世界银行-巴基斯坦贫困地图2019-2020（英）_43页_5mb

Pakistan Poverty Map 2019-2020 Summary

Core Content

This paper presents the methodology and results of the Pakistan Poverty Map 2019-2020, which estimates monetary poverty at the district level using small area estimation (SAE) techniques. The study combines data from two household surveys: the Household Income and Expenditure Survey (HIES) 2018-19 and the Pakistan Social and Living Standard Measurement Survey (PSLM) 2019-20. The approach improves upon previous methods by using the Census Empirical Best (Census EB) estimator, which allows for more accurate and precise poverty estimates at the district level.

Main Viewpoints

• Small Area Estimation (SAE) is used to estimate poverty at a finer geographic level than typical household surveys allow.
• The Census EB estimator is applied to model the relationship between household welfare and characteristics, using the HIES as training data and the PSLM as target data.
• The PSLM 2019-20 has a larger sample size and more detailed household characteristics than the HIES, making it suitable for district-level inference.
• The use of two surveys instead of a survey-to-census method provides more common variables for modeling but introduces sampling noise.
• The number of household members is a key variable that affects the accuracy of poverty estimates due to systematic measurement errors in the PSLM compared to the HIES and the 2017 Census.
• The discrepancy in household size between the surveys creates a bias in poverty estimates, which is addressed through two proposed solutions: cross-entropy weight calibration and rescaling of household members to match UN population projections.

Key Information

1. Methodology Overview

• The study uses the nested error model to estimate the relationship between household welfare and characteristics.
• The Census EB estimator is applied to estimate poverty rates for 126 districts in Punjab, Sindh, Khyber Pakhtunkhwa (KP), and Balochistan, including the former Federal Administered Tribal Areas (FATA) and Frontier Regions (FR).
• The linear model includes variables like education, health, housing, and consumption, and the model parameters are estimated using generalized least squares (GLS).
• Monte Carlo simulations are used to estimate the mean and variance of the locality effects, leading to more accurate poverty estimates.
• Parametric bootstraps are employed to estimate the mean squared error (MSE), improving the reliability of the results.

2. Data Sources

• HIES 2018-19 provides detailed consumption data and is representative at the provincial and rural-urban levels.
• PSLM 2019-20 has a larger sample size and is designed for district-level inference, including modules on education, ICT, health, disability, and food insecurity.
• The PSLM sample size was increased significantly to improve the precision of poverty estimates, especially in districts with high variability.
• The PSLM 2019-20 was conducted using computer-assisted personal interviews (CAPI), a new technology that improved data collection.

3. Data Matching and Variable Selection

• Area code matching between HIES and PSLM was necessary for the application of the Census EB method.
• Only Punjab, Sindh, and KP had complete district code matching, while Balochistan had some districts excluded from the PSLM.
• Variable consistency was ensured by selecting 305 variables that are common to both surveys and are relevant for explaining household consumption.
• Variables were filtered based on missing data and statistical consistency across provinces.
• A national model was chosen, with 106 variables, to ensure comparability across regions.

4. Challenges and Solutions

• The discrepancy in household size between HIES and PSLM led to a systematic measurement error, which affected poverty estimates.
• The average household size in PSLM 2019-20 was smaller than in HIES 2018-19, leading to higher per-adult-equivalent consumption and lower poverty rates.
• The error in household size varied across provinces and districts, with the largest differences in Balochistan.
• To correct for this bias, the study proposes:
  • Cross-entropy weight calibration to adjust the expansion factor.
  • Rescaling of household members using the 2017 Census district-level averages to align with United Nations population projections.

5. Results and Implications

• The Census EB method provides more accurate and precise poverty estimates than previous approaches, especially at the district level.
• The PSLM 2019-20 offers a more detailed and representative dataset for district-level poverty mapping.
• The poverty map helps in identifying districts with higher poverty rates, aiding in the targeting of policy interventions.
• The study highlights the importance of data quality and consistency in SAE applications and the need for methodological improvements to address sampling and measurement errors.