amsterdam (project 2019-CDE-10504).

An intercepts-and-slopes-as-outcomes multilevel model was fitted to test our research questions and corresponding hypotheses (Kreft and de Leeuw, 1998). This model takes the nested data structure of students within classrooms into account (Geiser, 2013). Accounting for the nested data allows for estimation of unbiased standard errors for the regression coefficients. Moreover, variables can be added to this model at both the student (student-specific TSE and SRL-skills) and the class level (aggregated TSE, teacher–student interactions, and class-average SRL-skills). Another advantage of the intercepts-and-slopes-as-outcomes model is that both the intercepts and slopes of student-level regressions can vary across classroom-level variables (Kreft and de Leeuw, 1998; Geiser, 2013). Any differences between classes in intercepts and slopes can, in turn, be explained by variables at the class level. Hence, this model enabled us to answer our first and second research question, that is, whether TSE for SRL-skills and the three interaction domains are associated with students’ SRL-skills at the student and at the class-level.

The third research question, whether teacher–student interactions moderated the relationship between student-specific TSE for SRL-skills and SRL-skills, was examined by adding a cross-level interaction to the intercepts-and-slopes-as-outcomes model. This interaction specifies that the relationship between the student level predictor TSE for SRL-skills and the outcome variables, that is students’ SRL-skills, is dependent on the level of the class-level predictors, that is the quality level of teacher–student interactions.

Maximum likelihood estimation with robust standard errors was employed to deal with potentially non-normal data (Kline, 2016). Missing data was handled using full information maximum likelihood (FIML; Muthén and Muthén, 2017). All student and class-level predictors were grand-mean centered to ease interpretation (Kreft and de Leeuw, 1998).

Data was screened and descriptive statistics were evaluated before the multilevel models were fitted to the data. Most of the scales were approximately normally distributed following a threshold of ±1.00 for skewness and kurtosis (Kline, 2016). However, for Classroom Organization, skewness (−2.06) and kurtosis (5.87) were considerably higher than ±1.00. To account for this non-normality, we used robust maximum likelihood estimation (Kline, 2016). In addition, no multivariate outliers were found using Cook’s Distance and there was no sign of multicollinearity using variance-infliction factors with a threshold of five (Tabachnick and Fidell, 2014). Probability plots were inspected and showed no violation of the assumption of multivariate normality.

ICC1 and ICC2 values are displayed in Table 2. ICC1 indicates the percentage of the total variance in a variable that can be ascribed to the class-level. These percentages were 7.5% for Metacognitive, 2.2% for Motivational, and 8.8% for Behavioral SRL-skills occurred between classes. Even though these intraclass correlations are relatively small (< 10%), it is still advised to use a multilevel modeling to account for the nested structure of the data (Kreft and de Leeuw, 1998).

In addition, we evaluated the reliability of the aggregated variables (ICC2; Marsh et al., 2012), including Metacognitive, Motivational, and Behavioral SRL-skills, and Student-Specific and Class-specific TSE. Metacognitive, Motivational, and Behavioral SRL-skills were based on the full sample of 1,278 students, whereas Class-specific TSE for SRL was based on eight randomly selected students per class. The ICC2 values were good for the Metacognitive (0.621) and Behavioral SRL-skills (0.664), and even excellent for student-specific (0.803) and Class-specific TSE for SRL (0.807; Cicchetti, 1994). The ICC2 value of 0.316 for the Motivational SRL-skills was not satisfactory. This value indicates that the class-average mean may be unreliable (Marsh et al., 2012). We therefore tested models with and without the Motivation SRL-skills, but did not find significant differences in coefficients at the class level or in model fit. Since the Motivational SRL-skills are considered to be important to the SRL process and also a reliable outcome variable at the within level, we retained this variable in the models that were tested.

Regarding the domains of teacher–student interactions, Instructional Support had the lowest overall mean, while Classroom Organization demonstrated the highest mean among the three interaction domains. This is in accordance with previous research (e.g., Schenke, 2018), and shows that the majority of classrooms are well-organized, while the average quality of instruction falls only in the lower midrange of the CLASS. The variance of Emotional Support shows that the quality varies more among teachers than in the other domains.

The student and class-level correlations are in Table 2. At the class level, the correlations between various domains of classroom interaction quality were high, particularly between Classroom Organization and Emotional Support (r = 0.71, p < 0.001), and Emotional Support and Instructional Support (r = 0.71, p < 0.001). These correlations imply a significant overlap among these interaction domains. The various components of SRL were also correlated at the class level, but, as compared to the student level, several correlations were much stronger, such as the correlations between the Motivational and Metacognitive SRL-skills (r = 0.83, p < 0.001) and Motivational and Behavioral SRL-skills (r = 0.72, p < 0.001). It appears that these SRL-skills are more strongly interconnected between classrooms than within classrooms.

Contrary to expectations, there were negative correlations between the domains of classroom interactions and SRL-skills, including significant negative correlations between the three domains and Metacognitive and Motivational SRL-skills. These negative correlations suggest that as the quality of teacher–student interactions increases, the level of SRL-skills in students decreases.

Our first model was the intercept-only model that served as a baseline model for model comparison. In step two, a random intercept model with the student-level predictor Student-Specific TSE and the classroom-level predictors Instructional Support, Classroom Organization, and Emotional Support and TSE was fitted to the data. This CLASS model with the three measures of the CLASS was a significant improvement over the intercept-only model, according to the Satorra-Bentler adjusted chi-square difference test (TRd = 2326.75 (27), p < 0.01). The standardized coefficients of this model are presented in Table 3. These coefficients demonstrate that student-specific TSE is significantly related to students’ Metacognitive (β = 0.17, p < 0.05), Motivational (β = 0.10, p < 0.05), and Behavioral SRL-skills (β = 0.38, p < 0.05).

Standardized coefficients for the relationship between the three domains of class-level interactions and SRL-skills were generally non-significant and predominantly close to zero (see Table 3). The three interaction domains did not have an independent association with each of the SRL-skills. Their effect was likely determined by what they collectively contributed to the dimension of SRL. Therefore, the three domains of teacher–student interactions were combined into a single measure. The combined measure had excellent internal consistency (Cronbach’s α = 0.91). The combined CLASS model appeared to have better model fit than the intercept-only model (TRd = 901.16 (17), p < 0.01). The results of the combined CLASS model showed significant negative standardized coefficients between teacher–student interactions and students’ Metacognitive and Motivational SRL-skills.

In step three, we added random slopes to the random intercept model for each combination of the student level predictor TSE with Metacognitive, Motivational, and Behavioral SRL-skills. Random slopes were non-significant for the association between TSE and students’ SRL-skills when they were separately added to the combined CLASS model. This implies that the variation among classes for students’ SRL-skills is not significant in our data. Hence, there is no indication for a moderating effect of the quality of teacher–student interactions on the relationship between TSE and SRL-skills, given that significant slope variation is a precondition for such an effect (Hox et al., 2018).

An extra step was added to our modeling procedure as we observed positive associations between TSE and SRL-skills at the student level, whereas at the class level the coefficients were near zero and non-significant. Contextual effects for aggregated TSE and the three SRL-skills in the combined CLASS model were therefore calculated, based on three newly created parameters. These parameters match with the differences between the corresponding student and class-level variables for the relationships between TSE at the class level and the three SRL-skills. The contextual effects then reflect the effect of the aggregated class-level variable of TSE, after controlling for the effects of student-specific TSE at the student level (Marsh et al., 2012). Results showed significant negative context effects for TSE at the class level and Metacognitive (B = −0.25, p < 0.001), Motivational (B = −0.21, p < 0.01), and Behavioral SRL-skills (B = −0.37, p < 0.001). These findings indicate that the relationship between student-specific TSE and the three SRL-skills is negatively moderated by TSE at the class level.

Relationships between TSE and students-SRL skills were examined at both the student and the class level. At the student level, teachers who reported higher levels of student-specific TSE for SRL had students who reported higher levels of metacognitive, motivational, and behavioral SRL-skills. These findings are in line with our hypothesis and with previous research in which the effects of student-specific TSE on student outcomes were examined (Zee et al., 2016b, 2018). Such a positive association between TSE and various SRL-skills is not surprising, as teachers may be better able to make judgments about their capacity to support a particular student when they consider the concrete needs, behaviors, and actions of individual students rather than an entire group (Zee et al., 2016b). Moreover, it seems that a domain and student-specific form of TSE is well-tailored to the outcome of interest, increasing the predictive merit of TSE in relation to self-regulation skills (Bandura, 1997). Our study is, to our knowledge, the first to demonstrate that judgments of teachers about whether they are able to foster SRL-skills for specific students are associated with higher levels of SRL-skills. It appears therefore important for teachers to consider how they may enhance their sense of self-efficacy for SRL-skills for individual students, as it seems that individual students’ SRL-skills benefit the most from this approach.

However, at the class level, the average level of TSE for SRL reported by teachers, based on their student-specific TSE, was not related to the average level of SRL of a class as reported by the students. This finding is not in line with the commonly recognized belief that higher levels of TSE lead to higher levels of SRL-skills in students (e.g., De Smul et al., 2018). A main explanation for why our results deviate from most previous findings, is that the level of SRL-skills is often not directly assessed in other studies (e.g., De Smul et al., 2018; Karlen et al., 2020). Rather, in most studies, teachers were asked to report on both their TSE for fostering SRL and SRL-skills in their students. In that case, teachers are the same informant for both the predictor and outcome variables. The use of the same informant for TSE and students’ SRL-skills seems to overestimate their relation (Vandevelde et al., 2013; De Smul et al., 2018). Our findings resonate with the results of a recent study in which the associations between teachers’ self-reported judgments and students’ perceived levels of SRL-skills were nonsignificant as well (Dignath, 2021). In this study and the current study, TSE was based on teacher reports and the level of SRL-skills was directly assessed through student reports.

Another reason for the absence of a relation between student-specific TSE and students’ SRL-skills at the class level, is that in earlier studies teachers rated their TSE based on general judgments about their self-efficacy towards an entire class. In the current study an aggregated measure of student-specific TSE was used. This aggregated measure was based on teachers’ judgments of a random sample of individual students within each class. Previous research has shown that general and aggregated TSE are only moderately correlated, which may result in different relationships with student outcomes (Zee et al., 2016b, 2018). An explanation for these differences is the distinction between aggregated TSE as a classroom context variable and general TSE as a classroom climate variable, which are both based on different referents (Marsh et al., 2012). Aggregated TSE is a representation of teachers’ feelings of self-efficacy towards specific students in a particular classroom, and therefore based on individual students who each receive an individual score. General TSE, on the other hand, uses the entire class as referent (Zee et al., 2018). Hence why students in the same class all receive the same score. The differences in referents may have a fundamental influence on the interpretation of both construct, which is why they may have different relationships with the outcome variables (Marsh et al., 2012; Morin et al., 2014).

The positive associations between TSE and SRL-skills at the student level and the lack of relationships at the class level, suggests a contextual effect of the class (Marsh et al., 2012). The contextual effect is the effect of the class-level variable of TSE, after controlling for the effects of student-specific TSE. Indeed, in our study the contextual effect shows that teachers who feel more confident about fostering SRL-skills in their classes are likely to teach students in those classes who perceive their own self-regulated learning skills as lower. In contrast, in classes where TSE for SRL-skills is reported to be lower, students perceive their SRL-skills on average to be higher. We had not anticipated this effect and it is not straightforward to provide an explanation. One possibility is that teachers use the teacher report scale somewhat differently, resulting in, for instance, one teacher consistently assigning higher ratings than another teacher about their TSE for SRL-skills. Consequently, the relationship between TSE and SRL-skills may be predominantly a within-class effect. Another possible explanation is that teachers’ efficacy beliefs directly influence teachers’ behavior and may therefore serve as a good proxy for how teachers support students within their classrooms (Bandura, 1997; Zee and Koomen, 2016). Thus, if teachers feel highly self-efficacious in teaching SRL, they may invest considerable time in this area and potentially scaffold many SRL-skills for their students. This may have two consequences. First, due to teacher’s emphasis on SRL-skills, students may become more aware of the strategies they do or do not employ, leading to rate themselves lower on these skills. Second, due to the focus on SRL-skills, the students may feel that they need to apply fewer skills themselves to regulate their own learning as the teacher is already doing it for them.

A main finding of this study is that in classes with higher quality teacher–student interactions students reported lower mean class-levels of their SRL-skills. This is particularly evident for students’ metacognitive and motivational SRL-skills. Based on previous research that used student and teacher reports, we had expected the opposite outcome (e.g., Rieser et al., 2013). That is, higher teacher-student interaction quality would, at least for the metacognitive SRL-skills, be associated with a higher mean class-level of SRL-skills. Our results deviate from these expectations.

Until now, there were four studies that reported no or weak correlations between domains of the CLASS and students’ SRL-skills (Rimm-Kaufman et al., 2014; Cadima et al., 2015; Schenke, 2018; van Dijk et al., 2019). There are two explanations for this findings. First, absent or weak correlations may be expected because the total score of the CLASS could be disproportionally influenced by a small number of students. Teacher–student interactions consist of both teacher and student behavior that are observed simultaneously. For example, observers score teacher behavior on the Instructional Learning Formats scale, such as promoting involvement and teacher interests, and at the same time also active participation and sustained attention of students (Pianta et al., 2012). Another example is when two students are arguing during the observation, while there is little else happening in the classroom. Their actions probably taint the scores, while this situation may not necessarily impact all students. Hence, it seems that the CLASS, although considered to be a classroom climate measure, also encompasses aspects related to specific student behaviors and actions in class.

Second, another explanation for this unexpected finding is that classroom interaction quality and students’ SRL-skills are measured at different levels of classroom ecology. Whereas interaction quality refers to the overall, more objective atmosphere of the classroom (i.e., contextual variable), the average SRL-skills of a class are based on the subjective ratings of all students in that particular class (i.e., climate variable). Each score is assigned equal weight when aggerated to the class level. In that case, the aggregated outcome measure is based on a different level of analysis than the overall classroom measure. This may potentially lead to correlations and coefficients that are lower than expected (Marsh et al., 2012; Morin et al., 2014). A suggestion for future research may therefore be to design and use an observation instrument that is aimed at the quality of dyadic, one-on-one, interactions instead of the classroom as a whole. The predictor and outcome variables are than both at the student level and reflect the same level of analysis and theory, also when aggregated to the class level (Marsh et al., 2012).

However, in the current study negative relationships between the quality of teacher–student interactions and students’ SRL-skills were found. This may not be fully explained by differences in the levels of analysis. There seem to be two main reasons for the negative relationships. First, it is possible that higher quality interactions made students believe that they have less SRL-skills, because they are provided with relatively better support from their teachers. Second, even though the quality of interactions are high and teachers provide numerous opportunities for learning, these high-quality interactions may also take away some of the necessity for students to regulate their own learning in their experience. For example, when teachers provide high-quality instructional support they demonstrate, guide, and model extensively, which might reduce the necessity for students to actively use their SRL-skills to some extent during instruction. Consequently, when asked to report about their SRL-skills, students may indicate a lower use of these skills. Further research investigating relationships between the domains of the CLASS and SRL-skills seems therefore needed to determine if the negative associations from the current study persist across different classroom and school contexts.

The third and final aim of the study was to examine the moderating role of instructional support, classroom organization, and emotional support on the relationship between student-specific TSE with the three SRL-skills. Notably, none of the random slopes for each combination of TSE with metacognitive, motivational, or behavioral SRL-skills were statically significant. As a result, we were unable to identify any moderation effect for one of the domains of teacher–student interactions. These findings do not align with our hypothesis that a moderation effect would be present. Rather, our findings demonstrate that relationships between TSE and SRL-skills are relatively less sensitive to the influence of the classroom ecology.

There are three limitations in the current study that need to be addressed. First, our study is cross-sectional and correlational in nature, preventing us from drawing causal conclusions. Although our findings offer insight into the complex relationships between TSE and SRL-skills at both the student and class level, it is possible that these associations are reciprocal (Bandura, 1997). This would be in line with social-cognitive theory suggesting that teachers’ personal factors, behaviors, and the classroom context interact with each other bidirectionally. Future research could employ longitudinal, cross-lagged multilevel designs to disentangle causal effects and examine whether these relationships are indeed reciprocal.

Second, we did not add a general measure of TSE for SRL to our study in addition to the aggregated TSE variable at the class level. Previous research has demonstrated that general TSE and aggregated TSE are only moderately correlated and may have different relationships with student outcomes (Zee et al., 2016b, 2018). Inclusion of a general measure of TSE would have made our results more comparable with those from previous studies. However, as seen in previous research, relatively weaker relationships are found between general TSE and student outcomes as compared to a specific measure of TSE (Zee and Koomen, 2016; Zee et al., 2018). This is the primary reason why we relied on specific-student measures for TSE.

Third, despite intensive training and a yearly recertification, inter-rater reliability was poor or just fair in two areas of the instructional support domain of the CLASS. These findings are in line with previous studies, in which variability among multiple raters was found (Hamre et al., 2013; Mashburn et al., 2013). However, Cronbach’s alpha for the total score of the instructional support domain appeared to be good in the current study, which is why we still included it and believe it is a reliable scale for the instructional support domain. Yet, it seems necessary for future research to explore methods to enhance inter-rater reliability of the CLASS, by for example increasing the number of raters per classroom or by improving existing observational measures.