#18 - Ofir Reich on using data science to end poverty & the spurious action-inaction distinction

Center for Effective Global Action (CEGA) Structure and Focus

• CEGA is based at UC Berkeley and has three primary activities:
  • Global Networks: Facilitates scholars from developing countries spending a semester at UC Berkeley to form academic ties, learn impact evaluation methods, and return to their home countries to lead prosperous careers.
  • Berkeley Initiative for Transparency in the Social Sciences (BITS): Addresses the "reproducibility crisis" by promoting open data, open code, and pre-registration to prevent p-hacking and false results in social science research.
  • Global Poverty Research: Operates as a network of over 70 academics (mostly on the US West Coast and Canada) conducting research on poverty alleviation, with CEGA directing funding and competing for top-tier academic research grants.

Ophir Reich's Current Data Science Projects

• Tax Evasion Detection in India:
  • Utilizes machine learning on historical value-added tax returns to identify fraudulent companies set up solely to evade taxes.
  • Targets "false" companies that exist only on paper to increase government revenue for spending on the poor.
  • Requires government partnership; uses labeled data from past manual inspections to train models for future inspections.
  • Reich remains cautious about the algorithm's real-world utility until predictive lists are tested via actual government inspections.
• Precision Agriculture (Precision Agricultural for Development):
  • Collaborates with organizations in India (Gujarat) and Ethiopia to improve agricultural extension via mobile phones (ICT).
  • Aims to increase farmer yields and profits by timing advice on fertilizer usage with current weather and soil conditions.
  • Uses Randomized Controlled Trials (RCTs) and A/B testing to measure outcomes; a previous RCT in India found returns of approximately 10x the monetary investment.
• Mobile Salary Payments in Afghanistan:
  • Evaluating the Afghan government's shift to mobile money systems (similar to M-Pesa) for teacher salaries.
  • Focuses on reducing transaction costs and improving payment reliability compared to current systems.

Data, Methodology, and Research Challenges

• Data Availability: Developing countries often lack comprehensive digital data; many transactions are physical or cash-based, limiting the scope of data science applications compared to the developed world.
• Machine Learning vs. Simple Correlation: Complex machine learning approaches outperform simple correlation checks in the tax evasion project by incorporating diverse features, though domain expertise remains crucial due to the anonymous nature of the data.
• Reproducibility Crisis:
  • Driven by practices like p-hacking (testing multiple hypotheses until a significant result is found) and selective reporting of results.
  • CEGA/BITS promotes solutions including open data, open code, pre-registration of analysis plans, and "results-blind" peer review.
  • Incentives are shifting as journals increasingly require open data/code, creating a "virtuous circle" where bad actors face the risk of being exposed.
• RCT Generalizability:
  • Meta-analyses suggest RCT results have a correlation of over 0.5, indicating moderate replication.
  • Experts argue that while specific effect sizes may not generalize, the underlying principles (e.g., teaching at the right level) do apply across contexts.
  • The cost of rigorous evaluation is tiny compared to the billions spent on social programs; investing in RCTs is viewed as high-value for policy efficacy.

Career Advice and Organizational Landscape

• Recommended Career Paths:
  • Private Sector First: Strongly recommended for data scientists to build "career capital" (technical skills, credentials) before moving to non-profits; private companies offer better data access and faster skill acquisition.
  • High-Income Earning: Suggests high-earning data science roles in the private sector as a viable path for "earnings to give," with a significant portion of income donated to effective charities.
  • Avoid Small NGOs: Early-career professionals are advised against starting or working at small, unknown non-profits; large, reputable organizations (e.g., GiveWell, Open Philanthropy, J-PAL, IPA, CEGA) offer better networking and credentialing.
• Organizational Comparisons:
  • CEGA vs. J-PAL: CEGA affiliates are not restricted to economists (including computer scientists and engineers) and are geographically focused on the US West Coast, whereas J-PAL affiliates are all economists and J-PAL has direct country offices for implementing trials.
  • Impact Attribution: Hard to quantify specific policy changes driven by research, but notable successes include the biometric smart card system in Andhra Pradesh, India, which reduced leakage significantly and gained over 90% public support.
• Skills for Effectiveness:
  • Data Science: Requires comfort with statistics, programming (Python), and dealing with messy real-world data rather than idealized models.
  • Product Management/Marketing: Modern product development practices (A/B testing, feedback loops) and lobbying skills are identified as high-leverage but underutilized in poverty alleviation.

Personal Insights and Philosophy

• Motivation: Driven by the "veil of ignorance" (treating others impartially) and the rejection of the "action vs. inaction" distinction; inaction is viewed as a responsible choice with consequences.
• Global Poverty Focus: Chosen because the scale of suffering and opportunity for improvement is greatest in developing countries compared to wealthy nations.
• Israel Defense Forces (IDF) Experience:
  • Reich spent six years in a mathematical research unit of the IDF, solving difficult problems with a team of top mathematicians.
  • Provides invaluable training in statistical reasoning, hypothesis generation, and resilience, which serves as a strong foundation for data science.
• Travel and Blogging: Maintains a travel blog (half Hebrew, half English) to immerse himself in different cultures in developing countries, debunking romanticized views of poverty and gaining on-the-ground context for research.
• Challenges:
  • Frustration: Success in policy influence often depends on external partners (governments) who may have different incentives or fail to act on data.
  • Personal Sacrifice: Living in the US (San Francisco Bay Area) away from family and home culture in Israel is a significant downside.
  • Burnout: Emphasizes that sustainability is critical for long-term impact; donors should not expect a linear trajectory of impact, and individuals must avoid burnout by choosing work they can sustain.