The Case: Who Should Get the Subsidy?

The thread that runs through every session Data: CASEN 2022 · Chile

Every session in this course works on the same problem and the same data. Each new method answers the next question that problem raises, and nothing gets used before it has been taught.

The problem

Governments in emerging economies spend a large share of their social budget on programmes targeted at poor households. To target, you need to know who is poor, and that usually means income. But income is hard to observe: much of it is informal, people have reasons to under-report it, and verifying it household by household is expensive.

The standard answer is a proxy means test: predict a household's welfare from characteristics that are easy to observe and hard to fake, such as schooling, household size, type of dwelling or health insurance. Chile's Registro Social de Hogares, Mexico's Progresa/Prospera and many World Bank–supported programmes use versions of this idea.

That is a machine learning problem, with all of its difficulties:

Question	Where we meet it
What kind of data do we have about a household?	Session 1 · Data types
How do we build one clean row per household?	Session 2 · Data manipulation
Can we predict income from what we observe? How wrong are we?	Session 3 · ML models
Is the error we report honest, or did we fool ourselves?	Session 4 · The ML pipeline
Who is poor? Which mistakes matter more?	Session 5 · Classification
Can more flexible models do better, and at what cost?	Session 6 · Trees and ensembles
Which variables do we really need?	Session 7 · Selection and shrinkage
What types of households are there, without labels?	Session 8 · PCA and clustering

The data

CASEN (Encuesta de Caracterización Socioeconómica Nacional) is Chile's main household survey, run by the Ministry of Social Development and Family. The 2022 round interviewed about 72,000 households. We use a random sample of 15,000 households, one row each, built from the record of the head of household.

⬇️ Download casen_households.csv

import pandas as pd
url = "https://raw.githubusercontent.com/hizocar/machine_learning_egei/main/docs/data/casen_households.csv"
df = pd.read_csv(url)

Data dictionary

Column	Type	Description
household_id	id	Course identifier (not the CASEN folio)
region	nominal	Region of residence (16 regions)
area	nominal	urban / rural
head_sex	nominal	Sex of the head of household
head_age	continuous	Age of the head of household
head_schooling_years	discrete	Years of schooling of the head (missing = doesn't know)
head_education	ordinal	Highest level reached, 0_none … 6_tertiary_complete
head_activity	nominal	employed / unemployed / inactive
head_job_type	nominal	Type of job; missing when the head does not work
head_weekly_hours	continuous	Usual weekly hours; missing when the head does not work, and for 87 employed heads who did not know
health_insurance	nominal	fonasa (public), isapre (private), armed_forces, none, other
head_nationality	nominal	chilean / dual / foreign
head_indigenous	nominal	Belongs to an indigenous people (yes / no)
household_size	discrete	Number of people in the household
has_children, has_elderly	binary	Anyone under 18 / over 60 in the household
household_type	nominal	Single person, nuclear, extended, …
overcrowding	ordinal	People per bedroom, 0_none … 3_critical
housing_tenure	nominal	owned / rented / lent / irregular
dwelling_type	nominal	house / apartment / precarious
household_income	continuous	⚠️ Total household income, CLP per month
income_per_capita	continuous	⚠️ Total income per person, CLP per month
income_decile	ordinal	⚠️ National income decile (1 = poorest)
poor	target	1 if the household is in income poverty
survey_weight	weight	Expansion factor (how many households each row represents)

The ⚠️ columns: target in one task, leak in another

Income is the target when we predict income (Sessions 3–4). When we predict poverty (Session 5 onwards), income is off limits: poverty is defined from income per capita, and the whole point of a proxy means test is that income is not observed. A model that uses it scores brilliantly and is useless. Learning to spot this is part of the course.

Two things to know about the numbers

• Poverty measure. poor uses the updated poverty methodology published in the current CASEN 2022 database: 23% of households in our sample and about 20% of people (weighted) are poor. Under the old 2013 methodology the official 2022 figure was 6.5%.
• Weights. CASEN is a complex survey. For official statistics you would use survey_weight. For our prediction exercises we treat each household as one example; when a result is meant to describe Chile as a whole, we say so and weight it.

Source: Ministerio de Desarrollo Social y Familia, Encuesta CASEN 2022, current Stata database (data catalogue). The sample is rebuilt with scripts/build_casen_sample.py in the course repository.