First, thanks to advances in classroom observations, we are now measuring many more aspects of the student-teacher relationship. Classroom observation tools (such as CLASS, Stallings and Teach) are designed to capture several micro-dimensions of the student-teacher interaction. Does a teacher ask questions? Do they acknowledge positive student behaviour? Do they make connections between lesson content and students’ daily lives? So now we know much more about what teachers do.

Second, we are getting more interested in what teachers believe. See, for instance, studies on Teacher Mindsets, the Teacher Cultural Beliefs Scale , Teachers’ gender stereotype scale towards mathematics, and the Teachers’ Conceptions of Assessment. And teacher beliefs have been shown to significantly impact student learning outcomes. For instance, teachers’ over-assessment of boys in a specific subject has a positive and significant effect on boys’ achievements and a significant negative effect on girls. Prior beliefs among teachers can predispose them to evaluating poorer students less favourably compared to their more affluent classmates, even when their scholastic aptitude and behaviour is the same. A study from across 9 countries and 20,000 teachers found that as many as 43 percent of teachers believe that there is little they can do to help a child learn if parents are uneducated.

Third, we are getting better at letting data speak. This is clearest in the use of Machine Learning in economics. Machine learning methods help because they allow us to analyse large numbers of teacher, school, and student variables (high-dimensional data) without worrying about ad-hoc variable or model selection by the researcher. Unlike past work that relied largely on linear modelling and various model-specific assumptions, it reduces the scope for researcher subjectivity and biases. Machine learning algorithms also help uncover generalisable patterns and discover complex relationships between variables that were not specified in advance, without jeopardising out-of-sample replicability.

In our paper, we put all these elements together to study teacher effectiveness in Tanzanian primary schools. We ask whether it is – who teachers are, what teachers know, what teachers do, or what teachers believe—that matters most for student learning. (For those who are interested, we do this using Conditional Inference Forests and Least Absolute Shrinkage and Selection Operator [LASSO] as our machine learning algorithms.)

We find that overall, teacher covariates matter more, and differently, for Math than Kiswahili. For Math, the teacher belief that they can help disadvantaged and struggling students learn; the teacher practice of providing clear and helpful written feedback (on homework and tests); and the teacher preparation in teaching foundational concepts are the three most predictive factors for student learning gains. For Kiswahili, the teacher belief that they have good relationships in schools; the teacher practice of providing extra help to struggling students; and (as in Math) teacher preparation in teaching foundational concepts are the three most predictive teacher covariates for student learning gains.

So ultimately, specific elements of what teachers know (teacher preparation); what teachers do (teacher practice); and what teachers believe (teacher beliefs) are more strongly predictive of student learning gains than other teacher, student, and school factors, especially in Math.

This is good news because it suggests that training teachers to engage in specific, well-defined practices could potentially make a big difference. While our results aren’t causal, they nevertheless point to the fact that ensuring that teachers provide written feedback to students, review key concepts at the end of class, and spend extra time with struggling students—will potentially make them better teachers. So will training them how to teach foundational skills (reading, numeracy).

Fostering specific beliefs in teachers may be harder, but does offer a promising new area of inquiry. The belief that all their students—even the ones who are disadvantaged or struggling—can learn could be a game-changer. Emerging insights from behavioural economics and social psychology demonstrate that such positive beliefs can be systematically fostered, but we need more evidence on how to do this sustainably at scale.

Millions of students are struggling to master basic skills, and COVID has only made the situation worse. At this urgent time, rather than chasing the elusive crux of what makes a good teacher, it might pay off to ensure that all teachers incorporate basic good practices in their teaching. And let’s foster in them optimism about the potential of their students and their role. Our paper suggests that we can be practical and focused—on what teachers do and what teachers believe.