An Empirically Driven Theory of Poverty Reduction
Download as PPTX, PDF0 likes672 views
The document outlines a project that aims to develop an empirically driven theory of poverty reduction using data from unconditional cash transfer impact evaluations in four sub-Saharan African countries. It employs machine learning tools to identify household characteristics and actions that lead to significant consumption gains among high-performing households, termed 'high flyers.' The findings aim to inform cash transfer policies and improve the effectiveness of poverty alleviation programs by understanding key factors for sustained poverty reduction.
![Example of impacts in Malawi (left) and Zambia (right)
Total consumption pc [24m]
[36m]
Food security scale (HFIAS) [24m]
[36m]
Overall asset index [24m]
[36m]
Relative poverty index [24m]
[36m]
Incomes & Revenues index (SD) [24m]
[36m]
Finance & Debt index (SD) [24m]
[36m]
Material needs index (5-17)[24m]
[36m]
Schooling index (11-17) [24m]
[36m]
Anthropometric index (0-59m) [24m]
[36m]
-.2 0 .2 .4 .6 .8
Effect size in SDs of the control group
Endlines 1 & 2 (24 & 36-months) at a glance
Intent-to-Treat effects (CGP) - indices](https://image.slidesharecdn.com/oxfordpresentationcedilpovertytheoryl-200226133735/85/An-Empirically-Driven-Theory-of-Poverty-Reduction-9-320.jpg)
An Empirically Driven Theory of Poverty Reduction
- 1. L.191- An Empirically Driven Theory of Poverty Reduction Clement Adamba – University of Ghana Sudhanshu Handa – University of North Carolina (presenting) Peter Mvula – University of Malawi Gelson Tembo-University of Zambia
- 2. Project Overview • Use data from four unconditional cash transfer (UCT) impact evaluations to build a theory of poverty reduction • Use Machine Learning (ML) tools to identify initial household and contextual characteristics that predict heterogeneous treatment effects (‘high flyers’) • Use ML tools to then identify what actions these high flyers took to realize large consumption gains • Based on initial characteristics and subsequent behaviors across these four countries, develop a theory of sustained poverty reduction or ‘graduation’
- 3. Transfer Project https://transfer.cpc.unc.edu/ Created 2009 as an Institutional Partnership between FAO, UNICEF, University of North Carolina at Chapel Hill Works with sub-Saharan African governments and research partners to: 1. Provide evidence on the effectiveness of cash transfers in achieving impacts for children and households 2. Inform the development and design of cash transfer policy and programs 3. Promote learning across the continent on the design and implementation of cash transfer evaluations and research
- 4. Transfer Project affiliated evaluations to date Country/Program IE Design Survey years Ethiopia Tigray (Bolsa) RDD 2012, 2014 Ethiopia Tigray II RDD 2016, 2018, 2020 Ghana LEAP Longitudinal PSM 2010, 2012, 2016 Ghana LEAP Phase 2 RDD 2017, 2019 Ghana LEAP 1000 RDD 2015, 2017 Kenya CT-OVC RCT 2007, 2009, 2011 Lesotho CGP RCT 2011, 2013 Malawi SCTP RCT 2013, 2014, 2015 South Africa CSG PSM 2010 Tanzania PSSN RCT 2015, 2017, 2020 Zambia CGP RCT 2010, 2012, 2013, 2014, 2017 Zambia MCP RCT 2011, 2013, 2014 Zimbabwe HSCT Longitudinal Ward Matching 2013, 2014, 2017
- 5. Sample sizes and survey years (first year is pre-treatment) Treatment Control Survey years Ghana LEAP 1,262 1,235 2015, 2017 Malawi SCTP 1,730 1,800 2013, 2014, 2015 Zambia CGP 1,260 1,252 2010, 2012, 2013, 2014, 2017 Zimbabwe HSCT 1,029 1,034 2013, 2014, 2017 Total 5,281 5,321
- 6. Key program features of these UCTs • Like all TP work, these are government programs implemented by Ministry of Social Welfare or equivalent • Benefit is paid bimonthly in cash • Value ranges from 15 (Ghana) to 25 (Zambia) percent of baseline consumption • Ranges from US$10-25 per family per month • Households are ultra-poor, but demographic composition varies across programs • All programs generated strong impacts on consumption or food security. The larger the transfer size, the larger the transformative impacts of the program
- 7. 7
- 8. Though all households are ultra-poor, variation in demographic composition may yield interesting results 0.02.04.06.08.1 Density 0 20 40 60 80 100 Age in years 0 .02.04.06 Density 0 20 40 60 80 100 Age in years SCT Eligibles Zambia GCP Malawi SCT
- 9. Example of impacts in Malawi (left) and Zambia (right) Total consumption pc [24m] [36m] Food security scale (HFIAS) [24m] [36m] Overall asset index [24m] [36m] Relative poverty index [24m] [36m] Incomes & Revenues index (SD) [24m] [36m] Finance & Debt index (SD) [24m] [36m] Material needs index (5-17)[24m] [36m] Schooling index (11-17) [24m] [36m] Anthropometric index (0-59m) [24m] [36m] -.2 0 .2 .4 .6 .8 Effect size in SDs of the control group Endlines 1 & 2 (24 & 36-months) at a glance Intent-to-Treat effects (CGP) - indices
- 10. Step One: Who is a high-flyer?
- 11. Livelihood diversification? Zambia: Top consumption decile at 84m
- 12. High flyer: Consumption growth higher than in C group (23% of sample). Tree based on baseline features High flyers: 29%Low flyers: 13%
- 13. Apply latest developments in ML and causal inference: Athey & Imbens PNAS 2016 • Regression classification tree (CARTs) based on treatment effect heterogeneity • Leafs are identified not based on actual outcome, but on actual outcome relative to potential outcome • Why use CART in the first place? • We have 100+ potential pre-treatment variables that could predict high flyers • We might miss something, or fail to identify interactions • ML built for exactly this scenario, let the data tell us what characteristics and combinations are linked to treatment heterogeneity • Domains include: demographics, financial, productive, human capital, psychological, contextual factors like land-use, soil fertility, temperature, rainfall, climate, infrastructure
- 14. Step 2: After we identify the high flyers, want to understand what they did to become high flyers • Here we use post-treatment choices made by households • Types of livestock, types of crop production, non-farm enterprise, migration • These can mix with contextual factors and psychological traits • Again, many, many potential combinations, let the data decide • Apply K-means clustering to the data • ML approach that looks for patterns in the data, puts ‘similar’ units together • Unsupervised: There is no outcome • Clusters are built by minimizing a distance function • Do this with high flyers • Can also do this with low flyers, do we get different behavioral choices? • This is a good check on the theory
- 15. Step 3: Build the theory! • If everything goes according to plan… • We will identify households that are high flyers, converted a small cash transfer into large consumption gains • Will be able to describe their characteristics, e.g. more able-bodied members, better educated, live close to roads or markets • Will characterize what they did with the cash transfer to attain large consumption gains • Move into cash cropping, diversify livelihoods, send out migrants • Will contrast these behaviors with those of low flyers • Use this evidence to characterize how households can ‘graduate’ out of poverty
- 16. Policy implications • Cash transfer programs throughout sub-Saharan Africa working on ‘cash+’ approaches to strengthen economic effects of programs • Whom to target? What types of complementary interventions or ‘+’ services to provide? • This study can identify households most likely (and least likely) to realize large consumption gains • For low flyer, what do they need to help strengthen their livelihoods. What can we learn from the high flyers?
- 17. Engagement Plan • All four original impact evaluations were commissioned by government • Study team is linked to policy process via UNICEF country offices and direct relationships with implementing agencies during original evaluations • Envisage disseminating results directly to implementing agencies ine ach country via national PIs. • Dedicated session at Transfer Project Research Workshop in 2021 or 2022 • Typical workshop attracts ~130 participants from 25+ SSA countries, primarily government staff
Editor's Notes
- #5: 13 rigorous large scale IEs over 13 countries Updated November 2015