Ultra Poor Graduation Pilot in Philippines

The Graduation pilot combines elements of social protection, livelihood development, financial inclusion, and social empowerment to deliver results by combining support for immediate needs with longer-term human capital and asset investments. The objective is to actively support participants in the short run while promoting sustainable and independent livelihoods in the long run. The components complement each other to help push participants out of poverty.

2019-02-04

2020-09-01

Consumption, food security, income, assets, productive time use, mental health, physical health, financial behavior, social capital, women empowerment

- Household consumption. The primary measure of living standards used in this study, will be consumption based. This will be measured by the following variables:
a. Per capita consumption
b. Food security

- Economic activity of all household members measured by:
a. Monthly income and revenues
b. Time spent on productive activities
c. Total asset holdings and value (including households’ dwellings, livestock and agriculture)

- Financial inclusion (savings, money borrowed and lent)
- Health, measured by:
a. Self-reported mental health
b. Physical health (number of accidents/illnesses among household members).

- Women’s empowerment measured by whether a woman has major say in decisions regarding food

- Social capital, measured by:
a. Trust and social networks among community members
b. Business relations and entrepreneurial activity among community members

The sample for the Ultra Poor Graduation Pilot in the Philippines consists of 2,400 ultra-poor households that were drawn from recipient households of the Philippine Department of Social Welfare and Development’s (DSWD) 4Ps conditional cash transfer program. In each of the 30 barangays, about 600 eligible households surveyed by IPA were randomly assigned to one of three treatment groups or the control group by computer. Randomization was conducted by dividing the barangay into 4 smaller geographic regions and assigning a treatment to
each, simplifying the logistics of implementation and reducing the risk of spillovers as rural barangays often cover extensive areas.

Random assignment was used to avoid selection and program placement bias. In each of the 30 barangays, about 80 eligible households surveyed by IPA were assigned to one of three treatment groups or the control group by computer. Randomization was conducted by dividing the barangay into 4 smaller geographic regions and assigning a treatment to each. This was done to simplify the logistics of implementation. Rural barangays often cover extensive areas. Had such barangays’ households been assigned to treatment groups completely at random, some participants would have had to travel far for trainings, meetings and coaching sessions, which would likely have reduced attendance. The geographical separation of different treatments and the control group also reduces the likelihood of spillovers. Households were randomized in two stages:

1. Each barangay was divided into four clusters based on GPS coordinates collected during surveying. Households were assigned points based on their distance to the nearest cluster's center minus the distance to the farthest and ranked accordingly. Subsequently, each of the four clusters was filled with the highest ranked households until 20 households were placed, at which point that cluster was taken out of the equation and the next cluster filled with the highest-ranking households.

This process was repeated until all 120 clusters of 20 households across the 30 sample barangays were filled.

2. Once all households had been assigned to clusters, randomized assignment to treatment clusters and control clusters began. For treatment groups to be comparable and the control group to be a valid counterfactual, all groups need to have, on average, statistically identical characteristics. The four clusters in all barangays were repeatedly assigned to different treatment groups, or control. Statistical tests were run on each configuration, comparing every treatment to control and the treatments to one another based on four key covariates. Finally, the statistically most balanced configuration was chosen.

Households assigned to treatment groups were eligible to receive treatment as soon as the randomization was complete. There is unlikely to be differential attrition: households are unlikely to drop out of the treatment group, and controls are unlikely to opt in to treatment. Further, IPA and BRAC will mitigate differential attrition by minimizing the link between the research and program implementation. This is done through carefully monitoring implementation to ensure assignment to treatment and control groups is maintained. If there are any crossovers, the original assignment will be used for analysis.

Computer.

Cluster level.

Yes

120 clusters of 20 households each across the 30 sample barangays.

2,400 households

600 households control, 600 individual livelihood and group coaching, 600 group livelihood and group coaching, 600 individual livelihood and individual coaching

Using baseline data, IPA calculates an intra-cluster correlation coefficient (ICC) on key variables of interest (total income, durable asset value, 30-day consumption) ranging from less than 0.001 to 0.0015 within the subdivided barangays. Conservatively using 0.0015, we estimate that we have 80% power to detect a minimum detectable effect size (MDE) of 0.16 standard deviations between any treatment group and the control group when we randomize by sub-barangay clusters, assuming full compliance and no attrition. Power improves for pooled comparisons: testing any group coaching (n = 1,200) vs. any individual coaching (n = 600) or any group livelihood (n = 600) vs. any individual livelihood (n = 1,200) yields estimated MDEs of 0.14 standard deviations. In Banerjee et al. (2016), treatment assignment, covariates, and stratification FE explain between 5% and 60% of total variation, depending on the outcome of interest. In the case of consumption, for example, the previously estimated R2 of 0.46 would provide 80% power to detect an MDE of 0.12 standard deviations between treatment and control, and an MDE of 0.10 standard deviations for the pooled comparisons described above .