Health and Development

Complements or Substitutes?

Carolina Torreblanca

University of Pennsylvania

Global Development: Intermediate Topics in Politics, Policy, and Data

PSCI 3200 - Spring 2026

Agenda

  1. Health and Development
  2. Chen et al. (2013)
  3. Regression Discontinuity Design

Your Research Design is due TONIGHT

Health and Development

The State Provides Public Goods

Last week: the state and security

  • Crime can mobilize citizens (Bateson)
  • But policing can alienate them (Weaver and Lerman)

Today: the state and health

  • Governments build hospitals, subsidize medicine, regulate pollution
  • But development policies can also harm health

We Live Much, Much Longer

Richer Countries Live Longer

The Dark Side

Development has undoubtedly prolonged life: vaccines, sanitation, medicine, nutrition, heating.

But development also brings negative externalities:

  • Air and water pollution
  • Industrial hazards
  • Environmental degradation
  • Inequality in who bears the costs

Today’s paper: a case where a development policy had unintended health consequences

Chen et al. (2013)

The State of the World

Air quality in China is notoriously poor

  • Between 1981–2001, particulate concentrations were more than double China’s national standard and five times the level in the pre-1970 US
  • Premier Zhu Rongji (1999): “If I work in your Beijing, I would shorten my life at least five years”
  • Was he right?

The Huai River Policy

During the 1950–1980 period of central planning, the Chinese government gave free coal winter heating

But due to budgetary constraints, this right was only extended to areas north of the Huai River.

The Huai River Policy

  • North: free coal → more coal burning → more air pollution
  • South: no free coal → less coal burning → less air pollution
  • This created a massive, arbitrary discontinuity in pollution levels

The Huai River

Cities north of the solid line were covered by the heating policy.

The Research Question

Does sustained exposure to air pollution reduce life expectancy?

  • Pollution correlates with everything: income, urbanization, industry, access to healthcare
  • A simple comparison of polluted vs. clean areas confounds pollution with all of these
  • The Huai River policy creates a plausible natural quasi-experiment
  • Why?

The Mechanism

Coal combustion releases Total Suspended Particulates (TSPs) into the air

  • TSPs penetrate the lungs and enter the bloodstream
  • Cause heart disease, stroke, lung cancer, respiratory illness
  • The medical evidence linking particulates to mortality is well established
  • The question is: how large is the effect at population scale?

Data and Measurement

Air pollution: TSP concentrations from 90 cities (1981–2000)

Health: Mortality data from China’s Disease Surveillance Points (DSPs)

  • 145 locations, nationally representative
  • Records all deaths and population counts
  • ~500,000 recorded deaths between 1991 and 2000

**Treatment: north or south of the Huai River

Regression Discontinuity Design

What Is a Regression Discontinuity?

Imagine you want to know: does passing a class improve your career?

  • Someone who scored 0 and someone who scored 100 are very different people
  • Comparing them confounds ability with passing
  • But what about someone who scored 59 vs. someone who scored 61?
  • Almost the same student. A bit of luck, random noise, one extra question
  • The only difference: one passed, the other did not

RDD: The Logic

  • A cutoff assigns treatment: score \(\geq\) 60 → pass
  • Units just below and just above the cutoff are nearly identical
  • So any jump in the outcome at the cutoff is the causal effect of treatment
  • The key assumption: nothing else changes abruptly at the cutoff

This is what Chen et al. exploit: the Huai River is a cutoff that assigns free coal.

Applying the RDD to the Huai River

  • Running variable: degrees of latitude (distance from the river)
  • Cutoff: the Huai River (0 degrees)
  • Treatment: free coal heating (north of the river)
  • A city 1 degree south vs. 1 degree north: similar climate, similar people, but very different pollution
  • Compare health outcomes just above and just below the river

Key concern: is there anything else that jumps at the river?

What Else Could Jump at the River?

  • Education, income, urbanization, tap water access: no discontinuity at the river
  • Temperature and climate: different north vs. south, but change smoothly — no jump
  • The Huai River cuts through provinces — it is not an administrative boundary
  • This is always the hardest part of an RDD: you can check observable things, but you can never be 100% sure

Findings

First Stage: Air Pollution

TSP concentrations jump by ~200 \(\mu g/m^3\) at the Huai River.

The Payoff: Life Expectancy

Life expectancy drops by 5.5 years north of the Huai River.

The Numbers

  • North of the river: TSPs are 184 \(\mu g/m^3\) higher (55% more pollution)
  • Life expectancy is 5.5 years lower (95% CI: 0.8, 10.2)
  • Almost entirely due to cardiorespiratory mortality (heart and lung disease)
  • Noncardiorespiratory mortality: no effect
  • An additional 100 \(\mu g/m^3\) of TSP → 3.0 fewer years of life expectancy at birth

The Scale

  • The population of Northern China between 1990 and 2000 exceeded 500 million
  • The Huai River policy led to a staggering loss of over 2.5 billion life years
  • Current particulate concentrations are 22.9 \(\mu g/m^3\) higher (26% higher north of the river)
  • Residents of the North continue to have shortened lifespans

Complements or Substitutes?

  • The policy was meant to improve welfare: free heating in cold winters
  • But it dramatically shortened lives — a negative externality, an unintended consequence
  • Nobody designed this policy to kill people. But it did.
  • Policy design is hard. Development and health can be complements, but only if you get the details right
  • “These results may help explain why China’s explosive economic growth has led to relatively anemic growth in life expectancy”

RDD in R

Using rdrobust

# install.packages("rdrobust")
library(rdrobust)

# Estimate the discontinuity
rd <- rdrobust(y = data$life_expectancy,
               x = data$latitude_from_river,
               c = 0)
summary(rd)
  • y: outcome variable
  • x: running variable (distance from cutoff)
  • c: the cutoff value
  • rdrobust automatically picks the optimal bandwidth and polynomial

Reading the Output

  • Coefficient: the estimated jump at the cutoff (the treatment effect)
  • Std. Error: precision of the estimate
  • p-value: is the jump statistically significant?
  • Bandwidth: how wide a window around the cutoff is used
  • N left / N right: how many observations on each side

Same interpretation as lm() — but the estimate is local to the cutoff

Simulating an RDD

Show code 
library(ggplot2)
set.seed(42)

n <- 200
latitude <- runif(n, -8, 8)
north <- as.numeric(latitude >= 0)

# True effect: 5.5 fewer years north of the river
life_exp <- 78 - 0.3 * latitude - 5.5 * north + rnorm(n, 0, 3)

sim_data <- data.frame(latitude, north, life_exp)

ggplot(sim_data, aes(x = latitude, y = life_exp, color = factor(north))) +
  geom_point(alpha = 0.6, size = 2) +
  geom_smooth(method = "lm", se = FALSE, linewidth = 1.2) +
  geom_vline(xintercept = 0, linetype = "dashed", linewidth = 1) +
  scale_color_manual(values = c("0" = "#011F5B", "1" = "red"),
                     labels = c("South", "North")) +
  annotate("label", x = 0, y = 85, label = "Huai River",
           size = 5, color = "grey30") +
  labs(x = "Degrees of latitude from Huai River",
       y = "Life expectancy (years)",
       title = "Simulated RDD: life expectancy drops north of the river",
       color = "") +
  theme_minimal(base_size = 16) +
  theme(legend.position = "bottom")

Running rdrobust on Simulated Data

library(rdrobust)

rd <- rdrobust(y = sim_data$life_exp,
               x = sim_data$latitude,
               c = 0)
summary(rd)
Sharp RD estimates using local polynomial regression.

Number of Obs.                  200
BW type                       mserd
Kernel                   Triangular
VCE method                       NN

Number of Obs.                   89          111
Eff. Number of Obs.              23           28
Order est. (p)                    1            1
Order bias  (q)                   2            2
BW est. (h)                   2.057        2.057
BW bias (b)                   3.581        3.581
rho (h/b)                     0.574        0.574
Unique Obs.                      89          111

=============================================================================
        Method     Coef. Std. Err.         z     P>|z|      [ 95% C.I. ]       
=============================================================================
  Conventional    -4.363     2.440    -1.788     0.074    [-9.145 , 0.419]     
        Robust         -         -    -1.188     0.235    [-9.613 , 2.357]     
=============================================================================

The RDD Plot

rdplot(y = sim_data$life_exp,
       x = sim_data$latitude,
       c = 0,
       title = "RDD Plot: Life Expectancy at the Huai River",
       x.label = "Degrees of latitude from Huai River",
       y.label = "Life expectancy (years)")