Long(er)-term Effects of the Targeting the Ultra Poor Program

Beneficiaries received an asset (e.g., livestock, non-farm microenterprise inventory),
weekly consumption support for 30-40 weeks, savings access, and 18 months of weekly
training visits. No further program contact occurred post-intervention, other than follow-up surveys.

2007-06-01

2009-09-01

The primary outcomes to be analyzed are regarding households’ economic well-being:
– Per capita consumption (total, food, non-food, durable goods)
– Income and revenue (wages, self-employment earnings, remittances)
– Asset ownership (livestock, non-farm assets, durable goods)
– Financial inclusion (savings, loans, credit access)
– Food security (household meals per day, instances of food shortage)
We track TUP program’s effects on adult household members’ physical health, mental
health, productive work, and political involvement.
• Physical and Mental Health:
– Physical health index (self-reported health, workdays missed due to illness, activities
of daily living score)
– Mental health index (life satisfaction, stress levels, sadness)
• Social Outcomes:
– Time spent in productive activities
– Political engagement (voting behavior, community involvement)
– Economic satisfaction (1-10)

We track economic outcomes for all household members and not just the TUP recipient.
These indices are constructed using the same methodology as in the previous follow-up
(Banerjee et al. 2021), which studied the TUP’s impact 10 years post the delivery of
assets. All indices are created by first constructing z-scores (i.e. subtracting the baseline
mean and dividing by the baseline standard deviation) for each variable, averaging over
all variables that comprise the index, and standardizing to the baseline value of the index.
Results are reported in units of baseline standard deviations of the index. One exception is
the income and revenue index, for which we do not have baseline information about some
sub-components; it is therefore standardized to the control mean and results are reported
in units of control group standard deviation.

Further, we identify channels of persistence with the following outcomes:
• Labour Markets and Migration
– Occupational shifts (livestock to microbusiness, wage employment, migration)
– Wage earnings (local vs. migrant earnings, remittance levels)
– Migration patterns (duration, destinations)

Adult job loss and job transitions. This will be separated by migrants and nonmigrants.
– Experienced earnings loss after the job loss
– Switched occupation (from blue-collar to white-collar or vice versa)
– Switched job industry
– Took up government work programs
– Passed away
• Migrants’ job loss and travel back to villages. We will report the share of migrants
who respond ”yes” to the following binary indicators:
– Lost job in migration destination
– Traveled back to home village
• Share of households that report any business loss, non-farm business closure during
COVID lockdowns.
• Share of households that report any sale of land or other assets during lockdown.
• Social protection services during COVID lockdowns. This will be reported as an
index, constructed as the normalized number of transfers that a household received
from:
– extra food ration
– bank deposits as part of existing social welfare schemes
– asset or cash transfers from the local government
– asset or cash transfers from local non-government bodies
• Social protection services in the last 12 months. Takeup will equal 1 if the HH applied
for OR received the scheme in the last 12 months. The index is a normalized count
summing the following indicators:
– work in an employment generating scheme
– old age or widow pension
– Indira Aawas housing plan
– other assets gifted by the Panchayat (village government)
– vocational training through the Panchayat
– Lakshmir Bhandar monthly bank transfer for women
– AAY card
– any ration card (BPL or Annapura rationing)
– Krishak Bandhu scheme
• Healthcare seeking behavior in last month.
• Index for extent that household delayed major events due to COVID. This index is
a normalized count summing the indicators:
– delayed wedding or engagement
– delayed funeral
– delayed opening new business
– delayed taking a large loan
– delayed a child starting or progressing in school
– delayed migration
• Index for household consumption smoothing during lockdown. The index is a normalized
count summing the following indicators:
– consumed goods that you were planning to sell or consume later
– take a loan to buy food or goods that are regularly consumed
– increase time spent foraging
– increase time spent begging
– delayed purchasing essential household items other than food
– sought a loan from anyone
– conditional on having savings, drew down savings to cover expenses

INTERGENERATIONAL OUTCOMES
We are interested in outcomes for beneficiaries’ children, who are now adults. Outcomes in this analysis include:
• educational attainment
• occupation and income
• access to credit and financial inclusion
• type of job (public sector, private white collar, private blue collar, farm wage employment,
business)
We will also investigate outcomes for children aged 3 to 16, disaggregated by whether they
are children or grandchildren of the beneficiary, where it is possible to distinguish. Child’s
outcomes are:
• height and weight
• age-adjusted middle-upper arm circumference percentile
• share that attend school or anganwadi
• child age distribution

Finally, if sufficient tracking information is available, we will investigate beneficiary mortality.
Specifically we will report the share of original beneficiaries who have died since the
original intervention in 2007, conditional on age at baseline.

COVID-19 pandemic outcomes:
There are few studies quantifying the long-run impacts of the COVID-19 pandemic on the
rural poor in a developing country. Therefore we are not only interested in estimating the
treatment effect, but also in describing the overall experiences of the sample during and in
between the 2020 and 2021 COVID-19 lockdowns. In order to publish the descriptives, we
will present tables with 4 columns: control group mean, treatment group mean, combined
mean, and p-value for difference between control and treatment group.

Intergenerational outcomes:
Secondary outcomes regarding intergenerational effects on beneficiary’s children and grandchildren
will be tested.
We are interested in outcomes for beneficiaries’ children, who are now adults. However,
we only observe beneficiaries’ adult children if the children still live in the same household as the beneficiary, or replaced a deceased beneficiary as household head. Estimates based
on coresident adult children may be biased if treatment changes coresidence patterns.
Therefore we will first test for a treatment effect on household division using the share of
beneficiaries’ daughters (and separately sons) who live with the beneficiary, conditional on
child age. We will estimate treatment effects separately for sons and daughters, and only
if we find no evidence of a treatment effect on probability of coresidence for that group.

The present study does not involve any intervention or experiment. It conducts an additional follow-up survey on treatment and control groups that received an intervention from 2007-2008.

In the original experiment, the poorest households were identified in two steps. First, residents across 120 village hamlets ranked households into five wealth quintiles. Among households ranked in the bottom quintile, Bandhan then verified eligibility per seven criteria: (i) presence of an able-
bodied female member (to manage the asset), (ii) no credit access, (iii) landholding below 0.2 acres, (iv) no ownership of productive assets, (v) no able-bodied male member, (vi) presence of school-aged children who
were working instead of attending school, and (vii) primary source of income being informal labor or begging. Households had to meet the first two criteria and at least three of the remaining five in order to be eligible for the TUP intervention. In total, 978 households were deemed eligible. Roughly half of these (514) were randomly assigned to receive the intervention, with stratification at the hamlet level. Of these, only 266 accepted treatment. All reported estimates are intent-to-treat estimates.

Households in the treatment group who chose to participate chose a productive asset from a menu of options (two cows, four goats, one cow and two goats, nonfarm microenterprise inventory, etc). About 82 percent chose livestock. In addition to the asset, they received weekly consumption support for 30–
40 weeks,3 access to savings, and weekly visits from Bandhan staff over a span of 18 months. These visits were designed to deliver training on generating income from the chosen asset, lifeskills coaching, and health information. Bandhan had no contact with beneficiary
households starting 18 months after the asset transfer.

To collect information on baseline household characteristics, the research team administered a survey prior to the distribution of assets in 2007–2008, recording household demographics, consumption, food security, asset ownership, income, income sources, financial inclusion, adult time use, and physical and mental well-being. We track economic and health outcomes for treated and control households
through four subsequent survey waves administered at 18 months, 3 years, 7 years,
and 10 years, and now 16 years after the transfer of productive assets.

Not available

Randomization done in office by a computer.

Randomization at the household level, stratified by hamlet (similar to a village).

Yes

978 households

2,300 people

514 households assigned to treatment, 464 assigned to control

The intervention has already happened, so the sample size is known. Table 1 below estimates the minimum detectable effect sizes (MDES) for the 9 primary outcomes from past follow ups, matching Table 1 of Banerjee, Duflo, and Sharma (2021). We consider three different rates of attrition from the baseline sample of 978 households. These are 10.5% (Row 1, attrition rate from the 10 year endline), 17% (Row 2, the max over four previous endlines), and 25% (Row 3, the upper bound if we experience extremely low rates of tracking). Column 1 reports the number of households remaining after dropping a randomly attrited sample (N HHs), and Column 2 (N adults) reports the number of adults. We employ the following data and assumptions. All calculations use the actual rate of take up of the program (52%). We assume significance level alpha=0.05, and power beta=0.8. Simulations employ actual data on the control group from the ten year follow up study, since outcomes have increased even in the control group since 2007. We loop over several effect sizes to determine the minimum detectable effect size for each outcome. For each effect size, we first construct a full study population resembling the control group. We then randomly assign treatment status, assign the treatment effect to the treated group, and evaluate whether we can reject the null hypothesis of zero treatment effect (repeating the whole exercise 1000 times for each effect size). Just as in the study, treatment is assigned stratifying by hamlet (PSU). The treatment effect is modeled as a normal distribution with mean equals MDES and standard deviation of 0.05. Attrition is simulated before testing the null hypothesis. The regression includes PSU fixed effects, but no baseline controls. For individual-level outcomes, we cluster regression standard errors at the household level. The minimum detectable effect (MDE) is the effect size for which we have 80% power to correctly reject the null hypothesis of zero treatment effect. In other words, where we correctly reject the null hypothesis in 80% of simulations. The MDEs represent upper bounds since the regressions do not control for the baseline value of a variable (as done in Banerjee, Duflo and Sharma 2023). Table 1 reports results. All MDEs are reported in standard deviation units relative to the baseline value in the control group. All estimates are intent-to-treat values. Loss Rate: 10.50% Number of Households: 875 Number of Adults: 1,957 Asset Index: 0.27 Per Capita Consumption: 0.30 Food Security Index: 0.14 Income and Revenues: 0.15 Financial Inclusion: 0.28 Physical Health: 0.06 Mental Health: 0.08 Productive Time Use: 0.10 Loss Rate: 17% Number of Households: 802 Number of Adults: 1,827 Asset Index: 0.27 Per Capita Consumption: 0.34 Food Security Index: 0.16 Income and Revenues: 0.15 Financial Inclusion: 0.30 Physical Health: 0.06 Mental Health: 0.08 Productive Time Use: 0.10 Loss Rate: 25% Number of Households: 734 Number of Adults: 1,676 Asset Index: 0.27 Per Capita Consumption: 0.36 Food Security Index: 0.16 Income and Revenues: 0.17 Financial Inclusion: 0.32 Physical Health: 0.08 Mental Health: 0.08 Productive Time Use: 0.10

Pre-Analysis Plan: Long(er)-term Effects of the Targeting the Ultra Poor Program

Revised Pre-Analysis Plan w/ Additional Outcomes: Long(er)-term Effects of the Targeting the Ultra Poor Program