Understanding gender dynamics: a tutoring intervention.

We will work with primary 3rd, 4th and 6th and middle school 1st grade students to allow us to see heterogeneous effects by age. We will choose 300 schools, 200 in Baja California and 100 in Durango. Randomization will be at the student level as we are not concerned of spillover effects, students that participated last semester in the RCT met with their teacher on average less than 5 hours per week through google meets. We will do a stratified randomization caring about the grade of students. We will determine whether tutors have a traditional view of gender roles using the test implemented by (Alan et al 2018), which additionally measures extrinsic vs intrinsic motivation, modern vs traditional teaching, and warm vs distanced teaching to understand whether biased teachers affect kids decisions and performance in the direction of the bias.
We will recruit tutors, volunteers of universities and normal schools seeking to meet their social service requirements of 480 hours to graduate. This same recruitment strategy was successfully implemented by our team in a past tutoring RCT. We will assign the tutors into WhatsApp groups of 5 kids that will meet 2 times a week. All tutors will receive an initial training on how to teach math, a psychopedagogy training, a training on how to handle difficult situations in case they find out of cases of abuse in the family of children and a training session on how to talk about sickness and death with children (because in past tutoring programs during the pandemic the topic has challenged tutors). Tutors will be given the list of subjects sorted in level of difficulty. They will be given the scores in the baseline evaluation of their students and be instructed to start working on the easiest material according to the list that any of their students does not master. After the initial training sessions, all tutors will receive a weekly training in math content, and we will provide office hours on zoom where they can log in to solve their math questions. Additionally, we will create Whatsapp groups (of around 50 tutors of the same treatment) that tutors can use to communicate among each other and share tips and questions. There will be two different ways to structure the groups, and on each kids will receive either a tutor with high stereotypes or low stereotypes. We will randomize children into having mixed gender groups or same gender groups. And then into having a tutor with traditional or non traditional gender beliefs.

2021-09-06

2022-07-15

Our main outcome is the academic learning of students measured by the math and spanish score on a subset of the questions of a standardized evaluation used in previous years to evaluate the educational system in Mexico(PLANEA and ENLACE). Students will answer this evaluation before they are assigned to treatment or control groups. We will apply this evaluation again after the intervention.
Additionally, we are interested in investigating the affective bond that students report for their tutors in three dimensions clossenes, instrumental help and conflict.

Another important outcome is the recording of the audio of tutoring sessions. We want to do some exploratory work using Natural Language Processing techniques to study the interactions between the students and tutors and see how these interactions relate to the success of the tutoring sessions. For instance, variables that we are considering are the tone in which the tutor speaks (encouraging, scolding...), the frequency of each student’s interventions among others. The literature of computer science on sentiment analysis using natural language processing of texts and recordings of conversations is growing and this is a good setting to use these tools. Currently, to study classroom interactions, we rely on the annotations done by observers or student surveys. For instance, the Gates Foundation has a big initiative, “Measures of Effective Teaching” that relies solely on observers and surveys. The problem is that both measurements are subjective and potentially biased: what if the observer didn’t pay attention to something, they consider unimportant? Each observer will add their own subjective beliefs to their observations so how reliable are these across observers? Student surveys are also biased to their own perception of the class. Additionally, when teachers know they are being observed by a third person they might act differently than they would otherwise also biasing the results. We want to
automate this task and put a “neutral” or at least “constant” algorithm to analyze what is happening in the tutoring session. Additionally, this observer is a less visible observer as we can truthfully tell the tutors that no one is going to listen to the recordings except that they will run through a program to help us understand teaching at an aggregate level. Other potential outcomes for which we cannot collect data at the moment are future earnings and college enrollment.
Another outcome is the aspirations of students whether they want to study a math related subject if they go to university and whether they want to study an undergraduate degree. We are interested in seeing whether this effects varies by gender.

We will look into the math confidence of students. We will construct this measurement by asking students to guess their grade by the end of the evaluation and see how this differs from their true grade.

We will randomize students into mixed gender or same gender groups,then we will randomize these groups into receiving a tutor with strong or weak traditional gender beliefs.

Will be done in an office by computer

The randomization will be done at the student level.

No

-

around 2000 students.

500 students wiith same gender groups and tutor with strong traditional beliefs
500 students wiith mixed gender groups and tutor with strong traditional beliefs
500 students wiith same gender groups and tutor with weak traditional beliefs
500 students wiith mixed gender groups and tutor with weak traditional beliefs

Given our projected population, we expect to correctly estimate the sign of the effect whenever it is larger than 0.07 standard deviations.