The previous module posed the problem: we need a credible counterfactual, and selection bias corrupts naive comparisons. This module gives the most powerful solution ever devised — the randomised controlled trial — which constructs a credible counterfactual by the simple, radical act of deciding who gets the treatment by chance. Understanding WHY randomisation works is the key to the whole field.
Why randomisation works
Chance breaks selection
In a randomised controlled trial, you take a pool of eligible units and assign the treatment AT RANDOM — by a coin flip, a lottery, a random-number generator — to a treatment group and a control group. Because assignment is random, the two groups are, in expectation, IDENTICAL in every respect except the treatment: same average motivation, same average ability, same average baseline outcomes, same average of every characteristic, OBSERVED AND UNOBSERVED. Randomisation makes treatment status STATISTICALLY INDEPENDENT of the potential outcomes — the treated and control groups would, on average, have had the same outcome without treatment. This is the magic: with randomisation, the control group's average outcome is a VALID counterfactual for the treated group (they were comparable to begin with), so the difference in average outcomes between treatment and control IS the average causal effect — selection bias is eliminated by design, because nothing about the units determined who got treated; only chance did. Crucially, randomisation balances UNOBSERVABLES too (motivation, ability, unmeasured confounders) — which no statistical adjustment of observational data can guarantee. That is why the RCT is the gold standard: it solves the selection problem at its root.
The credibility revolution
The rise of the RCT (and the quasi-experimental designs of later modules) is part of what Angrist and Pischke called the 'credibility revolution' in empirical economics — a shift from elaborate but assumption-heavy structural models toward DESIGN-BASED identification, where the credibility of a causal claim rests on the research DESIGN (a randomisation, a natural experiment) rather than on untestable modelling assumptions. The question became not 'what's your model?' but 'what's your source of identifying variation — and is it credibly as-good-as-random?' This reorientation transformed applied microeconomics and is the intellectual context for the whole course: we trust estimates that come from credible designs, and we are sceptical of those that rest on heroic assumptions about selection.
The development-RCT movement
The 2019 Nobel
From the late 1990s, Abhijit Banerjee, Esther Duflo, Michael Kremer, and a growing movement (institutionalised in J-PAL, the Abdul Latif Jameel Poverty Action Lab) brought the RCT into development economics, running field experiments to test anti-poverty interventions — deworming, microfinance, conditional cash transfers, textbooks, fertiliser, immunisation incentives. The approach broke big questions ('how do we end poverty?') into testable pieces ('does THIS intervention work, here, measured how?'), building an evidence base of what actually works. It won Banerjee, Duflo, and Kremer the 2019 Nobel Prize 'for their experimental approach to alleviating global poverty'. The movement reshaped development economics and aid practice (evidence-based policy, the rise of cost-effectiveness comparison — module 8), and it is the reason this course exists. Its canonical findings — that deworming is astonishingly cost-effective (Miguel-Kremer), that the impact of microfinance is far more modest than its hype, that conditional cash transfers work — came from RCTs that a naive comparison could never have established credibly.
Strengths and ethics
The RCT's great strength is INTERNAL VALIDITY — within the experiment, the causal estimate is clean and credible (selection bias eliminated). Its limits (external validity, the leap to policy) are the subject of module 8, and its threats (attrition, spillovers, non-compliance) of module 4. On ETHICS: randomisation is often challenged as unfair ('why deny the control group the treatment?'), but the ethical case rests on EQUIPOISE (genuine uncertainty about whether the treatment helps — if we KNEW it worked, we wouldn't need the trial, and if we don't know, randomising is not denying a known benefit) and SCARCITY (when resources can't reach everyone at once, a lottery is arguably the FAIREST way to allocate, and a randomised phase-in lets you evaluate at no extra cost). RCTs still require genuine ethical scrutiny (informed consent, do-no-harm, the ethics-review process — the Measurement course's final module), and not everything can or should be randomised — but the reflexive objection that experimentation is inherently unethical misunderstands both the equipoise condition and the alternative (rolling out unevaluated programmes that may not work, or may harm).
Exercise
Return to the microfinance evaluation from the last module, where selection bias made the naive comparison unreliable. A researcher proposes instead to randomly assign microloans: from a pool of eligible applicants, half are randomly offered loans (treatment) and half are not (control), and business growth is compared after two years. (1) Explain precisely how randomisation solves the selection-bias problem that ruined the naive comparison. (2) Explain why randomisation balances UNOBSERVABLES, and why that matters. (3) Explain what 'internal validity' means here and why the RCT has it. (4) Address the ethical objection that it is unfair to deny loans to the control group.