This final chapter brings together the book's threads — micro, macro, institutions, and empirics — to address the most consequential question in economics: why are some countries rich and others poor, and what can be done about it?
Development economics is not "applied growth theory." It deals with coordination failures, institutional traps, human capital deficits, and political economy that standard models abstract away. It also features the most dramatic methodological revolution in modern economics: the rise of randomized controlled trials as a tool for evaluating interventions — and, more recently, the counter-revolution in structural estimation that seeks to push beyond what any single experiment can tell us.
This chapter synthesizes the entire textbook. Growth theory (Ch 13) provides the framework. Institutions (Ch 18) provide the deep determinants. Econometrics (Ch 10) provides the identification tools — instrumental variables, regression discontinuity, and the logic of causal inference. Behavioral insights (Ch 19) inform the design of development interventions.
By the end of this chapter, you will be able to:
Describe the stylized facts of global income distribution and structural transformation
Formalize the Lewis dual-economy model and compute the Lewis turning point
Analyze poverty traps using phase diagrams and multiple-equilibrium models
Evaluate the institutions-vs-geography-vs-culture debate using IV identification
Explain the role of human capital and health as development drivers
Interpret RCT evidence on development interventions, including power calculations
Assess the external validity debate and the case for structural estimation
Connect development economics to contemporary policy frontiers
The richest countries — Norway, Switzerland, the United States — have GDP per capita above \$60,000 (PPP). The poorest — Burundi, South Sudan, the Central African Republic — have GDP per capita below \$500. A factor of over 100 separates the richest from the poorest, and this gap has widened dramatically over two centuries. In 1800, the richest-to-poorest ratio was approximately 5:1. By 2000, it exceeded 100:1. This "Great Divergence" is the central fact that development economics must explain.
GDP per capita map
The Penn World Table reveals several patterns. In the early 19th century, the distribution was approximately unimodal: nearly all countries were poor. The Industrial Revolution created a divergence that accelerated through the 20th century. By the 1970s–1980s, the distribution had become distinctly bimodal — "twin peaks" (Quah 1996). Since 2000, rapid growth in China and India has partially filled the gap, though Sub-Saharan Africa remains largely at the lower peak.
Kuznets facts.A set of stylized regularities about economic structure as countries develop: (i) agriculture's GDP share declines with income; (ii) manufacturing's share first rises then falls (inverted U); (iii) services' share rises monotonically; (iv) urbanization increases; (v) inequality first rises then falls (the Kuznets curve, though contested).
Structural transformation.The long-run reallocation of economic activity across three broad sectors — agriculture, manufacturing, and services. As economies develop, the agricultural share falls from 50–70% to below 5%, manufacturing rises and then falls, and services eventually dominate.
Dual economy.An economy characterized by the coexistence of a large, low-productivity traditional sector (typically agriculture) and a small, high-productivity modern sector (typically manufacturing or formal services). The analytical framework implies growth through transfer of surplus labor from traditional to modern sector.
Kaldor Facts vs. Development Facts
Kaldor Facts (Ch 13)
Development Facts (This Chapter)
Constant capital-output ratio
Rising capital-output ratio during industrialization
Constant labor share
Falling labor share in agriculture, rising in industry then services
Constant growth rate of output per worker
Highly variable growth; episodes of acceleration and stagnation
The Solow model (Ch 13) captures the Kaldor facts well. It does not capture the development facts — it has one sector, one type of labor, and smooth convergence. Development economics requires models with multiple sectors, heterogeneous labor, and the possibility of traps.
Figure 20.3. Global income distribution over time (stylized). Slide through decades to see the evolution from unimodal (1800) to twin peaks (1970s) to partial convergence (2000s). Use the slider or play button.
20.2 The Lewis Model
Lewis model.Arthur Lewis's (1954) two-sector model of economic development, in which a low-productivity subsistence sector coexists with a high-productivity modern sector. Growth occurs through the transfer of surplus labor from subsistence to the modern sector.
Formalizing the Dual Economy
The modern sector uses capital and labor in a Cobb-Douglas production function:
Surplus labor.Workers in the subsistence sector whose marginal product is zero (or below the subsistence wage). When $L_S > \bar{L}$, there are $L_S - \bar{L}$ surplus workers who can be reallocated to the modern sector without reducing agricultural output.
The modern sector hires workers as long as $MPL_M > \bar{w}$. During the surplus labor phase, the modern sector faces a perfectly elastic labor supply at wage $\bar{w}$. Profits ($\Pi_M = Y_M - \bar{w}L_M$) are reinvested, creating a virtuous cycle: capital accumulation raises $MPL_M$, absorbing more workers, generating more profits.
The Lewis Turning Point
Lewis turning point.The moment when surplus labor in the subsistence sector is exhausted. Beyond this point, further labor absorption requires pulling workers whose marginal product exceeds the subsistence wage, causing wages to rise. The economy transitions from "growth through labor reallocation" to "growth through productivity improvement."
Why it matters: A poor country has a bottomless pool of farm workers whose extra output is essentially nil — pull one off the land and nothing is lost. So the modern factory sector can hire as many as it wants at a flat subsistence wage, reinvest the profits, and grow by absorbing workers rather than bidding up pay. That free ride lasts until the pool runs dry — the Lewis turning point — after which extra workers come only by raising wages, and growth has to shift to making each worker more productive. China between 1980 and 2010 is the textbook case: hundreds of millions moved from fields to coastal factories, with wages staying flat until they finally surged around 2010–2015. The slider figure below lets you watch the modern sector swell and the turning point arrive.
China is the most dramatic modern illustration. Between 1980 and 2010, China transferred hundreds of millions of workers from rural agriculture to urban manufacturing, generating growth rates of 10% per year. Economists debate whether China crossed its Lewis turning point around 2010–2015, evidenced by rapidly rising wages in coastal manufacturing zones.
Figure 20.2. Lewis dual-economy model. Left: modern sector MPL curve and subsistence wage. Right: output by sector. Increase capital to absorb labor; watch for the Lewis turning point. Drag the sliders to explore.
Example 20.1 — Lewis Model Computation
The Kaelani Republic has 10 million workers. Currently, 7 million work in subsistence with surplus labor of 3 million ($\bar{L} = 4$ million). Modern sector: $A_M = 2$, $K_M = 100$, $\alpha = 0.4$.
(a) Current modern output ($L_M = 3$M): $Y_M^{\text{before}} = 2 \times 100^{0.4} \times 3^{0.6} \approx 24.40$. After reallocating 1M workers ($L_M = 4$M): $Y_M^{\text{after}} = 2 \times 100^{0.4} \times 4^{0.6} \approx 28.99$. Output gain = 4.59 units (18.8% increase), with zero subsistence loss since transferred workers were surplus.
(b) At the turning point, $L_M = L - \bar{L} = 6$M. Setting $MPL_M = \bar{w} = 1$: $K_M^* \approx 3.80$ — a low threshold reflecting the abundance of surplus labor and modest subsistence wage.
20.3 Poverty Traps and the Big Push
Poverty trap.A self-reinforcing mechanism that causes poverty to persist. The economy has multiple steady states, and without a sufficiently large intervention, it remains stuck at a low-level equilibrium. Traps arise from coordination failures, threshold effects, credit constraints, or institutional feedback loops.
The S-Shaped Production Function
The standard Solow model features a concave production function guaranteeing a unique stable steady state. Poverty traps require an S-shaped (locally convex) production function creating multiple crossings between $sf(k)$ and $(n+\delta)k$.
Figure 20.1. Poverty trap diagram. The S-shaped $sf(k)$ curve crosses the $(n+\delta)k$ line at up to three points. Drag the dot to see convergence to the low trap or high equilibrium. Adjust saving rate and curvature with the sliders. Drag the initial condition dot to explore.
The Big Push
Big push.A coordinated, large-scale investment program designed to push an economy from a low-level equilibrium past the unstable threshold and onto a convergent path to the high equilibrium. Originates with Rosenstein-Rodan (1943).
Multiple equilibria.A situation in which an economy can settle at more than one self-sustaining outcome. Both $k_L^*$ and $k_H^*$ are equilibria — the initial conditions or a sufficiently large shock determine which is reached.
The Murphy-Shleifer-Vishny Model
Murphy-Shleifer-Vishny model.A formalization of the big push idea in which industrialization in one sector generates demand spillovers to other sectors. Each sector can use a traditional technology (constant returns) or a modern technology (increasing returns, but requiring fixed cost $F$). Whether it is profitable to modernize depends on how many other sectors have already done so.
Coordination failure.A situation in which all agents would be better off if they could simultaneously change their behavior, but no individual has an incentive to change alone. In the MSV model, each firm profits from industrializing only if others also industrialize.
$$\pi_i(n) = \alpha\!\left(\frac{n}{N}\right)L - F \tag{Eq. 20.5}$$(Eq. 20.5)
Intuition
Why it matters: When the production function bends the wrong way at low capital — each early dollar adds little, but past some threshold it pays off — an economy can have two resting points: a poverty trap and a prosperous equilibrium, with an unstable tipping point between them. The reason no single factory modernizes on its own is that modernizing only pays once enough other firms have done it too: a steel mill needs customers with money, who need jobs at other modern firms. Everyone waiting on everyone else is a coordination failure, and it locks the economy at the low point. A coordinated “big push” — investing across many sectors at once — jumps the whole economy over the tipping point together. Drag the starting-capital dot in the figure below across the unstable threshold and watch the economy fall toward the trap or climb to prosperity.
Here $n$ is the number of other sectors that have industrialized, $N$ is the total number of sectors, $L$ is the labor force, and $F$ is the fixed cost of adopting modern technology. The key feature: $\pi_i$ is increasing in $n$ — demand spillovers from industrialized sectors raise profits for each firm that modernizes. When $n = 0$, $\pi_i(0) = -F < 0$: no firm wants to industrialize alone. When $n = N$, $\pi_i(N) = \alpha L - F > 0$ if $L$ is large enough. The MSV model thus generates two Nash equilibria: no industrialization ($n = 0$, the poverty trap) and full industrialization ($n = N$, the developed equilibrium). A government can serve as the coordinating mechanism — subsidizing simultaneous investment across sectors to push the economy from the low to the high equilibrium.
When Do Traps Exist?
Not all poor countries are trapped. Kraay and McKenzie (2014) find limited evidence for poverty traps at the household level. At the country level, persistent underdevelopment in parts of Sub-Saharan Africa is more consistent with trap dynamics, particularly when combined with institutional failure and conflict.
Example 20.2 — Poverty Trap Steady States
Given $f(k) = k^2/(1+k^2)$ (S-shaped), $s = 0.20$, $n+\delta = 0.10$. Setting $sf(k) = (n+\delta)k$ and solving yields $k = 0$ and $k = 1$ (repeated root — the trap is on the verge of existence).
For a richer example, $f(k) = k^{3}/(1+k^{3})$ with $s = 0.25$, $n+\delta = 0.10$ yields three solutions: $k_L^* \approx 0$ (poverty trap), $k_U \approx 0.76$ (unstable threshold), $k_H^* \approx 2.31$ (high equilibrium). At $k_U$, the production function is locally convex so $g'(k_U) > 0$ — unstable. The big push requires injecting $\Delta k \approx 0.76$ per worker.
20.4 Institutions and Development
The Institutions Hypothesis
Extractive institutions (review from Ch 18).Political and economic institutions that concentrate power and wealth in the hands of a narrow elite, creating poor incentives for broad-based investment and innovation.
Inclusive institutions (review from Ch 18).Political and economic institutions that distribute power broadly, enforce property rights, provide public goods, and create a level playing field for economic activity.
The AJR Identification Strategy
Settler mortality instrument.An IV for institutional quality based on European settlers' mortality rates in colonial territories. Where settlers survived, they built inclusive institutions; where they died quickly, they built extractive ones. Settler mortality centuries ago is plausibly unrelated to current income except through its effect on institutions.
The fundamental challenge is endogeneity: rich countries can afford better institutions. AJR (2001) proposed an IV strategy using settler mortality. The first-stage coefficient $\beta$ is negative and highly significant (F-statistic > 20). The 2SLS estimate $\hat{\delta} \approx 0.94$ exceeds OLS ($\approx 0.52$) — consistent with attenuation bias from measurement error.
Intuition
Why it matters: You can’t prove institutions cause wealth just by noticing that rich countries have good institutions — rich countries can afford good institutions, so the arrow might run the other way. Acemoglu, Johnson and Robinson found a natural experiment: where European colonizers faced deadly disease they couldn’t settle, so they set up purely extractive states; where they survived, they built the inclusive institutions they knew. Those centuries-old death rates can only affect a country’s income today through the institutions they shaped — which lets them isolate the causal channel. The instrument estimate comes out larger than the raw correlation not by magic but because mismeasured institutions blur the simple comparison. The scatter figure below lets you switch between settler mortality, latitude, and rule-of-law on the horizontal axis and watch how tightly each tracks income.
Atlantic trade map
The Slave Trade and Long-Run Development
Slave trade instrument (Nunn).Nunn (2008) uses historical slave export data as a source of variation in institutional quality, showing that more heavily affected regions have worse institutions and lower income today. Complementary to AJR's identification strategy.
Geography vs. Institutions vs. Culture
Natural experiments reinforce the institutions hypothesis: North vs. South Korea, East vs. West Germany, pre- vs. post-reform China, and Botswana vs. its neighbors all illustrate how institutional divergence drives income divergence.
Figure 20.4. Institutions vs. geography scatter. Toggle the x-axis variable to compare settler mortality, latitude, and rule of law as predictors of income. Use the dropdown to switch views.
Example 20.3 — AJR IV Interpretation
Results: First-stage F = 22.9, $\hat{\beta} = -0.61$, 2SLS $\hat{\delta} = 0.94$ (SE = 0.16), OLS = 0.52. (a) A one-unit increase in institutional quality causes a 0.94 log-point increase in GDP/capita. Moving from 25th percentile (score 5) to 75th (score 8) predicts a \$1 \times 0.94 = 2.82$ log-point increase — roughly 16.8x.
(b) Exclusion restriction threats: settler mortality may proxy for current disease environment (directly reducing productivity); Europeans may have invested differently in infrastructure beyond institutions. (c) IV > OLS likely due to attenuation bias: if the reliability ratio is ~0.55, then \$1.52/0.55 \approx 0.94$.
20.5 Human Capital and Health
The Mincer Equation
Mincer equation.A regression of log wages on years of schooling, experience, and experience squared: $\ln w_i = \alpha + \rho S_i + \beta_1 \text{Exp}_i + \beta_2 \text{Exp}_i^2 + u_i$. The coefficient $\rho$ is the return to an additional year of education.
Returns to education.The percentage increase in wages from one additional year of schooling. Typical estimates: 10–14% in low-income countries, 5–7% in high-income countries, reflecting scarcity of educated workers in developing economies.
Returns to Education Across Development Levels
Income Group
Average Return ($\hat{\rho}$)
Low-income countries
10.5%
Lower-middle-income
8.7%
Upper-middle-income
7.2%
High-income countries
5.4%
Health as Human Capital
Health as human capital.Physical health — freedom from disease, adequate nutrition, cognitive development — is a form of human capital affecting productivity and earnings. Health investments (clean water, vaccination, deworming) have returns comparable to education investments.
$$Y = A(H) \cdot K^\alpha \cdot (h \cdot L)^{1-\alpha}, \quad h = e^{\phi S + \psi \text{Health}} \tag{Eq. 20.8}$$(Eq. 20.8)
Intuition
Why it matters: Each extra year of school raises a worker’s pay by a roughly constant percentage — and that percentage is larger where educated workers are scarce. So the return is around 10–14% in poor countries and only 5–7% in rich ones, simply because scarcity commands a premium. Health is human capital in the same way: a child who is dewormed, fed, and free of chronic disease learns more in school and earns more as an adult, with returns that rival schooling — which is why a few dollars of deworming can be one of the most cost-effective things a development budget buys. The figure below lets you slide schooling years and the return rate to trace out the wage profile.
Empirical Evidence on Health and Development
Bleakley (2007) exploited geographic variation in hookworm prevalence to show a 17% income increase per SD reduction. Miguel & Kremer (2004) found deworming reduced school absenteeism by 25% with large spillovers — approximately \$3.50 per additional year of attendance, among the most cost-effective development interventions known.
Figure 20.5. Mincer equation explorer. Adjust schooling years and returns to see how the log-wage profile shifts. The dashed line shows the premium from 4 additional years. Drag the sliders to explore.
Example 20.4 — Mincer Regression
Country A (low-income): $\hat{\rho} = 0.10$, $\hat{\beta}_1 = 0.03$, $\hat{\beta}_2 = -0.0005$. Country B (high-income): $\hat{\rho} = 0.05$, $\hat{\beta}_1 = 0.05$, $\hat{\beta}_2 = -0.0008$. The education premium for 4 extra years: Country A = $e^{0.40}-1 = 49.2\%$; Country B = $e^{0.20}-1 = 22.1\%$.
Wage peaks at $\text{Exp}^* = \beta_1 / (2|\beta_2|)$: Country A at 30 years, Country B at 31.25 years. Returns differ due to scarcity, ability bias, credit constraints, school quality, and signaling vs. human capital effects.
20.6 The RCT Revolution
Background
Randomized controlled trial (RCT).An experimental design in which units are randomly assigned to treatment and control groups. Randomization ensures the groups are identical in expectation on all characteristics, so any outcome difference is causally attributable to the treatment.
Banerjee, Duflo, and Kremer received the 2019 Nobel Prize for their experimental approach to alleviating global poverty. Key findings: cash transfers work and do not reduce effort; microfinance is not transformative; deworming is extraordinarily cost-effective. The RCT revolution's greatest contribution was replacing prior beliefs with evidence.
The ATE Estimator
Average treatment effect (ATE).The expected difference in outcomes between treatment and control groups, estimated as the simple difference in sample means under random assignment. No regression adjustment is needed for unbiasedness.
Intent-to-treat (ITT).The causal effect of being assigned to treatment, regardless of whether treatment was received. Always identified by randomization; the most policy-relevant estimand since governments cannot force participation.
Local average treatment effect (LATE).The causal effect of receiving treatment for compliers. LATE = ITT / compliance rate. This is the IV estimand using assignment as an instrument for receipt (Ch 10).
Power Calculations
Statistical power.The probability of detecting an effect when it truly exists. Conventionally 80%. Underpowered studies are unlikely to detect real effects and contribute to the file drawer problem.
Why it matters: Flip a coin to decide who gets a program and who doesn’t, and the two groups end up identical in expectation on everything — rich and poor, motivated and not. So any difference you see afterward must be the program’s doing; you don’t need a model of human behavior to believe it. That credibility is what won Banerjee, Duflo and Kremer the 2019 Nobel. The catch is sample size: with too few people, a real effect hides inside ordinary noise. The power formula just turns that worry into a number — how many people (or villages) you must enroll to be confident of spotting an effect of a given size. The figure below lets you dial the effect size, variability, and cluster design and watch the required sample swing.
Key RCT Results
Intervention
Finding
Study
Deworming
25% reduction in absenteeism; large spillovers
Miguel & Kremer (2004)
Bed nets
Free distribution yields much higher adoption than cost-sharing
Cohen & Dupas (2010)
Microfinance
Modest effects on business income; no transformative poverty reduction
Banerjee et al. (2015)
Cash transfers (UCT)
Recipients invest productively; effects persist
GiveDirectly (Haushofer & Shapiro 2016)
Cash transfers (CCT, Progresa)
+8pp school enrollment, improved nutrition
Schultz (2004)
Teacher incentives
Incentive pay raises test scores; design details matter
Muralidharan & Sundararaman (2011)
Figure 20.6. RCT power calculator. See how effect size, variance, significance level, and clustering affect the required sample size. The dashed line marks 80% power. Drag the sliders to explore.
Example 20.5 — RCT Power Calculation
Kaelani's ministry expects a \$30/month income effect ($\sigma = 120$). At $\alpha = 0.05$, 80% power: $N = 2 \times 120^2 \times (1.96+0.84)^2 / 30^2 \approx 251$ per arm. With cluster randomization (42 villages, 60 households each, ICC = 0.04): design effect = 3.36, effective sample = 744 — well above 251.
If budget allows only 1,500 per arm: effective sample $\approx 446$. MDE $= \sqrt{2 \times 14400 \times 7.84 / 446} \approx$ \$22.50/month — smaller than the expected \$30 effect, so the study remains adequately powered.
Take
'Aid is not just ineffective — it's actively destructive' — Dambisa Moyo, Dead Aid (2009)
Zambian economist Dambisa Moyo's Dead Aid and her TED talk made the incendiary case: over \$1 trillion in aid to Africa hadn't just failed — it had "created dependency, fueled corruption, and killed African entrepreneurship." Bill Gates publicly called the book "evil." Jeffrey Sachs accused Moyo of advocating policies that would "lead to the deaths of millions." Moyo fired back that Sachs's own Millennium Villages Project was the real failure. The debate went nuclear. But who was actually right about the evidence?
Advanced
The popular version
Moyo's TED talk version ("aid is the problem, trade is the solution, end of story") is a mirror image of Sachs's equally crude "if rich countries just gave 0.7% of GDP, we could end poverty." Both are cases of the tradition being weak here, not just the person. Moyo treats all aid as identical: she lumps Cold War bribes to Mobutu together with malaria bed nets, budget support together with GiveDirectly cash transfers. This is like evaluating "medicine" by averaging chemotherapy with homeopathy. But Sachs's tradition has its own problem: decades of aid have not produced the convergence his models predicted, and the Millennium Villages Project produced no detectable effect in rigorous evaluation. Both popular positions reflect genuine weaknesses in their parent intellectual traditions; the smart money is genuinely confused. Gates calling the book "evil" was telling: it was an emotional response to a structural argument he couldn't fully rebut.
The strongest version of Moyo's argument
If Moyo were pushing her argument to its strongest form, she would lean harder on Angus Deaton and William Easterly rather than broad polemic. The structural mechanism is serious: aid undermines the accountability relationship between governments and citizens. When governments are funded by external donors rather than domestic taxation, they answer to donors, not citizens. This breaks the social contract that drives institutional development, the same mechanism that makes resource-cursed petrostates dysfunctional. The countries that developed successfully (Western Europe, East Asia, Botswana) did so primarily through domestic resource mobilization and institutional evolution, not through aid. Easterly extends this: the aid industry creates its own demand through bureaucratic incentives that systematically favor visible, measurable, short-term projects over the institutional development that actually matters. Moyo's strongest data point: Africa received over \$1 trillion in development aid between 1960 and 2010, and the aggregate statistical relationship between aid and growth is indistinguishable from zero. Not slightly positive. Zero. That is a devastating empirical fact that the pro-aid camp has never adequately explained at the macro level.
The strongest case against
The aggregate aid-growth regression that Moyo relies on is meaningless because it treats all aid as identical. Cold War-era strategic aid (US aid to Mobutu's Zaire, Soviet aid to client states) was geopolitical bribery, never intended to promote development. When you disaggregate and look at specific, well-designed interventions, the evidence is strong: insecticide-treated bed nets reduce malaria mortality by 50–70% and cost roughly \$5 per DALY averted (GiveWell estimates). Deworming programs show persistent effects on earnings a decade later (Baird, Hicks, Kremer, Miguel). Oral rehydration therapy has saved tens of millions of children's lives. The RCT revolution (Banerjee, Duflo, Kremer, Nobel 2019) has given us tools to identify which interventions work and which do not. Moyo's "trade not aid" alternative also has problems: trade access requires infrastructure, institutions, and human capital that the poorest countries lack, and building those things often requires external funding. Her prescription for bond markets and FDI as aid replacements ignores that the countries most dependent on aid are precisely the ones that cannot access capital markets.
The judgment
So was Moyo right that aid is "actively destructive"? The strong version of her claim fails. Targeted health and education interventions demonstrably work and should be funded generously: insecticide-treated bed nets, mass deworming, childhood vaccination programs. Calling this "destructive" is empirically wrong. Moyo's structural critique, inherited from Deaton and Easterly, has not been answered by the RCT movement: even successful micro-interventions may perpetuate a dependency structure that prevents the institutional development that is the actual path out of poverty. A country that relies on external bed net distribution may never develop the domestic public health capacity to solve malaria itself. The real answer is the one neither Moyo nor Sachs fully embraces. The aggregate question "does aid work?" is the wrong question, because it combines fundamentally different activities. Targeted health interventions work. Governance and institutional aid mostly doesn't, and may be counterproductive. The best aid is aid that builds domestic capacity to solve problems without aid, which is precisely the kind of aid that is hardest to design, measure, and get funded. Gates was wrong to call the book "evil." Moyo was wrong to call all aid destructive. The truth requires a disaggregation that neither side's soundbites permit.
20.7 External Validity and Structural Estimation
The External Validity Problem
External validity.The extent to which a causal effect estimated in one context applies to other contexts. An RCT may have perfect internal validity but zero external validity if the effect depends on features specific to the study site.
Internal validity.The extent to which the estimated causal effect is unbiased for the study population. Randomization guarantees internal validity. Necessary but not sufficient for policy guidance.
Site selection bias.The tendency for RCTs to be conducted where implementation is easiest and expected effects are largest, creating an upward-biased picture of effectiveness. Allcott (2015) documented declining effects as programs scale to less favorable sites.
Structural Estimation as a Complement
Structural estimation.An approach in which the researcher specifies a theoretical model of behavior, estimates its parameters, and uses the model to simulate counterfactual policies or predict outcomes in new settings. Makes assumptions explicit and enables extrapolation beyond observed data.
Todd and Wolpin (2006) validated a structural model against the Progresa RCT, then used it to simulate untested counterfactuals. Attanasio et al. (2012) showed the CCT worked primarily by reducing opportunity costs of schooling rather than relaxing budget constraints — a mechanism-based understanding that enables transportability.
Neither Approach Dominates
The resolution combines structural and reduced-form approaches. RCTs provide credible causal estimates; structural models provide frameworks for generalization. The ideal workflow: use an RCT to identify parameters, feed them into a structural model, validate against experimental data, then extrapolate with honest uncertainty bounds.
Figure 20.8. Structural vs. reduced-form comparison. The left panel shows the original RCT estimate; the right shows predictions for a new site. As contexts diverge, the structural model adjusts honestly while naive extrapolation stays falsely precise. Use the toggle to switch scenarios.
Example 20.6 — External Validity Analysis
Miguel & Kremer found 25% absenteeism reduction in Kenya; a replication in India found ~3pp (not significant). Key structural differences: helminth prevalence 75% (Kenya) vs. 20–30% (India); different school quality and access; different opportunity costs of child labor; smaller spillover effects.
A structural model of schooling with health inputs, calibrated to Kenya, predicts 7pp. Recalibrated with Indian parameters: 2–3pp — consistent with the replication. The model "knows what it doesn't know": it adjusts predictions and widens confidence intervals rather than falsely extrapolating.
20.8 Contemporary Development
Industrial Policy Revival
Industrial policy.Government intervention to promote specific industries through subsidies, trade protection, or public investment. The new case (Lin 2012, Rodrik) differs from old import substitution: facilitate latent comparative advantage rather than fight it.
The new structural economics (Lin) argues governments should identify industries consistent with latent comparative advantage. Rodrik extends this to green industrial policy: the clean energy transition requires coordinated public investment because carbon externalities are underpriced and learning-by-doing spillovers are not internalized.
Trade and Development
Climate Adaptation and Development
Conditional Cash Transfers
Conditional cash transfer (CCT).A social protection program providing cash to poor households conditional on specific behaviors — typically children's school attendance and health visits. Operating in 60+ countries, CCTs consistently boost enrollment (5–10pp), nutrition (0.2–0.5 SD), and health utilization.
The debate between conditional and unconditional cash transfers (UCTs) is central to contemporary policy. GiveDirectly's programs show UCTs work well — recipients invest productively and effects persist. Conditionality may matter when behavioral biases prevent optimal investment (connecting to Ch 19), but may be unnecessary when households already want to invest in children's human capital.
Figure 20.7. Cash transfer RCT simulator. Adjust transfer amount, duration, and conditionality to see how treatment effects vary across outcomes. Significance stars appear when the CI excludes zero. Drag the sliders to explore.
Historical Lens
The colonial era (pre-1945) created the institutional foundations. The post-independence era (1945–1980) was dominated by big push thinking. The Washington Consensus (1980–2000) promoted markets. The RCT revolution (2000–2019) shifted focus to micro-level evidence. The post-2015 era synthesizes: big questions need structural thinking; specific policy questions need experimental evidence.
The Kaelani Republic — CCT Evaluation
Kaelani implements a CCT: \$50/month to 2,500 randomly selected rural households, conditional on 80%+ school attendance, for 18 months. Control group: 2,500 households. Power calculation (Eq. 20.10): with $\sigma = 120$ and effective sample = 744 per arm (after cluster adjustment), the MDE is \$17/month at 80% power. The expected \$30–35 effect is well above this threshold.
Results after 18 months: Monthly consumption +\$32 (p < 0.01), school enrollment +8pp (p = 0.01), dietary diversity +0.4 SD (p < 0.01), savings +\$15 (p = 0.02), adult labor supply −2 hrs/wk (p = 0.27, not significant). Compliance 94%; labor supply concern dismissed. The \$50 transfer generates \$32 in consumption gains, suggesting local spending multipliers.
Institutional analysis (Ch 18): The CCT builds state capacity — payment systems, monitoring infrastructure, bureaucratic accountability. The school attendance condition works because Kaelani invested in school construction during its 2005 reform. Without schools, conditionality is meaningless.
External validity (Sec 20.7): The Talani Republic wants to replicate. Reduced-form: naive extrapolation ignores Talani's weaker institutions and different demographics. Structural model: predicts +5pp enrollment (vs. Kaelani's +8pp) and +\$28 consumption (vs. \$32), with 90% interval [+1pp, +9pp] for enrollment. The Deaton critique applies: RCTs answer "did it work here?" but not "will it work there?"
The textbook's threads converge: Kaelani's development depends on institutions (Ch 18), growth fundamentals (Ch 13), macroeconomic stability (Chs 14–16), behavioral insights (Ch 19), and evidence-based evaluation (this chapter).
Summary
The central fact of development is the 100x income gap between richest and poorest countries. The global distribution evolved from unimodal (1800) to twin peaks (1970s–1980s) to partial convergence (2000s).
Structural transformation — agriculture to manufacturing to services — is the defining feature of development. The Lewis model formalizes labor reallocation from surplus subsistence to the modern sector.
The Lewis turning point marks the exhaustion of surplus labor and the transition from labor-reallocation growth to productivity-driven growth.
Poverty traps arise from S-shaped production functions and coordination failures. The MSV model shows demand spillovers creating multiple equilibria, with a big push required to escape.
Institutions are the fundamental cause (AJR). The settler mortality IV identifies the causal effect; the slave trade (Nunn) provides complementary evidence. IV > OLS due to attenuation bias.
Human capital and health are central drivers. Mincer returns are higher in developing countries (10–14% vs. 5–7%). Health interventions (deworming) are extraordinarily cost-effective.
The RCT revolution provided credible causal evidence: cash transfers work; microfinance is not transformative; deworming is cost-effective. Power calculations discipline study design.
External validity is the central limitation. Structural estimation complements RCTs by providing mechanism-based extrapolation with honest uncertainty.
Contemporary development features industrial policy revival, trade through GVCs, climate adaptation challenges, and scaling of CCTs. The best policy combines experimental evidence with structural understanding.
Key Equations
Label
Equation
Description
Eq. 20.1
$Y_M = A_M K_M^\alpha L_M^{1-\alpha}$
Modern sector Cobb-Douglas production
Eq. 20.2
$Y_S = A_S \min(L_S, \bar{L})$
Subsistence sector with surplus labor
Eq. 20.3
Lewis turning point: $MPL_S = \bar{w} \Rightarrow L_S^* = \bar{L}$
$Y = A(H)K^\alpha(hL)^{1-\alpha}$, $h = e^{\phi S + \psi\text{Health}}$
Augmented production (health + education)
Eq. 20.9
$\hat{\tau}_{ATE} = \bar{Y}_T - \bar{Y}_C$
ATE estimator under randomization
Eq. 20.10
$N = 2\sigma^2(z_{\alpha/2}+z_\beta)^2 / \tau^2$
Minimum sample size for power \$1-\beta$
Practice
An economy has 7M workers: 5M in subsistence (surplus = 2M, $\bar{L} = 3$M) and 2M in modern. Modern: $\alpha = 0.3$, $K = 50$, $A_M = 1$. (a) Compute current modern output. (b) Reallocate 1M surplus workers; compute new output and gain. (c) Is there any subsistence loss?
Find steady states of $\dot{k} = 0.15 \cdot k^{1.5}/(1+k^{1.5}) - 0.08k$ numerically. Classify each as stable or unstable. What is the minimum big push?
An RCT has 8 treatment schools (scores: 5.2, 3.8, 6.1, 4.5, 7.0, 3.2, 5.5, 4.7) and 8 control schools (2.1, 3.5, 1.8, 2.9, 4.0, 1.5, 3.3, 2.7). (a) Compute ATE. (b) Compute pooled SE. (c) Test at 5% level.
Mincer equation with $\rho = 0.08$, $\beta_1 = 0.04$, $\beta_2 = -0.0006$. (a) Compute log-wage for Worker A ($S=16$, Exp=10) and B ($S=12$, Exp=14). (b) Who earns more? Decompose. (c) Peak experience?
Compare CCTs vs. UCTs for school enrollment, nutrition, and labor supply. When does conditionality matter? What role do behavioral biases (Ch 19) play?
Design an RCT for a school feeding program. Choose individual vs. cluster randomization. Compute sample size ($\sigma = 0.8$ SD, $\tau = 0.15$ SD, $\alpha = 0.05$, 80% power). If cluster-randomized (30 students/school, ICC = 0.10), how many schools per arm? Two internal validity threats.
Nunn's slave trade instrument: state relevance and exclusion restriction. Compare to AJR. Could both be used simultaneously (overidentified IV)? What test applies?
Challenge
Formalize MSV: $N$ sectors, each uses traditional ($y_T = 1$) or modern ($y_M = \alpha > 1$, fixed cost $F$). (a) Profit as function of $n$ industrialized sectors. (b) Show both $n=0$ and $n=N$ can be Nash equilibria. (c) When is the big push welfare-improving?
A structural model finds 60% of apparent returns to education is ability sorting ($\rho_{\text{causal}} = 0.04$, OLS $= 0.10$). An RCT finds 8% per year for scholarship recipients. (a) Reconcile using LATE vs. ATE. (b) For whom does each apply? (c) Which guides national expansion?
Critique "institutions cause growth": (a) When is settler mortality weak? Consequences for 2SLS? (b) If reliability ratio = 0.6, what is OLS bias? Could IV overcorrect? (c) Propose an alternative channel and test.
Climate change reduces tropical agricultural productivity 10–25% by 2050. (a) Analyze a 20% $A_S$ decline using the Lewis model. (b) Distinguish two scenarios for $\bar{L}$. (c) Propose an adaptation strategy using institutional reform, human capital, and CCTs.
Sources
Lewis (1954); Rosenstein-Rodan (1943); Murphy, Shleifer & Vishny (1989); Acemoglu, Johnson & Robinson (2001); Nunn (2008); Mincer (1974); Bleakley (2007); Miguel & Kremer (2004); Banerjee, Duflo & Kremer (Nobel 2019); Todd & Wolpin (2006); Attanasio, Meghir & Santiago (2012); Deaton (2010); Allcott (2015); Lin (2012); Rodrik (2004).
