Field experiments in behavioural econ

Field experiments — randomised controlled trials run in real-world settings — have produced most of the rigorous empirical evidence in behavioural-development economics over the past 20 years. This module covers the methodology, the canonical findings, and the open questions that the field is still working through.

Field-experimental designs

• Individual-level RCT — random assignment at the individual level. Best for individually-administered interventions (commitment products, savings reminders)
• Cluster-randomised — assignment at group level (villages, schools, classes). Used when spillover or coordination matters within the cluster
• Stepped-wedge — different clusters start treatment at different times; allows everyone to eventually receive treatment. Useful when ethical or political concerns prevent permanent treatment-control split
• Encouragement design — random encouragement to take up a product/programme, not random treatment itself. Used when actual treatment can't be randomised but encouragement can (intent-to-treat analysis)

Efficacy vs effectiveness

• Efficacy studies — does the intervention work under ideal conditions? Implemented by experienced researchers, in selected high-quality contexts. Internal validity is high; external validity is uncertain
• Effectiveness studies — does the intervention work when implemented at scale by typical operators? Implemented by government agencies, NGOs at scale. External validity is the target

Most behavioural-finance field experiments are efficacy studies — researchers in carefully selected contexts. Effectiveness at scale through government implementation is a separate question, often with reduced effect sizes.

Canonical behavioural-finance field experiments

Karlan-McConnell-Mullainathan-Zinman (2016) — Savings reminders

Six countries (Bolivia, Peru, Philippines, India, Indonesia, etc.). Random assignment of borrowers to receive periodic SMS reminders about savings goals. Result: 6-16% increase in savings rates over 12 months among those who received reminders. Modest individual effect; large aggregate effect over millions of customers.

Beaman-Karlan-Thuysbaert-Udry (2014) — Cash transfers + financial literacy

Mali. Random assignment of agricultural inputs + financial-literacy training. Treatment-group households increased income 30%+. Decomposition: most of the gain was from inputs, not from financial-literacy training per se.

Dupas-Robinson (2013) — Mental accounting and commitment savings

Western Kenya. RCT of simple lockbox savings devices. Treatment increased savings rates by 66% and led to more business investment. Mechanism: the device created a mental-accounting compartment and a commitment device simultaneously.

Brune-Giné-Goldberg-Yang (2016) — Commitment savings

Malawi. RCT of commitment savings accounts tied to harvest income. Treatment increased savings by 60% and led to 22% higher subsequent harvest because farmers could afford more inputs.

Karlan-Mullainathan-Roth (2019) — Meta-analysis

Meta-analysis of 22 commitment-savings RCTs across 9 countries (mostly developing). Common findings:

• Take-up rates 20-40% across studies
• Among takers, savings increases 50-90%
• Sustained effects: 1-3 year persistence after intervention
• Heterogeneous effects: men benefit less than women in some studies; effects stronger for sophisticated time-inconsistent agents

Open methodological questions

• External validity — does an effect demonstrated in Kenya replicate in Tanzania? In urban vs rural? In a different macroeconomic environment? The literature is converging on 'often yes, sometimes no' — context matters, but mechanisms can be transported
• Implementation fidelity at scale — government-run programmes don't always deliver the same effect as researcher-supervised pilots. Microcredit scaled imperfectly; deworming scaled with some implementation degradation
• Heterogeneous effects — average treatment effect can mask substantial heterogeneity. Knowing the average doesn't tell you which subset of users benefit
• Long-run effects — most studies have 1-3 year follow-up. Longer-run effects (5-10 years) are less studied but often substantial. Miguel-Kremer deworming showed 10-15 year wage effects

Applications to African behavioural-finance design

Practical implications for product design and policy:

• Commitment savings products work — at scale, well-designed commitment savings products produce 30-60% savings increases in target populations
• Default-design captures most users — auto-enrolment with opt-out captures 70-90% of eligible users
• Savings reminders are cheap and modestly effective — should be part of any savings-product design
• Insurance subsidies + bundled enrolment work better than voluntary insurance markets at low income levels
• Goal-anchored savings products outperform unlabelled savings — make the labelling explicit and visible

What the field experiments don't answer

Important limitations:

• Macroeconomic effects — RCTs measure individual-level impact. Aggregate effects (would universal HSNP raise inflation? Would universal SHIF improve population health?) require structural macroeconomic modelling, not RCTs
• Political economy — what level of public spending is the right level? RCTs can show this programme works at this scale; can't tell you about budget-share priority across competing uses
• Institutional change — institutional reforms can't be randomly assigned. Some of the most important development outcomes (rule of law, regulatory quality) are studied with non-experimental methods
• Cultural and contextual variation — RCTs are field-specific. Whether findings generalise across cultural contexts is empirical question, not theoretical certainty