« Back to Results

Education Technology and Inequality: Randomized Trials of Tech to Enable Children's Learning across 10 Countries

Paper Session

Sunday, Jan. 5, 2025 1:00 PM - 3:00 PM (PST)

Hilton San Francisco Union Square, Union Square 22
Hosted By: American Economic Association
  • Chair: Susanna Loeb, Stanford University

The Effect of Tutor CoPilot for Virtual Tutoring Sessions: Testing an Intervention to Improve Tutor Instruction with Expert-Guided LLM-Generated Remediation Language

Susanna Loeb
,
Stanford University
Carly Robinson
,
Stanford University
Rose Wang
,
Stanford University
Dora Demszky
,
Stanford University
Ana Ribeiro
,
Stanford University

Abstract

Tutoring is one of the most common and effective approach for accelerating student learning (Fryer Jr and Howard-Noveck, 2020; Nickow et al., 2023; Robinson and Loeb, 2021). To accommodate growing demand for tutoring, many providers engage novice tutors, who lack the specialized training of professional educators. One area where novice tutors in mathematics struggle is in responding to student mistakes, a prime learning opportunity to address misconceptions (Boaler, 2013). Our study builds on previous research, showing that responses to student mistakes generated by a large language model (LLM) guided by experienced educators were higher rated than responses written by tutors alone in a blind rating test (Wang et al., 2024). To evaluate whether an expert-guided LLM assistant providing curated pedagogical strategies can improve tutor responses to student mistakes, we partnered with a chat-based tutoring provider operating at a large school district in Texas to conduct an RCT. We randomized tutors into having access to the LLM assistant as an embedded feature into the tutoring chat platform or “business-as-usual” without access to the LLM assistant. The provider’s algorithm for matching students and tutors each session is largely random, so student exposure to the treatment can be inferred by the assignment of the tutor matched for each session individually or as a proportion of all sessions. Beyond measuring tutor’s activation of the LLM assistant, we evaluate impact on the variety and frequency of high-rated pedagogical strategies used by tutors. We also evaluate the potential for the tool to improve student experience in the chat-based tutoring setting through student session ratings, and student learning as measured by end-of-the-year achievement scores. The study contributes to our understanding of how to use new technology to support a broader-than-typical range of educators.

Call Me, Maybe? Technology and Teacher Support in Niger

Jenny Aker
,
Tufts University
Josue Awonon
,
Tufts University

Abstract

Teacher absenteeism is a pervasive problem in lower-income countries. While monitoring teachers should, in theory, address shirking, the effects of monitoring interventions on student learning are mixed, in part due to heterogeneous teaching quality. We report the results from a randomized evaluation of a mobile phone-based intervention designed to improve both teacher accountability and teaching quality in Niger. Teachers in primary schools were randomly assigned to one of three interventions: monitoring phone calls to teachers and the community, texts and phone calls providing pedagogical support to teachers, or both. We find that the mobile monitoring intervention alone increased teachers’ attendance and motivation in the short- and medium-term, without crowding out their intrinsic motivation. The monitoring calls also increased parental engagement in students’ schooling. While there were no effects of the individual interventions on student learning outcomes, the combined interventions increased math test scores. Yet in those same schools, teachers felt less appreciated as compared with control schools or schools with the individual interventions. These findings suggest that the design of teacher support and monitoring programs, especially via “low cost” technological solutions, must consider potential inflection points in the intensity of such interventions.

Optimizing Tech Tutoring Programs for Scale: Evidence from 10 Randomized A/B Tests

Noam Angrist
,
Oxford University and Youth Impact
Claire Cullen
,
Youth Impact
Janica Magat
,
Youth Impact

Abstract

A growing literature explores the effectiveness of various types of education technologies, yet few directly compare them to identify the most productive ones. We study a wide array of technology-enabled education interventions in a unified framework, ranging from low to high tech options, in randomized trials across six countries. Interventions include radio content, SMS-based nudges, as well as numeracy content and information on learning levels, live phone calls, and comparison of non-tech interventions. We study two critical outcomes: access and learning. We explore heterogeneity along multiple dimensions to understand not only how technology improves average outcomes, but also how it exacerbates or closes inequalities. Results inform which types of educational technologies can have the highest productivity, in which settings, and for whom.

Free to Choose: Introducing Technology in a Public Schooling System

Tahir Andrabi
,
Pomona College and CERP
Juan Baron
,
World Bank
Isabel Macdonald
,
Harvard University
Zainab Qureshi
,
Harvard Kennedy School

Abstract

In the context of over 5000 public primary school teachers in the northwestern Khyber Pakhtunkhwa province of Pakistan, we develop a smart phone based technology support tool that helps teachers in implementing an targeted instruction program that sorts (k-5) students into four groups according to levels of test scores in English, Urdu and Math. The instruction is carried over a 40 day program with a different set of lessons and activities for each subject-group level. Teachers show varying degrees of confidence and comfort in using technology based on gender and age. We experimentally test whether technology is an “experience good” and whether self-selection into technology vs a blanket mandate results in more effective usage. To test take up and utilization of the technology, we randomly place teachers in an optional technology usage group, a mandated technology usage group and a group where after an initial mandate to use technology, teachers then have the option to continue or not. Preliminary results show that while in the baseline, women are less likely than men to state confidence in using technology and opt less into technology in the optional treatment plan. However after the mandated trial period, women end up utilizing the technology more. This confirms that there is gender based status quo bias in technology and experiencing technology can over turn it. We find small effects of the trial period on age.

Delivering Remote Learning in Developing Countries using a Low-tech Solution: Evidence from Bangladesh

Paul Glewwe
,
University of Minnesota
Asadul Islam
,
Monash University
Khandker Wahedur Rahman
,
University of Oxford
Shwetlena Sabarwal
,
World Bank
Xing Xia
,
Yale University

Abstract

We conduct a randomized controlled trial in Bangladesh to test the effectiveness of an educational intervention designed to address the learning losses among secondary school students that have resulted from prolonged school closures during the COVID-19 pandemic. Secondary school students got access to a series of pre-recorded math and English audio lessons delivered via basic-feature mobile phones using Interactive Voice Response (IVR) technology. One treatment arm (T1), the self-help group, offered access to the IVR lessons for 24 weeks. A second treatment arm (T2), the assisted group, offered biweekly phone calls from a tutor to discuss the content covered in the IVR lesson in addition to the IVR lessons. We find learning outcomes improved moderately after six months in T1 (0.153 SD for math and 0.133 SD for English) and significantly in T2 (0.267 SD for math and 0.198 SD for English). Students in T2 also had higher aspirations and improved growth mindset compared to the control group.
JEL Classifications
  • I2 - Education and Research Institutions
  • O0 - General