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, and political economy that standard models abstract away. It also features the most dramatic empirical revolution in modern economics: the rise of randomized controlled trials as a tool for evaluating interventions.
The richest countries have GDP per capita above \$10,000. The poorest are below \$100. A factor of over 100 separates them — and this gap has widened over two centuries. In 1800, the ratio was about 5:1. By 2000, it exceeded 100:1.
As countries develop, agricultural employment falls from 50–70% to below 5%, driven by Engel's law and rising agricultural productivity.
As the modern sector expands, it absorbs surplus labor at a constant wage. Profits are reinvested, driving further expansion. Growth continues until the surplus is exhausted and wages begin rising — the Lewis turning point.
The modern sector expands by accumulating capital and absorbing labor from the subsistence sector. Watch as the MPL in the modern sector declines with each additional worker. At the Lewis turning point, surplus labor is exhausted and wages begin rising.
Figure 18.1. The Lewis model. Left panel: modern sector MPL curve with labor demand. Right: GDP decomposition. As capital rises, the modern sector absorbs more labor, pushing the economy toward the Lewis turning point where all surplus labor is absorbed. Drag the capital slider to simulate industrialization.
Kaelani has 5M workers: 3.5M in subsistence agriculture ($MPL = \\$100$/year) and 1.5M in the modern sector ($Y_{modern} = A K^{0.5}L^{0.5}$ with $A = 100$, $K = 50{,}000$).
Step 1: Current modern-sector output: $Y_m = 100 \times 50{,}000^{0.5} \times 1{,}500{,}000^{0.5} = 100 \times 223.6 \times 1224.7 = \\$17.4B$.
Step 2: Move 500,000 workers from subsistence to modern sector ($L_m = 2M$): $Y_m' = 100 \times 223.6 \times 1414.2 = \\$11.6B$. Gain in modern output: $\\$1.2B$.
Step 3: Loss in subsistence output: \$100{,}000 \times \\$100 = \\$150M$. But if these workers had $MPL \approx 0$ (surplus labor), the actual loss is near zero.
Step 4: Net GDP gain: $\\$1.2B - \\$1.25B = \\$1.95B$, a 14% increase in GDP from pure labor reallocation — no new investment required.
Key insight: With surplus labor, the Lewis model predicts "free" growth from structural transformation. This is the mechanism behind China's 10% annual growth during 1980–2010.
With an S-shaped production function $f(k)$, the Solow equation $\dot{k} = sf(k) - (n+\delta)k$ has three intersections: $k_L^*$ (low steady state), $k_U$ (unstable threshold), and $k_H^*$ (high steady state). The big push is the investment needed to cross $k_U$.
Drag the initial capital level to see where the economy converges. Below the unstable threshold $k_U$, it falls back to the poverty trap. Above $k_U$, it converges to the high steady state. The "big push" is the investment needed to cross $k_U$.
Figure 18.2. Poverty trap with S-shaped production function. The green dots are stable steady states; the red dot is the unstable threshold. Starting below $k_U$, the economy falls back to $k_L^*$. Starting above it, the economy reaches $k_H^*$. The "big push" arrow shows the investment jump required to escape the trap. Drag the slider to change initial capital.
Consider an S-shaped production function: $f(k) = k^{0.3}$ for $k < 4$ and $f(k) = 0.5(k-2)^{0.6} + 1.5$ for $k \geq 4$. Saving rate $s = 0.20$, depreciation $(n+\delta) = 0.05$.
Step 1: Find where $sf(k) = (n+\delta)k$, i.e., \$1.2f(k) = 0.05k$, or $f(k) = 0.25k$.
Step 2: Low steady state ($k < 4$): $k^{0.3} = 0.25k$, so $k^{-0.7} = 0.25$. $k_L^* = 0.25^{-1/0.7} = 0.25^{-1.43} = 7.1$. But this exceeds 4, so the low branch gives $k_L^* \approx 1.5$ (by numerical root-finding).
Step 3: High steady state ($k \geq 4$): the S-shape creates a second crossing at $k_H^* \approx 12$.
Step 4: Unstable threshold: $k_U \approx 5$ (where $sf(k)$ crosses $(n+\delta)k$ from above). Below $k_U$, the economy converges to $k_L^*$ (poverty trap). Above $k_U$, it converges to $k_H^*$ (development).
Step 5 (Big push): An economy at $k_L^* = 1.5$ needs investment of $\Delta k = k_U - k_L^* = 5 - 1.5 = 3.5$ units of capital per worker to escape the trap. This must be delivered as a coordinated lump-sum — gradual investment is absorbed by the trap's gravitational pull.
Geography (Sachs, 2001): Tropical climates cause disease, reduce agricultural productivity, and create trade barriers. Strong latitude-income correlation.
Institutions (Acemoglu, Johnson & Robinson, 2001): Property rights, rule of law, and checks on power are the fundamental cause. The AJR IV strategy (settler mortality → institutional type → income) shows causal impact.
Culture (Landes, 1998): Values like trust, work ethic, and attitudes toward education shape behavior. Hard to test rigorously since culture is endogenous.
The emerging consensus is interactionist: institutions are the proximate cause, geography shapes institutions historically, and culture shapes informal institutions. All three interact in feedback loops.
Toggle the x-axis variable to explore the development debate visually. Each point is a country. How does the relationship change when you switch from institutional quality to latitude to settler mortality?
Figure 18.5. The institutions-vs-geography-vs-culture debate, visualized. Toggle the x-axis to see how different fundamental causes correlate with income. Institutional quality shows the strongest relationship. Hover over points for country names. Click buttons to switch variables.
Banerjee, Duflo, and Kremer received the 2019 Nobel Prize for "their experimental approach to alleviating global poverty." The RCT brings the randomized controlled trial from medicine to development economics.
Power calculation — minimum sample size to detect an effect of size $\tau$:
Statistical power is the probability of detecting a true effect. Underpowered studies miss real effects; overpowered ones waste resources. Adjust sample size, effect size, and variance to see how power responds.
Figure 18.3. Power curve: probability of detecting the effect as a function of sample size. The dashed line marks the conventional 80% threshold. The red dot shows your current design. Increasing sample size or effect size raises power; increasing variance reduces it. Drag sliders to design your study.
An RCT gives \$10/month to 2,500 randomly selected households for 12 months. Control: 2,500 households. Results after 12 months:
| Outcome | Control mean | Treatment mean | Difference | SE | p-value |
|---|---|---|---|---|---|
| Monthly income ($) | 120 | 148 | +28 | 1.27 | <0.001 |
| School enrollment (%) | 62 | 70 | +8pp | 0.51 | <0.001 |
| Meals per day | 2.1 | 2.5 | +0.4 | 0.017 | <0.001 |
| Business assets (%) | 15 | 26 | +11pp | 0.34 | <0.001 |
ITT vs TOT: Compliance was 94% (47 of 50 assigned households actually received transfers). $TOT = ITT / 0.94$. For income: $TOT = 28/0.94 = \\$19.8$/month. With high compliance, ITT and TOT are similar.
Practical significance: The \$18 income gain exceeds the \$10 transfer? No — the gain is in total household income, which includes both the transfer and any additional earnings from investment of the transfer (e.g., buying inventory for a small business). The marginal propensity to earn from the transfer is $(28-50 \times 0.94)/120 \approx -0.16$, meaning households save and invest some of the transfer rather than consuming it all.
You want to detect a $\tau = 0.20$ standard deviation effect on household income, with significance $\alpha = 0.05$ and power \$1-\beta = 0.80$.
Step 1: Formula: $n = \frac{(z_{\alpha/2} + z_\beta)^2 \cdot 2\sigma^2}{\tau^2}$.
Step 2: Plug in: $z_{0.025} = 1.96$, $z_{0.20} = 0.84$. With standardized outcomes ($\sigma = 1$): $n = \frac{(1.96 + 0.84)^2 \times 2 \times 1}{0.20^2} = \frac{7.84 \times 2}{0.04} = \frac{15.68}{0.04} = 392$ per group.
Step 3: Total sample: \$1 \times 392 = 784$ households. With 10% expected attrition: recruit \$184/0.9 = 871$ total.
Step 4 (sensitivity): If the true effect is only $\tau = 0.10$ SD (half as large), the required sample quadruples: $n = 392 \times 4 = 1,568$ per group. Small effects require large samples — this is why many development RCTs enroll thousands of participants.
Step 5 (cost): With per-household cost of \$100/year (transfer) + \$100 (survey): total budget = \$171 \times 700 = \\$109,700$. The cost of answering "does this work?" is itself a significant development expenditure.
| Intervention | Finding | Study |
|---|---|---|
| Deworming | Massive effects on school attendance; spillovers | Miguel & Kremer (2004) |
| Free bed nets | Free > subsidized for adoption | Cohen & Dupas (2010) |
| Microfinance | Modest business effects; no poverty reduction | Banerjee et al. (2015) |
| Cash transfers | Surprisingly effective; not "wasted" | Haushofer & Shapiro (2016) |
| Conditional transfers | Increased schooling and health visits | Schultz (2004) |
Watch the world income distribution evolve from 1800 to 2020. In 1800, nearly everyone was poor. By 1960, a bimodal "twin peaks" emerged. Since 1990, the middle has filled in as Asia industrialized.
Figure 18.4. Global income distribution over time. Each bar represents the share of world population in an income bracket. The distribution shifts from unimodal (1800) to bimodal (1960) to a long-tailed distribution with a growing middle class (2020). Drag the year slider to watch history unfold.
Angus Deaton (2010) offered the sharpest critique: (1) Context-dependence: results from Kenya may not apply in India. (2) General equilibrium effects: scaling up changes wages, prices, and politics. (3) Site selection bias: RCTs are conducted where they're likely to work. (4) The "gold standard" label is misleading.
The resolution is not RCTs vs theory — it's RCTs and theory. Causal identification tells you what works. Theory tells you why — under what conditions, at what scale, through which mechanisms.
Kaelani runs an RCT on \$10/month cash transfers to 2,500 rural households for 12 months (total: \$1.5M). Results: income +23%, school enrollment +8pp, meals +0.4/day, business assets +11pp. All significant at 5%.
External validity concerns: Scaling to all 5M citizens would cost \$1B/year (30% of GDP). At that scale, general equilibrium effects (inflation, labor supply changes) would emerge. Rural Kaelani has an active informal economy — results may differ in urban settings or arid regions.
Simulate treatment vs. control outcomes for a cash transfer program. Adjust the transfer amount and duration to see how outcomes and cost-effectiveness change. Treatment effects exhibit diminishing returns in transfer size and partial persistence over duration.
Figure 18.A. Simulated RCT outcomes for a cash transfer program (N = 2,500 per arm). Blue bars are control group means; green bars are treatment group means. Error bars show 95% confidence intervals. Stars (*) indicate statistical significance at the 5% level. Treatment effects scale with transfer amount (with diminishing returns) and duration (with partial persistence). Adjust sliders to explore cost-effectiveness tradeoffs.
For: Coordination failures (big push), learning by doing, credit market failures. South Korea's success. Against: Information problems, rent-seeking, survivorship bias. Consensus: provide public goods and correct market failures, but be cautious about "picking winners."
Developing countries face a double burden: most vulnerable to climate change (agriculture-dependent) and under pressure to limit carbon-intensive growth. Dell, Jones, and Olken (2012) estimate a 1°C increase reduces GDP growth by 1.3pp in poor countries with no effect in rich ones — climate change could widen the global income gap.
| Label | Equation | Description |
|---|---|---|
| Eq. 18.1 | Agriculture → Manufacturing → Services | Structural transformation |
| Eq. 18.2 | Lewis surplus labor transfer | Dual-economy reallocation |
| Eq. 18.3 | $\dot{k} = sf(k) - (n+\delta)k$ with S-shaped $f$ | Poverty trap model |
| Eq. 18.4 | $\hat{\tau} = E[Y|T=1] - E[Y|T=0]$ | ATE from RCT |
| Eq. 18.5 | $n = (z_{\alpha/2}+z_\beta)^2 \cdot 2\sigma^2/\tau^2$ | Power calculation |