Teacher Self-Efficacy,Instructional Practice,and Student Outcomes: Evidence from the TALIS Video Study

Theoretical Background

The theoretical background underpinning research into self-efficacy stems from the pioneering work of Bandura (1977), building on Rotter’s (1966) prior research into locus of control. In particular, Bandura identified a distinct psychological construct that reflects the belief an individual has in their capacity to achieve a specific task (Hussain et al., 2022). Individuals who believe that they can succeed at a task are more likely to do so, with their positive mindset leading them to become more motivated, committed, and interested in achieving their goal. They are also more likely to recover from setbacks and overcome obstacles that stand in their way.

Most studies into teacher self-efficacy have been based on these foundational ideas, focusing on the extent that teachers believe that they can achieve specific tasks (e.g., controlling classroom behavior) and how this then relates to student outcomes. The first way that this may occur is via a role-modeling process—what Jerrim et al. (2023) described as its direct path to student outcomes. Here greater levels of teacher self-efficacy rub off onto their students, who go on to develop greater levels of self-efficacy themselves (along with, potentially, other areas such as subject interest and confidence). In other words, students internalize their teachers’ beliefs that they can improve learning outcomes and thus work harder toward their (shared) goal. This will feed through into improved test scores.

However, perhaps the most obvious driver of any relationship between teacher self-efficacy and student outcomes occurs through the instruction that teachers provide. This is what Jerrim et al. (2023) describe as the indirect path. Within this study, we drawn on the three basic dimensions framework (Praetorius et al., 2018) to focus on three specific aspects of teacher instructional practice, considering how these may mediate the relationship between teacher self-efficacy and student outcomes.

The first of these three dimensions is classroom management. This broadly refers to the extent that teachers can maintain order in the classroom. It is supposed that teachers who are more confident in their skills—particularly in their ability to control behavior—are more likely to take the appropriate action to do so and persist if students attempt to challenge their authority. This then translates into a more conducive learning environment, with students’ skills developing more rapidly as a result.

The second dimension—student support—captures social interactions between teachers and students in the classroom. This includes showing one another warmth, encouragement, and respect. In situations where teachers lack confidence in their abilities, it is likely to be more challenging for them to gain the respect of students in their classes. Previous research also has suggested that teachers with higher levels of self-efficacy may help to improve relationships between students by—for instance—effectively managing instances of bullying (Van Aalst et al., 2021). Efficacious teachers also may be more persistent and tolerant with students (Zee & Koomen, 2016) and thus go further to support their learning. Praetorius et al. (2018, p. 409) note how “these factors are mostly emotional or social and thus are expected to trigger socio-emotional outcomes such as student well-being and learning motivation.” It is thus likely that it is mainly through boosting students’ socioemotional outcomes that the socioemotional support offered by teachers leads to eventual gains in student achievement.

The final dimension within this framework is cognitive activation, capturing the extent to which teachers set challenging tasks to students, motivating them to think deeply about questions and problems, which, in turn, will deepen their understanding. Teachers who have greater confidence in their teaching abilities may be more likely to set their students such challenging tasks and be more persistent in helping them to overcome difficulties they face in finding the correct solution. However, while previous research has considered this link between teacher self-efficacy and cognitive activation, results have been somewhat inconclusive (Holzberger & Prestele, 2021; Holzberger et al., 2013). Yet, if a link between teacher self-efficacy and cognitive activation does exist, then this is likely to support student interest and motivation in a subject (e.g., students will be motivated and enjoy the need to think deeply about problems) as well as boosting their knowledge and achievement.

An overview of these hypothesized relationships is presented in Figure 1. The figure illustrates how, although teacher self-efficacy may be directly related to student's socioemotional competencies, its link with the three basic dimensions of teaching quality also may play a key mediating role. We note, however, that several of the paths presented in the figure have not found consistent support in the existing literature. Indeed, the existing evidence on many of the associations presented is somewhat mixed. This holds true both for the links between teacher self-efficacy and instructional practices (e.g., inconsistent findings regarding whether teacher self-efficacy is positively related with cognitive activation) and for how the three basic dimensions of instructional quality are associated with student outcomes. For instance, in their analysis of this framework, Praetorius et al. (2018, p. 419) concluded that it is “only partly supported by the current evidence.” This, in turn, highlights the need for further research into the relationships between teacher self-efficacy, instructional practices, and student outcomes displayed in Figure 1.

OPEN IN VIEWER

Figure 1.

Hypothesized relationship between teacher self-efficacy and student outcomes.

Although the relationships in Figure 1 are hypothesized to work in this way, previous literature has “struggle[d] to establish a reliable empirical link between teachers’ competence beliefs and students’ academic outcomes” (Lauermann & ten Hagen, 2021, p. 265). Lauermann and ten Hagen (2021) highlight several possible explanations that we review here, including (a) variation across raters of teaching quality, (b) predictor-outcome specificity—or generality—correspondence, (c) context- and situation-specific influences, and (d) further methodologic and design considerations.

Research Questions

The theoretical background outlined earlier—along with what is known from the previous literature—leads us to the following research questions. Our broad epistemological method is to use an argument-based approach to validity, which is “to validate an interpretation . . . to evaluate the rationale, or argument, for the claims being made” (Kane, 2006, p. 17). In reference to the literature on TSE, the claim often made is that teachers with higher levels of self-efficacy deliver higher-quality instruction to their students, whose socioemotional and learning outcomes benefit as a result (recall Figure 1). In this paper we seek to evaluate this argument. In particular, we seek to establish whether the supposed relationship between TSE and the three basic dimensions of instructional quality holds regardless of the source of information on the latter (i.e., student, teacher, or expert observers).

To begin, we explore the relationship between TSE and the three basic dimensions of lesson quality—the key mediating channel depicted in Figure 1. As noted by Lauermann and ten Hagen (2021), there is a dearth of evidence on this issue, particularly when the outcome (lesson quality) is not solely based on teacher reports. Indeed, their review stated how “there is an urgent need to account for rater-specific effects in analyses of the implications of teachers’ competence beliefs for their students” (Lauermann & ten Hagen, 2021, p. 279). We thus advance the literature by investigating how the relationship between TSE and the three basic dimensions of lesson quality varies across the views of teachers, students, and expert observers. Thus, we ask

We then proceed to investigate the link between TSE and students’ socioemotional competencies—the other key intermediate outcome depicted in Figure 1. As noted earlier, it has been suggested previously that higher self-efficacy levels among teachers may directly rub off on their students, who then become more confident in dealing with subject tasks as a result. These intermediate student outcomes, of course, also may be conditioned by improved lesson quality. Under either scenario, the association between TSE and student self-efficacy (and broader socioemotional competencies) should be positive. We test these associations by asking

Next, we turn to how TSE relates to student test scores. If we identify a link between TSE and our intermediate outcomes (i.e., improved lesson quality, student self-efficacy, and other socioemotional outcomes), we might anticipate this to translate into learning gains. Moreover, if such a relationship is found, it would be of interest to know the extent that it is being driven by enhanced lesson quality (as compared with possible alternative channels). Thus, we ask

We then explore the issue of nonlinearities—a somewhat underexplored topic across the TSE literature. In particular, if a teacher has very low levels of self-efficacy, is this particularly detrimental to the quality of their teaching and their student's educational and socioemotional outcomes? As noted by Lauermann and ten Hagen (2021), one possible reason why such nonlinear associations may exist is because “a teacher's (displayed) lack of confidence in being able to teach effectively might undermine students’ confidence in their own capability to learn, especially for those students who struggle academically and rely on their teacher's support” (p. 271). Moreover, knowing the answer to this question is important from a policy perspective: Should school leaders attempt to increase the self-efficacy of their staff across the board or only focus on those who lack any real conviction in their pedagogic skill? Thus, we ask

Measurement of TSE

As part of the baseline questionnaire, teachers were asked 12 questions using a four-point scale (1 = “not at all” to 4 = “a lot”) capturing their self-efficacy (e.g., the extent that they can “control disruptive behavior in this classroom”). In our main analysis, we use the general TSE scale designed by the OECD and standardized to mean zero and standard deviation one. Standardization is done within countries in country-specific analyses and across countries in pooled analyses. The same questions were included in another, larger cross-national study of teachers (the “main” TALIS study) with a metric level of measurement invariance found to hold (OECD, 2019, Table 11.6). The internal consistency (Cronbach’s alpha) of this scale is 0.89. However, in online Appendix B we test the robustness of a selection of our results using domain-specific measures of self-efficacy instead.

Table 1 provides the distribution of responses for three of the TSE questions (one from each of the three domains). The table suggests that relatively few teachers in the sample had very low levels of self-efficacy—for example, in most countries only around 5%–10% of the responding teachers felt unable to control disruptive behavior in the classroom or could not provide alternative explanations when their students were confused. However, in all countries, a nontrivial proportion of teachers did feel unable to motivate students who lacked interest in their schoolwork. There are also some clear differences across countries in responses—for example, Japanese teachers were the most likely to disagree with the statements—which may at least partially reflect differences in interpretation. Thus, while there is clear variation in the responses teachers provided to these questions, the sample only contains a limited number with a severe lack confidence in their teaching skills.

OPEN IN VIEWER

Table 1

Distribution of Selected Questions from the Teacher Self-Efficacy Scale (Column Percentages)

Factor	Chile	Colombia	England	Germany	Japan	Madrid	Mexico	Shanghai	Pooled
Control disruptive behavior in this classroom
Strongly disagree/disagree	5	10	10	8	9	7	18	0	9
Agree	43	37	28	58	63	58	45	25	43
Strongly agree	52	54	62	34	28	36	37	75	48
Total	100	100	100	100	100	100	100	100	100
Provide an alternative explanation for examples in this class when students are confused
Strongly disagree/disagree	3	5	3	6	24	0	6	4	7
Agree	43	38	36	62	60	39	40	46	45
Strongly agree	54	57	60	32	16	61	54	51	49
Total	100	100	100	100	100	100	100	100	100
Motivate students in this class who show low interest in schoolwork
Strongly disagree/disagree	24	15	26	46	66	37	28	16	31
Agree	45	49	53	54	30	49	41	48	45
Strongly agree	31	36	21	0	5	14	31	35	23
Total	100	100	100	100	100	100	100	100	100

Lesson Quality

To define instructional quality, the OECD started by drawing on measures included within the TALIS and Programme for International Student Assessment frameworks. These were then supplemented by information from (a) international research literature and (b) countries own conceptualizations of teaching quality. The latter critically included how countries saw various teaching activities being linked to student outcomes. Hence, the operationalization of instructional quality used in this study was based on both prior research evidence and expert input from the participating countries. OECD (2021, Chap. 2) provides further details.

This process resulted in six aspects of teaching practices being measured as part of the TALIS Video Study, which have been grouped by the OECD into the three broad domains (OECD, 2021, Chap. 19). In this study, we focus on the following:

Each of these aspects of teaching practice has been independently judged by three separate groups: students (via endpoint questionnaire), teachers (via endpoint questionnaire), and expert observers (via ratings awarded to video-taped lessons). An overview of the information reported by these groups is provided below, with further details (including the exact wording of questions used to form each scale) provided in online Appendix C.

Teacher Views

In the endpoint questionnaire, teachers self-reported their reflections on the lessons they conducted concerning quadratic equations. From their responses, three scales are formed, with details provided in online Appendix Table C1. The first scale was classroom management, formed of 10 questions with Cronbach's alpha = 0.82 (e.g., “I lost quite a lot of time because of students interrupting the lessons”). The second was socioemotional support, formed of 16 questions, with Cronbach's alpha = 0.78 (e.g., “I showed interest in these students’ well-being”). The third was cognitive activation, formed of four questions, with Cronbach's alpha = 0.72 (e.g., “I gave tasks that required these students to think critically”).

Student Views

Within the endpoint questionnaire, students reported their views on the practices, procedures, and approaches their teacher used during the quadratic equation lessons. Three scales are again formed, based on similarly worded questions to those posed to the teachers. The classroom management scale was formed of 10 questions, with Cronbach's alpha = 0.77 (e.g., “When the lesson began, our mathematics teacher has to wait quite a long time for us to quieten down”). The socioemotional support scale encompassed 16 questions, with Cronbach's alpha = 0.95 (e.g., “My mathematics teacher was interested in my well-being”). Cognitive activation was measured via 4 questions, with Cronbach's alpha = 0.73 (e.g., “Our mathematics teacher gave tasks that required us to think critically”). The specific wording of questions can be found in online Appendix Table C2. Online Appendix D provides details about the reliability of student-reported measures of instructional quality, including the internal consistency of responses (Cronbach’s alpha), the extent that students in the same class give similar ratings of instructional quality (intracluster correlations), and the extent that their responses are correlated with the views of their teachers and expert observers (interrater reliability).

Expert Observers

The expert observers used the video recordings to assess the different domains of lesson quality. We use the scales constructed by the OECD. These used the average rating of two experts per lesson segment, with two lessons recorded per teacher (OECD, 2021). Classroom management covered three areas (i.e., disruptions, classroom monitoring, and routines)—an example being how quickly and effectively the teacher dealt with classroom disruptions. Socioemotional support covered two areas (i.e., respect and encouragement/warmth), with an example being how frequently and consistently teachers and students demonstrated respect for one another. Cognitive activation covered three areas (i.e., engagement in cognitively demanding subject matter, multiple approaches to and perspectives on reasoning, and understanding of subject matter procedures and processes)—an example being whether students in the class regularly engage in analyses, creation, or evaluation work that is cognitively rich and requires thoughtfulness. See online Appendix Table C3 for further details. Each scale is standardized to mean zero and standard deviation one.

Correspondence Across Raters

An interesting feature of the data is that the correspondence between teacher and student reports is relatively modest, with the Pearson correlation sitting around 0.35 for each of the three instructional quality measures (see online Appendix Table D4). The magnitude of these correlations is consistent with those reported elsewhere in the literature. For instance, Wisniewski et al. (2022) investigated the correlation between teacher and student perceptions of the three basic dimensions of instructional quality across 171 German classrooms. They found the correlation to stand at 0.35 for cognitive activation, 0.40 for student support, and 0.52 for classroom management. Undertaking a similar analysis, Kunter and Baumert (2006) found the correlation between student and teacher reports to stand at 0.25 for teacher-provided social support, 0.64 for classroom management, and 0.24 for supporting cognitive autonomy. Similarly, Wagner et al. (2016) reported class-level correlations between student and teacher reports of around 0.35 for goal clarity, 0.7 for classroom management, and 0.1 for support for autonomy. The same authors noted how similar findings were reported by Clausen (2002), stating that the “relative agreement between teacher and student ratings . . . ranged from –.28 to .42 for 12 different instructional quality dimensions” (Wagner et al., 2016, p. 706). The correlation between student- and teacher-reported measures of instructional quality within the TALIS Video Study thus appear to be of a similar magnitude to those reported elsewhere in the literature.

However, the correlation with the views of expert observers is much lower—often sitting close to or below zero (see online Appendix Table D4). In Appendix J, we provide an overview of the existing literature on the consistency of reports of lesson quality across different observers, noting that many other studies have found correlations with the views of expert observers to be relatively low (although usually greater than zero). We go on to explore potential reasons for the near-zero correlations between student/teacher views and those of expert observers within the TALIS Video Study, including possible floor and ceiling effects, lack of variation in the scales, issues surrounding question wording, investigating the multilevel reliability, and exploring the strength of the association in the views expressed by different expert observers. From these investigations, we conclude that our greatest concern centers around the reliability and validity of the information provided by expert observers. This is based on the following:

Methodology

Research Question 1

To address the first research question, we estimate a set of ordinary least squares regression models. These are estimated at the teacher/class level when using teacher/observer reports of lesson quality and at the student level when using student reports (with standard errors clustered at the class/teacher level). In Online Appendix E, we test the robustness of our findings to estimating multilevel models instead, finding little substantive difference to our results. The specification of the expert/teacher models is as follows:

L Q jk = α + β · TS E jk + γ · T C jk + δ · SC ¯ jk + μ k + ε j

(1)

where L Q jk is lesson quality of the teacher in each domain (i.e., classroom management, socioemotional activation, and cognitive activation) according to student views, teacher views, and expert observers (separate models are estimated using the ratings of each expert observer); TS E jk is the TSE scale (standardized to mean zero and standard deviation one); T C jk is a vector of teacher characteristics (e.g., teacher gender and experience); SC ¯ jk is a vector of class-average student characteristics; j is teacher j; k is country k; μ k is country fixed effects; and ε j is a random error term.

Our variable of interest is TSE. The β parameter reveals whether teachers who have higher levels of self-efficacy deliver higher-quality lessons based on student views, teacher views, and expert observer views. Within our theoretical model (outlined in Figure 1), these estimates attempt to measure channel A. Because all outcome measures are standardized, estimates can be interpreted in terms of standardized differences (e.g., the standard deviation change in outcome per each standard deviation increase in the TSE scale).

Research Questions 2 and 3

Similar models are used to address our second and third research questions, although with each now estimated at the student level. Formally, the model is specified as

O ijk = α + β · TS E jk + γ · T C jk + δ · S ijk + μ k + ε ij

(2a)

where O ijk is one of our outcomes of interest (e.g., student test scores); S_ijk is a vector of student background characteristics; i is student i; j is teacher j; and k is country k.

With all other variables defined as earlier. The parameter of interest is again β , capturing the standardized change in outcome for each standard deviation increase in TSE. Returning to the theoretical model presented in Figure 1, the β parameter will capture the joint effect of channels B and A + C (for socioemotional outcomes) and all the channels combined for test score outcomes.

As Abadie et al. (2023) noteed, in cases of cluster samples, standard errors should be clustered at the level of the sampling unit. With respect to the TALIS Video Study, this implies that standard errors should be clustered at the school/teacher level. (Note that because only one teacher participated per school, clustering the sample at the teacher level is equivalent to clustering the standard errors at the school level.) All standard errors are hence clustered by teacher to account for the hierarchical nature of the data.

If the β parameter from the model specified in equation (2a) is statistically significant (i.e., there is a relationship between TSE and pupil outcomes), we will then estimate the following supplementary model:

O ijk = α + β · TS E jk + γ · T C jk + δ · S ijk + θ · L Q ijk + μ k + ε ij

(2b)

where LQ_ijk is teacher and student measures of lesson quality.

The change in the β parameter between equation (2a) and equation (2b) will reflect the extent that the link between TSE and outcome can be explained (in a statistical sense) by the role TSE plays in improving instructional quality. For instance, with respect to students’ socioemotional outcomes, these models will now hold constant the channels A + C, thus isolating the role played by channel B.

Pooled Versus Country-Level Estimates

Because the TALIS Video Study data are drawn from across eight countries, the analyses can be conducted in two ways. The first is to pool the data from across the eight countries into a single dataset and then create all scales and conduct all analyses using this international sample. This has the advantage of maximizing statistical power. However, one may question the wisdom of such an approach due to—for instance—possible cross-national differences in interpretation of the survey questions. The alternative is to treat the data from the eight countries as eight separate samples, investigating whether the sample substantive answers to our research questions emerge in each. The statistical power in any of these individual country samples, of course, will then be far more limited.

In most of our main results tables, we have chosen to present estimates from both approaches. The main exception is Research Question 4, where we only present results based on the pooled sample. This is due to the limited number of teachers in individual countries, meaning that there is not sufficient statistical power to divide national samples into TSE tertiles. Note that when we do present separate estimates for each country, we make no attempt to compare them, which we consider to be unwise (e.g., due to limited statistical power, differences in response rates and representativeness, and potential issues of measurement invariance). Rather, our focus is on whether analysis of each separate sample produces substantive findings that are consistent with those from the analysis pooled across countries.

Research Question 1: How strong is the association between TSE and teacher, student, and expert reports of lesson quality?

Table 2 begins by presenting the relationship between TSE and the three dimensions of lesson quality. These estimates thus capture channel A within the theoretical framework set out in Figure 1. Results are presented when information on these outcomes is drawn from different sources—teachers (left-hand column), students (middle column), and expert observers (right-hand column). All estimates capture the standardized change in the outcome per standard deviation increase in the TSE scale.

OPEN IN VIEWER

Table 2

Association Between Teacher Self-Efficacy and Measures of Lesson Quality

Country/area	Teacher		Student		Video
	Beta	SE	Beta	SE	Beta	SE
(a) Classroom management
Chile	0.30**	0.12	0.11*	0.06	0.01	0.14
Colombia	0.52**	0.11	0.14**	0.05	−0.10	0.14
England	0.26*	0.15	0.10	0.08	−0.05	0.14
Germany	0.25*	0.14	0.25**	0.07	0.02	0.15
Japan	0.26**	0.10	0.12**	0.05	0.03	0.06
Madrid	0.32**	0.15	0.08	0.08	0.27**	0.10
Mexico	0.55**	0.08	0.04	0.05	0.12	0.13
Shanghai	0.23*	0.13	0.03	0.04	0.05	0.04
Pooled analysis	0.39**	0.04	0.09**	0.02	0.04	0.04
Country/area	Teacher		Student		Video
	Beta	SE	Beta	SE	Beta	SE
(b) Socioemotional support
Chile	0.25**	0.12	0.02	0.06	0.24**	0.09
Colombia	0.35**	0.13	−0.02	0.04	−0.19	0.12
England	0.43**	0.13	0.05	0.05	0.06	0.12
Germany	0.59**	0.17	0.15**	0.06	0.26*	0.15
Japan	0.66**	0.08	0.09**	0.04	−0.07	0.08
Madrid	0.45**	0.13	−0.01	0.05	0.05	0.13
Mexico	0.58**	0.09	0.11**	0.04	0.10	0.10
Shanghai	0.57**	0.11	0.05	0.04	0.06	0.07
Pooled analysis	0.49**	0.03	0.06**	0.02	0.05	0.03
	Teacher		Student		Video
	Beta	SE	Beta	SE	Beta	SE
(c) Cognitive activation
Chile	0.16	0.12	−0.01	0.04	0.00	0.03
Colombia	0.26**	0.11	−0.01	0.03	−0.05*	0.03
England	0.20	0.13	0.03	0.04	0.01	0.04
Germany	0.35	0.20	0.02	0.05	−0.14**	0.06
Japan	0.38**	0.09	0.06	0.05	−0.03	0.04
Madrid	0.31**	0.14	0.04	0.04	0.05	0.04
Mexico	0.40**	0.09	0.13**	0.04	0.02	0.03
Shanghai	0.30**	0.10	0.08**	0.03	0.04	0.03
Pooled analysis	0.29**	0.04	0.05**	0.01	−0.01	0.04
Country/area	Teacher		Student		Video
	Beta	SE	Beta	SE	Beta	SE
(b) Socioemotional support.
Chile	0.25**	0.12	0.02	0.06	0.24**	0.09
Colombia	0.35**	0.13	−0.02	0.04	−0.19	0.12
England	0.43**	0.13	0.05	0.05	0.06	0.12
Germany	0.59**	0.17	0.15**	0.06	0.26*	0.15
Japan	0.66**	0.08	0.09**	0.04	−0.07	0.08
Madrid	0.45**	0.13	−0.01	0.05	0.05	0.13
Mexico	0.58**	0.09	0.11**	0.04	0.10	0.10
Shanghai	0.57**	0.11	0.05	0.04	0.06	0.07
Pooled analysis	0.49**	0.03	0.06**	0.02	0.05	0.03
	Teacher		Student		Video
	Beta	SE	Beta	SE	Beta	SE
(c) Cognitive activation.
Chile	0.16	0.12	−0.01	0.04	0.00	0.03
Colombia	0.26**	0.11	−0.01	0.03	−0.05*	0.03
England	0.20	0.13	0.03	0.04	0.01	0.04
Germany	0.35	0.20	0.02	0.05	−0.14**	0.06
Japan	0.38**	0.09	0.06	0.05	−0.03	0.04
Madrid	0.31**	0.14	0.04	0.04	0.05	0.04
Mexico	0.40**	0.09	0.13**	0.04	0.02	0.03
Shanghai	0.30**	0.10	0.08**	0.03	0.04	0.03
Pooled analysis	0.29**	0.04	0.05**	0.01	−0.01	0.04

Notes. Pooled analysis refers to whether analysis based on the international sample includes all countries. Estimates refer to the standardized change in the outcome per one standard deviation increase in the TSE scale. Models control for teacher gender, experience, class average socioeconomic status, class average pretest score, grade, proportion of immigrant students, how long the students have been taught by the teacher, students’ interest and self-efficacy in mathematics under their previous teacher, test effort and motivation, the proportion of students in the class the teacher deemed to have characteristics limiting their instruction, and country fixed effects. Standard errors in student-level estimations are clustered at the teacher level.

Source. Authors’ own calculations using data from OECD (2019).

*

Statistical significance at the 10% level; **statistical significance at the 5% level.

When teacher reports are used to measure lesson quality, sizable and statistically significant estimates are produced. For instance, in the pooled sample, each standard deviation increase in TSE is associated with an ~0.5 standard deviation increase in teachers’ views of the socioemotional support they provided to their students, with an ~0.3 standard deviation increase in the classroom management and cognitive activation scale. These large, standardized differences are consistent with previous research into TSE using teacher-reported outcome measures (Lauermann & ten Hagen, 2021). If taken at face value, these results seem to suggest that TSE is an important antecedent of instructional quality.

Rather different findings emerge, however, in the middle column of Table 2 when the information on lesson quality is captured from the perspective of students. Although still reaching statistical significance in the pooled sample at the 5% level, the estimated standardized differences are now much more modest. For instance, in the pooled analysis, each standard deviation increase in TSE is associated with a 0.05 standard deviation increase in the (student-reported) cognitive activation scale—a small effect.

Table 3 extends these findings to encompass a broader set of student-reported outcome measures. Consistent with the results presented in Table 2, statistically significant associations are found in the pooled analysis, but all the estimated standardized differences (standing around 0.05 standard deviations) are relatively small.

OPEN IN VIEWER

Table 3

Association Between Teacher Self-Efficacy and Additional Measures of Lesson Quality from the Student Perspective

Country/area	Teacher enthusiasm		Student engagement		Student being on task		Teacher use of feedback
	Beta	SE	Beta	SE	Beta	SE	Beta	SE
Chile	0.03	0.06	0.00	0.02	0.09**	0.03	0.00	0.05
Colombia	0.02	0.05	−0.07**	0.03	−0.02	0.03	−0.09**	0.04
England	0.04	0.05	0.01	0.04	0.04	0.04	0.12**	0.04
Germany	0.16**	0.06	0.09**	0.04	0.10	0.05*	0.14**	0.04
Japan	0.09**	0.03	0.06**	0.02	0.06**	0.03	0.04	0.04
Madrid	−0.01	0.05	0.03	0.03	0.01	0.05	−0.02	0.04
Mexico	0.12**	0.04	0.07**	0.03	0.03	0.03	0.13**	0.04
Shanghai	0.05	0.04	0.05**	0.02	0.01	0.02	0.04	0.03
Pooled analysis	0.07**	0.02	0.03**	0.01	0.05**	0.01	0.05**	0.02

Notes. Estimates refer to the standardized change in the outcome per one standard deviation increase in the TSE scale. Models control for teacher gender, experience, socioeconomic status, pretest score, grade, immigrant status, how long the students have been taught by the teacher, students’ interest and self-efficacy in mathematics under their previous teacher, test effort and motivation, the proportion of students in the class the teacher deemed to have characteristics limiting their instruction, and country fixed effects. Standard errors are clustered at the teacher level.

Source. Authors’ own calculations using data from OECD (2019).

*

Statistical significance at the 10% level; **statistical significance at the 5% level.

The results based on the opinion of expert observers (right-hand column of Table 2) are even starker, with null effects typically found. Now, in the pooled analysis, all the estimates are small (~0.05 standard deviations or less) and statistically indistinguishable from zero.

The evidence with respect to Research Question 1 is hence clearly rather mixed and largely depends on the source of the outcome measure used. Given this, one may ask which—if any—of the reports (i.e., teacher, student, or expert) should be preferred? One logical way to decide is to consider which has the strongest association with students’ outcomes. We present evidence on this matter in Table 4, where we illustrate the standard deviation change in various student-level outcomes associated with a one standard deviation increase in the measure of lesson quality, based on teacher reports (column 1), student reports (column 2), and expert observers (column 3). These estimates are conditional on baseline measures of the respective outcome along with a set of additional background covariates (see table notes for further details) and based on the sample that has been pooled across countries. Online Appendix F provides alternative estimates with fewer background controls.

OPEN IN VIEWER

Table 4

Association Between Teacher, Student and Expert Reports of Lesson Quality and Student Outcomes

Factor	Teacher reports		Student reports		Expert reports
	Standardized difference	SE	Standardized difference	SE	Standardized difference	SE
Classroom management
General mathematics self-efficacy	0.02*	0.01	0.08**	0.01	0.01	0.01
Self-efficacy mathematics tasks	0.02**	0.01	0.13**	0.01	0.03**	0.01
Personal interest	0.03**	0.01	0.22**	0.01	0.02	0.02
Self-confidence	0.02**	0.01	0.11**	0.01	0.03**	0.01
Post-test scores	0.01	0.01	0.03**	0.01	0.02**	0.01
Socioemotional support
General mathematics self-efficacy	0.03**	0.01	0.17**	0.01	0.01	0.01
Self-efficacy mathematics tasks	0.03**	0.01	0.23**	0.01	0.05**	0.01
Personal interest	0.05**	0.01	0.38**	0.01	0.06**	0.02
Self-confidence	0.04**	0.01	0.22**	0.01	0.04**	0.01
Post-test scores	0.01	0.01	0.04**	0.01	0.03**	0.01
Cognitive activation
General mathematics self-efficacy	0.01	0.01	0.11**	0.01	0.01	0.01
Self-efficacy mathematics tasks	0.01	0.01	0.15**	0.01	0.02	0.01
Personal interest	0.03*	0.01	0.23**	0.01	0.02	0.01
Self-confidence	0.00	0.01	0.12**	0.01	0.00	0.01
Post-test scores	0.01	0.01	0.01**	0.01	0.03**	0.01

Notes. Results based on analysis pooled across countries. Controls include a baseline measure of the outcome along with teacher gender, experience, socioeconomic status, pretest score, grade, immigrant status, how long the students have been taught by the teacher, students’ interest and self-efficacy in mathematics under their previous teacher, test effort and motivation, and the proportion of students in the class the teacher deemed to have characteristics limiting their instruction. All figures refer to the change in the outcome measure per one standard deviation increase in the lesson quality scale.

Source. Authors’ own calculations using data from OECD (2019).

*

Statistical significance at the 10% level; **statistical significance at the 5% level.

The results from Table 4 indicate that teacher and expert reports of lesson quality are only weakly related to student outcomes. For instance, a one standard deviation increase in teacher or expert reports of socioemotional support provided by teachers is associated with a 0.04 standard deviation increase in students’ mathematics self-confidence. However, the analogous association with the student-reported measures is notably stronger, standing at 0.22 standard deviations. Hence, of the three raters of instructional quality, it is the information reported by students that is most strongly associated with student outcomes.

Thus, our overall attempt to synthesize the findings presented across Tables 2–4 is that they generally point toward there being some positive association between TSE and instructional quality, although the strength of this link is rather weak.

Research Question 2: How strong is the association between TSE and students’ interest in a subject?

Table 5 addresses our second research question, where we explore the link between TSE and students’ socioemotional outcomes, including their self-confidence, self-efficacy, and subject interest. Referring to Figure 1, these estimates capture both the direct association between TSE and these outcomes (channel B) as well as any potential indirect association that occurs through lesson quality (channels A + C).

OPEN IN VIEWER

Table 5

Association Between Teacher Self-Efficacy and Student's Socioemotional Outcomes

	Student self-confidence				General math self-efficacy
	Total effect		Direct effect		Total effect		Direct effect
Country/area	Beta	SE	Beta	SE	Beta	SE	Beta	SE
Chile	0.07**	0.03	0.07**	0.03	0.05*	0.03	0.05*	0.03
Colombia	−0.04	0.04	−0.03	0.03	−0.05*	0.03	−0.05**	0.02
England	0.01	0.04	0.00	0.04	0.06*	0.03	0.04	0.03
Germany	0.19**	0.07	0.15**	0.06	0.15**	0.06	0.12**	0.06
Japan	0.01	0.02	−0.01	0.02	0.01	0.02	0.00	0.02
Madrid	−0.04	0.04	−0.05	0.04	−0.04	0.03	−0.04*	0.03
Mexico	0.07*	0.04	0.03	0.03	0.06**	0.03	0.03	0.03
Shanghai	0.03	0.02	0.01	0.02	0.03	0.02	0.01	0.02
Pooled analysis	0.03**	0.01	0.00	0.02	0.03**	0.01	0.01	0.01
	Self-efficacy math task				Interest
	Total effect		Direct effect		Total effect		Direct effect
Country/area	Beta	SE	Beta	SE	Beta	SE	Beta	SE
Chile	0.03	0.03	0.03	0.02	0.04	0.03	0.04	0.03
Colombia	−0.02	0.04	−0.01	0.03	0.01	0.04	0.00	0.02
England	0.05	0.04	0.03	0.04	0.01	0.04	−0.01	0.03
Germany	0.05	0.05	0.00	0.05	0.12*	0.06	0.05	0.05
Japan	0.00	0.02	−0.02	0.02	0.03	0.02	0.00	0.02
Madrid	−0.02	0.03	−0.04	0.03	−0.01	0.04	−0.02	0.03
Mexico	0.10**	0.03	0.06*	0.03	0.07*	0.04	0.02	0.03
Shanghai	0.05*	0.03	0.03	0.02	0.01	0.02	−0.02	0.02
Pooled analysis	0.03**	0.01	0.01	0.01	0.03**	0.01	0.00	0.01

Notes. Estimates refer to the standardized change in the outcome per one standard deviation increase in the TSE scale. Total effect models control for teacher gender, experience, socioeconomic status, pretest score, grade, immigrant status, how long the students have been taught by the teacher, students’ interest and self-efficacy in mathematics under their previous teacher, test effort and motivation, the proportion of students in the class the teacher deemed to have characteristics limiting their instruction, and country fixed effects. Direct effect additionally adds controls for student and teacher views of lesson quality. Standard errors are clustered at the teacher level.

Source. Authors’ own calculations using data from OECD (2019).

*

Statistical significance at the 10% level; **statistical significance at the 5% level.

The results in the “Total effect” columns of Table 5 are in many ways similar to those produced under Research Question 1 when using student reports of lesson quality. While all estimates within the pooled sample are statistically significant at the 5% level (in part due to the large student sample size), the standardized differences are consistently very small (~0.03 standard deviations). Consequently, while there appears to be some modest association between TSE and students’ socioemotional competencies, any such link is—at best—weak.

The “Direct effect” columns in Table 5 illustrate the link between TSE and socioemotional outcomes, controlling for the three basic dimensions of teaching quality in students’ and teachers’ views. Relative to Figure 1, these estimates attempt to capture the direct association between TSE and these outcomes (channel B). Results show that the direct association between TSE and students’ interest in a subject is essentially zero. Thus, we do not find evidence that there may be a direct influence of TSE on students’ socioemotional outcomes. This implies that higher TSE leads to better instruction, which ultimately fosters students’ interest in a subject (although the magnitude of these associations is consistently small).

We extend this part of our analysis in online Appendix G using a structural equation model to estimate the total, direct, and indirect associations between TSE and student outcomes using a mediation model. Results from this SEM analysis support the findings reported in Table 5; there is a small total association between TSE and students’ socioemotional outcomes, which almost entirely operates through the indirect channel of lesson quality.

Research Question 3. Is there a relationship between TSE and students’ test scores?

Table 6 turns to the total association between TSE and student test scores. It clearly points toward a null result. The point estimate is essentially zero (0.01 standard deviations in the pooled analysis) and not statistically significant. Given the findings from our first two research questions—weak associations between TSE and intermediate outcomes (lesson quality and students’ socioemotional competencies)—it is perhaps unsurprising that we then find little association with student test scores.

OPEN IN VIEWER

Table 6

Association Between Teacher Self-Efficacy and Student's Test Scores

Country/area	Beta	SE
Chile	0.05**	0.02
Colombia	−0.01	0.01
England	0.02	0.02
Germany	0.02	0.02
Japan	0.01	0.03
Madrid	−0.01	0.02
Mexico	0.00	0.01
Shanghai	0.05	0.03
Pooled analysis	0.01	0.01

Notes. Estimates refer to the standardized change in the outcome per one standard deviation increase in the TSE scale. Models control for teacher gender, experience, socioeconomic status, pretest score, grade, immigrant status, how long the students have been taught by the teacher, students’ interest and self-efficacy in mathematics under their previous teacher, test effort and motivation, the proportion of students in the class the teacher deemed to have characteristics limiting their instruction, and country fixed effects. Standard errors are clustered at the teacher level.

Source. Authors’ own calculations using data from OECD (2019).

**

Statistical significance at the 5% level.

Research Question 4: Is there any evidence of nonlinearity in these relationships?

Thus far our models have implicitly assumed that the relationship between TSE and our outcomes is linear. We now relax this assumption in Table 7, where we investigate how each outcome differs across teachers with low, average, and high self-efficacy levels (with the average category as the reference group).

OPEN IN VIEWER

Table 7

Exploration of Potential Nonlinearities in the Relationship Between Teacher Self-Efficacy with (a) Lesson Quality and (b) Pupil Outcomes

Factor	Teacher views		Student views		Expert observers
(a) Measures of lesson quality.
Classroom management
Low level	−0.320**		−0.177**		−0.047
High level	0.590**		0.035		0.007
Socioemotional support
Low level	−0.418**		−0.107**		−0.101
High level	0.617**		0.007		0.061
Cognitive activation
Low level	−0.171**		−.098**		0.124
High level	0.433**		0.002		0.081
Teacher enthusiasm
Low level	—		−0.098**		—
High level	—		0.010		—
Cognitive engagement
Low level	—		−0.065**		—
High level	—		−0.029		—
Pupil on task
Low level	—		−0.029		—
High level	—		0.080**		—
Use of feedback
Low level	—		−0.131**		—
High level	—		−0.031		—
Factor	Self- confidence	General self-efficacy in math	Self-efficacy in math tasks	Interest	Test scores
(b) Pupil outcomes.
Teacher self-efficacy (ref: average level)
Low level	−0.061*	−0.059**	−0.062**	−0.065**	−0.022
High level	−0.001	0.015	0.014	0.000	0.015
Observations	15,096	15,078	14,497	15,091	15,085

Notes. Estimates based on the pooled sample across all countries. Estimates refer to the standardized change in the outcome in comparison with teachers with average self-efficacy levels (middle third of the TSE distribution) as the reference group. Models control for teacher gender, experience, socioeconomic status, pretest score, grade, immigrant status, how long the students have been taught by the teacher, students’ interest and self-efficacy in mathematics under their previous teacher, test effort and motivation, the proportion of students in the class the teacher deemed to have characteristics limiting their instruction, and country fixed effects. Standard errors in student-level estimations are clustered at the teacher level.

Source. Authors’ own calculations using data from OECD (2019).

*

Statistical significance at the 10% level; **statistical significance at the 5% level.

There are three key points of note. First, similar broad substantive findings hold, as reported under Research Questions 1 to 3. That is, there are strong associations between TSE and teacher reports of lesson quality, only weak relationships with student reports of lesson quality, and essentially no link with the views of expert observers.

Second, to the extent that there is an association between TSE and student reports of lesson quality, it is being driven by differences between the low versus middle TSE groups. In particular, note how most of the standardized differences for the low-TSE group in the middle column of Table 7 are around 0.1, with most being statistically significant at (at least) the 10% level. In contrast, the standardized differences for the high-TSE group versus the average group are smaller (typically standing at 0.05 or below) with only one reaching statistical significance.

Finally, the results in panel (b)—for student outcomes—exhibit a similar pattern. To the extent that there is any link between TSE and student outcomes, it is restricted to the low- versus middle-TSE groups and only for socioemotional competencies (not test scores). For instance, there is around a 0.06 standard deviation difference in student self-confidence and self-efficacy when they are taught by a teacher with average compared with low levels of self-efficacy but no difference between the average- and high-TSE groups.

Overview of Results from Alternative Analytic Approaches

The overarching principle we have followed has been to apply widely used statistical techniques that are understood across disciplinary audiences. This has the advantage of making the analysis as transparent and accessible as possible. We appreciate, however, that other methodologies can be applied that may offer some technical advantages. A series of alternative estimates is thus presented in the online supplementary material.

An assumption implicitly made within the regression analyses presented earlier is that all variables are measured without error. If this assumption does not hold, then it may bias the estimated associations between TSE and student outcomes. In particular, if there is random noise in our measure of TSE, then estimates of its association with later outcomes are likely to be attenuated (i.e., bias downward). Structural equation modeling can be used to relax this assumption. This approach treats the measure(s) in question (e.g., TSE) as a latent variable. The correlation between the questions used to capture this latent indicator provides an indication of how well it is measured. This can be used to deattenuate estimates of its association with our outcomes.

Online Appendix H consequently presents alternative estimates using a structural equation model (SEM). The use of SEM instead of ordinary least squares typically leads to a slight increase in the estimated standardized differences. Indeed, the only potential exception is with respect to the association between TSE and teacher reports of socioemotional support and cognitive activation during their lessons, where the association becomes somewhat larger. This further strengthens our finding that the link between TSE and the three basic dimensions of teaching quality is much stronger when using teacher self-reports. Otherwise, the use of SEM only leads to rather modest changes to our parameter estimates.

Another limitation with the TALIS Video Study data is that in some participating countries response rates were low (e.g., England), whereas others deviated from the intended sample design (e.g., in Germany, teachers were not randomly selected within schools). This may limit the national representativeness of the sample and thus the external validity (generalizability) of our results. One way to test the sensitivity of estimates to this issue is to construct a set of weights to improve the representativeness of the data. In essence, this approach increases the influence of some students/teachers on the estimates if they are underrepresented in the sample (compared with the broader population). In online Appendix I, we follow the approach of Jerrim (2024), developing and applying a set of weights that attempt to enhance the representativeness of the data. The application of these weights leads to little change to the substantive conclusions drawn.

In our primary model specifications, we also have used missing dummy flags to limit the number of observations dropped from the analysis due to missing data. There are, however, known challenges with this approach that could bias estimates of the association between TSE and our outcomes of interest (Groenwold et al., 2012). In particular, missing data may bring selection bias into the sample. In online Appendix A, we use multiple imputation as an alternative approach. This has the advantaged of predicting—as far as possible—the likely values of the data that are missing, with the standard errors adjusted upward to reflect the extra uncertainty that missing data bring into the analysis. As online Appendix A illustrates, this also does not lead to any substantive change to our conclusions.