Development Economics Guide

Fundamental question: What is the causal effect of a program/policy?

Challenge: We observe outcomes for treated individuals, but we cannot
observe what would have happened to them without treatment
(the counterfactual).

Solutions (from strongest to weakest causal identification):
  1. Randomized Controlled Trials (RCTs / field experiments)
  2. Regression Discontinuity Design (RDD)
  3. Instrumental Variables (IV)
  4. Difference-in-Differences (DiD)
  5. Matching / Propensity Score Methods
  6. Cross-sectional regression with controls (weakest)

def design_field_experiment(intervention: str,
                             unit: str,
                             clusters: int,
                             expected_effect: float) -> dict:
    """
    Design a cluster-randomized field experiment.

    Args:
        intervention: Description of the program/policy
        unit: Unit of randomization (individual, household, village, school)
        clusters: Number of clusters available
        expected_effect: Expected effect size (standard deviations)
    """
    return {
        "intervention": intervention,
        "randomization_unit": unit,
        "design_considerations": {
            "cluster_vs_individual": (
                "Cluster randomization when intervention operates at group level "
                "or to avoid spillovers between treated and control within clusters."
            ),
            "stratification": (
                "Stratify randomization by baseline covariates (e.g., region, "
                "baseline outcome) to improve balance and statistical power."
            ),
            "sample_size": {
                "clusters": clusters,
                "note": (
                    "With cluster randomization, power depends more on number "
                    "of clusters than individuals per cluster. Aim for 20+ "
                    "clusters per arm. Account for ICC (intracluster correlation)."
                )
            },
            "expected_effect": expected_effect,
            "pre_registration": "Register at AEA RCT Registry (socialscienceregistry.org)"
        },
        "threats": [
            "Attrition (differential dropout between arms)",
            "Non-compliance (some treated do not take up, some controls do)",
            "Spillovers (treatment affects control units)",
            "Hawthorne effects (behavior changes from being observed)",
            "Ethical concerns (withholding a beneficial intervention)"
        ]
    }

Setup:
  Treatment group and control group
  Observed before and after the intervention

Estimator:
  DiD = (Y_treat_after - Y_treat_before) - (Y_control_after - Y_control_before)

Key assumption: Parallel trends
  In the absence of treatment, treatment and control groups would have
  followed the same trajectory over time.

Validation:
  - Plot pre-treatment trends for both groups
  - Test for pre-treatment differences in trends
  - Consider event-study specification with leads and lags

import pandas as pd


def estimate_did(df: pd.DataFrame, outcome: str,
                 treatment_col: str, post_col: str) -> dict:
    """
    Estimate a Difference-in-Differences model.

    Args:
        df: Panel DataFrame
        outcome: Outcome variable name
        treatment_col: Binary treatment indicator
        post_col: Binary post-period indicator
    """
    from statsmodels.formula.api import ols

    df["treat_post"] = df[treatment_col] * df[post_col]

    model = ols(
        f"{outcome} ~ {treatment_col} + {post_col} + treat_post",
        data=df
    ).fit(cov_type="cluster", cov_kwds={"groups": df["cluster_id"]})

    return {
        "did_estimate": model.params["treat_post"],
        "std_error": model.bse["treat_post"],
        "p_value": model.pvalues["treat_post"],
        "ci_95": model.conf_int().loc["treat_post"].tolist(),
        "note": "Standard errors clustered at the cluster level"
    }

Poverty measurement:
  - Consumption vs. income (consumption preferred in developing countries)
  - Purchasing power parity (PPP) adjustments
  - Poverty line selection ($2.15/day international line)

Survey design:
  - Sampling frame may miss mobile/nomadic populations
  - Recall period affects consumption estimates
  - Sensitive questions (income, violence) require careful design
  - Translation and cultural adaptation of instruments

Administrative data:
  - Often incomplete or of variable quality
  - Can complement survey data for larger populations
  - Satellite imagery increasingly used as proxy (nighttime lights, rooftop material)